1.

Exploratory Data Analysis

1. Exploratory Data Analysis

This chapter presents the assumptions, principles, and techniques necessary to gain
insight into data via EDA--exploratory data analysis.

1. EDA Introduction 2. EDA Assumptions

1. What is EDA? 1. Underlying Assumptions
2. EDA vs Classical & Bayesian 2. Importance
3. EDA vs Summary 3. Techniques for Testing
4. EDA Goals Assumptions
5. The Role of Graphics 4. Interpretation of 4-Plot
6. An EDA/Graphics Example 5. Consequences
7. General Problem Categories

3. EDA Techniques 4. EDA Case Studies

1. Introduction 1. Introduction
2. Analysis Questions 2. By Problem Category
3. Graphical Techniques: Alphabetical
4. Graphical Techniques: By Problem
Category
5. Quantitative Techniques
6. Probability Distributions

Detailed Chapter Table of Contents

References
Dataplot Commands for EDA Techniques

http://www.itl.nist.gov/div898/handbook/eda/eda.htm [5/7/2002 3:14:01 PM]

1. Exploratory Data Analysis

1. Exploratory Data Analysis - Detailed Table of

Contents [1.]

This chapter presents the assumptions, principles, and techniques necessary to gain insight into
data via EDA--exploratory data analysis.
1. EDA Introduction [1.1.]
1. What is EDA? [1.1.1.]
2. How Does Exploratory Data Analysis differ from Classical Data Analysis? [1.1.2.]
1. Model [1.1.2.1.]
2. Focus [1.1.2.2.]
3. Techniques [1.1.2.3.]
4. Rigor [1.1.2.4.]
5. Data Treatment [1.1.2.5.]
6. Assumptions [1.1.2.6.]
3. How Does Exploratory Data Analysis Differ from Summary Analysis? [1.1.3.]
4. What are the EDA Goals? [1.1.4.]
5. The Role of Graphics [1.1.5.]
6. An EDA/Graphics Example [1.1.6.]
7. General Problem Categories [1.1.7.]

2. EDA Assumptions [1.2.]

1. Underlying Assumptions [1.2.1.]
2. Importance [1.2.2.]
3. Techniques for Testing Assumptions [1.2.3.]
4. Interpretation of 4-Plot [1.2.4.]
5. Consequences [1.2.5.]
1. Consequences of Non-Randomness [1.2.5.1.]
2. Consequences of Non-Fixed Location Parameter [1.2.5.2.]

http://www.itl.nist.gov/div898/handbook/eda/eda_d.htm (1 of 8) [5/7/2002 3:13:51 PM]

1. Exploratory Data Analysis

3. Consequences of Non-Fixed Variation Parameter [1.2.5.3.]

4. Consequences Related to Distributional Assumptions [1.2.5.4.]

3. EDA Techniques [1.3.]

1. Introduction [1.3.1.]
2. Analysis Questions [1.3.2.]
3. Graphical Techniques: Alphabetic [1.3.3.]
1. Autocorrelation Plot [1.3.3.1.]
1. Autocorrelation Plot: Random Data [1.3.3.1.1.]
2. Autocorrelation Plot: Moderate Autocorrelation [1.3.3.1.2.]
3. Autocorrelation Plot: Strong Autocorrelation and Autoregressive
Model [1.3.3.1.3.]
4. Autocorrelation Plot: Sinusoidal Model [1.3.3.1.4.]
2. Bihistogram [1.3.3.2.]
3. Block Plot [1.3.3.3.]
4. Bootstrap Plot [1.3.3.4.]
5. Box-Cox Linearity Plot [1.3.3.5.]
6. Box-Cox Normality Plot [1.3.3.6.]
7. Box Plot [1.3.3.7.]
8. Complex Demodulation Amplitude Plot [1.3.3.8.]
9. Complex Demodulation Phase Plot [1.3.3.9.]
10. Contour Plot [1.3.3.10.]
1. DEX Contour Plot [1.3.3.10.1.]
11. DEX Scatter Plot [1.3.3.11.]
12. DEX Mean Plot [1.3.3.12.]
13. DEX Standard Deviation Plot [1.3.3.13.]
14. Histogram [1.3.3.14.]
1. Histogram Interpretation: Normal [1.3.3.14.1.]
2. Histogram Interpretation: Symmetric, Non-Normal,
Short-Tailed [1.3.3.14.2.]
3. Histogram Interpretation: Symmetric, Non-Normal,
Long-Tailed [1.3.3.14.3.]
4. Histogram Interpretation: Symmetric and Bimodal [1.3.3.14.4.]
5. Histogram Interpretation: Bimodal Mixture of 2 Normals [1.3.3.14.5.]

http://www.itl.nist.gov/div898/handbook/eda/eda_d.htm (2 of 8) [5/7/2002 3:13:51 PM]

1. Exploratory Data Analysis

6. Histogram Interpretation: Skewed (Non-Normal) Right [1.3.3.14.6.]

7. Histogram Interpretation: Skewed (Non-Symmetric) Left [1.3.3.14.7.]
8. Histogram Interpretation: Symmetric with Outlier [1.3.3.14.8.]
15. Lag Plot [1.3.3.15.]
1. Lag Plot: Random Data [1.3.3.15.1.]
2. Lag Plot: Moderate Autocorrelation [1.3.3.15.2.]
3. Lag Plot: Strong Autocorrelation and Autoregressive
Model [1.3.3.15.3.]
4. Lag Plot: Sinusoidal Models and Outliers [1.3.3.15.4.]
16. Linear Correlation Plot [1.3.3.16.]
17. Linear Intercept Plot [1.3.3.17.]
18. Linear Slope Plot [1.3.3.18.]
19. Linear Residual Standard Deviation Plot [1.3.3.19.]
20. Mean Plot [1.3.3.20.]
21. Normal Probability Plot [1.3.3.21.]
1. Normal Probability Plot: Normally Distributed Data [1.3.3.21.1.]
2. Normal Probability Plot: Data Have Short Tails [1.3.3.21.2.]
3. Normal Probability Plot: Data Have Long Tails [1.3.3.21.3.]
4. Normal Probability Plot: Data are Skewed Right [1.3.3.21.4.]
22. Probability Plot [1.3.3.22.]
23. Probability Plot Correlation Coefficient Plot [1.3.3.23.]
24. Quantile-Quantile Plot [1.3.3.24.]
25. Run-Sequence Plot [1.3.3.25.]
26. Scatter Plot [1.3.3.26.]
1. Scatter Plot: No Relationship [1.3.3.26.1.]
2. Scatter Plot: Strong Linear (positive correlation)
Relationship [1.3.3.26.2.]
3. Scatter Plot: Strong Linear (negative correlation)
Relationship [1.3.3.26.3.]
4. Scatter Plot: Exact Linear (positive correlation)
Relationship [1.3.3.26.4.]
5. Scatter Plot: Quadratic Relationship [1.3.3.26.5.]
6. Scatter Plot: Exponential Relationship [1.3.3.26.6.]
7. Scatter Plot: Sinusoidal Relationship (damped) [1.3.3.26.7.]

http://www.itl.nist.gov/div898/handbook/eda/eda_d.htm (3 of 8) [5/7/2002 3:13:51 PM]

1. Exploratory Data Analysis

8. Scatter Plot: Variation of Y Does Not Depend on X

(homoscedastic) [1.3.3.26.8.]
9. Scatter Plot: Variation of Y Does Depend on X
(heteroscedastic) [1.3.3.26.9.]
10. Scatter Plot: Outlier [1.3.3.26.10.]
11. Scatterplot Matrix [1.3.3.26.11.]
12. Conditioning Plot [1.3.3.26.12.]
27. Spectral Plot [1.3.3.27.]
1. Spectral Plot: Random Data [1.3.3.27.1.]
2. Spectral Plot: Strong Autocorrelation and Autoregressive
Model [1.3.3.27.2.]
3. Spectral Plot: Sinusoidal Model [1.3.3.27.3.]
28. Standard Deviation Plot [1.3.3.28.]
29. Star Plot [1.3.3.29.]
30. Weibull Plot [1.3.3.30.]
31. Youden Plot [1.3.3.31.]
1. DEX Youden Plot [1.3.3.31.1.]
32. 4-Plot [1.3.3.32.]
33. 6-Plot [1.3.3.33.]
4. Graphical Techniques: By Problem Category [1.3.4.]
5. Quantitative Techniques [1.3.5.]
1. Measures of Location [1.3.5.1.]
2. Confidence Limits for the Mean [1.3.5.2.]
3. Two-Sample t-Test for Equal Means [1.3.5.3.]
1. Data Used for Two-Sample t-Test [1.3.5.3.1.]
4. One-Factor ANOVA [1.3.5.4.]
5. Multi-factor Analysis of Variance [1.3.5.5.]
6. Measures of Scale [1.3.5.6.]
7. Bartlett's Test [1.3.5.7.]
8. Chi-Square Test for the Standard Deviation [1.3.5.8.]
1. Data Used for Chi-Square Test for the Standard Deviation [1.3.5.8.1.]
9. F-Test for Equality of Two Standard Deviations [1.3.5.9.]
10. Levene Test for Equality of Variances [1.3.5.10.]
11. Measures of Skewness and Kurtosis [1.3.5.11.]

http://www.itl.nist.gov/div898/handbook/eda/eda_d.htm (4 of 8) [5/7/2002 3:13:51 PM]

1. Exploratory Data Analysis

12. Autocorrelation [1.3.5.12.]

13. Runs Test for Detecting Non-randomness [1.3.5.13.]
14. Anderson-Darling Test [1.3.5.14.]
15. Chi-Square Goodness-of-Fit Test [1.3.5.15.]
16. Kolmogorov-Smirnov Goodness-of-Fit Test [1.3.5.16.]
17. Grubbs' Test for Outliers [1.3.5.17.]
18. Yates Analysis [1.3.5.18.]
1. Defining Models and Prediction Equations [1.3.5.18.1.]
2. Important Factors [1.3.5.18.2.]
6. Probability Distributions [1.3.6.]
1. What is a Probability Distribution [1.3.6.1.]
2. Related Distributions [1.3.6.2.]
3. Families of Distributions [1.3.6.3.]
4. Location and Scale Parameters [1.3.6.4.]
5. Estimating the Parameters of a Distribution [1.3.6.5.]
1. Method of Moments [1.3.6.5.1.]
2. Maximum Likelihood [1.3.6.5.2.]
3. Least Squares [1.3.6.5.3.]
4. PPCC and Probability Plots [1.3.6.5.4.]
6. Gallery of Distributions [1.3.6.6.]
1. Normal Distribution [1.3.6.6.1.]
2. Uniform Distribution [1.3.6.6.2.]
3. Cauchy Distribution [1.3.6.6.3.]
4. t Distribution [1.3.6.6.4.]
5. F Distribution [1.3.6.6.5.]
6. Chi-Square Distribution [1.3.6.6.6.]
7. Exponential Distribution [1.3.6.6.7.]
8. Weibull Distribution [1.3.6.6.8.]
9. Lognormal Distribution [1.3.6.6.9.]
10. Fatigue Life Distribution [1.3.6.6.10.]
11. Gamma Distribution [1.3.6.6.11.]
12. Double Exponential Distribution [1.3.6.6.12.]
13. Power Normal Distribution [1.3.6.6.13.]

http://www.itl.nist.gov/div898/handbook/eda/eda_d.htm (5 of 8) [5/7/2002 3:13:51 PM]

1. Exploratory Data Analysis

14. Power Lognormal Distribution [1.3.6.6.14.]

15. Tukey-Lambda Distribution [1.3.6.6.15.]
16. Extreme Value Type I Distribution [1.3.6.6.16.]
17. Binomial Distribution [1.3.6.6.17.]
18. Poisson Distribution [1.3.6.6.18.]
7. Tables for Probability Distributions [1.3.6.7.]
1. Cumulative Distribution Function of the Standard Normal
Distribution [1.3.6.7.1.]
2. Upper Critical Values of the Student's-t Distribution [1.3.6.7.2.]
3. Upper Critical Values of the F Distribution [1.3.6.7.3.]
4. Critical Values of the Chi-Square Distribution [1.3.6.7.4.]
5. Critical Values of the t* Distribution [1.3.6.7.5.]
6. Critical Values of the Normal PPCC Distribution [1.3.6.7.6.]

4. EDA Case Studies [1.4.]

1. Case Studies Introduction [1.4.1.]
2. Case Studies [1.4.2.]
1. Normal Random Numbers [1.4.2.1.]
1. Background and Data [1.4.2.1.1.]
2. Graphical Output and Interpretation [1.4.2.1.2.]
3. Quantitative Output and Interpretation [1.4.2.1.3.]
4. Work This Example Yourself [1.4.2.1.4.]
2. Uniform Random Numbers [1.4.2.2.]
1. Background and Data [1.4.2.2.1.]
2. Graphical Output and Interpretation [1.4.2.2.2.]
3. Quantitative Output and Interpretation [1.4.2.2.3.]
4. Work This Example Yourself [1.4.2.2.4.]
3. Random Walk [1.4.2.3.]
1. Background and Data [1.4.2.3.1.]
2. Test Underlying Assumptions [1.4.2.3.2.]
3. Develop A Better Model [1.4.2.3.3.]
4. Validate New Model [1.4.2.3.4.]
5. Work This Example Yourself [1.4.2.3.5.]
4. Josephson Junction Cryothermometry [1.4.2.4.]

http://www.itl.nist.gov/div898/handbook/eda/eda_d.htm (6 of 8) [5/7/2002 3:13:51 PM]

1. Exploratory Data Analysis

1. Background and Data [1.4.2.4.1.]

2. Graphical Output and Interpretation [1.4.2.4.2.]
3. Quantitative Output and Interpretation [1.4.2.4.3.]
4. Work This Example Yourself [1.4.2.4.4.]
5. Beam Deflections [1.4.2.5.]
1. Background and Data [1.4.2.5.1.]
2. Test Underlying Assumptions [1.4.2.5.2.]
3. Develop a Better Model [1.4.2.5.3.]
4. Validate New Model [1.4.2.5.4.]
5. Work This Example Yourself [1.4.2.5.5.]
6. Filter Transmittance [1.4.2.6.]
1. Background and Data [1.4.2.6.1.]
2. Graphical Output and Interpretation [1.4.2.6.2.]
3. Quantitative Output and Interpretation [1.4.2.6.3.]
4. Work This Example Yourself [1.4.2.6.4.]
7. Standard Resistor [1.4.2.7.]
1. Background and Data [1.4.2.7.1.]
2. Graphical Output and Interpretation [1.4.2.7.2.]
3. Quantitative Output and Interpretation [1.4.2.7.3.]
4. Work This Example Yourself [1.4.2.7.4.]
8. Heat Flow Meter 1 [1.4.2.8.]
1. Background and Data [1.4.2.8.1.]
2. Graphical Output and Interpretation [1.4.2.8.2.]
3. Quantitative Output and Interpretation [1.4.2.8.3.]
4. Work This Example Yourself [1.4.2.8.4.]
9. Airplane Glass Failure Time [1.4.2.9.]
1. Background and Data [1.4.2.9.1.]
2. Graphical Output and Interpretation [1.4.2.9.2.]
3. Weibull Analysis [1.4.2.9.3.]
4. Lognormal Analysis [1.4.2.9.4.]
5. Gamma Analysis [1.4.2.9.5.]
6. Power Normal Analysis [1.4.2.9.6.]
7. Power Lognormal Analysis [1.4.2.9.7.]

http://www.itl.nist.gov/div898/handbook/eda/eda_d.htm (7 of 8) [5/7/2002 3:13:51 PM]

1. Exploratory Data Analysis

8. Work This Example Yourself [1.4.2.9.8.]

10. Ceramic Strength [1.4.2.10.]
1. Background and Data [1.4.2.10.1.]
2. Analysis of the Response Variable [1.4.2.10.2.]
3. Analysis of the Batch Effect [1.4.2.10.3.]
4. Analysis of the Lab Effect [1.4.2.10.4.]
5. Analysis of Primary Factors [1.4.2.10.5.]
6. Work This Example Yourself [1.4.2.10.6.]
3. References For Chapter 1: Exploratory Data Analysis [1.4.3.]

http://www.itl.nist.gov/div898/handbook/eda/eda_d.htm (8 of 8) [5/7/2002 3:13:51 PM]

1.1. EDA Introduction

1. Exploratory Data Analysis

1.1. EDA Introduction

Summary What is exploratory data analysis? How did it begin? How and where
did it originate? How is it differentiated from other data analysis
approaches, such as classical and Bayesian? Is EDA the same as
statistical graphics? What role does statistical graphics play in EDA? Is
statistical graphics identical to EDA?
These questions and related questions are dealt with in this section. This
section answers these questions and provides the necessary frame of
reference for EDA assumptions, principles, and techniques.

Table of 1. What is EDA?

Contents for 2. EDA versus Classical and Bayesian
Section 1
1. Models
2. Focus
3. Techniques
4. Rigor
5. Data Treatment
6. Assumptions
3. EDA vs Summary
4. EDA Goals
5. The Role of Graphics
6. An EDA/Graphics Example
7. General Problem Categories

http://www.itl.nist.gov/div898/handbook/eda/section1/eda1.htm [5/7/2002 3:14:02 PM]

1.1.1. What is EDA?

1. Exploratory Data Analysis

1.1. EDA Introduction

1.1.1. What is EDA?

Approach Exploratory Data Analysis (EDA) is an approach/philosophy for data
analysis that employs a variety of techniques (mostly graphical) to
1. maximize insight into a data set;
2. uncover underlying structure;
3. extract important variables;
4. detect outliers and anomalies;
5. test underlying assumptions;
6. develop parsimonious models; and
7. determine optimal factor settings.

Focus The EDA approach is precisely that--an approach--not a set of

techniques, but an attitude/philosophy about how a data analysis should
be carried out.

Philosophy EDA is not identical to statistical graphics although the two terms are
used almost interchangeably. Statistical graphics is a collection of
techniques--all graphically based and all focusing on one data
characterization aspect. EDA encompasses a larger venue; EDA is an
approach to data analysis that postpones the usual assumptions about
what kind of model the data follow with the more direct approach of
allowing the data itself to reveal its underlying structure and model.
EDA is not a mere collection of techniques; EDA is a philosophy as to
how we dissect a data set; what we look for; how we look; and how we
interpret. It is true that EDA heavily uses the collection of techniques
that we call "statistical graphics", but it is not identical to statistical
graphics per se.

http://www.itl.nist.gov/div898/handbook/eda/section1/eda11.htm (1 of 2) [5/7/2002 3:14:02 PM]

1.1.1. What is EDA?

History The seminal work in EDA is Exploratory Data Analysis, Tukey, (1977).
Over the years it has benefitted from other noteworthy publications such
as Data Analysis and Regression, Mosteller and Tukey (1977),
Interactive Data Analysis, Hoaglin (1977), The ABC's of EDA,
Velleman and Hoaglin (1981) and has gained a large following as "the"
way to analyze a data set.

Techniques Most EDA techniques are graphical in nature with a few quantitative
techniques. The reason for the heavy reliance on graphics is that by its
very nature the main role of EDA is to open-mindedly explore, and
graphics gives the analysts unparalleled power to do so, enticing the
data to reveal its structural secrets, and being always ready to gain some
new, often unsuspected, insight into the data. In combination with the
natural pattern-recognition capabilities that we all possess, graphics
provides, of course, unparalleled power to carry this out.
The particular graphical techniques employed in EDA are often quite
simple, consisting of various techniques of:
1. Plotting the raw data (such as data traces, histograms,
bihistograms, probability plots, lag plots, block plots, and Youden
plots.
2. Plotting simple statistics such as mean plots, standard deviation
plots, box plots, and main effects plots of the raw data.
3. Positioning such plots so as to maximize our natural
pattern-recognition abilities, such as using multiple plots per
page.

http://www.itl.nist.gov/div898/handbook/eda/section1/eda11.htm (2 of 2) [5/7/2002 3:14:02 PM]

1.1.2. How Does Exploratory Data Analysis differ from Classical Data Analysis?

1. Exploratory Data Analysis

1.1. EDA Introduction

1.1.2. How Does Exploratory Data Analysis

differ from Classical Data Analysis?
Data EDA is a data analysis approach. What other data analysis approaches
Analysis exist and how does EDA differ from these other approaches? Three
Approaches popular data analysis approaches are:
1. Classical
2. Exploratory (EDA)
3. Bayesian

Paradigms These three approaches are similar in that they all start with a general
for Analysis science/engineering problem and all yield science/engineering
Techniques conclusions. The difference is the sequence and focus of the
intermediate steps.
For classical analysis, the sequence is
Problem => Data => Model => Analysis => Conclusions
For EDA, the sequence is
Problem => Data => Analysis => Model => Conclusions
For Bayesian, the sequence is
Problem => Data => Model => Prior Distribution => Analysis =>
Conclusions

http://www.itl.nist.gov/div898/handbook/eda/section1/eda12.htm (1 of 2) [5/7/2002 3:14:02 PM]

1.1.2. How Does Exploratory Data Analysis differ from Classical Data Analysis?

Method of Thus for classical analysis, the data collection is followed by the
dealing with imposition of a model (normality, linearity, etc.) and the analysis,
underlying estimation, and testing that follows are focused on the parameters of
model for that model. For EDA, the data collection is not followed by a model
the data imposition; rather it is followed immediately by analysis with a goal of
distinguishes inferring what model would be appropriate. Finally, for a Bayesian
the 3 analysis, the analyst attempts to incorporate scientific/engineering
approaches knowledge/expertise into the analysis by imposing a data-independent
distribution on the parameters of the selected model; the analysis thus
consists of formally combining both the prior distribution on the
parameters and the collected data to jointly make inferences and/or test
assumptions about the model parameters.
In the real world, data analysts freely mix elements of all of the above
three approaches (and other approaches). The above distinctions were
made to emphasize the major differences among the three approaches.

Further Focusing on EDA versus classical, these two approaches differ as

discussion of follows:
the 1. Models
distinction
between the 2. Focus
classical and 3. Techniques
EDA 4. Rigor
approaches
5. Data Treatment
6. Assumptions

http://www.itl.nist.gov/div898/handbook/eda/section1/eda12.htm (2 of 2) [5/7/2002 3:14:02 PM]

1.1.2.1. Model

1. Exploratory Data Analysis

1.1. EDA Introduction
1.1.2. How Does Exploratory Data Analysis differ from Classical Data Analysis?

1.1.2.1. Model
Classical The classical approach imposes models (both deterministic and
probabilistic) on the data. Deterministic models include, for example,
regression models and analysis of variance (ANOVA) models. The most
common probabilistic model assumes that the errors about the
deterministic model are normally distributed--this assumption affects the
validity of the ANOVA F tests.

Exploratory The Exploratory Data Analysis approach does not impose deterministic
or probabilistic models on the data. On the contrary, the EDA approach
allows the data to suggest admissible models that best fit the data.

http://www.itl.nist.gov/div898/handbook/eda/section1/eda121.htm [5/7/2002 3:14:02 PM]

1.1.2.2. Focus

1. Exploratory Data Analysis

1.1. EDA Introduction
1.1.2. How Does Exploratory Data Analysis differ from Classical Data Analysis?

1.1.2.2. Focus
Classical The two approaches differ substantially in focus. For classical analysis,
the focus is on the model--estimating parameters of the model and
generating predicted values from the model.

Exploratory For exploratory data analysis, the focus is on the data--its structure,
outliers, and models suggested by the data.

http://www.itl.nist.gov/div898/handbook/eda/section1/eda122.htm [5/7/2002 3:14:02 PM]

1.1.2.3. Techniques

1. Exploratory Data Analysis

1.1. EDA Introduction
1.1.2. How Does Exploratory Data Analysis differ from Classical Data Analysis?

1.1.2.3. Techniques
Classical Classical techniques are generally quantitative in nature. They include
ANOVA, t tests, chi-squared tests, and F tests.

Exploratory EDA techniques are generally graphical. They include scatter plots,
character plots, box plots, histograms, bihistograms, probability plots,
residual plots, and mean plots.

http://www.itl.nist.gov/div898/handbook/eda/section1/eda123.htm [5/7/2002 3:14:02 PM]

1.1.2.4. Rigor

1. Exploratory Data Analysis

1.1. EDA Introduction
1.1.2. How Does Exploratory Data Analysis differ from Classical Data Analysis?

1.1.2.4. Rigor
Classical Classical techniques serve as the probabilistic foundation of science and
engineering; the most important characteristic of classical techniques is
that they are rigorous, formal, and "objective".

Exploratory EDA techniques do not share in that rigor or formality. EDA techniques
make up for that lack of rigor by being very suggestive, indicative, and
insightful about what the appropriate model should be.
EDA techniques are subjective and depend on interpretation which may
differ from analyst to analyst, although experienced analysts commonly
arrive at identical conclusions.

http://www.itl.nist.gov/div898/handbook/eda/section1/eda124.htm [5/7/2002 3:14:03 PM]

1.1.2.5. Data Treatment

1. Exploratory Data Analysis

1.1. EDA Introduction
1.1.2. How Does Exploratory Data Analysis differ from Classical Data Analysis?

1.1.2.5. Data Treatment

Classical Classical estimation techniques have the characteristic of taking all of
the data and mapping the data into a few numbers ("estimates"). This is
both a virtue and a vice. The virtue is that these few numbers focus on
important characteristics (location, variation, etc.) of the population. The
vice is that concentrating on these few characteristics can filter out other
characteristics (skewness, tail length, autocorrelation, etc.) of the same
population. In this sense there is a loss of information due to this
"filtering" process.

Exploratory The EDA approach, on the other hand, often makes use of (and shows)
all of the available data. In this sense there is no corresponding loss of
information.

http://www.itl.nist.gov/div898/handbook/eda/section1/eda125.htm [5/7/2002 3:14:03 PM]

1.1.2.6. Assumptions

1. Exploratory Data Analysis

1.1. EDA Introduction
1.1.2. How Does Exploratory Data Analysis differ from Classical Data Analysis?

1.1.2.6. Assumptions
Classical The "good news" of the classical approach is that tests based on
classical techniques are usually very sensitive--that is, if a true shift in
location, say, has occurred, such tests frequently have the power to
detect such a shift and to conclude that such a shift is "statistically
significant". The "bad news" is that classical tests depend on underlying
assumptions (e.g., normality), and hence the validity of the test
conclusions becomes dependent on the validity of the underlying
assumptions. Worse yet, the exact underlying assumptions may be
unknown to the analyst, or if known, untested. Thus the validity of the
scientific conclusions becomes intrinsically linked to the validity of the
underlying assumptions. In practice, if such assumptions are unknown
or untested, the validity of the scientific conclusions becomes suspect.

Exploratory Many EDA techniques make little or no assumptions--they present and

show the data--all of the data--as is, with fewer encumbering
assumptions.

http://www.itl.nist.gov/div898/handbook/eda/section1/eda126.htm [5/7/2002 3:14:03 PM]

1.1.3. How Does Exploratory Data Analysis Differ from Summary Analysis?

1. Exploratory Data Analysis

1.1. EDA Introduction

1.1.3. How Does Exploratory Data Analysis

Differ from Summary Analysis?
Summary A summary analysis is simply a numeric reduction of a historical data
set. It is quite passive. It's focus is in the past. Quite commonly, its
purpose is to simply arrive at a few key statistics (for example, mean
and standard deviation) which may then either replace the data set or be
added to the data set in the form of a summary table.

Exploratory In contrast, EDA has as its broadest goal the desire to gain insight into
the engineering/scientific process behind the data. Whereas summary
statistics are passive and historical, EDA is active and futuristic. In an
attempt to "understand" the process and improve it in the future, EDA
uses the data as a "window" to peer into the heart of the process that
generated the data. There is an archival role in the research and
manufacturing world for summary statistics, but there is an enormously
larger role for the EDA approach.

http://www.itl.nist.gov/div898/handbook/eda/section1/eda13.htm [5/7/2002 3:14:03 PM]

1.1.4. What are the EDA Goals?

1. Exploratory Data Analysis

1.1. EDA Introduction

1.1.4. What are the EDA Goals?

Primary and The primary goal of EDA is to maximize the analyst's insight into a data
Secondary set and into the underlying structure of a data set, while providing all of
Goals the specific items that an analyst would want to extract from a data set,
such as:
1. a good-fitting, parsimonious model
2. a list of outliers
3. a sense of robustness of conclusions
4. estimates for parameters
5. uncertainties for those estimates
6. a ranked list of important factors
7. conclusions as to whether individual factors are statistically
significant
8. optimal settings

Insight into Insight implies detecting and uncovering underlying structure in the
the Data data. Such underlying structure may not be encapsulated in the list of
items above; such items serve as the specific targets of an analysis, but
the real insight and "feel" for a data set comes as the analyst judiciously
probes and explores the various subtleties of the data. The "feel" for the
data comes almost exclusively from the application of various graphical
techniques, the collection of which serves as the window into the
essence of the data. Graphics are irreplaceable--there are no quantitative
analogues that will give the same insight as well-chosen graphics.
To get a "feel" for the data, it is not enough for the analyst to know what
is in the data; the analyst also must know what is not in the data, and the
only way to do that is to draw on our own human pattern-recognition
and comparative abilities in the context of a series of judicious graphical
techniques applied to the data.

http://www.itl.nist.gov/div898/handbook/eda/section1/eda14.htm [5/7/2002 3:14:03 PM]

1.1.5. The Role of Graphics

1. Exploratory Data Analysis

1.1. EDA Introduction

1.1.5. The Role of Graphics

Quantitative/ Statistics and data analysis procedures can broadly be split into two
Graphical parts:
● quantitative

● graphical

Quantitative Quantitative techniques are the set of statistical procedures that yield
numeric or tabular output. Examples of quantitative techniques include:
● hypothesis testing

● analysis of variance
● point estimates and confidence intervals
● least squares regression
These and similar techniques are all valuable and are mainstream in
terms of classical analysis.

Graphical On the other hand, there is a large collection of statistical tools that we
generally refer to as graphical techniques. These include:
● scatter plots

● histograms
● probability plots
● residual plots
● box plots
● block plots

http://www.itl.nist.gov/div898/handbook/eda/section1/eda15.htm (1 of 2) [5/7/2002 3:14:03 PM]

1.1.5. The Role of Graphics

EDA The EDA approach relies heavily on these and similar graphical
Approach techniques. Graphical procedures are not just tools that we could use in
Relies an EDA context, they are tools that we must use. Such graphical tools
Heavily on are the shortest path to gaining insight into a data set in terms of
Graphical ● testing assumptions
Techniques
● model selection

● model validation

● estimator selection

● relationship identification

● factor effect determination

● outlier detection

If one is not using statistical graphics, then one is forfeiting insight into
one or more aspects of the underlying structure of the data.

http://www.itl.nist.gov/div898/handbook/eda/section1/eda15.htm (2 of 2) [5/7/2002 3:14:03 PM]

1.1.6. An EDA/Graphics Example

1. Exploratory Data Analysis

1.1. EDA Introduction

1.1.6. An EDA/Graphics Example

Anscombe A simple, classic (Anscombe) example of the central role that graphics
Example play in terms of providing insight into a data set starts with the
following data set:

Data
X Y
10.00 8.04
8.00 6.95
13.00 7.58
9.00 8.81
11.00 8.33
14.00 9.96
6.00 7.24
4.00 4.26
12.00 10.84
7.00 4.82
5.00 5.68

Summary If the goal of the analysis is to compute summary statistics plus

Statistics determine the best linear fit for Y as a function of X, the results might
be given as:
N = 11
Mean of X = 9.0
Mean of Y = 7.5
Intercept = 3
Slope = 0.5
Residual standard deviation = 1.237
Correlation = 0.816
The above quantitative analysis, although valuable, gives us only
limited insight into the data.

http://www.itl.nist.gov/div898/handbook/eda/section1/eda16.htm (1 of 5) [5/7/2002 3:14:04 PM]

1.1.6. An EDA/Graphics Example

Scatter Plot In contrast, the following simple scatter plot of the data

suggests the following:

1. The data set "behaves like" a linear curve with some scatter;
2. there is no justification for a more complicated model (e.g.,
quadratic);
3. there are no outliers;
4. the vertical spread of the data appears to be of equal height
irrespective of the X-value; this indicates that the data are
equally-precise throughout and so a "regular" (that is,
equi-weighted) fit is appropriate.

Three This kind of characterization for the data serves as the core for getting
Additional insight/feel for the data. Such insight/feel does not come from the
Data Sets quantitative statistics; on the contrary, calculations of quantitative
statistics such as intercept and slope should be subsequent to the
characterization and will make sense only if the characterization is
true. To illustrate the loss of information that results when the graphics
insight step is skipped, consider the following three data sets
[Anscombe data sets 2, 3, and 4]:

X2 Y2 X3 Y3 X4 Y4
10.00 9.14 10.00 7.46 8.00 6.58
8.00 8.14 8.00 6.77 8.00 5.76
13.00 8.74 13.00 12.74 8.00 7.71

http://www.itl.nist.gov/div898/handbook/eda/section1/eda16.htm (2 of 5) [5/7/2002 3:14:04 PM]

1.1.6. An EDA/Graphics Example

9.00 8.77 9.00 7.11 8.00 8.84

11.00 9.26 11.00 7.81 8.00 8.47
14.00 8.10 14.00 8.84 8.00 7.04
6.00 6.13 6.00 6.08 8.00 5.25
4.00 3.10 4.00 5.39 19.00 12.50
12.00 9.13 12.00 8.15 8.00 5.56
7.00 7.26 7.00 6.42 8.00 7.91
5.00 4.74 5.00 5.73 8.00 6.89

Quantitative A quantitative analysis on data set 2 yields

Statistics for N = 11
Data Set 2 Mean of X = 9.0
Mean of Y = 7.5
Intercept = 3
Slope = 0.5
Residual standard deviation = 1.237
Correlation = 0.816
which is identical to the analysis for data set 1. One might naively
assume that the two data sets are "equivalent" since that is what the
statistics tell us; but what do the statistics not tell us?

Quantitative Remarkably, a quantitative analysis on data sets 3 and 4 also yields

Statistics for N = 11
Data Sets 3 Mean of X = 9.0
and 4 Mean of Y = 7.5
Intercept = 3
Slope = 0.5
Residual standard deviation = 1.236
Correlation = 0.816 (0.817 for data set 4)
which implies that in some quantitative sense, all four of the data sets
are "equivalent". In fact, the four data sets are far from "equivalent"
and a scatter plot of each data set, which would be step 1 of any EDA
approach, would tell us that immediately.

http://www.itl.nist.gov/div898/handbook/eda/section1/eda16.htm (3 of 5) [5/7/2002 3:14:04 PM]

1.1.6. An EDA/Graphics Example

Scatter Plots

Interpretation Conclusions from the scatter plots are:

of Scatter 1. data set 1 is clearly linear with some scatter.
Plots
2. data set 2 is clearly quadratic.
3. data set 3 clearly has an outlier.
4. data set 4 is obviously the victim of a poor experimental design
with a single point far removed from the bulk of the data
"wagging the dog".

Importance These points are exactly the substance that provide and define "insight"
of and "feel" for a data set. They are the goals and the fruits of an open
Exploratory exploratory data analysis (EDA) approach to the data. Quantitative
Analysis statistics are not wrong per se, but they are incomplete. They are
incomplete because they are numeric summaries which in the
summarization operation do a good job of focusing on a particular
aspect of the data (e.g., location, intercept, slope, degree of relatedness,
etc.) by judiciously reducing the data to a few numbers. Doing so also
filters the data, necessarily omitting and screening out other sometimes
crucial information in the focusing operation. Quantitative statistics
focus but also filter; and filtering is exactly what makes the
quantitative approach incomplete at best and misleading at worst.
The estimated intercepts (= 3) and slopes (= 0.5) for data sets 2, 3, and
4 are misleading because the estimation is done in the context of an
assumed linear model and that linearity assumption is the fatal flaw in
this analysis.

http://www.itl.nist.gov/div898/handbook/eda/section1/eda16.htm (4 of 5) [5/7/2002 3:14:04 PM]

1.1.6. An EDA/Graphics Example

The EDA approach of deliberately postponing the model selection until

further along in the analysis has many rewards, not the least of which is
the ultimate convergence to a much-improved model and the
formulation of valid and supportable scientific and engineering
conclusions.

http://www.itl.nist.gov/div898/handbook/eda/section1/eda16.htm (5 of 5) [5/7/2002 3:14:04 PM]

1.1.7. General Problem Categories

1. Exploratory Data Analysis

1.1. EDA Introduction

1.1.7. General Problem Categories

Problem The following table is a convenient way to classify EDA problems.
Classification

Univariate
UNIVARIATE CONTROL
and Control
Data: Data:
A single column of A single column of
numbers, Y. numbers, Y.
Model: Model:
y = constant + error y = constant + error
Output: Output:
1. A number (the estimated A "yes" or "no" to the
constant in the model). question "Is the system
2. An estimate of uncertainty out of control?".
for the constant. Techniques:
3. An estimate of the ● Control Charts
distribution for the error.
Techniques:
● 4-Plot

● Probability Plot
● PPCC Plot

http://www.itl.nist.gov/div898/handbook/eda/section1/eda17.htm (1 of 4) [5/7/2002 3:14:04 PM]

1.1.7. General Problem Categories

Comparative
COMPARATIVE SCREENING
and
Screening Data: Data:
A single response variable A single response variable
and k independent and k independent
variables (Y, X1, X2, ... , variables (Y, X1, X2, ... ,
Xk), primary focus is on Xk).
one (the primary factor) of Model:
these independent
y = f(x1, x2, ..., xk) + error
variables.
Model: Output:
y = f(x1, x2, ..., xk) + error 1. A ranked list (from most
important to least
Output: important) of factors.
A "yes" or "no" to the 2. Best settings for the
question "Is the primary factors.
factor significant?".
3. A good model/prediction
Techniques: equation relating Y to the
● Block Plot factors.
● Scatter Plot Techniques:
● Box Plot ● Block Plot

● Probability Plot
● Bihistogram

Optimization
OPTIMIZATION REGRESSION
and
Regression Data: Data:
A single response variable A single response variable
and k independent and k independent
variables (Y, X1, X2, ... , variables (Y, X1, X2, ... ,
Xk). Xk). The independent
Model: variables can be
continuous.
y = f(x1, x2, ..., xk) + error
Model:
Output:
y = f(x1, x2, ..., xk) + error
Best settings for the factor
variables. Output:
Techniques: A good model/prediction
equation relating Y to the
● Block Plot
factors.

http://www.itl.nist.gov/div898/handbook/eda/section1/eda17.htm (2 of 4) [5/7/2002 3:14:04 PM]

1.1.7. General Problem Categories

● Least Squares Fitting Techniques:

● Contour Plot ● Least Squares Fitting

● Scatter Plot
● 6-Plot

Time Series
TIME SERIES MULTIVARIATE
and
Multivariate Data: Data:
A column of time k factor variables (X1, X2, ... ,
dependent numbers, Y. Xk).
In addition, time is an
indpendent variable. Model:
The time variable can The model is not explicit.
be either explicit or Output:
implied. If the data are Identify underlying
not equi-spaced, the correlation structure in the
time variable should be data.
explicitly provided.
Techniques:
Model:
● Star Plot
yt = f(t) + error
● Scatter Plot Matrix
The model can be either
a time domain based or ● Conditioning Plot
frequency domain ● Profile Plot
based.
● Principal Components
Output:
Clustering
●
A good
● Discrimination/Classification
model/prediction
equation relating Y to Note that multivarate analysis is
previous values of Y. only covered lightly in this
Techniques: Handbook.
● Autocorrelation Plot

● Spectrum
● Complex Demodulation
Amplitude Plot
● Complex Demodulation
Phase Plot
● ARIMA Models

http://www.itl.nist.gov/div898/handbook/eda/section1/eda17.htm (3 of 4) [5/7/2002 3:14:04 PM]

1.1.7. General Problem Categories

http://www.itl.nist.gov/div898/handbook/eda/section1/eda17.htm (4 of 4) [5/7/2002 3:14:04 PM]

1.2. EDA Assumptions

1. Exploratory Data Analysis

1.2. EDA Assumptions

Summary The gamut of scientific and engineering experimentation is virtually
limitless. In this sea of diversity is there any common basis that allows
the analyst to systematically and validly arrive at supportable, repeatable
research conclusions?
Fortunately, there is such a basis and it is rooted in the fact that every
measurement process, however complicated, has certain underlying
assumptions. This section deals with what those assumptions are, why
they are important, how to go about testing them, and what the
consequences are if the assumptions do not hold.

Table of 1. Underlying Assumptions

Contents for 2. Importance
Section 2
3. Testing Assumptions
4. Importance of Plots
5. Consequences

http://www.itl.nist.gov/div898/handbook/eda/section2/eda2.htm [5/7/2002 3:14:04 PM]

1.2.1. Underlying Assumptions

1. Exploratory Data Analysis

1.2. EDA Assumptions

1.2.1. Underlying Assumptions

Assumptions There are four assumptions that typically underlie all measurement
Underlying a processes; namely, that the data from the process at hand "behave
Measurement like":
Process 1. random drawings;
2. from a fixed distribution;
3. with the distribution having fixed location; and
4. with the distribution having fixed variation.

Univariate or The "fixed location" referred to in item 3 above differs for different
Single problem types. The simplest problem type is univariate; that is, a
Response single variable. For the univariate problem, the general model
Variable response = deterministic component + random component
becomes
response = constant + error

Assumptions For this case, the "fixed location" is simply the unknown constant. We
for Univariate can thus imagine the process at hand to be operating under constant
Model conditions that produce a single column of data with the properties
that
● the data are uncorrelated with one another;

● the random component has a fixed distribution;

● the deterministic component consists of only a constant; and

● the random component has fixed variation.

Extrapolation The universal power and importance of the univariate model is that it
to a Function can easily be extended to the more general case where the
of Many deterministic component is not just a constant, but is in fact a function
Variables of many variables, and the engineering objective is to characterize and
model the function.

http://www.itl.nist.gov/div898/handbook/eda/section2/eda21.htm (1 of 2) [5/7/2002 3:14:04 PM]

1.2.1. Underlying Assumptions

Residuals Will The key point is that regardless of how many factors there are, and
Behave regardless of how complicated the function is, if the engineer succeeds
According to in choosing a good model, then the differences (residuals) between the
Univariate raw response data and the predicted values from the fitted model
Assumptions should themselves behave like a univariate process. Furthermore, the
residuals from this univariate process fit will behave like:
● random drawings;

● from a fixed distribution;

● with fixed location (namely, 0 in this case); and

● with fixed variation.

Validation of Thus if the residuals from the fitted model do in fact behave like the
Model ideal, then testing of underlying assumptions becomes a tool for the
validation and quality of fit of the chosen model. On the other hand, if
the residuals from the chosen fitted model violate one or more of the
above univariate assumptions, then the chosen fitted model is
inadequate and an opportunity exists for arriving at an improved
model.

http://www.itl.nist.gov/div898/handbook/eda/section2/eda21.htm (2 of 2) [5/7/2002 3:14:04 PM]

1.2.2. Importance

1. Exploratory Data Analysis

1.2. EDA Assumptions

1.2.2. Importance
Predictability Predictability is an all-important goal in science and engineering. If the
and four underlying assumptions hold, then we have achieved probabilistic
Statistical predictability--the ability to make probability statements not only
Control about the process in the past, but also about the process in the future.
In short, such processes are said to be "in statistical control".

Validity of Moreover, if the four assumptions are valid, then the process is
Engineering amenable to the generation of valid scientific and engineering
Conclusions conclusions. If the four assumptions are not valid, then the process is
drifting (with respect to location, variation, or distribution),
unpredictable, and out of control. A simple characterization of such
processes by a location estimate, a variation estimate, or a distribution
"estimate" inevitably leads to engineering conclusions that are not
valid, are not supportable (scientifically or legally), and which are not
repeatable in the laboratory.

http://www.itl.nist.gov/div898/handbook/eda/section2/eda22.htm [5/7/2002 3:14:04 PM]

1.2.3. Techniques for Testing Assumptions

1. Exploratory Data Analysis

1.2. EDA Assumptions

1.2.3. Techniques for Testing Assumptions

Testing Because the validity of the final scientific/engineering conclusions
Underlying is inextricably linked to the validity of the underlying univariate
Assumptions assumptions, it naturally follows that there is a real necessity that
Helps Assure the each and every one of the above four assumptions be routinely
Validity of tested.
Scientific and
Engineering
Conclusions

Four Techniques The following EDA techniques are simple, efficient, and powerful
to Test for the routine testing of underlying assumptions:
Underlying 1. run sequence plot (Yi versus i)
Assumptions
2. lag plot (Yi versus Yi-1)
3. histogram (counts versus subgroups of Y)
4. normal probability plot (ordered Y versus theoretical ordered
Y)

Plot on a Single The four EDA plots can be juxtaposed for a quick look at the
Page for a characteristics of the data. The plots below are ordered as follows:
Quick 1. Run sequence plot - upper left
Characterization
2. Lag plot - upper right
of the Data
3. Histogram - lower left
4. Normal probability plot - lower right

http://www.itl.nist.gov/div898/handbook/eda/section2/eda23.htm (1 of 3) [5/7/2002 3:14:05 PM]

1.2.3. Techniques for Testing Assumptions

Sample Plot:
Assumptions
Hold

This 4-plot reveals a process that has fixed location, fixed variation,
is random, apparently has a fixed approximately normal
distribution, and has no outliers.

Sample Plot: If one or more of the four underlying assumptions do not hold, then
Assumptions Do it will show up in the various plots as demonstrated in the following
Not Hold example.

http://www.itl.nist.gov/div898/handbook/eda/section2/eda23.htm (2 of 3) [5/7/2002 3:14:05 PM]

1.2.3. Techniques for Testing Assumptions

This 4-plot reveals a process that has fixed location, fixed variation,
is non-random (oscillatory), has a non-normal, U-shaped
distribution, and has several outliers.

http://www.itl.nist.gov/div898/handbook/eda/section2/eda23.htm (3 of 3) [5/7/2002 3:14:05 PM]

1.2.4. Interpretation of 4-Plot

1. Exploratory Data Analysis

1.2. EDA Assumptions

1.2.4. Interpretation of 4-Plot

Interpretation The four EDA plots discussed on the previous page are used to test the
of EDA Plots: underlying assumptions:
Flat and 1. Fixed Location:
Equi-Banded, If the fixed location assumption holds, then the run sequence
Random, plot will be flat and non-drifting.
Bell-Shaped,
and Linear 2. Fixed Variation:
If the fixed variation assumption holds, then the vertical spread
in the run sequence plot will be the approximately the same over
the entire horizontal axis.
3. Randomness:
If the randomness assumption holds, then the lag plot will be
structureless and random.
4. Fixed Distribution:
If the fixed distribution assumption holds, in particular if the
fixed normal distribution holds, then
1. the histogram will be bell-shaped, and
2. the normal probability plot will be linear.

Plots Utilized Conversely, the underlying assumptions are tested using the EDA
to Test the plots:
Assumptions ● Run Sequence Plot:
If the run sequence plot is flat and non-drifting, the
fixed-location assumption holds. If the run sequence plot has a
vertical spread that is about the same over the entire plot, then
the fixed-variation assumption holds.
● Lag Plot:
If the lag plot is structureless, then the randomness assumption
holds.
● Histogram:
If the histogram is bell-shaped, the underlying distribution is
symmetric and perhaps approximately normal.
● Normal Probability Plot:

http://www.itl.nist.gov/div898/handbook/eda/section2/eda24.htm (1 of 2) [5/7/2002 3:14:05 PM]

1.2.4. Interpretation of 4-Plot

If the normal probability plot is linear, the underlying

distribution is approximately normal.
If all four of the assumptions hold, then the process is said
definitionally to be "in statistical control".

http://www.itl.nist.gov/div898/handbook/eda/section2/eda24.htm (2 of 2) [5/7/2002 3:14:05 PM]

1.2.5. Consequences

1. Exploratory Data Analysis

1.2. EDA Assumptions

1.2.5. Consequences
What If If some of the underlying assumptions do not hold, what can be done
Assumptions about it? What corrective actions can be taken? The positive way of
Do Not Hold? approaching this is to view the testing of underlying assumptions as a
framework for learning about the process. Assumption-testing
promotes insight into important aspects of the process that may not
have surfaced otherwise.

Primary Goal The primary goal is to have correct, validated, and complete
is Correct and scientific/engineering conclusions flowing from the analysis. This
Valid usually includes intermediate goals such as the derivation of a
Scientific good-fitting model and the computation of realistic parameter
Conclusions estimates. It should always include the ultimate goal of an
understanding and a "feel" for "what makes the process tick". There is
no more powerful catalyst for discovery than the bringing together of
an experienced/expert scientist/engineer and a data set ripe with
intriguing "anomalies" and characteristics.

Consequences The following sections discuss in more detail the consequences of

of Invalid invalid assumptions:
Assumptions 1. Consequences of non-randomness
2. Consequences of non-fixed location parameter
3. Consequences of non-fixed variation
4. Consequences related to distributional assumptions

http://www.itl.nist.gov/div898/handbook/eda/section2/eda25.htm [5/7/2002 3:14:05 PM]

1.2.5.1. Consequences of Non-Randomness

1. Exploratory Data Analysis

1.2. EDA Assumptions
1.2.5. Consequences

1.2.5.1. Consequences of Non-Randomness

Randomness There are four underlying assumptions:
Assumption 1. randomness;
2. fixed location;
3. fixed variation; and
4. fixed distribution.
The randomness assumption is the most critical but the least tested.

Consequeces of If the randomness assumption does not hold, then

Non-Randomness 1. All of the usual statistical tests are invalid.
2. The calculated uncertainties for commonly used statistics
become meaningless.
3. The calculated minimal sample size required for a
pre-specified tolerance becomes meaningless.
4. The simple model: y = constant + error becomes invalid.
5. The parameter estimates become suspect and
non-supportable.

Non-Randomness One specific and common type of non-randomness is

Due to autocorrelation. Autocorrelation is the correlation between Yt and
Autocorrelation Yt-k, where k is an integer that defines the lag for the
autocorrelation. That is, autocorrelation is a time dependent
non-randomness. This means that the value of the current point is
highly dependent on the previous point if k = 1 (or k points ago if k
is not 1). Autocorrelation is typically detected via an
autocorrelation plot or a lag plot.

If the data are not random due to autocorrelation, then

1. Adjacent data values may be related.
2. There may not be n independent snapshots of the
phenomenon under study.

http://www.itl.nist.gov/div898/handbook/eda/section2/eda251.htm (1 of 2) [5/7/2002 3:14:05 PM]

1.2.5.1. Consequences of Non-Randomness

3. There may be undetected "junk"-outliers.

4. There may be undetected "information-rich"-outliers.

http://www.itl.nist.gov/div898/handbook/eda/section2/eda251.htm (2 of 2) [5/7/2002 3:14:05 PM]

1.2.5.2. Consequences of Non-Fixed Location Parameter

1. Exploratory Data Analysis

1.2. EDA Assumptions
1.2.5. Consequences

1.2.5.2. Consequences of Non-Fixed

Location Parameter
Location The usual estimate of location is the mean
Estimate

from N measurements Y1, Y2, ... , YN.

Consequences If the run sequence plot does not support the assumption of fixed
of Non-Fixed location, then
Location 1. The location may be drifting.
2. The single location estimate may be meaningless (if the process
is drifting).
3. The choice of location estimator (e.g., the sample mean) may be
sub-optimal.
4. The usual formula for the uncertainty of the mean:

may be invalid and the numerical value optimistically small.

5. The location estimate may be poor.
6. The location estimate may be biased.

http://www.itl.nist.gov/div898/handbook/eda/section2/eda252.htm [5/7/2002 3:14:05 PM]

1.2.5.3. Consequences of Non-Fixed Variation Parameter

1. Exploratory Data Analysis

1.2. EDA Assumptions
1.2.5. Consequences

1.2.5.3. Consequences of Non-Fixed

Variation Parameter
Variation The usual estimate of variation is the standard deviation
Estimate

from N measurements Y1, Y2, ... , YN.

Consequences If the run sequence plot does not support the assumption of fixed
of Non-Fixed variation, then
Variation 1. The variation may be drifting.
2. The single variation estimate may be meaningless (if the process
variation is drifting).
3. The variation estimate may be poor.
4. The variation estimate may be biased.

http://www.itl.nist.gov/div898/handbook/eda/section2/eda253.htm [5/7/2002 3:14:12 PM]

1.2.5.4. Consequences Related to Distributional Assumptions

1. Exploratory Data Analysis

1.2. EDA Assumptions
1.2.5. Consequences

1.2.5.4. Consequences Related to

Distributional Assumptions
Distributional Scientists and engineers routinely use the mean (average) to estimate
Analysis the "middle" of a distribution. It is not so well known that the
variability and the noisiness of the mean as a location estimator are
intrinsically linked with the underlying distribution of the data. For
certain distributions, the mean is a poor choice. For any given
distribution, there exists an optimal choice-- that is, the estimator
with minimum variability/noisiness. This optimal choice may be, for
example, the median, the midrange, the midmean, the mean, or
something else. The implication of this is to "estimate" the
distribution first, and then--based on the distribution--choose the
optimal estimator. The resulting engineering parameter estimators
will have less variability than if this approach is not followed.

Case Studies The airplane glass failure case study gives an example of determining
an appropriate distribution and estimating the parameters of that
distribution. The uniform random numbers case study gives an
example of determining a more appropriate centrality parameter for a
non-normal distribution.

Other consequences that flow from problems with distributional

assumptions are:

Distribution 1. The distribution may be changing.

2. The single distribution estimate may be meaningless (if the
process distribution is changing).
3. The distribution may be markedly non-normal.
4. The distribution may be unknown.
5. The true probability distribution for the error may remain
unknown.

http://www.itl.nist.gov/div898/handbook/eda/section2/eda254.htm (1 of 2) [5/7/2002 3:14:13 PM]

1.2.5.4. Consequences Related to Distributional Assumptions

Model 1. The model may be changing.

2. The single model estimate may be meaningless.
3. The default model
Y = constant + error
may be invalid.
4. If the default model is insufficient, information about a better
model may remain undetected.
5. A poor deterministic model may be fit.
6. Information about an improved model may go undetected.

Process 1. The process may be out-of-control.

2. The process may be unpredictable.
3. The process may be un-modelable.

http://www.itl.nist.gov/div898/handbook/eda/section2/eda254.htm (2 of 2) [5/7/2002 3:14:13 PM]

1.3. EDA Techniques

1. Exploratory Data Analysis

1.3. EDA Techniques

Summary After you have collected a set of data, how do you do an exploratory
data analysis? What techniques do you employ? What do the various
techniques focus on? What conclusions can you expect to reach?
This section provides answers to these kinds of questions via a gallery
of EDA techniques and a detailed description of each technique. The
techniques are divided into graphical and quantitative techniques. For
exploratory data analysis, the emphasis is primarily on the graphical
techniques.

Table of 1. Introduction
Contents for 2. Analysis Questions
Section 3
3. Graphical Techniques: Alphabetical
4. Graphical Techniques: By Problem Category
5. Quantitative Techniques: Alphabetical
6. Probability Distributions

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3.htm [5/7/2002 3:14:13 PM]

1.3.1. Introduction

1. Exploratory Data Analysis

1.3. EDA Techniques

1.3.1. Introduction
Graphical This section describes many techniques that are commonly used in
and exploratory and classical data analysis. This list is by no means meant
Quantitative to be exhaustive. Additional techniques (both graphical and
Techniques quantitative) are discussed in the other chapters. Specifically, the
product comparisons chapter has a much more detailed description of
many classical statistical techniques.
EDA emphasizes graphical techniques while classical techniques
emphasize quantitative techniques. In practice, an analyst typically
uses a mixture of graphical and quantitative techniques. In this section,
we have divided the descriptions into graphical and quantitative
techniques. This is for organizational clarity and is not meant to
discourage the use of both graphical and quantitiative techniques when
analyzing data.

Use of This section emphasizes the techniques themselves; how the graph or
Techniques test is defined, published references, and sample output. The use of the
Shown in techniques to answer engineering questions is demonstrated in the case
Case Studies studies section. The case studies do not demonstrate all of the
techniques.

Availability The sample plots and output in this section were generated with the
in Software Dataplot software program. Other general purpose statistical data
analysis programs can generate most of the plots, intervals, and tests
discussed here, or macros can be written to acheive the same result.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda31.htm [5/7/2002 3:14:13 PM]

1.3.2. Analysis Questions

1. Exploratory Data Analysis

1.3. EDA Techniques

1.3.2. Analysis Questions

EDA Some common questions that exploratory data analysis is used to
Questions answer are:
1. What is a typical value?
2. What is the uncertainty for a typical value?
3. What is a good distributional fit for a set of numbers?
4. What is a percentile?
5. Does an engineering modification have an effect?
6. Does a factor have an effect?
7. What are the most important factors?
8. Are measurements coming from different laboratories equivalent?
9. What is the best function for relating a response variable to a set
of factor variables?
10. What are the best settings for factors?
11. Can we separate signal from noise in time dependent data?
12. Can we extract any structure from multivariate data?
13. Does the data have outliers?

Analyst A critical early step in any analysis is to identify (for the engineering
Should problem at hand) which of the above questions are relevant. That is, we
Identify need to identify which questions we want answered and which questions
Relevant have no bearing on the problem at hand. After collecting such a set of
Questions questions, an equally important step, which is invaluable for maintaining
for his focus, is to prioritize those questions in decreasing order of importance.
Engineering EDA techniques are tied in with each of the questions. There are some
Problem EDA techniques (e.g., the scatter plot) that are broad-brushed and apply
almost universally. On the other hand, there are a large number of EDA
techniques that are specific and whose specificity is tied in with one of
the above questions. Clearly if one chooses not to explicitly identify
relevant questions, then one cannot take advantage of these
question-specific EDA technqiues.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda32.htm (1 of 2) [5/7/2002 3:14:13 PM]

1.3.2. Analysis Questions

EDA Most of these questions can be addressed by techniques discussed in this

Approach chapter. The process modeling and process improvement chapters also
Emphasizes address many of the questions above. These questions are also relevant
Graphics for the classical approach to statistics. What distinguishes the EDA
approach is an emphasis on graphical techniques to gain insight as
opposed to the classical approach of quantitative tests. Most data
analysts will use a mix of graphical and classical quantitative techniques
to address these problems.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda32.htm (2 of 2) [5/7/2002 3:14:13 PM]

1.3.3. Graphical Techniques: Alphabetic

1. Exploratory Data Analysis

1.3. EDA Techniques

1.3.3. Graphical Techniques: Alphabetic

This section provides a gallery of some useful graphical techniques. The
techniques are ordered alphabetically, so this section is not intended to
be read in a sequential fashion. The use of most of these graphical
techniques is demonstrated in the case studies in this chapter. A few of
these graphical techniques are demonstrated in later chapters.

Autocorrelation Bihistogram: Block Plot: 1.3.3.3 Bootstrap Plot:

Plot: 1.3.3.1 1.3.3.2 1.3.3.4

Box-Cox Linearity Box-Cox Box Plot: 1.3.3.7 Complex

Plot: 1.3.3.5 Normality Plot: Demodulation
1.3.3.6 Amplitude Plot:
1.3.3.8

Complex Contour Plot: DEX Scatter Plot: DEX Mean Plot:

Demodulation 1.3.3.10 1.3.3.11 1.3.3.12
Phase Plot: 1.3.3.9

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33.htm (1 of 3) [5/7/2002 3:14:14 PM]

1.3.3. Graphical Techniques: Alphabetic

DEX Standard Histogram: Lag Plot: 1.3.3.15 Linear Correlation

Deviation Plot: 1.3.3.14 Plot: 1.3.3.16
1.3.3.13

Linear Intercept Linear Slope Plot: Linear Residual Mean Plot: 1.3.3.20
Plot: 1.3.3.17 1.3.3.18 Standard Deviation
Plot: 1.3.3.19

Normal Probability Probability Plot: Probability Plot Quantile-Quantile

Plot: 1.3.3.21 1.3.3.22 Correlation Plot: 1.3.3.24
Coefficient Plot:
1.3.3.23

Run Sequence Scatter Plot: Spectrum: 1.3.3.27 Standard Deviation

Plot: 1.3.3.25 1.3.3.26 Plot: 1.3.3.28

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33.htm (2 of 3) [5/7/2002 3:14:14 PM]

1.3.3. Graphical Techniques: Alphabetic

Star Plot: 1.3.3.29 Weibull Plot: Youden Plot: 4-Plot: 1.3.3.32

1.3.3.30 1.3.3.31

6-Plot: 1.3.3.33

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33.htm (3 of 3) [5/7/2002 3:14:14 PM]

1.3.3.1. Autocorrelation Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.1. Autocorrelation Plot

Purpose: Autocorrelation plots (Box and Jenkins, pp. 28-32) are a
Check commonly-used tool for checking randomness in a data set. This
Randomness randomness is ascertained by computing autocorrelations for data
values at varying time lags. If random, such autocorrelations should
be near zero for any and all time-lag separations. If non-random,
then one or more of the autocorrelations will be significantly
non-zero.
In addition, autocorrelation plots are used in the model identification
stage for Box-Jenkins autoregressive, moving average time series
models.

Sample Plot:
Autocorrelations
should be
near-zero for
randomness.
Such is not the
case in this
example and
thus the
randomness
assumption fails

This sample autocorrelation plot shows that the time series is not
random, but rather has a high degree of autocorrelation between
adjacent and near-adjacent observations.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda331.htm (1 of 4) [5/7/2002 3:14:15 PM]

1.3.3.1. Autocorrelation Plot

Definition: Autocorrelation plots are formed by

r(h) versus h ● Vertical axis: Autocorrelation coefficient

where Ch is the autocovariance function

and C0 is the variance function

Note--Rh is between -1 and +1.

● Horizontal axis: Time lag h (h = 1, 2, 3, ...)

Questions The autocorrelation plot can provide answers to the following

questions:
1. Are the data random?
2. Is an observation related to an adjacent observation?
3. Is an observation related to an observation twice-removed?
(etc.)
4. Is the observed time series white noise?
5. Is the observed time series sinusoidal?
6. Is the observed time series autoregressive?
7. What is an appropriate model for the observed time series?
8. Is the model
Y = constant + error
valid and sufficient?
9. Is the formula valid?

http://www.itl.nist.gov/div898/handbook/eda/section3/eda331.htm (2 of 4) [5/7/2002 3:14:15 PM]

1.3.3.1. Autocorrelation Plot

Importance: Randomness (along with fixed model, fixed variation, and fixed
Ensure validity distribution) is one of the four assumptions that typically underlie all
of engineering measurement processes. The randomness assumption is critically
conclusions important for the following three reasons:
1. Most standard statistical tests depend on randomness. The
validity of the test conclusions is directly linked to the
validity of the randomness assumption.
2. Many commonly-used statistical formulae depend on the
randomness assumption, the most common formula being the
formula for determining the standard deviation of the sample
mean:

where is the standard deviation of the data. Although

heavily used, the results from using this formula are of no
value unless the randomness assumption holds.
3. For univariate data, the default model is
Y = constant + error
If the data are not random, this model is incorrect and invalid,
and the estimates for the parameters (such as the constant)
become nonsensical and invalid.
In short, if the analyst does not check for randomness, then the
validity of many of the statistical conclusions becomes suspect. The
autocorrelation plot is an excellent way of checking for such
randomness.

Examples Examples of the autocorrelation plot for several common situations

are given in the following pages.
1. Random (= White Noise)
2. Weak autocorrelation
3. Strong autocorrelation and autoregressive model
4. Sinusoidal model

Related Partial Autocorrelation Plot

Techniques Lag Plot
Spectral Plot
Seasonal Subseries Plot

Case Study The autocorrelation plot is demonstrated in the beam deflection data
case study.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda331.htm (3 of 4) [5/7/2002 3:14:15 PM]

1.3.3.1. Autocorrelation Plot

Software Autocorrelation plots are available in most general purpose

statistical software programs including Dataplot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda331.htm (4 of 4) [5/7/2002 3:14:15 PM]

1.3.3.1.1. Autocorrelation Plot: Random Data

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.1. Autocorrelation Plot

1.3.3.1.1. Autocorrelation Plot: Random

Data
Autocorrelation The following is a sample autocorrelation plot.
Plot

Conclusions We can make the following conclusions from this plot.

1. There are no significant autocorrelations.
2. The data are random.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3311.htm (1 of 2) [5/7/2002 3:14:15 PM]

1.3.3.1.1. Autocorrelation Plot: Random Data

Discussion Note that with the exception of lag 0, which is always 1 by

definition, almost all of the autocorrelations fall within the 95%
confidence limits. In addition, there is no apparent pattern (such as
the first twenty-five being positive and the second twenty-five being
negative). This is the abscence of a pattern we expect to see if the
data are in fact random.
A few lags slightly outside the 95% and 99% confidence limits do
not neccessarily indicate non-randomness. For a 95% confidence
interval, we might expect about one out of twenty lags to be
statistically significant due to random fluctuations.
There is no associative ability to infer from a current value Yi as to
what the next value Yi+1 will be. Such non-association is the essense
of randomness. In short, adjacent observations do not "co-relate", so
we call this the "no autocorrelation" case.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3311.htm (2 of 2) [5/7/2002 3:14:15 PM]

1.3.3.1.2. Autocorrelation Plot: Moderate Autocorrelation

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.1. Autocorrelation Plot

1.3.3.1.2. Autocorrelation Plot: Moderate

Autocorrelation
Autocorrelation The following is a sample autocorrelation plot.
Plot

Conclusions We can make the following conclusions from this plot.

1. The data come from an underlying autoregressive model with
moderate positive autocorrelation.

Discussion The plot starts with a moderately high autocorrelation at lag 1

(approximately 0.75) that gradually decreases. The decreasing
autocorrelation is generally linear, but with significant noise. Such a
pattern is the autocorrelation plot signature of "moderate
autocorrelation", which in turn provides moderate predictability if
modeled properly.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3312.htm (1 of 2) [5/7/2002 3:14:15 PM]

1.3.3.1.2. Autocorrelation Plot: Moderate Autocorrelation

Recommended The next step would be to estimate the parameters for the
Next Step autoregressive model:

Such estimation can be performed by using least squares linear

regression or by fitting a Box-Jenkins autoregressive (AR) model.

The randomness assumption for least squares fitting applies to the

residuals of the model. That is, even though the original data exhibit
randomness, the residuals after fitting Yi against Yi-1 should result in
random residuals. Assessing whether or not the proposed model in
fact sufficiently removed the randomness is discussed in detail in the
Process Modeling chapter.

The residual standard deviation for this autoregressive model will be

much smaller than the residual standard deviation for the default
model

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3312.htm (2 of 2) [5/7/2002 3:14:15 PM]

1.3.3.1.3. Autocorrelation Plot: Strong Autocorrelation and Autoregressive Model

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.1. Autocorrelation Plot

1.3.3.1.3. Autocorrelation Plot: Strong

Autocorrelation and
Autoregressive Model
Autocorrelation The following is a sample autocorrelation plot.
Plot for Strong
Autocorrelation

Conclusions We can make the following conclusions from the above plot.
1. The data come from an underlying autoregressive model with
strong positive autocorrelation.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3313.htm (1 of 2) [5/7/2002 3:14:16 PM]

1.3.3.1.3. Autocorrelation Plot: Strong Autocorrelation and Autoregressive Model

Discussion The plot starts with a high autocorrelation at lag 1 (only slightly less
than 1) that slowly declines. It continues decreasing until it becomes
negative and starts showing an incresing negative autocorrelation.
The decreasing autocorrelation is generally linear with little noise.
Such a pattern is the autocorrelation plot signature of "strong
autocorrelation", which in turn provides high predictability if
modeled properly.

Recommended The next step would be to estimate the parameters for the
Next Step autoregressive model:

Such estimation can be performed by using least squares linear

regression or by fitting a Box-Jenkins autoregressive (AR) model.

The randomness assumption for least squares fitting applies to the

residuals of the model. That is, even though the original data exhibit
randomness, the residuals after fitting Yi against Yi-1 should result in
random residuals. Assessing whether or not the proposed model in
fact sufficiently removed the randomness is discussed in detail in the
Process Modeling chapter.

The residual standard deviation for this autoregressive model will be

much smaller than the residual standard deviation for the default
model

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3313.htm (2 of 2) [5/7/2002 3:14:16 PM]

1.3.3.1.4. Autocorrelation Plot: Sinusoidal Model

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.1. Autocorrelation Plot

1.3.3.1.4. Autocorrelation Plot: Sinusoidal

Model
Autocorrelation The following is a sample autocorrelation plot.
Plot for
Sinusoidal
Model

Conclusions We can make the following conclusions from the above plot.
1. The data come from an underlying sinusoidal model.

Discussion The plot exhibits an alternating sequence of positive and negative

spikes. These spikes are not decaying to zero. Such a pattern is the
autocorrelation plot signature of a sinusoidal model.

Recommended The beam deflection case study gives an example of modeling a

Next Step sinusoidal model.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3314.htm (1 of 2) [5/7/2002 3:14:16 PM]

1.3.3.1.4. Autocorrelation Plot: Sinusoidal Model

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3314.htm (2 of 2) [5/7/2002 3:14:16 PM]

1.3.3.2. Bihistogram

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.2. Bihistogram
Purpose: The bihistogram is an EDA tool for assessing whether a
Check for a before-versus-after engineering modification has caused a change in
change in ● location;
location,
● variation; or
variation, or
distribution ● distribution.

It is a graphical alternative to the two-sample t-test. The bihistogram

can be more powerful than the t-test in that all of the distributional
features (location, scale, skewness, outliers) are evident on a single plot.
It is also based on the common and well-understood histogram.

Sample Plot:
This
bihistogram
reveals that
there is a
significant
difference in
ceramic
breaking
strength
between
batch 1
(above) and
batch 2
(below)

From the above bihistogram, we can see that batch 1 is centered at a

ceramic strength value of approximately 725 while batch 2 is centered
at a ceramic strength value of approximately 625. That indicates that
these batches are displaced by about 100 strength units. Thus the batch

http://www.itl.nist.gov/div898/handbook/eda/section3/eda332.htm (1 of 3) [5/7/2002 3:14:16 PM]

1.3.3.2. Bihistogram

factor has a significant effect on the location (typical value) for strength
and hence batch is said to be "significant" or to "have an effect". We
thus see graphically and convincingly what a t-test or analysis of
variance would indicate quantitatively.

With respect to variation, note that the spread (variation) of the

above-axis batch 1 histogram does not appear to be that much different
from the below-axis batch 2 histogram. With respect to distributional
shape, note that the batch 1 histogram is skewed left while the batch 2
histogram is more symmetric with even a hint of a slight skewness to
the right.
Thus the bihistogram reveals that there is a clear difference between the
batches with respect to location and distribution, but not in regard to
variation. Comparing batch 1 and batch 2, we also note that batch 1 is
the "better batch" due to its 100-unit higher average strength (around
725).

Definition: Bihistograms are formed by vertically juxtaposing two histograms:

Two ● Above the axis: Histogram of the response variable for condition
adjoined 1
histograms
● Below the axis: Histogram of the response variable for condition
2

Questions The bihistogram can provide answers to the following questions:

1. Is a (2-level) factor significant?
2. Does a (2-level) factor have an effect?
3. Does the location change between the 2 subgroups?
4. Does the variation change between the 2 subgroups?
5. Does the distributional shape change between subgroups?
6. Are there any outliers?

Importance: The bihistogram is an important EDA tool for determining if a factor

Checks 3 out "has an effect". Since the bihistogram provides insight into the validity
of the 4 of three (location, variation, and distribution) out of the four (missing
underlying only randomness) underlying assumptions in a measurement process, it
assumptions is an especially valuable tool. Because of the dual (above/below) nature
of a of the plot, the bihistogram is restricted to assessing factors that have
measurement only two levels. However, this is very common in the
process before-versus-after character of many scientific and engineering
experiments.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda332.htm (2 of 3) [5/7/2002 3:14:16 PM]

1.3.3.2. Bihistogram

Related t test (for shift in location)

Techniques F test (for shift in variation)
Kolmogorov-Smirnov test (for shift in distribution)
Quantile-quantile plot (for shift in location and distribution)

Case Study The bihistogram is demonstrated in the ceramic strength data case
study.

Software The bihistogram is not widely available in general purpose statistical

software programs. Bihistograms can be generated using Dataplot

http://www.itl.nist.gov/div898/handbook/eda/section3/eda332.htm (3 of 3) [5/7/2002 3:14:16 PM]

1.3.3.3. Block Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.3. Block Plot

Purpose: The block plot (Filliben 1993) is an EDA tool for assessing whether the
Check to factor of interest (the primary factor) has a statistically significant effect
determine if on the response, and whether that conclusion about the primary factor
a factor of effect is valid robustly over all other nuisance or secondary factors in
interest has the experiment.
an effect
robust over It replaces the analysis of variance test with a less
all other assumption-dependent binomial test and should be routinely used
factors whenever we are trying to robustly decide whether a primary factor has
an effect.

Sample
Plot:
Weld
method 2 is
lower
(better) than
weld method
1 in 10 of 12
cases

This block plot reveals that in 10 of the 12 cases (bars), weld method 2
is lower (better) than weld method 1. From a binomial point of view,
weld method is statistically significant.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda333.htm (1 of 4) [5/7/2002 3:14:16 PM]

1.3.3.3. Block Plot

Definition Block Plots are formed as follows:

● Vertical axis: Response variable Y

● Horizontal axis: All combinations of all levels of all nuisance

(secondary) factors X1, X2, ...
● Plot Character: Levels of the primary factor XP

Discussion: Average number of defective lead wires per hour from a study with four
Primary factors,
factor is 1. weld strength (2 levels)
denoted by
2. plant (2 levels)
plot
character: 3. speed (2 levels)
within-bar 4. shift (3 levels)
plot are shown in the plot above. Weld strength is the primary factor and the
character. other three factors are nuisance factors. The 12 distinct positions along
the horizontal axis correspond to all possible combinations of the three
nuisance factors, i.e., 12 = 2 plants x 2 speeds x 3 shifts. These 12
conditions provide the framework for assessing whether any conclusions
about the 2 levels of the primary factor (weld method) can truly be
called "general conclusions". If we find that one weld method setting
does better (smaller average defects per hour) than the other weld
method setting for all or most of these 12 nuisance factor combinations,
then the conclusion is in fact general and robust.

Ordering In the above chart, the ordering along the horizontal axis is as follows:
along the ● The left 6 bars are from plant 1 and the right 6 bars are from plant
horizontal 2.
axis
● The first 3 bars are from speed 1, the next 3 bars are from speed
2, the next 3 bars are from speed 1, and the last 3 bars are from
speed 2.
● Bars 1, 4, 7, and 10 are from the first shift, bars 2, 5, 8, and 11 are
from the second shift, and bars 3, 6, 9, and 12 are from the third
shift.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda333.htm (2 of 4) [5/7/2002 3:14:16 PM]

1.3.3.3. Block Plot

Setting 2 is In the block plot for the first bar (plant 1, speed 1, shift 1), weld method
better than 1 yields about 28 defects per hour while weld method 2 yields about 22
setting 1 in defects per hour--hence the difference for this combination is about 6
10 out of 12 defects per hour and weld method 2 is seen to be better (smaller number
cases of defects per hour).
Is "weld method 2 is better than weld method 1" a general conclusion?
For the second bar (plant 1, speed 1, shift 2), weld method 1 is about 37
while weld method 2 is only about 18. Thus weld method 2 is again seen
to be better than weld method 1. Similarly for bar 3 (plant 1, speed 1,
shift 3), we see weld method 2 is smaller than weld method 1. Scanning
over all of the 12 bars, we see that weld method 2 is smaller than weld
method 1 in 10 of the 12 cases, which is highly suggestive of a robust
weld method effect.

An event What is the chance of 10 out of 12 happening by chance? This is

with chance probabilistically equivalent to testing whether a coin is fair by flipping it
probability and getting 10 heads in 12 tosses. The chance (from the binomial
of only 2% distribution) of getting 10 (or more extreme: 11, 12) heads in 12 flips of
a fair coin is about 2%. Such low-probability events are usually rejected
as untenable and in practice we would conclude that there is a difference
in weld methods.

Advantage: The advantages of the block plot are as follows:

Graphical ● A quantitative procedure (analysis of variance) is replaced by a
and graphical procedure.
binomial
● An F-test (analysis of variance) is replaced with a binomial test,
which requires fewer assumptions.

Questions The block plot can provide answers to the following questions:
1. Is the factor of interest significant?
2. Does the factor of interest have an effect?
3. Does the location change between levels of the primary factor?
4. Has the process improved?
5. What is the best setting (= level) of the primary factor?
6. How much of an average improvement can we expect with this
best setting of the primary factor?
7. Is there an interaction between the primary factor and one or more
nuisance factors?
8. Does the effect of the primary factor change depending on the
setting of some nuisance factor?

http://www.itl.nist.gov/div898/handbook/eda/section3/eda333.htm (3 of 4) [5/7/2002 3:14:16 PM]

1.3.3.3. Block Plot

9. Are there any outliers?

Importance: The block plot is a graphical technique that pointedly focuses on

Robustly whether or not the primary factor conclusions are in fact robustly
checks the general. This question is fundamentally different from the generic
significance multi-factor experiment question where the analyst asks, "What factors
of the factor are important and what factors are not" (a screening problem)? Global
of interest data analysis techniques, such as analysis of variance, can potentially be
improved by local, focused data analysis techniques that take advantage
of this difference.

Related t test (for shift in location for exactly 2 levels)

Techniques ANOVA (for shift in location for 2 or more levels)
Bihistogram (for shift in location, variation, and distribution for exactly
2 levels).

Case Study The block plot is demonstrated in the ceramic strength data case study.

Software Block plots can be generated with the Dataplot software program. They
are not currently available in other statistical software programs.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda333.htm (4 of 4) [5/7/2002 3:14:16 PM]

1.3.3.4. Bootstrap Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.4. Bootstrap Plot

Purpose: The bootstrap (Efron and Gong) plot is used to estimate the uncertainty
Estimate of a statistic.
uncertainty

Generate To generate a bootstrap uncertainty estimate for a given statistic from a

subsamples set of data, a subsample of a size less than or equal to the size of the data
with set is generated from the data, and the statistic is calculated. This
replacement subsample is generated with replacement so that any data point can be
sampled multiple times or not sampled at all. This process is repeated
for many subsamples, typically between 500 and 1000. The computed
values for the statistic form an estimate of the sampling distribution of
the statistic.
For example, to estimate the uncertainty of the median from a dataset
with 50 elements, we generate a subsample of 50 elements and calculate
the median. This is repeated at least 500 times so that we have at least
500 values for the median. Although the number of bootstrap samples to
use is somewhat arbitrary, 500 subsamples is usually sufficient. To
calculate a 90% confidence interval for the median, the sample medians
are sorted into ascending order and the value of the 25th median
(assuming exactly 500 subsamples were taken) is the lower confidence
limit while the value of the 475th median (assuming exactly 500
subsamples were taken) is the upper confidence limit.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda334.htm (1 of 3) [5/7/2002 3:14:17 PM]

1.3.3.4. Bootstrap Plot

Sample
Plot:

This bootstrap plot was generated from 500 uniform random numbers.
Bootstrap plots and corresponding histograms were generated for the
mean, median, and mid-range. The histograms for the corresponding
statistics clearly show that for uniform random numbers the mid-range
has the smallest variance and is, therefore, a superior location estimator
to the mean or the median.

Definition The bootstrap plot is formed by:

● Vertical axis: Computed value of the desired statistic for a given
subsample.
● Horizontal axis: Subsample number.

The bootstrap plot is simply the computed value of the statistic versus
the subsample number. That is, the bootstrap plot generates the values
for the desired statistic. This is usually immediately followed by a
histogram or some other distributional plot to show the location and
variation of the sampling distribution of the statistic.

Questions The bootstrap plot is used to answer the following questions:

● What does the sampling distribution for the statistic look like?

● What is a 95% confidence interval for the statistic?

● Which statistic has a sampling distribution with the smallest

variance? That is, which statistic generates the narrowest
confidence interval?

http://www.itl.nist.gov/div898/handbook/eda/section3/eda334.htm (2 of 3) [5/7/2002 3:14:17 PM]

1.3.3.4. Bootstrap Plot

Importance The most common uncertainty calculation is generating a confidence

interval for the mean. In this case, the uncertainty formula can be
derived mathematically. However, there are many situations in which
the uncertainty formulas are mathematically intractable. The bootstrap
provides a method for calculating the uncertainty in these cases.

Cautuion on The bootstrap is not appropriate for all distributions and statistics (Efron
use of the and Tibrashani). For example, because of the shape of the uniform
bootstrap distribution, the bootstrap is not appropriate for estimating the
distribution of statistics that are heavily dependent on the tails, such as
the range.

Related Histogram
Techniques Jackknife
The jacknife is a technique that is closely related to the bootstrap. The
jackknife is beyond the scope of this handbook. See the Efron and Gong
article for a discussion of the jackknife.

Case Study The bootstrap plot is demonstrated in the uniform random numbers case
study.

Software The bootstrap is becoming more common in general purpose statistical

software programs. However, it is still not supported in many of these
programs. Dataplot supports a bootstrap capability.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda334.htm (3 of 3) [5/7/2002 3:14:17 PM]

1.3.3.5. Box-Cox Linearity Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.5. Box-Cox Linearity Plot

Purpose: When performing a linear fit of Y against X, an appropriate
Find the transformation of X can often significantly improve the fit. The
transformation Box-Cox transformation (Box and Cox, 1964) is a particularly useful
of the X family of transformations. It is defined as:
variable that
maximizes the
correlation where X is the variable being transformed and is the transformation
between a Y parameter. For = 0, the natural log of the data is taken instead of
and an X using the above formula.
variable
The Box-Cox linearity plot is a plot of the correlation between Y and
the transformed X for given values of . That is, is the coordinate
for the horizontal axis variable and the value of the correlation
between Y and the transformed X is the coordinate for the vertical
axis of the plot. The value of corresponding to the maximum
correlation (or minimum for negative correlation) on the plot is then
the optimal choice for .
Transforming X is used to improve the fit. The Box-Cox
transformation applied to Y can be used as the basis for meeting the
error assumptions. That case is not covered here. See page 225 of
(Draper and Smith, 1981) or page 77 of (Ryan, 1997) for a discussion
of this case.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda335.htm (1 of 3) [5/7/2002 3:14:17 PM]

1.3.3.5. Box-Cox Linearity Plot

Sample Plot

The plot of the original data with the predicted values from a linear fit
indicate that a quadratic fit might be preferable. The Box-Cox
linearity plot shows a value of = 2.0. The plot of the transformed
data with the predicted values from a linear fit with the transformed
data shows a better fit (verified by the significant reduction in the
residual standard deviation).

Definition Box-Cox linearity plots are formed by

● Vertical axis: Correlation coefficient from the transformed X
and Y
● Horizontal axis: Value for

Questions The Box-Cox linearity plot can provide answers to the following
questions:
1. Would a suitable transformation improve my fit?
2. What is the optimal value of the transformation parameter?

Importance: Transformations can often significantly improve a fit. The Box-Cox

Find a linearity plot provides a convenient way to find a suitable
suitable transformation without engaging in a lot of trial and error fitting.
transformation

Related Linear Regression

Techniques Box-Cox Normality Plot

http://www.itl.nist.gov/div898/handbook/eda/section3/eda335.htm (2 of 3) [5/7/2002 3:14:17 PM]

1.3.3.5. Box-Cox Linearity Plot

Case Study The Box-Cox linearity plot is demonstrated in the Alaska pipeline
data case study.

Software Box-Cox linearity plots are not a standard part of most general
purpose statistical software programs. However, the underlying
technique is based on a transformation and computing a correlation
coefficient. So if a statistical program supports these capabilities,
writing a macro for a Box-Cox linearity plot should be feasible.
Dataplot supports a Box-Cox linearity plot directly.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda335.htm (3 of 3) [5/7/2002 3:14:17 PM]

1.3.3.6. Box-Cox Normality Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.6. Box-Cox Normality Plot

Purpose: Many statistical tests and intervals are based on the assumption of
Find normality. The assumption of normality often leads to tests that are
transformation simple, mathematically tractable, and powerful compared to tests that
to normalize do not make the normality assumption. Unfortunately, many real data
data sets are in fact not approximately normal. However, an appropriate
transformation of a data set can often yield a data set that does follow
approximately a normal distribution. This increases the applicability
and usefulness of statistical techniques based on the normality
assumption.
The Box-Cox transformation is a particulary useful family of
transformations. It is defined as:

where Y is the response variable and is the transformation

parameter. For = 0, the natural log of the data is taken instead of
using the above formula.
Given a particular transformation such as the Box-Cox transformation
defined above, it is helpful to define a measure of the normality of the
resulting transformation. One measure is to compute the correlation
coefficient of a normal probability plot. The correlation is computed
between the vertical and horizontal axis variables of the probability
plot and is a convenient measure of the linearity of the probability plot
(the more linear the probability plot, the better a normal distribution
fits the data).
The Box-Cox normality plot is a plot of these correlation coefficients
for various values of the parameter. The value of corresponding
to the maximum correlation on the plot is then the optimal choice for
.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda336.htm (1 of 3) [5/7/2002 3:14:17 PM]

1.3.3.6. Box-Cox Normality Plot

Sample Plot

The histogram in the upper left-hand corner shows a data set that has
significant right skewness (and so does not follow a normal
distribution). The Box-Cox normality plot shows that the maximum
value of the correlation coefficient is at = -0.3. The histogram of the
data after applying the Box-Cox transformation with = -0.3 shows a
data set for which the normality assumption is reasonable. This is
verified with a normal probability plot of the transformed data.

Definition Box-Cox normality plots are formed by:

● Vertical axis: Correlation coefficient from the normal
probability plot after applying Box-Cox transformation
● Horizontal axis: Value for

Questions The Box-Cox normality plot can provide answers to the following
questions:
1. Is there a transformation that will normalize my data?
2. What is the optimal value of the transformation parameter?

Importance: Normality assumptions are critical for many univariate intervals and
Normalization hypothesis tests. It is important to test the normality assumption. If the
Improves data are in fact clearly not normal, the Box-Cox normality plot can
Validity of often be used to find a transformation that will approximately
Tests normalize the data.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda336.htm (2 of 3) [5/7/2002 3:14:17 PM]

1.3.3.6. Box-Cox Normality Plot

Related Normal Probability Plot

Techniques Box-Cox Linearity Plot

Software Box-Cox normality plots are not a standard part of most general
purpose statistical software programs. However, the underlying
technique is based on a normal probability plot and computing a
correlation coefficient. So if a statistical program supports these
capabilities, writing a macro for a Box-Cox normality plot should be
feasible. Dataplot supports a Box-Cox normality plot directly.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda336.htm (3 of 3) [5/7/2002 3:14:17 PM]

1.3.3.7. Box Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.7. Box Plot

Purpose: Box plots (Chambers 1983) are an excellent tool for conveying location
Check and variation information in data sets, particularly for detecting and
location and illustrating location and variation changes between different groups of
variation data.
shifts

Sample
Plot:
This box
plot reveals
that
machine has
a significant
effect on
energy with
respect to
location and
possibly
variation

This box plot, comparing four machines for energy output, shows that
machine has a significant effect on energy with respect to both location
and variation. Machine 3 has the highest energy response (about 72.5);
machine 4 has the least variable energy response with about 50% of its
readings being within 1 energy unit.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda337.htm (1 of 3) [5/7/2002 3:14:18 PM]

1.3.3.7. Box Plot

Definition Box plots are formed by

Vertical axis: Response variable
Horizontal axis: The factor of interest
More specifically, we
1. Calculate the median and the quartiles (the lower quartile is the
25th percentile and the upper quartile is the 75th percentile).
2. Plot a symbol at the median (or draw a line) and draw a box
(hence the name--box plot) between the lower and upper
quartiles; this box represents the middle 50% of the data--the
"body" of the data.
3. Draw a line from the lower quartile to the minimum point and
another line from the upper quartile to the maximum point.
Typically a symbol is drawn at these minimum and maximum
points, although this is optional.
Thus the box plot identifies the middle 50% of the data, the median, and
the extreme points.

Single or A single box plot can be drawn for one batch of data with no distinct
multiple box groups. Alternatively, multiple box plots can be drawn together to
plots can be compare multiple data sets or to compare groups in a single data set. For
drawn a single box plot, the width of the box is arbitrary. For multiple box
plots, the width of the box plot can be set proportional to the number of
points in the given group or sample (some software implementations of
the box plot simply set all the boxes to the same width).

Box plots There is a useful variation of the box plot that more specifically
with fences identifies outliers. To create this variation:
1. Calculate the median and the lower and upper quartiles.
2. Plot a symbol at the median and draw a box between the lower
and upper quartiles.
3. Calculate the interquartile range (the difference between the upper
and lower quartile) and call it IQ.
4. Calculate the following points:
L1 = lower quartile - 1.5*IQ
L2 = lower quartile - 3.0*IQ
U1 = upper quartile + 1.5*IQ
U2 = upper quartile + 3.0*IQ
5. The line from the lower quartile to the minimum is now drawn
from the lower quartile to the smallest point that is greater than
L1. Likewise, the line from the upper quartile to the maximum is
now drawn to the largest point smaller than U1.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda337.htm (2 of 3) [5/7/2002 3:14:18 PM]

1.3.3.7. Box Plot

6. Points between L1 and L2 or between U1 and U2 are drawn as

small circles. Points less than L2 or greater than U2 are drawn as
large circles.

Questions The box plot can provide answers to the following questions:
1. Is a factor significant?
2. Does the location differ between subgroups?
3. Does the variation differ between subgroups?
4. Are there any outliers?

Importance: The box plot is an important EDA tool for determining if a factor has a
Check the significant effect on the response with respect to either location or
significance variation.
of a factor
The box plot is also an effective tool for summarizing large quantities of
information.

Related Mean Plot

Techniques Analysis of Variance

Case Study The box plot is demonstrated in the ceramic strength data case study.

Software Box plots are available in most general purpose statistical software
programs, including Dataplot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda337.htm (3 of 3) [5/7/2002 3:14:18 PM]

1.3.3.8. Complex Demodulation Amplitude Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.8. Complex Demodulation Amplitude

Plot
Purpose: In the frequency analysis of time series models, a common model is the
Detect sinusoidal model:
Changing
Amplitude in
Sinusoidal In this equation, is the amplitude, is the phase shift, and is the
Models dominant frequency. In the above model, and are constant, that is
they do not vary with time, ti.

The complex demodulation amplitude plot (Granger, 1964) is used to

determine if the assumption of constant amplitude is justifiable. If the
slope of the complex demodulation amplitude plot is zero, then the
above model is typically replaced with the model:

where is some type of linear model fit with standard least squares.
The most common case is a linear fit, that is the model becomes

Quadratic models are sometimes used. Higher order models are

relatively rare.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda338.htm (1 of 3) [5/7/2002 3:14:18 PM]

1.3.3.8. Complex Demodulation Amplitude Plot

Sample
Plot:

This complex demodulation amplitude plot shows that:

● the amplitude is fixed at approximately 390;

● there is a start-up effect; and

● there is a change in amplitude at around x = 160 that should be

investigated for an outlier.

Definition: The complex demodulation amplitude plot is formed by:

● Vertical axis: Amplitude

● Horizontal axis: Time

The mathematical computations for determining the amplitude are

beyond the scope of the Handbook. Consult Granger (Granger, 1964)
for details.

Questions The complex demodulation amplitude plot answers the following

questions:
1. Does the amplitude change over time?
2. Are there any outliers that need to be investigated?
3. Is the amplitude different at the beginning of the series (i.e., is
there a start-up effect)?

http://www.itl.nist.gov/div898/handbook/eda/section3/eda338.htm (2 of 3) [5/7/2002 3:14:18 PM]

1.3.3.8. Complex Demodulation Amplitude Plot

Importance: As stated previously, in the frequency analysis of time series models, a

Assumption common model is the sinusoidal model:
Checking

In this equation, is assumed to be constant, that is it does not vary

with time. It is important to check whether or not this assumption is
reasonable.
The complex demodulation amplitude plot can be used to verify this
assumption. If the slope of this plot is essentially zero, then the
assumption of constant amplitude is justified. If it is not, should be
replaced with some type of time-varying model. The most common
cases are linear (B0 + B1*t) and quadratic (B0 + B1*t + B2*t2).

Related Spectral Plot

Techniques Complex Demodulation Phase Plot
Non-Linear Fitting

Case Study The complex demodulation amplitude plot is demonstrated in the beam
deflection data case study.

Software Complex demodulation amplitude plots are available in some, but not
most, general purpose statistical software programs. Dataplot supports
complex demodulation amplitude plots.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda338.htm (3 of 3) [5/7/2002 3:14:18 PM]

1.3.3.9. Complex Demodulation Phase Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.9. Complex Demodulation Phase Plot

Purpose: As stated previously, in the frequency analysis of time series models, a
Improve the common model is the sinusoidal model:
estimate of
frequency in
sinusoidal In this equation, is the amplitude, is the phase shift, and is the
time series dominant frequency. In the above model, and are constant, that is
models they do not vary with time ti.

The complex demodulation phase plot (Granger, 1964) is used to

improve the estimate of the frequency (i.e., ) in this model.
If the complex demodulation phase plot shows lines sloping from left to
right, then the estimate of the frequency should be increased. If it shows
lines sloping right to left, then the frequency should be decreased. If
there is essentially zero slope, then the frequency estimate does not need
to be modified.

Sample
Plot:

http://www.itl.nist.gov/div898/handbook/eda/section3/eda339.htm (1 of 3) [5/7/2002 3:14:19 PM]

1.3.3.9. Complex Demodulation Phase Plot

This complex demodulation phase plot shows that:

● the specified demodulation frequency is incorrect;

● the demodulation frequency should be increased.

Definition The complex demodulation phase plot is formed by:

● Vertical axis: Phase

● Horizontal axis: Time

The mathematical computations for the phase plot are beyond the scope
of the Handbook. Consult Granger (Granger, 1964) for details.

Questions The complex demodulation phase plot answers the following question:
Is the specified demodulation frequency correct?

Importance The non-linear fitting for the sinusoidal model:

of a Good
Initial
Estimate for is usually quite sensitive to the choice of good starting values. The
the initial estimate of the frequency, , is obtained from a spectral plot. The
Frequency complex demodulation phase plot is used to assess whether this estimate
is adequate, and if it is not, whether it should be increased or decreased.
Using the complex demodulation phase plot with the spectral plot can
significantly improve the quality of the non-linear fits obtained.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda339.htm (2 of 3) [5/7/2002 3:14:19 PM]

1.3.3.9. Complex Demodulation Phase Plot

Related Spectral Plot

Techniques Complex Demodulation Phase Plot
Non-Linear Fitting

Case Study The complex demodulation amplitude plot is demonstrated in the beam
deflection data case study.

Software Complex demodulation phase plots are available in some, but not most,
general purpose statistical software programs. Dataplot supports
complex demodulation phase plots.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda339.htm (3 of 3) [5/7/2002 3:14:19 PM]

1.3.3.10. Contour Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.10. Contour Plot

Purpose: A contour plot is a graphical technique for representing a
Display 3-d 3-dimensional surface by plotting constant z slices, called contours, on
surface on a 2-dimensional format. That is, given a value for z, lines are drawn for
2-d plot connecting the (x,y) coordinates where that z value occurs.
The contour plot is an alternative to a 3-D surface plot.

Sample Plot:

This contour plot shows that the surface is symmetric and peaks in the
center.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33a.htm (1 of 3) [5/7/2002 3:14:19 PM]

1.3.3.10. Contour Plot

Definition The contour plot is formed by:

● Vertical axis: Independent variable 2

● Horizontal axis: Independent variable 1

● Lines: iso-response values

The independent variables are usually restricted to a regular grid. The

actual techniques for determining the correct iso-response values are
rather complex and are almost always computer generated.
An additional variable may be required to specify the Z values for
drawing the iso-lines. Some software packages require explicit values.
Other software packages will determine them automatically.
If the data (or function) do not form a regular grid, you typically need
to perform a 2-D interpolation to form a regular grid.

Questions The contour plot is used to answer the question

How does Z change as a function of X and Y?

Importance: For univariate data, a run sequence plot and a histogram are considered
Visualizing necessary first steps in understanding the data. For 2-dimensional data,
3-dimensional a scatter plot is a necessary first step in understanding the data.
data
In a similar manner, 3-dimensional data should be plotted. Small data
sets, such as result from designed experiments, can typically be
represented by block plots, dex mean plots, and the like (here, "DEX"
stands for "Design of Experiments"). For large data sets, a contour plot
or a 3-D surface plot should be considered a necessary first step in
understanding the data.

DEX Contour The dex contour plot is a specialized contour plot used in the design of
Plot experiments. In particular, it is useful for full and fractional designs.

Related 3-D Plot

Techniques

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33a.htm (2 of 3) [5/7/2002 3:14:19 PM]

1.3.3.10. Contour Plot

Software Contour plots are available in most general purpose statistical software
programs. They are also available in many general purpose graphics
and mathematics programs. These programs vary widely in the
capabilities for the contour plots they generate. Many provide just a
basic contour plot over a rectangular grid while others permit color
filled or shaded contours. Dataplot supports a fairly basic contour plot.

Most statistical software programs that support design of experiments

will provide a dex contour plot capability.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33a.htm (3 of 3) [5/7/2002 3:14:19 PM]

1.3.3.10.1. DEX Contour Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.10. Contour Plot

1.3.3.10.1. DEX Contour Plot

DEX Contour The dex contour plot is a specialized contour plot used in the analysis of
Plot: full and fractional experimental designs. These designs often have a low
Introduction level, coded as "-1" or "-", and a high level, coded as "+1" or "+" for each
factor. In addition, there can optionally be one or more center points.
Center points are at the mid-point between the low and high level for each
factor and are coded as "0".
The dex contour plot is generated for two factors. Typically, this would be
the two most important factors as determined by previous analyses (e.g.,
through the use of the dex mean plots and a Yates analysis). If more than
two factors are important, you may want to generate a series of dex
contour plots, each of which is drawn for two of these factors. You can
also generate a matrix of all pairwise dex contour plots for a number of
important factors (similar to the scatter plot matrix for scatter plots).

The typical application of the dex contour plot is in determining settings

that will maximize (or minimize) the response variable. It can also be
helpful in determining settings that result in the response variable hitting a
pre-determined target value. The dex contour plot plays a useful role in
determining the settings for the next iteration of the experiment. That is,
the initial experiment is typically a fractional factorial design with a fairly
large number of factors. After the most important factors are determined,
the dex contour plot can be used to help define settings for a full factorial
or response surface design based on a smaller number of factors.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33a1.htm (1 of 4) [5/7/2002 3:14:20 PM]

1.3.3.10.1. DEX Contour Plot

Construction The following are the primary steps in the construction of the dex contour
of DEX plot.
Contour Plot 1. The x and y axes of the plot represent the values of the first and
second factor (independent) variables.
2. The four vertex points are drawn. The vertex points are (-1,-1),
(-1,1), (1,1), (1,-1). At each vertex point, the average of all the
response values at that vertex point is printed.
3. Similarly, if there are center points, a point is drawn at (0,0) and the
average of the response values at the center points is printed.
4. The linear dex contour plot assumes the model:

where is the overall mean of the response variable. The

values of , , , and are estimated from the vertex points

using a Yates analysis (the Yates analysis utilizes the special
structure of the 2-level full and fractional factorial designs to
simplify the computation of these parameter estimates). Note that
for the dex contour plot, a full Yates analysis does not need to
performed, simply the calculations for generating the parameter
estimates.
In order to generate a single contour line, we need a value for Y, say
Y0. Next, we solve for U2 in terms of U1 and, after doing the
algebra, we have the equation:

We generate a sequence of points for U1 in the range -2 to 2 and

compute the corresponding values of U2. These points constitute a
single contour line corresponding to Y = Y0.

The user specifies the target values for which contour lines will be
generated.
The above algorithm assumes a linear model for the design. Dex contour
plots can also be generated for the case in which we assume a quadratic
model for the design. The algebra for solving for U2 in terms of U1
becomes more complicated, but the fundamental idea is the same.
Quadratic models are needed for the case when the average for the center
points does not fall in the range defined by the vertex point (i.e., there is
curvature).

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33a1.htm (2 of 4) [5/7/2002 3:14:20 PM]

1.3.3.10.1. DEX Contour Plot

Sample DEX The following is a dex contour plot for the data used in the Eddy current
Contour Plot case study. The analysis in that case study demonstrated that X1 and X2
were the most important factors.

Interpretation From the above dex contour plot we can derive the following information.
of the Sample 1. Interaction significance;
DEX Contour
2. Best (data) setting for these 2 dominant factors;
Plot

Interaction Note the appearance of the contour plot. If the contour curves are linear,
Significance then that implies that the interaction term is not significant; if the contour
curves have considerable curvature, then that implies that the interaction
term is large and important. In our case, the contour curves do not have
considerable curvature, and so we conclude that the X1*X2 term is not
significant.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33a1.htm (3 of 4) [5/7/2002 3:14:20 PM]

1.3.3.10.1. DEX Contour Plot

Best Settings To determine the best factor settings for the already-run experiment, we
first must define what "best" means. For the Eddy current data set used to
generate this dex contour plot, "best" means to maximize (rather than
minimize or hit a target) the response. Hence from the contour plot we
determine the best settings for the two dominant factors by simply
scanning the four vertices and choosing the vertex with the largest value
(= average response). In this case, it is (X1 = +1, X2 = +1).
As for factor X3, the contour plot provides no best setting information, and
so we would resort to other tools: the main effects plot, the interaction
effects matrix, or the ordered data to determine optimal X3 settings.

Case Study The Eddy current case study demonstrates the use of the dex contour plot
in the context of the analysis of a full factorial design.

Software DEX contour plots are available in many statistical software programs that
analyze data from designed experiments. Dataplot supports a linear dex
contour plot and it provides a macro for generating a quadratic dex contour
plot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33a1.htm (4 of 4) [5/7/2002 3:14:20 PM]

1.3.3.11. DEX Scatter Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.11. DEX Scatter Plot

Purpose: The dex scatter plot shows the response values for each level of each
Determine factor (i.e., independent) variable. This graphically shows how the
Important location and scale vary for both within a factor variable and between
Factors with different factor variables. This graphically shows which are the
Respect to important factors and can help provide a ranked list of important
Location and factors from a designed experiment. The dex scatter plot is a
Scale complement to the traditional analyis of variance of designed
experiments.
Dex scatter plots are typically used in conjunction with the dex mean
plot and the dex standard deviation plot. The dex mean plot replaces
the raw response values with mean response values while the dex
standard deviation plot replaces the raw response values with the
standard deviation of the response values. There is value in generating
all 3 of these plots. The dex mean and standard deviation plots are
useful in that the summary measures of location and spread stand out
(they can sometimes get lost with the raw plot). However, the raw data
points can reveal subtleties, such as the presence of outliers, that might
get lost with the summary statistics.

Sample Plot:
Factors 4, 2,
3, and 7 are
the Important
Factors.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33b.htm (1 of 5) [5/7/2002 3:14:20 PM]

1.3.3.11. DEX Scatter Plot

Description For this sample plot, there are seven factors and each factor has two
of the Plot levels. For each factor, we define a distinct x coordinate for each level
of the factor. For example, for factor 1, level 1 is coded as 0.8 and level
2 is coded as 1.2. The y coordinate is simply the value of the response
variable. The solid horizontal line is drawn at the overall mean of the
response variable. The vertical dotted lines are added for clarity.
Although the plot can be drawn with an arbitrary number of levels for a
factor, it is really only useful when there are two or three levels for a
factor.

Conclusions This sample dex scatter plot shows that:

1. there does not appear to be any outliers;
2. the levels of factors 2 and 4 show distinct location differences;
and
3. the levels of factor 1 show distinct scale differences.

Definition: Dex scatter plots are formed by:

Response ● Vertical axis: Value of the response variable
Values
● Horizontal axis: Factor variable (with each level of the factor
Versus
coded with a slightly offset x coordinate)
Factor
Variables

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33b.htm (2 of 5) [5/7/2002 3:14:20 PM]

1.3.3.11. DEX Scatter Plot

Questions The dex scatter plot can be used to answer the following questions:
1. Which factors are important with respect to location and scale?
2. Are there outliers?

Importance: The goal of many designed experiments is to determine which factors

Identify are important with respect to location and scale. A ranked list of the
Important important factors is also often of interest. Dex scatter, mean, and
Factors with standard deviation plots show this graphically. The dex scatter plot
Respect to additionally shows if outliers may potentially be distorting the results.
Location and
Scale Dex scatter plots were designed primarily for analyzing designed
experiments. However, they are useful for any type of multi-factor data
(i.e., a response variable with 2 or more factor variables having a small
number of distinct levels) whether or not the data were generated from
a designed experiment.

Extension for
Interaction
Effects

Using the concept of the scatterplot matrix, the dex scatter plot can be
extended to display first order interaction effects.
Specifically, if there are k factors, we create a matrix of plots with k
rows and k columns. On the diagonal, the plot is simply a dex scatter
plot with a single factor. For the off-diagonal plots, we multiply the
values of Xi and Xj. For the common 2-level designs (i.e., each factor
has two levels) the values are typically coded as -1 and 1, so the
multiplied values are also -1 and 1. We then generate a dex scatter plot
for this interaction variable. This plot is called a dex interaction effects
plot and an example is shown below.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33b.htm (3 of 5) [5/7/2002 3:14:20 PM]

1.3.3.11. DEX Scatter Plot

Interpretation We can first examine the diagonal elements for the main effects. These
of the Dex diagonal plots show a great deal of overlap between the levels for all
Interaction three factors. This indicates that location and scale effects will be
Effects Plot relatively small.
We can then examine the off-diagonal plots for the first order
interaction effects. For example, the plot in the first row and second
column is the interaction between factors X1 and X2. As with the main
effect plots, no clear patterns are evident.

Related Dex mean plot

Techniques Dex standard deviation plot
Block plot
Box plot
Analysis of variance

Case Study The dex scatter plot is demonstrated in the ceramic strength data case
study.

Software Dex scatter plots are available in some general purpose statistical
software programs, although the format may vary somewhat between
these programs. They are essentially just scatter plots with the X
variable defined in a particular way, so it should be feasible to write
macros for dex scatter plots in most statistical software programs.
Dataplot supports a dex scatter plot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33b.htm (4 of 5) [5/7/2002 3:14:20 PM]

1.3.3.11. DEX Scatter Plot

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33b.htm (5 of 5) [5/7/2002 3:14:20 PM]

1.3.3.12. DEX Mean Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.12. DEX Mean Plot

Purpose: The dex mean plot is appropriate for analyzing data from a designed
Detect experiment, with respect to important factors, where the factors are at
Important two or more levels. The plot shows mean values for the two or more
Factors with levels of each factor plotted by factor. The means for a single factor are
Respect to connected by a straight line. The dex mean plot is a complement to the
Location traditional analysis of variance of designed experiments.

This plot is typically generated for the mean. However, it can be

generated for other location statistics such as the median.

Sample
Plot:
Factors 4, 2,
and 1 are
the Most
Important
Factors

This sample dex mean plot shows that:

1. factor 4 is the most important;
2. factor 2 is the second most important;
3. factor 1 is the third most important;

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33c.htm (1 of 3) [5/7/2002 3:14:20 PM]

1.3.3.12. DEX Mean Plot

4. factor 7 is the fourth most important;

5. factor 6 is the fifth most important;
6. factors 3 and 5 are relatively unimportant.
In summary, factors 4, 2, and 1 seem to be clearly important, factors 3
and 5 seem to be clearly unimportant, and factors 6 and 7 are borderline
factors whose inclusion in any subsequent models will be determined by
further analyses.

Definition: Dex mean plots are formed by:

Mean ● Vertical axis: Mean of the response variable for each level of the
Response factor
Versus
● Horizontal axis: Factor variable
Factor
Variables

Questions The dex mean plot can be used to answer the following questions:
1. Which factors are important? The dex mean plot does not provide
a definitive answer to this question, but it does help categorize
factors as "clearly important", "clearly not important", and
"borderline importance".
2. What is the ranking list of the important factors?

Importance: The goal of many designed experiments is to determine which factors

Determine are significant. A ranked order listing of the important factors is also
Significant often of interest. The dex mean plot is ideally suited for answering these
Factors types of questions and we recommend its routine use in analyzing
designed experiments.

Extension Using the concept of the scatter plot matrix, the dex mean plot can be
for extended to display first-order interaction effects.
Interaction
Effects Specifically, if there are k factors, we create a matrix of plots with k
rows and k columns. On the diagonal, the plot is simply a dex mean plot
with a single factor. For the off-diagonal plots, measurements at each
level of the interaction are plotted versus level, where level is Xi times
Xj and Xi is the code for the ith main effect level and Xj is the code for
the jth main effect. For the common 2-level designs (i.e., each factor has
two levels) the values are typically coded as -1 and 1, so the multiplied
values are also -1 and 1. We then generate a dex mean plot for this
interaction variable. This plot is called a dex interaction effects plot and
an example is shown below.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33c.htm (2 of 3) [5/7/2002 3:14:20 PM]

1.3.3.12. DEX Mean Plot

DEX
Interaction
Effects Plot

This plot shows that the most significant factor is X1 and the most
significant interaction is between X1 and X3.

Related Dex scatter plot

Techniques Dex standard deviation plot
Block plot
Box plot
Analysis of variance

Case Study The dex mean plot and the dex interaction effects plot are demonstrated
in the ceramic strength data case study.

Software Dex mean plots are available in some general purpose statistical
software programs, although the format may vary somewhat between
these programs. It may be feasible to write macros for dex mean plots in
some statistical software programs that do not support this plot directly.
Dataplot supports both a dex mean plot and a dex interaction effects
plot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33c.htm (3 of 3) [5/7/2002 3:14:20 PM]

1.3.3.13. DEX Standard Deviation Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.13. DEX Standard Deviation Plot

Purpose: The dex standard deviation plot is appropriate for analyzing data from a
Detect designed experiment, with respect to important factors, where the
Important factors are at two or more levels and there are repeated values at each
Factors with level. The plot shows standard deviation values for the two or more
Respect to levels of each factor plotted by factor. The standard deviations for a
Scale single factor are connected by a straight line. The dex standard deviation
plot is a complement to the traditional analysis of variance of designed
experiments.
This plot is typically generated for the standard deviation. However, it
can also be generated for other scale statistics such as the range, the
median absolute deviation, or the average absolute deviation.

Sample Plot

This sample dex standard deviation plot shows that:

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33d.htm (1 of 3) [5/7/2002 3:14:21 PM]

1.3.3.13. DEX Standard Deviation Plot

1. factor 1 has the greatest difference in standard deviations between

factor levels;
2. factor 4 has a significantly lower average standard deviation than
the average standard deviations of other factors (but the level 1
standard deviation for factor 1 is about the same as the level 1
standard deviation for factor 4);
3. for all factors, the level 1 standard deviation is smaller than the
level 2 standard deviation.

Definition: Dex standard deviation plots are formed by:

Response ● Vertical axis: Standard deviation of the response variable for each
Standard level of the factor
Deviations
● Horizontal axis: Factor variable
Versus
Factor
Variables

Questions The dex standard deviation plot can be used to answer the following
questions:
1. How do the standard deviations vary across factors?
2. How do the standard deviations vary within a factor?
3. Which are the most important factors with respect to scale?
4. What is the ranked list of the important factors with respect to
scale?

Importance: The goal with many designed experiments is to determine which factors
Assess are significant. This is usually determined from the means of the factor
Variability levels (which can be conveniently shown with a dex mean plot). A
secondary goal is to assess the variability of the responses both within a
factor and between factors. The dex standard deviation plot is a
convenient way to do this.

Related Dex scatter plot

Techniques Dex mean plot
Block plot
Box plot
Analysis of variance

Case Study The dex standard deviation plot is demonstrated in the ceramic strength
data case study.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33d.htm (2 of 3) [5/7/2002 3:14:21 PM]

1.3.3.13. DEX Standard Deviation Plot

Software Dex standard deviation plots are not available in most general purpose
statistical software programs. It may be feasible to write macros for dex
standard deviation plots in some statistical software programs that do
not support them directly. Dataplot supports a dex standard deviation
plot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33d.htm (3 of 3) [5/7/2002 3:14:21 PM]

1.3.3.14. Histogram

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.14. Histogram
Purpose: The purpose of a histogram (Chambers) is to graphically summarize the
Summarize distribution of a univariate data set.
a Univariate
Data Set The histogram graphically shows the following:
1. center (i.e., the location) of the data;
2. spread (i.e., the scale) of the data;
3. skewness of the data;
4. presence of outliers; and
5. presence of multiple modes in the data.
These features provide strong indications of the proper distributional
model for the data. The probability plot or a goodness-of-fit test can be
used to verify the distributional model.
The examples section shows the appearance of a number of common
features revealed by histograms.

Sample Plot

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e.htm (1 of 4) [5/7/2002 3:14:21 PM]

1.3.3.14. Histogram

Definition The most common form of the histogram is obtained by splitting the
range of the data into equal-sized bins (called classes). Then for each
bin, the number of points from the data set that fall into each bin are
counted. That is
● Vertical axis: Frequency (i.e., counts for each bin)

● Horizontal axis: Response variable

The classes can either be defined arbitrarily by the user or via some
systematic rule. A number of theoretically derived rules have been
proposed by Scott (Scott 1992).

The cumulative histogram is a variation of the histogram in which the

vertical axis gives not just the counts for a single bin, but rather gives
the counts for that bin plus all bins for smaller values of the response
variable.
Both the histogram and cumulative histogram have an additional variant
whereby the counts are replaced by the normalized counts. The names
for these variants are the relative histogram and the relative cumulative
histogram.
There are two common ways to normalize the counts.
1. The normalized count is the count in a class divided by the total
number of observations. In this case the relative counts are
normalized to sum to one (or 100 if a percentage scale is used).
This is the intuitive case where the height of the histogram bar
represents the proportion of the data in each class.
2. The normalized count is the count in the class divided by the

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e.htm (2 of 4) [5/7/2002 3:14:21 PM]

1.3.3.14. Histogram

number of observations times the class width. For this

normalization, the area (or integral) under the histogram is equal
to one. From a probabilistic point of view, this normalization
results in a relative histogram that is most akin to the probability
density function and a relative cumulative histogram that is most
akin to the cumulative distribution function. If you want to
overlay a probability density or cumulative distribution function
on top of the histogram, use this normalization. Although this
normalization is less intuitive (relative frequencies greater than 1
are quite permissible), it is the appropriate normalization if you
are using the histogram to model a probability density function.

Questions The histogram can be used to answer the following questions:

1. What kind of population distribution do the data come from?
2. Where are the data located?
3. How spread out are the data?
4. Are the data symmetric or skewed?
5. Are there outliers in the data?

Examples 1. Normal
2. Symmetric, Non-Normal, Short-Tailed
3. Symmetric, Non-Normal, Long-Tailed
4. Symmetric and Bimodal
5. Bimodal Mixture of 2 Normals
6. Skewed (Non-Symmetric) Right
7. Skewed (Non-Symmetric) Left
8. Symmetric with Outlier

Related Box plot

Techniques Probability plot

The techniques below are not discussed in the Handbook. However,

they are similar in purpose to the histogram. Additional information on
them is contained in the Chambers and Scott references.

Frequency Plot
Stem and Leaf Plot
Density Trace

Case Study The histogram is demonstrated in the heat flow meter data case study.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e.htm (3 of 4) [5/7/2002 3:14:21 PM]

1.3.3.14. Histogram

Software Histograms are available in most general purpose statistical software

programs. They are also supported in most general purpose charting,
spreadsheet, and business graphics programs. Dataplot supports
histograms.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e.htm (4 of 4) [5/7/2002 3:14:21 PM]

1.3.3.14.1. Histogram Interpretation: Normal

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.14. Histogram

1.3.3.14.1. Histogram Interpretation: Normal

Symmetric,
Moderate-
Tailed
Histogram

Note the classical bell-shaped, symmetric histogram with most of the

frequency counts bunched in the middle and with the counts dying off
out in the tails. From a physical science/engineering point of view, the
normal distribution is that distribution which occurs most often in
nature (due in part to the central limit theorem).

Recommended If the histogram indicates a symmetric, moderate tailed distribution,

Next Step then the recommended next step is to do a normal probability plot to
confirm approximate normality. If the normal probability plot is linear,
then the normal distribution is a good model for the data.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e1.htm (1 of 2) [5/7/2002 3:14:21 PM]

1.3.3.14.1. Histogram Interpretation: Normal

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e1.htm (2 of 2) [5/7/2002 3:14:21 PM]

1.3.3.14.2. Histogram Interpretation: Symmetric, Non-Normal, Short-Tailed

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.14. Histogram

1.3.3.14.2. Histogram Interpretation:

Symmetric, Non-Normal,
Short-Tailed
Symmetric,
Short-Tailed
Histogram

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e2.htm (1 of 3) [5/7/2002 3:14:22 PM]

1.3.3.14.2. Histogram Interpretation: Symmetric, Non-Normal, Short-Tailed

Description of For a symmetric distribution, the "body" of a distribution refers to the

What "center" of the distribution--commonly that region of the distribution
Short-Tailed where most of the probability resides--the "fat" part of the distribution.
Means The "tail" of a distribution refers to the extreme regions of the
distribution--both left and right. The "tail length" of a distribution is a
term that indicates how fast these extremes approach zero.
For a short-tailed distribution, the tails approach zero very fast. Such
distributions commonly have a truncated ("sawed-off") look. The
classical short-tailed distribution is the uniform (rectangular)
distribution in which the probability is constant over a given range and
then drops to zero everywhere else--we would speak of this as having
no tails, or extremely short tails.
For a moderate-tailed distribution, the tails decline to zero in a
moderate fashion. The classical moderate-tailed distribution is the
normal (Gaussian) distribution.
For a long-tailed distribution, the tails decline to zero very slowly--and
hence one is apt to see probability a long way from the body of the
distribution. The classical long-tailed distribution is the Cauchy
distribution.
In terms of tail length, the histogram shown above would be
characteristic of a "short-tailed" distribution.
The optimal (unbiased and most precise) estimator for location for the
center of a distribution is heavily dependent on the tail length of the
distribution. The common choice of taking N observations and using
the calculated sample mean as the best estimate for the center of the
distribution is a good choice for the normal distribution (moderate
tailed), a poor choice for the uniform distribution (short tailed), and a
horrible choice for the Cauchy distribution (long tailed). Although for
the normal distribution the sample mean is as precise an estimator as
we can get, for the uniform and Cauchy distributions, the sample mean
is not the best estimator.
For the uniform distribution, the midrange
midrange = (smallest + largest) / 2
is the best estimator of location. For a Cauchy distribution, the median
is the best estimator of location.

Recommended If the histogram indicates a symmetric, short-tailed distribution, the

Next Step recommended next step is to generate a uniform probability plot. If the
uniform probability plot is linear, then the uniform distribution is an
appropriate model for the data.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e2.htm (2 of 3) [5/7/2002 3:14:22 PM]

1.3.3.14.2. Histogram Interpretation: Symmetric, Non-Normal, Short-Tailed

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e2.htm (3 of 3) [5/7/2002 3:14:22 PM]

1.3.3.14.3. Histogram Interpretation: Symmetric, Non-Normal, Long-Tailed

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.14. Histogram

1.3.3.14.3. Histogram Interpretation:

Symmetric, Non-Normal,
Long-Tailed
Symmetric,
Long-Tailed
Histogram

Description of The previous example contains a discussion of the distinction between

Long-Tailed short-tailed, moderate-tailed, and long-tailed distributions.

In terms of tail length, the histogram shown above would be

characteristic of a "long-tailed" distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e3.htm (1 of 2) [5/7/2002 3:14:22 PM]

1.3.3.14.3. Histogram Interpretation: Symmetric, Non-Normal, Long-Tailed

Recommended If the histogram indicates a symmetric, long tailed distribution, the

Next Step recommended next step is to do a Cauchy probability plot. If the
Cauchy probability plot is linear, then the Cauchy distribution is an
appropriate model for the data. Alternatively, a Tukey Lambda PPCC
plot may provide insight into a suitable distributional model for the
data.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e3.htm (2 of 2) [5/7/2002 3:14:22 PM]

1.3.3.14.4. Histogram Interpretation: Symmetric and Bimodal

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.14. Histogram

1.3.3.14.4. Histogram Interpretation:

Symmetric and Bimodal
Symmetric,
Bimodal
Histogram

Description of The mode of a distribution is that value which is most frequently

Bimodal occurring or has the largest probability of occurrence. The sample
mode occurs at the peak of the histogram.
For many phenomena, it is quite common for the distribution of the
response values to cluster around a single mode (unimodal) and then
distribute themselves with lesser frequency out into the tails. The
normal distribution is the classic example of a unimodal distribution.
The histogram shown above illustrates data from a bimodal (2 peak)
distribution. The histogram serves as a tool for diagnosing problems
such as bimodality. Questioning the underlying reason for
distributional non-unimodality frequently leads to greater insight and

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e4.htm (1 of 2) [5/7/2002 3:14:22 PM]

1.3.3.14.4. Histogram Interpretation: Symmetric and Bimodal

improved deterministic modeling of the phenomenon under study. For

example, for the data presented above, the bimodal histogram is
caused by sinusoidality in the data.

Recommended If the histogram indicates a symmetric, bimodal distribution, the

Next Step recommended next steps are to:
1. Do a run sequence plot or a scatter plot to check for
sinusoidality.
2. Do a lag plot to check for sinusoidality. If the lag plot is
elliptical, then the data are sinusoidal.
3. If the data are sinusoidal, then a spectral plot is used to
graphically estimate the underlying sinusoidal frequency.
4. If the data are not sinusoidal, then a Tukey Lambda PPCC plot
may determine the best-fit symmetric distribution for the data.
5. The data may be fit with a mixture of two distributions. A
common approach to this case is to fit a mixture of 2 normal or
lognormal distributions. Further discussion of fitting mixtures of
distributions is beyond the scope of this Handbook.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e4.htm (2 of 2) [5/7/2002 3:14:22 PM]

1.3.3.14.5. Histogram Interpretation: Bimodal Mixture of 2 Normals

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.14. Histogram

1.3.3.14.5. Histogram Interpretation:

Bimodal Mixture of 2 Normals
Histogram
from Mixture
of 2 Normal
Distributions

Discussion of The histogram shown above illustrates data from a bimodal (2 peak)
Unimodal and distribution.
Bimodal
In contrast to the previous example, this example illustrates bimodality
due not to an underlying deterministic model, but bimodality due to a
mixture of probability models. In this case, each of the modes appears
to have a rough bell-shaped component. One could easily imagine the
above histogram being generated by a process consisting of two
normal distributions with the same standard deviation but with two
different locations (one centered at approximately 9.17 and the other
centered at approximately 9.26). If this is the case, then the research
challenge is to determine physically why there are two similar but
separate sub-processes.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e5.htm (1 of 2) [5/7/2002 3:14:23 PM]

1.3.3.14.5. Histogram Interpretation: Bimodal Mixture of 2 Normals

Recommended If the histogram indicates that the data might be appropriately fit with
Next Steps a mixture of two normal distributions, the recommended next step is:
Fit the normal mixture model using either least squares or maximum
likelihood. The general normal mixing model is

where p is the mixing proportion (between 0 and 1) and and are

normal probability density functions with location and scale
parameters , , , and , respectively. That is, there are 5
parameters to estimate in the fit.
Whether maximum likelihood or least squares is used, the quality of
the fit is sensitive to good starting values. For the mixture of two
normals, the histogram can be used to provide initial estimates for the
location and scale parameters of the two normal distributions.
Dataplot can generate a least squares fit of the mixture of two normals
with the following sequence of commands:
RELATIVE HISTOGRAM Y
LET Y2 = YPLOT
LET X2 = XPLOT
RETAIN Y2 X2 SUBSET TAGPLOT = 1
LET U1 = <estimated value from histogram>
LET SD1 = <estimated value from histogram>
LET U2 = <estimated value from histogram>
LET SD2 = <estimated value from histogram>
LET P = 0.5
FIT Y2 = NORMXPDF(X2,U1,S1,U2,S2,P)

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e5.htm (2 of 2) [5/7/2002 3:14:23 PM]

1.3.3.14.6. Histogram Interpretation: Skewed (Non-Normal) Right

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.14. Histogram

1.3.3.14.6. Histogram Interpretation:

Skewed (Non-Normal) Right
Right-Skewed
Histogram

Discussion of A symmetric distribution is one in which the 2 "halves" of the

Skewness histogram appear as mirror-images of one another. A skewed
(non-symmetric) distribution is a distribution in which there is no such
mirror-imaging.
For skewed distributions, it is quite common to have one tail of the
distribution considerably longer or drawn out relative to the other tail.
A "skewed right" distribution is one in which the tail is on the right
side. A "skewed left" distribution is one in which the tail is on the left
side. The above histogram is for a distribution that is skewed right.
Skewed distributions bring a certain philosophical complexity to the
very process of estimating a "typical value" for the distribution. To be

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e6.htm (1 of 3) [5/7/2002 3:14:36 PM]

1.3.3.14.6. Histogram Interpretation: Skewed (Non-Normal) Right

specific, suppose that the analyst has a collection of 100 values

randomly drawn from a distribution, and wishes to summarize these
100 observations by a "typical value". What does typical value mean?
If the distribution is symmetric, the typical value is unambiguous-- it is
a well-defined center of the distribution. For example, for a
bell-shaped symmetric distribution, a center point is identical to that
value at the peak of the distribution.
For a skewed distribution, however, there is no "center" in the usual
sense of the word. Be that as it may, several "typical value" metrics are
often used for skewed distributions. The first metric is the mode of the
distribution. Unfortunately, for severely-skewed distributions, the
mode may be at or near the left or right tail of the data and so it seems
not to be a good representative of the center of the distribution. As a
second choice, one could conceptually argue that the mean (the point
on the horizontal axis where the distributiuon would balance) would
serve well as the typical value. As a third choice, others may argue
that the median (that value on the horizontal axis which has exactly
50% of the data to the left (and also to the right) would serve as a good
typical value.
For symmetric distributions, the conceptual problem disappears
because at the population level the mode, mean, and median are
identical. For skewed distributions, however, these 3 metrics are
markedly different. In practice, for skewed distributions the most
commonly reported typical value is the mean; the next most common
is the median; the least common is the mode. Because each of these 3
metrics reflects a different aspect of "centerness", it is recommended
that the analyst report at least 2 (mean and median), and preferably all
3 (mean, median, and mode) in summarizing and characterizing a data
set.

Some Causes Skewed data often occur due to lower or upper bounds on the data.
for Skewed That is, data that have a lower bound are often skewed right while data
Data that have an upper bound are often skewed left. Skewness can also
result from start-up effects. For example, in reliability applications
some processes may have a large number of initial failures that could
cause left skewness. On the other hand, a reliability process could
have a long start-up period where failures are rare resulting in
right-skewed data.
Data collected in scientific and engineering applications often have a
lower bound of zero. For example, failure data must be non-negative.
Many measurement processes generate only positive data. Time to
occurence and size are common measurements that cannot be less than
zero.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e6.htm (2 of 3) [5/7/2002 3:14:36 PM]

1.3.3.14.6. Histogram Interpretation: Skewed (Non-Normal) Right

Recommended If the histogram indicates a right-skewed data set, the recommended

Next Steps next steps are to:
1. Quantitatively summarize the data by computing and reporting
the sample mean, the sample median, and the sample mode.
2. Determine the best-fit distribution (skewed-right) from the
❍ Weibull family (for the maximum)

❍ Gamma family
❍ Chi-square family
❍ Lognormal family
❍ Power lognormal family
3. Consider a normalizing transformation such as the Box-Cox
transformation.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e6.htm (3 of 3) [5/7/2002 3:14:36 PM]

1.3.3.14.7. Histogram Interpretation: Skewed (Non-Symmetric) Left

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.14. Histogram

1.3.3.14.7. Histogram Interpretation:

Skewed (Non-Symmetric) Left
Skewed Left
Histogram

The issues for skewed left data are similar to those for skewed right
data.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e7.htm [5/7/2002 3:14:36 PM]

1.3.3.14.8. Histogram Interpretation: Symmetric with Outlier

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.14. Histogram

1.3.3.14.8. Histogram Interpretation:

Symmetric with Outlier
Symmetric
Histogram
with Outlier

Discussion of A symmetric distribution is one in which the 2 "halves" of the

Outliers histogram appear as mirror-images of one another. The above example
is symmetric with the exception of outlying data near Y = 4.5.
An outlier is a data point that comes from a distribution different (in
location, scale, or distributional form) from the bulk of the data. In the
real world, outliers have a range of causes, from as simple as
1. operator blunders
2. equipment failures
3. day-to-day effects
4. batch-to-batch differences
5. anomalous input conditions

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e8.htm (1 of 2) [5/7/2002 3:14:37 PM]

1.3.3.14.8. Histogram Interpretation: Symmetric with Outlier

6. warm-up effects
to more subtle causes such as
1. A change in settings of factors that (knowingly or unknowingly)
affect the response.
2. Nature is trying to tell us something.

Outliers All outliers should be taken seriously and should be investigated

Should be thoroughly for explanations. Automatic outlier-rejection schemes
Investigated (such as throw out all data beyond 4 sample standard deviations from
the sample mean) are particularly dangerous.
The classic case of automatic outlier rejection becoming automatic
information rejection was the South Pole ozone depletion problem.
Ozone depletion over the South Pole would have been detected years
earlier except for the fact that the satellite data recording the low
ozone readings had outlier-rejection code that automatically screened
out the "outliers" (that is, the low ozone readings) before the analysis
was conducted. Such inadvertent (and incorrect) purging went on for
years. It was not until ground-based South Pole readings started
detecting low ozone readings that someone decided to double-check as
to why the satellite had not picked up this fact--it had, but it had gotten
thrown out!
The best attitude is that outliers are our "friends", outliers are trying to
tell us something, and we should not stop until we are comfortable in
the explanation for each outlier.

Recommended If the histogram shows the presence of outliers, the recommended next
Next Steps steps are:
1. Graphically check for outliers (in the commonly encountered
normal case) by generating a box plot. In general, box plots are
a much better graphical tool for detecting outliers than are
histograms.
2. Quantitatively check for outliers (in the commonly encountered
normal case) by carrying out Grubbs test which indicates how
many sample standard deviations away from the sample mean
are the data in question. Large values indicate outliers.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33e8.htm (2 of 2) [5/7/2002 3:14:37 PM]

1.3.3.15. Lag Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.15. Lag Plot

Purpose: A lag plot checks whether a data set or time series is random or not.
Check for Random data should not exhibit any identifiable structure in the lag plot.
randomness Non-random structure in the lag plot indicates that the underlying data
are not random. Several common patterns for lag plots are shown in the
examples below.

Sample Plot

This sample lag plot exhibits a linear pattern. This shows that the data
are strongly non-random and further suggests that an autoregressive
model might be appropriate.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33f.htm (1 of 2) [5/7/2002 3:14:37 PM]

1.3.3.15. Lag Plot

Definition A lag is a fixed time displacement. For example, given a data set Y1, Y2
..., Yn, Y2 and Y7 have lag 5 since 7 - 2 = 5. Lag plots can be generated
for any arbitrary lag, although the most commonly used lag is 1.
A plot of lag 1 is a plot of the values of Yi versus Yi-1
● Vertical axis: Yi for all i
● Horizontal axis: Yi-1 for all i

Questions Lag plots can provide answers to the following questions:

1. Are the data random?
2. Is there serial correlation in the data?
3. What is a suitable model for the data?
4. Are there outliers in the data?

Importance Inasmuch as randomness is an underlying assumption for most statistical

estimation and testing techniques, the lag plot should be a routine tool
for researchers.

Examples ● Random (White Noise)

● Weak autocorrelation
● Strong autocorrelation and autoregressive model
● Sinusoidal model and outliers

Related Autocorrelation Plot

Techniques Spectrum
Runs Test

Case Study The lag plot is demonstrated in the beam deflection data case study.

Software Lag plots are not directly available in most general purpose statistical
software programs. Since the lag plot is essentially a scatter plot with
the 2 variables properly lagged, it should be feasible to write a macro for
the lag plot in most statistical programs. Dataplot supports a lag plot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33f.htm (2 of 2) [5/7/2002 3:14:37 PM]

1.3.3.15.1. Lag Plot: Random Data

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.15. Lag Plot

1.3.3.15.1. Lag Plot: Random Data

Lag Plot

Conclusions We can make the following conclusions based on the above plot.
1. The data are random.
2. The data exhibit no autocorrelation.
3. The data contain no outliers.

Discussion The lag plot shown above is for lag = 1. Note the absence of structure.
One cannot infer, from a current value Yi-1, the next value Yi. Thus for a
known value Yi-1 on the horizontal axis (say, Yi-1 = +0.5), the Yi-th
value could be virtually anything (from Yi = -2.5 to Yi = +1.5). Such
non-association is the essence of randomness.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33f1.htm (1 of 2) [5/7/2002 3:14:38 PM]

1.3.3.15.1. Lag Plot: Random Data

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33f1.htm (2 of 2) [5/7/2002 3:14:38 PM]

1.3.3.15.2. Lag Plot: Moderate Autocorrelation

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.15. Lag Plot

1.3.3.15.2. Lag Plot: Moderate

Autocorrelation
Lag Plot

Conclusions We can make the conclusions based on the above plot.

1. The data are from an underlying autoregressive model with
moderate positive autocorrelation
2. The data contain no outliers.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33f2.htm (1 of 2) [5/7/2002 3:14:38 PM]

1.3.3.15.2. Lag Plot: Moderate Autocorrelation

Discussion In the plot above for lag = 1, note how the points tend to cluster (albeit
noisily) along the diagonal. Such clustering is the lag plot signature of
moderate autocorrelation.
If the process were completely random, knowledge of a current
observation (say Yi-1 = 0) would yield virtually no knowledge about
the next observation Yi. If the process has moderate autocorrelation, as
above, and if Yi-1 = 0, then the range of possible values for Yi is seen
to be restricted to a smaller range (.01 to +.01). This suggests
prediction is possible using an autoregressive model.

Recommended Estimate the parameters for the autoregressive model:

Next Step
Since Yi and Yi-1 are precisely the axes of the lag plot, such estimation
is a linear regression straight from the lag plot.

The residual standard deviation for the autoregressive model will be

much smaller than the residual standard deviation for the default
model

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33f2.htm (2 of 2) [5/7/2002 3:14:38 PM]

1.3.3.15.3. Lag Plot: Strong Autocorrelation and Autoregressive Model

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.15. Lag Plot

1.3.3.15.3. Lag Plot: Strong Autocorrelation

and Autoregressive Model
Lag Plot

Conclusions We can make the following conclusions based on the above plot.
1. The data come from an underlying autoregressive model with
strong positive autocorrelation
2. The data contain no outliers.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33f3.htm (1 of 2) [5/7/2002 3:14:38 PM]

1.3.3.15.3. Lag Plot: Strong Autocorrelation and Autoregressive Model

Discussion Note the tight clustering of points along the diagonal. This is the lag
plot signature of a process with strong positive autocorrelation. Such
processes are highly non-random--there is strong association between
an observation and a succeeding observation. In short, if you know
Yi-1 you can make a strong guess as to what Yi will be.

If the above process were completely random, the plot would have a
shotgun pattern, and knowledge of a current observation (say Yi-1 = 3)
would yield virtually no knowledge about the next observation Yi (it
could here be anywhere from -2 to +8). On the other hand, if the
process had strong autocorrelation, as seen above, and if Yi-1 = 3, then
the range of possible values for Yi is seen to be restricted to a smaller
range (2 to 4)--still wide, but an improvement nonetheless (relative to
-2 to +8) in predictive power.

Recommended When the lag plot shows a strongly autoregressive pattern and only
Next Step successive observations appear to be correlated, the next steps are to:
1. Extimate the parameters for the autoregressive model:

Since Yi and Yi-1 are precisely the axes of the lag plot, such
estimation is a linear regression straight from the lag plot.

The residual standard deviation for this autoregressive model

will be much smaller than the residual standard deviation for the
default model

2. Reexamine the system to arrive at an explanation for the strong

autocorrelation. Is it due to the
1. phenomenon under study; or
2. drifting in the environment; or
3. contamination from the data acquisition system?
Sometimes the source of the problem is contamination and
carry-over from the data acquisition system where the system
does not have time to electronically recover before collecting
the next data point. If this is the case, then consider slowing
down the sampling rate to achieve randomness.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33f3.htm (2 of 2) [5/7/2002 3:14:38 PM]

1.3.3.15.4. Lag Plot: Sinusoidal Models and Outliers

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.15. Lag Plot

1.3.3.15.4. Lag Plot: Sinusoidal Models and

Outliers
Lag Plot

Conclusions We can make the following conclusions based on the above plot.
1. The data come from an underlying single-cycle sinusoidal
model.
2. The data contain three outliers.

Discussion In the plot above for lag = 1, note the tight elliptical clustering of
points. Processes with a single-cycle sinusoidal model will have such
elliptical lag plots.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33f4.htm (1 of 3) [5/7/2002 3:14:39 PM]

1.3.3.15.4. Lag Plot: Sinusoidal Models and Outliers

Consequences If one were to naively assume that the above process came from the
of Ignoring null model
Cyclical
Pattern
and then estimate the constant by the sample mean, then the analysis
would suffer because
1. the sample mean would be biased and meaningless;
2. the confidence limits would be meaningless and optimistically
small.
The proper model

(where is the amplitude, is the frequency--between 0 and .5

cycles per observation--, and is the phase) can be fit by standard
non-linear least squares, to estimate the coefficients and their
uncertainties.
The lag plot is also of value in outlier detection. Note in the above plot
that there appears to be 4 points lying off the ellipse. However, in a lag
plot, each point in the original data set Y shows up twice in the lag
plot--once as Yi and once as Yi-1. Hence the outlier in the upper left at
Yi = 300 is the same raw data value that appears on the far right at Yi-1
= 300. Thus (-500,300) and (300,200) are due to the same outlier,
namely the 158th data point: 300. The correct value for this 158th
point should be approximately -300 and so it appears that a sign got
dropped in the data collection. The other two points lying off the
ellipse, at roughly (100,100) and at (0,-50), are caused by two faulty
data values: the third data point of -15 should be about +125 and the
fourth data point of +141 should be about -50, respectively. Hence the
4 apparent lag plot outliers are traceable to 3 actual outliers in the
original run sequence: at points 4 (-15), 5 (141) and 158 (300). In
retrospect, only one of these (point 158 (= 300)) is an obvious outlier
in the run sequence plot.

Unexpected Frequently a technique (e.g., the lag plot) is constructed to check one
Value of EDA aspect (e.g., randomness) which it does well. Along the way, the
technique also highlights some other anomaly of the data (namely, that
there are 3 outliers). Such outlier identification and removal is
extremely important for detecting irregularities in the data collection
system, and also for arriving at a "purified" data set for modeling. The
lag plot plays an important role in such outlier identification.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33f4.htm (2 of 3) [5/7/2002 3:14:39 PM]

1.3.3.15.4. Lag Plot: Sinusoidal Models and Outliers

Recommended When the lag plot indicates a sinusoidal model with possible outliers,
Next Step the recommended next steps are:
1. Do a spectral plot to obtain an initial estimate of the frequency
of the underlying cycle. This will be helpful as a starting value
for the subsequent non-linear fitting.
2. Omit the outliers.
3. Carry out a non-linear fit of the model to the 197 points.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33f4.htm (3 of 3) [5/7/2002 3:14:39 PM]

1.3.3.16. Linear Correlation Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.16. Linear Correlation Plot

Purpose: Linear correlation plots are used to assess whether or not correlations
Detect are consistent across groups. That is, if your data is in groups, you may
changes in want to know if a single correlation can be used across all the groups or
correlation whether separate correlations are required for each group.
between
groups Linear correlation plots are often used in conjunction with linear slope,
linear intercept, and linear residual standard deviation plots. A linear
correlation plot could be generated intially to see if linear fitting would
be a fruitful direction. If the correlations are high, this implies it is
worthwhile to continue with the linear slope, intercept, and residual
standard deviation plots. If the correlations are weak, a different model
needs to be pursued.
In some cases, you might not have groups. Instead you may have
different data sets and you want to know if the same correlation can be
adequately applied to each of the data sets. In this case, simply think of
each distinct data set as a group and apply the linear slope plot as for
groups.

Sample Plot

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33g.htm (1 of 3) [5/7/2002 3:14:39 PM]

1.3.3.16. Linear Correlation Plot

This linear correlation plot shows that the correlations are high for all
groups. This implies that linear fits could provide a good model for
each of these groups.

Definition: Linear correlation plots are formed by:

Group ● Vertical axis: Group correlations
Correlations
● Horizontal axis: Group identifier
Versus
Group ID A reference line is plotted at the correlation between the full data sets.

Questions The linear correlation plot can be used to answer the following
questions.
1. Are there linear relationships across groups?
2. Are the strength of the linear relationships relatively constant
across the groups?

Importance: For grouped data, it may be important to know whether the different
Checking groups are homogeneous (i.e., similar) or heterogeneous (i.e., different).
Group Linear correlation plots help answer this question in the context of
Homogeneity linear fitting.

Related Linear Intercept Plot

Techniques Linear Slope Plot
Linear Residual Standard Deviation Plot
Linear Fitting

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33g.htm (2 of 3) [5/7/2002 3:14:39 PM]

1.3.3.16. Linear Correlation Plot

Case Study The linear correlation plot is demonstrated in the Alaska pipeline data
case study.

Software Most general purpose statistical software programs do not support a

linear correlation plot. However, if the statistical program can generate
correlations over a group, it should be feasible to write a macro to
generate this plot. Dataplot supports a linear correlation plot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33g.htm (3 of 3) [5/7/2002 3:14:39 PM]

1.3.3.17. Linear Intercept Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.17. Linear Intercept Plot

Purpose: Linear intercept plots are used to graphically assess whether or not
Detect linear fits are consistent across groups. That is, if your data have
changes in groups, you may want to know if a single fit can be used across all the
linear groups or whether separate fits are required for each group.
intercepts
between Linear intercept plots are typically used in conjunction with linear slope
groups and linear residual standard deviation plots.

In some cases you might not have groups. Instead, you have different
data sets and you want to know if the same fit can be adequately applied
to each of the data sets. In this case, simply think of each distinct data
set as a group and apply the linear intercept plot as for groups.

Sample Plot

This linear intercept plot shows that there is a shift in intercepts.

Specifically, the first three intercepts are lower than the intercepts for

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33h.htm (1 of 2) [5/7/2002 3:14:39 PM]

1.3.3.17. Linear Intercept Plot

the other groups. Note that these are small differences in the intercepts.

Definition: Linear intercept plots are formed by:

Group ● Vertical axis: Group intercepts from linear fits
Intercepts
● Horizontal axis: Group identifier
Versus
Group ID A reference line is plotted at the intercept from a linear fit using all the
data.

Questions The linear intercept plot can be used to answer the following questions.
1. Is the intercept from linear fits relatively constant across groups?
2. If the intercepts vary across groups, is there a discernible pattern?

Importance: For grouped data, it may be important to know whether the different
Checking groups are homogeneous (i.e., similar) or heterogeneous (i.e., different).
Group Linear intercept plots help answer this question in the context of linear
Homogeneity fitting.

Related Linear Correlation Plot

Techniques Linear Slope Plot
Linear Residual Standard Deviation Plot
Linear Fitting

Case Study The linear intercept plot is demonstrated in the Alaska pipeline data
case study.

Software Most general purpose statistical software programs do not support a

linear intercept plot. However, if the statistical program can generate
linear fits over a group, it should be feasible to write a macro to
generate this plot. Dataplot supports a linear intercept plot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33h.htm (2 of 2) [5/7/2002 3:14:39 PM]

1.3.3.18. Linear Slope Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.18. Linear Slope Plot

Purpose: Linear slope plots are used to graphically assess whether or not linear
Detect fits are consistent across groups. That is, if your data have groups, you
changes in may want to know if a single fit can be used across all the groups or
linear slopes whether separate fits are required for each group.
between
groups Linear slope plots are typically used in conjunction with linear intercept
and linear residual standard deviation plots.

In some cases you might not have groups. Instead, you have different
data sets and you want to know if the same fit can be adequately applied
to each of the data sets. In this case, simply think of each distinct data
set as a group and apply the linear slope plot as for groups.

Sample Plot

This linear slope plot shows that the slopes are about 0.174 (plus or
minus 0.002) for all groups. There does not appear to be a pattern in the
variation of the slopes. This implies that a single fit may be adequate.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33i.htm (1 of 2) [5/7/2002 3:14:39 PM]

1.3.3.18. Linear Slope Plot

Definition: Linear slope plots are formed by:

Group ● Vertical axis: Group slopes from linear fits
Slopes
● Horizontal axis: Group identifier
Versus
Group ID A reference line is plotted at the slope from a linear fit using all the
data.

Questions The linear slope plot can be used to answer the following questions.
1. Do you get the same slope across groups for linear fits?
2. If the slopes differ, is there a discernible pattern in the slopes?

Importance: For grouped data, it may be important to know whether the different
Checking groups are homogeneous (i.e., similar) or heterogeneous (i.e., different).
Group Linear slope plots help answer this question in the context of linear
Homogeneity fitting.

Related Linear Intercept Plot

Techniques Linear Correlation Plot
Linear Residual Standard Deviation Plot
Linear Fitting

Case Study The linear slope plot is demonstrated in the Alaska pipeline data case
study.

Software Most general purpose statistical software programs do not support a

linear slope plot. However, if the statistical program can generate linear
fits over a group, it should be feasible to write a macro to generate this
plot. Dataplot supports a linear slope plot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33i.htm (2 of 2) [5/7/2002 3:14:39 PM]

1.3.3.19. Linear Residual Standard Deviation Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.19. Linear Residual Standard

Deviation Plot
Purpose: Linear residual standard deviation (RESSD) plots are used to
Detect graphically assess whether or not linear fits are consistent across
Changes in groups. That is, if your data have groups, you may want to know if a
Linear single fit can be used across all the groups or whether separate fits are
Residual required for each group.
Standard
Deviation The residual standard deviation is a goodness-of-fit measure. That is,
Between the smaller the residual standard deviation, the closer is the fit to the
Groups data.
Linear RESSD plots are typically used in conjunction with linear
intercept and linear slope plots. The linear intercept and slope plots
convey whether or not the fits are consistent across groups while the
linear RESSD plot conveys whether the adequacy of the fit is consistent
across groups.
In some cases you might not have groups. Instead, you have different
data sets and you want to know if the same fit can be adequately applied
to each of the data sets. In this case, simply think of each distinct data
set as a group and apply the linear RESSD plot as for groups.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33j.htm (1 of 3) [5/7/2002 3:14:40 PM]

1.3.3.19. Linear Residual Standard Deviation Plot

Sample Plot

This linear RESSD plot shows that the residual standard deviations
from a linear fit are about 0.0025 for all the groups.

Definition: Linear RESSD plots are formed by:

Group ● Vertical axis: Group residual standard deviations from linear fits
Residual
● Horizontal axis: Group identifier
Standard
Deviation A reference line is plotted at the residual standard deviation from a
Versus linear fit using all the data. This reference line will typically be much
Group ID greater than any of the individual residual standard deviations.

Questions The linear RESSD plot can be used to answer the following questions.
1. Is the residual standard deviation from a linear fit constant across
groups?
2. If the residual standard deviations vary, is there a discernible
pattern across the groups?

Importance: For grouped data, it may be important to know whether the different
Checking groups are homogeneous (i.e., similar) or heterogeneous (i.e., different).
Group Linear RESSD plots help answer this question in the context of linear
Homogeneity fitting.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33j.htm (2 of 3) [5/7/2002 3:14:40 PM]

1.3.3.19. Linear Residual Standard Deviation Plot

Related Linear Intercept Plot

Techniques Linear Slope Plot
Linear Correlation Plot
Linear Fitting

Case Study The linear residual standard deviation plot is demonstrated in the
Alaska pipeline data case study.

Software Most general purpose statistical software programs do not support a

linear residual standard deviation plot. However, if the statistical
program can generate linear fits over a group, it should be feasible to
write a macro to generate this plot. Dataplot supports a linear residual
standard deviation plot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33j.htm (3 of 3) [5/7/2002 3:14:40 PM]

1.3.3.20. Mean Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.20. Mean Plot

Purpose: Mean plots are used to see if the mean varies between different groups
Detect of the data. The grouping is determined by the analyst. In most cases,
changes in the data set contains a specific grouping variable. For example, the
location groups may be the levels of a factor variable. In the sample plot below,
between the months of the year provide the grouping.
groups
Mean plots can be used with ungrouped data to determine if the mean is
changing over time. In this case, the data are split into an arbitrary
number of equal-sized groups. For example, a data series with 400
points can be divided into 10 groups of 40 points each. A mean plot can
then be generated with these groups to see if the mean is increasing or
decreasing over time.
Although the mean is the most commonly used measure of location, the
same concept applies to other measures of location. For example,
instead of plotting the mean of each group, the median or the trimmed
mean might be plotted instead. This might be done if there were
significant outliers in the data and a more robust measure of location
than the mean was desired.
Mean plots are typically used in conjunction with standard deviation
plots. The mean plot checks for shifts in location while the standard
deviation plot checks for shifts in scale.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33k.htm (1 of 3) [5/7/2002 3:14:40 PM]

1.3.3.20. Mean Plot

Sample Plot

This sample mean plot shows a shift of location after the 6th month.

Definition: Mean plots are formed by:

Group ● Vertical axis: Group mean
Means
● Horizontal axis: Group identifier
Versus
Group ID A reference line is plotted at the overall mean.

Questions The mean plot can be used to answer the following questions.
1. Are there any shifts in location?
2. What is the magnitude of the shifts in location?
3. Is there a distinct pattern in the shifts in location?

Importance: A common assumption in 1-factor analyses is that of constant location.

Checking That is, the location is the same for different levels of the factor
Assumptions variable. The mean plot provides a graphical check for that assumption.
A common assumption for univariate data is that the location is
constant. By grouping the data into equal intervals, the mean plot can
provide a graphical test of this assumption.

Related Standard Deviation Plot

Techniques Dex Mean Plot
Box Plot

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33k.htm (2 of 3) [5/7/2002 3:14:40 PM]

1.3.3.20. Mean Plot

Software Most general purpose statistical software programs do not support a

mean plot. However, if the statistical program can generate the mean
over a group, it should be feasible to write a macro to generate this plot.
Dataplot supports a mean plot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33k.htm (3 of 3) [5/7/2002 3:14:40 PM]

1.3.3.21. Normal Probability Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.21. Normal Probability Plot

Purpose: The normal probability plot (Chambers 1983) is a graphical technique
Check If Data for assessing whether or not a data set is approximately normally
Are distributed.
Approximately
Normally The data are plotted against a theoretical normal distribution in such a
Distributed way that the points should form an approximate straight line.
Departures from this straight line indicate departures from normality.
The normal probability plot is a special case of the probability plot.
We cover the normal probability plot separately due to its importance
in many applications.

Sample Plot

The points on this plot form a nearly linear pattern, which indicates
that the normal distribution is a good model for this data set.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33l.htm (1 of 3) [5/7/2002 3:14:40 PM]

1.3.3.21. Normal Probability Plot

Definition: The normal probability plot is formed by:

Ordered ● Vertical axis: Ordered response values
Response
● Horizontal axis: Normal order statistic medians
Values Versus
Normal Order The observations are plotted as a function of the corresponding normal
Statistic order statistic medians which are defined as:
Medians N(i) = G(U(i))
where U(i) are the uniform order statistic medians (defined below) and
G is the percent point function of the normal distribution. The percent
point function is the inverse of the cumulative distribution function
(probability that x is less than or equal to some value). That is, given a
probability, we want the corresponding x of the cumulative
distribution function.
The uniform order statistic medians are defined as:
m(i) = 1 - m(n) for i = 1
m(i) = (i - 0.3175)/(n + 0.365) for i = 2, 3, ..., n-1
m(i) = 0.5(1/n) for i = n
In addition, a straight line can be fit to the points and added as a
reference line. The further the points vary from this line, the greater
the indication of departures from normality.
Probability plots for distributions other than the normal are computed
in exactly the same way. The normal percent point function (the G) is
simply replaced by the percent point function of the desired
distribution. That is, a probability plot can easily be generated for any
distribution for which you have the percent point function.
One advantage of this method of computing probability plots is that
the intercept and slope estimates of the fitted line are in fact estimates
for the location and scale parameters of the distribution. Although this
is not too important for the normal distribution since the location and
scale are estimated by the mean and standard deviation, respectively, it
can be useful for many other distributions.
The correlation coefficient of the points on the normal probability plot
can be compared to a table of critical values to provide a formal test of
the hypothesis that the data come from a normal distribution.

Questions The normal probability plot is used to answer the following questions.
1. Are the data normally distributed?
2. What is the nature of the departure from normality (data
skewed, shorter than expected tails, longer than expected tails)?

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33l.htm (2 of 3) [5/7/2002 3:14:40 PM]

1.3.3.21. Normal Probability Plot

Importance: The underlying assumptions for a measurement process are that the
Check data should behave like:
Normality 1. random drawings;
Assumption
2. from a fixed distribution;
3. with fixed location;
4. with fixed scale.
Probability plots are used to assess the assumption of a fixed
distribution. In particular, most statistical models are of the form:
response = deterministic + random
where the deterministic part is the fit and the random part is error. This
error component in most common statistical models is specifically
assumed to be normally distributed with fixed location and scale. This
is the most frequent application of normal probability plots. That is, a
model is fit and a normal probability plot is generated for the residuals
from the fitted model. If the residuals from the fitted model are not
normally distributed, then one of the major assumptions of the model
has been violated.

Examples 1. Data are normally distributed

2. Data have fat tails
3. Data have short tails
4. Data are skewed right

Related Histogram
Techniques Probability plots for other distributions (e.g., Weibull)
Probability plot correlation coefficient plot (PPCC plot)
Anderson-Darling Goodness-of-Fit Test
Chi-Square Goodness-of-Fit Test
Kolmogorov-Smirnov Goodness-of-Fit Test

Case Study The normal probability plot is demonstrated in the heat flow meter
data case study.

Software Most general purpose statistical software programs can generate a

normal probability plot. Dataplot supports a normal probability plot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33l.htm (3 of 3) [5/7/2002 3:14:40 PM]

1.3.3.21.1. Normal Probability Plot: Normally Distributed Data

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.21. Normal Probability Plot

1.3.3.21.1. Normal Probability Plot:

Normally Distributed Data
Normal The following normal probability plot is from the heat flow meter data.
Probability
Plot

Conclusions We can make the following conclusions from the above plot.
1. The normal probability plot shows a strongly linear pattern. There
are only minor deviations from the line fit to the points on the
probability plot.
2. The normal distribution appears to be a good model for these
data.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33l1.htm (1 of 2) [5/7/2002 3:14:41 PM]

1.3.3.21.1. Normal Probability Plot: Normally Distributed Data

Discussion Visually, the probability plot shows a strongly linear pattern. This is
verified by the correlation coefficient of 0.9989 of the line fit to the
probability plot. The fact that the points in the lower and upper extremes
of the plot do not deviate significantly from the straight-line pattern
indicates that there are not any significant outliers (relative to a normal
distribution).
In this case, we can quite reasonably conclude that the normal
distribution provides an excellent model for the data. The intercept and
slope of the fitted line give estimates of 9.26 and 0.023 for the location
and scale parameters of the fitted normal distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33l1.htm (2 of 2) [5/7/2002 3:14:41 PM]

1.3.3.21.2. Normal Probability Plot: Data Have Short Tails

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.21. Normal Probability Plot

1.3.3.21.2. Normal Probability Plot: Data

Have Short Tails
Normal The following is a normal probability plot for 500 random numbers
Probability generated from a Tukey-Lambda distribution with the parameter equal
Plot for to 1.1.
Data with
Short Tails

Conclusions We can make the following conclusions from the above plot.
1. The normal probability plot shows a non-linear pattern.
2. The normal distribution is not a good model for these data.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33l2.htm (1 of 2) [5/7/2002 3:14:41 PM]

1.3.3.21.2. Normal Probability Plot: Data Have Short Tails

Discussion For data with short tails relative to the normal distribution, the
non-linearity of the normal probability plot shows up in two ways. First,
the middle of the data shows an S-like pattern. This is common for both
short and long tails. Second, the first few and the last few points show a
marked departure from the reference fitted line. In comparing this plot
to the long tail example in the next section, the important difference is
the direction of the departure from the fitted line for the first few and
last few points. For short tails, the first few points show increasing
departure from the fitted line above the line and last few points show
increasing departure from the fitted line below the line. For long tails,
this pattern is reversed.
In this case, we can reasonably conclude that the normal distribution
does not provide an adequate fit for this data set. For probability plots
that indicate short-tailed distributions, the next step might be to generate
a Tukey Lambda PPCC plot. The Tukey Lambda PPCC plot can often
be helpful in identifying an appropriate distributional family.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33l2.htm (2 of 2) [5/7/2002 3:14:41 PM]

1.3.3.21.3. Normal Probability Plot: Data Have Long Tails

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.21. Normal Probability Plot

1.3.3.21.3. Normal Probability Plot: Data

Have Long Tails
Normal The following is a normal probability plot of 500 numbers generated
Probability from a double exponential distribution. The double exponential
Plot for distribution is symmetric, but relative to the normal it declines rapidly
Data with and has longer tails.
Long Tails

Conclusions We can make the following conclusions from the above plot.
1. The normal probability plot shows a reasonably linear pattern in
the center of the data. However, the tails, particularly the lower
tail, show departures from the fitted line.
2. A distribution other than the normal distribution would be a good
model for these data.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33l3.htm (1 of 2) [5/7/2002 3:14:41 PM]

1.3.3.21.3. Normal Probability Plot: Data Have Long Tails

Discussion For data with long tails relative to the normal distribution, the
non-linearity of the normal probability plot can show up in two ways.
First, the middle of the data may show an S-like pattern. This is
common for both short and long tails. In this particular case, the S
pattern in the middle is fairly mild. Second, the first few and the last few
points show marked departure from the reference fitted line. In the plot
above, this is most noticeable for the first few data points. In comparing
this plot to the short-tail example in the previous section, the important
difference is the direction of the departure from the fitted line for the
first few and the last few points. For long tails, the first few points show
increasing departure from the fitted line below the line and last few
points show increasing departure from the fitted line above the line. For
short tails, this pattern is reversed.
In this case we can reasonably conclude that the normal distribution can
be improved upon as a model for these data. For probability plots that
indicate long-tailed distributions, the next step might be to generate a
Tukey Lambda PPCC plot. The Tukey Lambda PPCC plot can often be
helpful in identifying an appropriate distributional family.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33l3.htm (2 of 2) [5/7/2002 3:14:41 PM]

1.3.3.21.4. Normal Probability Plot: Data are Skewed Right

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.21. Normal Probability Plot

1.3.3.21.4. Normal Probability Plot: Data are

Skewed Right
Normal
Probability
Plot for
Data that
are Skewed
Right

Conclusions We can make the following conclusions from the above plot.
1. The normal probability plot shows a strongly non-linear pattern.
Specifically, it shows a quadratic pattern in which all the points
are below a reference line drawn between the first and last points.
2. The normal distribution is not a good model for these data.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33l4.htm (1 of 2) [5/7/2002 3:14:41 PM]

1.3.3.21.4. Normal Probability Plot: Data are Skewed Right

Discussion This quadratic pattern in the normal probability plot is the signature of a
significantly right-skewed data set. Similarly, if all the points on the
normal probability plot fell above the reference line connecting the first
and last points, that would be the signature pattern for a significantly
left-skewed data set.
In this case we can quite reasonably conclude that we need to model
these data with a right skewed distribution such as the Weibull or
lognormal.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33l4.htm (2 of 2) [5/7/2002 3:14:41 PM]

1.3.3.22. Probability Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.22. Probability Plot

Purpose: The probability plot (Chambers 1983) is a graphical technique for
Check If assessing whether or not a data set follows a given distribution such as
Data Follow the normal or Weibull.
a Given
Distribution The data are plotted against a theoretical distribution in such a way that
the points should form approximately a straight line. Departures from
this straight line indicate departures from the specified distribution.
The correlation coefficient associated with the linear fit to the data in
the probability plot is a measure of the goodness of the fit. Estimates of
the location and scale parameters of the distribution are given by the
intercept and slope. Probability plots can be generated for several
competing distributions to see which provides the best fit, and the
probability plot generating the highest correlation coefficient is the best
choice since it generates the straightest probability plot.
For distributions with shape parameters (not counting location and
scale parameters), the shape parameters must be known in order to
generate the probability plot. For distributions with a single shape
parameter, the probability plot correlation coefficient (PPCC) plot
provides an excellent method for estimating the shape parameter.
We cover the special case of the normal probability plot separately due
to its importance in many statistical applications.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33m.htm (1 of 4) [5/7/2002 3:14:42 PM]

1.3.3.22. Probability Plot

Sample Plot

This data is a set of 500 Weibull random numbers with a shape

parameter = 2, location parameter = 0, and scale parameter = 1. The
Weibull probability plot indicates that the Weibull distribution does in
fact fit these data well.

Definition: The probability plot is formed by:

Ordered ● Vertical axis: Ordered response values
Response
● Horizontal axis: Order statistic medians for the given distribution
Values
Versus Order The order statistic medians are defined as:
Statistic N(i) = G(U(i))
Medians for
the Given where the U(i) are the uniform order statistic medians (defined below)
Distribution and G is the percent point function for the desired distribution. The
percent point function is the inverse of the cumulative distribution
function (probability that x is less than or equal to some value). That is,
given a probability, we want the corresponding x of the cumulative
distribution function.
The uniform order statistic medians are defined as:
m(i) = 1 - m(n) for i = 1
m(i) = (i - 0.3175)/(n + 0.365) for i = 2, 3, ..., n-1
m(i) = 0.5**(1/n) for i = n
In addition, a straight line can be fit to the points and added as a
reference line. The further the points vary from this line, the greater the

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33m.htm (2 of 4) [5/7/2002 3:14:42 PM]

1.3.3.22. Probability Plot

indication of a departure from the specified distribution.

This definition implies that a probability plot can be easily generated
for any distribution for which the percent point function can be
computed.
One advantage of this method of computing proability plots is that the
intercept and slope estimates of the fitted line are in fact estimates for
the location and scale parameters of the distribution. Although this is
not too important for the normal distribution (the location and scale are
estimated by the mean and standard deviation, respectively), it can be
useful for many other distributions.

Questions The probability plot is used to answer the following questions:

● Does a given distribution, such as the Weibull, provide a good fit
to my data?
● What distribution best fits my data?

● What are good estimates for the location and scale parameters of
the chosen distribution?

Importance: The discussion for the normal probability plot covers the use of
Check probability plots for checking the fixed distribution assumption.
distributional
assumption Some statistical models assume data have come from a population with
a specific type of distribution. For example, in reliability applications,
the Weibull, lognormal, and exponential are commonly used
distributional models. Probability plots can be useful for checking this
distributional assumption.

Related Histogram
Techniques Probability Plot Correlation Coefficient (PPCC) Plot
Hazard Plot
Quantile-Quantile Plot
Anderson-Darling Goodness of Fit
Chi-Square Goodness of Fit
Kolmogorov-Smirnov Goodness of Fit

Case Study The probability plot is demonstrated in the airplane glass failure time
data case study.

Software Most general purpose statistical software programs support probability

plots for at least a few common distributions. Dataplot supports
probability plots for a large number of distributions.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33m.htm (3 of 4) [5/7/2002 3:14:42 PM]

1.3.3.22. Probability Plot

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33m.htm (4 of 4) [5/7/2002 3:14:42 PM]

1.3.3.23. Probability Plot Correlation Coefficient Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.23. Probability Plot Correlation

Coefficient Plot
Purpose: The probability plot correlation coefficient (PPCC) plot (Filliben
Graphical 1975) is a graphical technique for identifying the shape parameter for
Technique for a distributional family that best describes the data set. This technique
Finding the is appropriate for families, such as the Weibull, that are defined by a
Shape single shape parameter and location and scale parameters, and it is not
Parameter of appropriate for distributions, such as the normal, that are defined only
a by location and scale parameters.
Distributional
Family that The PPCC plot is generated as follows. For a series of values for the
Best Fits a shape parameter, the correlation coefficient is computed for the
Data Set probability plot associated with a given value of the shape parameter.
These correlation coefficients are plotted against their corresponding
shape parameters. The maximum correlation coefficient corresponds
to the optimal value of the shape parameter. For better precision, two
iterations of the PPCC plot can be generated; the first is for finding
the right neighborhood and the second is for fine tuning the estimate.
The PPCC plot is used first to find a good value of the shape
parameter. The probability plot is then generated to find estimates of
the location and scale parameters and in addition to provide a
graphical assessment of the adequacy of the distributional fit.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33n.htm (1 of 4) [5/7/2002 3:14:42 PM]

1.3.3.23. Probability Plot Correlation Coefficient Plot

Compare In addition to finding a good choice for estimating the shape

Distributions parameter of a given distribution, the PPCC plot can be useful in
deciding which distributional family is most appropriate. For example,
given a set of reliabilty data, you might generate PPCC plots for a
Weibull, lognormal, gamma, and inverse Gaussian distributions, and
possibly others, on a single page. This one page would show the best
value for the shape parameter for several distributions and would
additionally indicate which of these distributional families provides
the best fit (as measured by the maximum probability plot correlation
coefficient). That is, if the maximum PPCC value for the Weibull is
0.99 and only 0.94 for the lognormal, then we could reasonably
conclude that the Weibull family is the better choice.

Tukey-Lambda The Tukey Lambda PPCC plot, with shape parameter , is

PPCC Plot for particularly useful for symmetric distributions. It indicates whether a
Symmetric distribution is short or long tailed and it can further indicate several
Distributions common distributions. Specifically,
1. = -1: distribution is approximately Cauchy
2. = 0: distribution is exactly logistic
3. = 0.14: distribution is approximately normal
4. = 0.5: distribution is U-shaped
5. = 1: distribution is exactly uniform
If the Tukey Lambda PPCC plot gives a maximum value near 0.14,
we can reasonably conclude that the normal distribution is a good
model for the data. If the maximum value is less than 0.14, a
long-tailed distribution such as the double exponential or logistic
would be a better choice. If the maximum value is near -1, this implies
the selection of very long-tailed distribution, such as the Cauchy. If
the maximum value is greater than 0.14, this implies a short-tailed
distribution such as the Beta or uniform.
The Tukey-Lambda PPCC plot is used to suggest an appropriate
distribution. You should follow-up with PPCC and probability plots of
the appropriate alternatives.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33n.htm (2 of 4) [5/7/2002 3:14:42 PM]

1.3.3.23. Probability Plot Correlation Coefficient Plot

Use When comparing distributional models, do not simply choose the one
Judgement with the maximum PPCC value. In many cases, several distributional
When fits provide comparable PPCC values. For example, a lognormal and
Selecting An Weibull may both fit a given set of reliability data quite well.
Appropriate Typically, we would consider the complexity of the distribution. That
Distributional is, a simpler distribution with a marginally smaller PPCC value may
Family be preferred over a more complex distribution. Likewise, there may be
theoretical justification in terms of the underlying scientific model for
preferring a distribution with a marginally smaller PPCC value in
some cases. In other cases, we may not need to know if the
distributional model is optimal, only that it is adequate for our
purposes. That is, we may be able to use techniques designed for
normally distributed data even if other distributions fit the data
somewhat better.

Sample Plot The following is a PPCC plot of 100 normal random numbers. The
maximum value of the correlation coefficient = 0.997 at = 0.099.

This PPCC plot shows that:

1. the best-fit symmetric distribution is nearly normal;
2. the data are not long tailed;
3. the sample mean would be an appropriate estimator of location.
We can follow-up this PPCC plot with a normal probability plot to
verify the normality model for the data.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33n.htm (3 of 4) [5/7/2002 3:14:42 PM]

1.3.3.23. Probability Plot Correlation Coefficient Plot

Definition: The PPCC plot is formed by:

● Vertical axis: Probability plot correlation coefficient;

● Horizontal axis: Value of shape parameter.

Questions The PPCC plot answers the following questions:

1. What is the best-fit member within a distributional family?
2. Does the best-fit member provide a good fit (in terms of
generating a probability plot with a high correlation
coefficient)?
3. Does this distributional family provide a good fit compared to
other distributions?
4. How sensitive is the choice of the shape parameter?

Importance Many statistical analyses are based on distributional assumptions

about the population from which the data have been obtained.
However, distributional families can have radically different shapes
depending on the value of the shape parameter. Therefore, finding a
reasonable choice for the shape parameter is a necessary step in the
analysis. In many analyses, finding a good distributional model for the
data is the primary focus of the analysis. In both of these cases, the
PPCC plot is a valuable tool.

Related Probability Plot

Techniques Maximum Likelihood Estimation
Least Squares Estimation
Method of Moments Estimation

Case Study The PPCC plot is demonstrated in the airplane glass failure data case
study.

Software PPCC plots are currently not available in most common general
purpose statistical software programs. However, the underlying
technique is based on probability plots and correlation coefficients, so
it should be possible to write macros for PPCC plots in statistical
programs that support these capabilities. Dataplot supports PPCC
plots.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33n.htm (4 of 4) [5/7/2002 3:14:42 PM]

1.3.3.24. Quantile-Quantile Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.24. Quantile-Quantile Plot

Purpose: The quantile-quantile (q-q) plot is a graphical technique for determining
Check If if two data sets come from populations with a common distribution.
Two Data
Sets Can Be A q-q plot is a plot of the quantiles of the first data set against the
Fit With the quantiles of the second data set. By a quantile, we mean the fraction (or
Same percent) of points below the given value. That is, the 0.3 (or 30%)
Distribution quantile is the point at which 30% percent of the data fall below and
70% fall above that value.
A 45-degree reference line is also plotted. If the two sets come from a
population with the same distribution, the points should fall
approximately along this reference line. The greater the departure from
this reference line, the greater the evidence for the conclusion that the
two data sets have come from populations with different distributions.
The advantages of the q-q plot are:
1. The sample sizes do not need to be equal.
2. Many distributional aspects can be simultaneously tested. For
example, shifts in location, shifts in scale, changes in symmetry,
and the presence of outliers can all be detected from this plot. For
example, if the two data sets come from populations whose
distributions differ only by a shift in location, the points should lie
along a straight line that is displaced either up or down from the
45-degree reference line.
The q-q plot is similar to a probability plot. For a probability plot, the
quantiles for one of the data samples are replaced with the quantiles of a
theoretical distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33o.htm (1 of 3) [5/7/2002 3:14:43 PM]

1.3.3.24. Quantile-Quantile Plot

Sample Plot

This q-q plot shows that

1. These 2 batches do not appear to have come from populations
with a common distribution.
2. The batch 1 values are significantly higher than the corresponding
batch 2 values.
3. The differences are increasing from values 525 to 625. Then the
values for the 2 batches get closer again.

Definition: The q-q plot is formed by:

Quantiles ● Vertical axis: Estimated quantiles from data set 1
for Data Set
● Horizontal axis: Estimated quantiles from data set 2
1 Versus
Quantiles of Both axes are in units of their respective data sets. That is, the actual
Data Set 2 quantile level is not plotted. For a given point on the q-q plot, we know
that the quantile level is the same for both points, but not what that
quantile level actually is.
If the data sets have the same size, the q-q plot is essentially a plot of
sorted data set 1 against sorted data set 2. If the data sets are not of equal
size, the quantiles are usually picked to correspond to the sorted values
from the smaller data set and then the quantiles for the larger data set are
interpolated.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33o.htm (2 of 3) [5/7/2002 3:14:43 PM]

1.3.3.24. Quantile-Quantile Plot

Questions The q-q plot is used to answer the following questions:

● Do two data sets come from populations with a common
distribution?
● Do two data sets have common location and scale?

● Do two data sets have similar distributional shapes?

● Do two data sets have similar tail behavior?

Importance: When there are two data samples, it is often desirable to know if the
Check for assumption of a common distribution is justified. If so, then location and
Common scale estimators can pool both data sets to obtain estimates of the
Distribution common location and scale. If two samples do differ, it is also useful to
gain some understanding of the differences. The q-q plot can provide
more insight into the nature of the difference than analytical methods
such as the chi-square and Kolmogorov-Smirnov 2-sample tests.

Related Bihistogram
Techniques T Test
F Test
2-Sample Chi-Square Test
2-Sample Kolmogorov-Smirnov Test

Case Study The quantile-quantile plot is demonstrated in the ceramic strength data
case study.

Software Q-Q plots are available in some general purpose statistical software
programs, including Dataplot. If the number of data points in the two
samples are equal, it should be relatively easy to write a macro in
statistical programs that do not support the q-q plot. If the number of
points are not equal, writing a macro for a q-q plot may be difficult.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33o.htm (3 of 3) [5/7/2002 3:14:43 PM]

1.3.3.25. Run-Sequence Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.25. Run-Sequence Plot

Purpose: Run sequence plots (Chambers 1983) are an easy way to graphically
Check for summarize a univariate data set. A common assumption of univariate
Shifts in data sets is that they behave like:
Location 1. random drawings;
and Scale
and Outliers 2. from a fixed distribution;
3. with a common location; and
4. with a common scale.
With run sequence plots, shifts in location and scale are typically quite
evident. Also, outliers can easily be detected.

Sample
Plot:
Last Third
of Data
Shows a
Shift of
Location

This sample run sequence plot shows that the location shifts up for the
last third of the data.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33p.htm (1 of 2) [5/7/2002 3:14:43 PM]

1.3.3.25. Run-Sequence Plot

Definition: Run sequence plots are formed by:

y(i) Versus i ● Vertical axis: Response variable Y(i)

● Horizontal axis: Index i (i = 1, 2, 3, ... )

Questions The run sequence plot can be used to answer the following questions
1. Are there any shifts in location?
2. Are there any shifts in variation?
3. Are there any outliers?
The run sequence plot can also give the analyst an excellent feel for the
data.

Importance: For univariate data, the default model is

Check Y = constant + error
Univariate
where the error is assumed to be random, from a fixed distribution, and
Assumptions
with constant location and scale. The validity of this model depends on
the validity of these assumptions. The run sequence plot is useful for
checking for constant location and scale.
Even for more complex models, the assumptions on the error term are
still often the same. That is, a run sequence plot of the residuals (even
from very complex models) is still vital for checking for outliers and for
detecting shifts in location and scale.

Related Scatter Plot

Techniques Histogram
Autocorrelation Plot
Lag Plot

Case Study The run sequence plot is demonstrated in the Filter transmittance data
case study.

Software Run sequence plots are available in most general purpose statistical
software programs, including Dataplot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33p.htm (2 of 2) [5/7/2002 3:14:43 PM]

1.3.3.26. Scatter Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.26. Scatter Plot

Purpose: A scatter plot (Chambers 1983) reveals relationships or association
Check for between two variables. Such relationships manifest themselves by any
Relationship non-random structure in the plot. Various common types of patterns are
demonstrated in the examples.

Sample
Plot:
Linear
Relationship
Between
Variables Y
and X

This sample plot reveals a linear relationship between the two variables
indicating that a linear regression model might be appropriate.

Definition: A scatter plot is a plot of the values of Y versus the corresponding

Y Versus X values of X:
● Vertical axis: variable Y--usually the response variable

● Horizontal axis: variable X--usually some variable we suspect

may ber related to the response

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33q.htm (1 of 3) [5/7/2002 3:14:43 PM]

1.3.3.26. Scatter Plot

Questions Scatter plots can provide answers to the following questions:

1. Are variables X and Y related?
2. Are variables X and Y linearly related?
3. Are variables X and Y non-linearly related?
4. Does the variation in Y change depending on X?
5. Are there outliers?

Examples 1. No relationship
2. Strong linear (positive correlation)
3. Strong linear (negative correlation)
4. Exact linear (positive correlation)
5. Quadratic relationship
6. Exponential relationship
7. Sinusoidal relationship (damped)
8. Variation of Y doesn't depend on X (homoscedastic)
9. Variation of Y does depend on X (heteroscedastic)
10. Outlier

Combining Scatter plots can also be combined in multiple plots per page to help
Scatter Plots understand higher-level structure in data sets with more than two
variables.
The scatterplot matrix generates all pairwise scatter plots on a single
page. The conditioning plot, also called a co-plot or subset plot,
generates scatter plots of Y versus X dependent on the value of a third
variable.

Causality Is The scatter plot uncovers relationships in data. "Relationships" means

Not Proved that there is some structured association (linear, quadratic, etc.) between
By X and Y. Note, however, that even though
Association
causality implies association
association does NOT imply causality.
Scatter plots are a useful diagnostic tool for determining association, but
if such association exists, the plot may or may not suggest an underlying
cause-and-effect mechanism. A scatter plot can never "prove" cause and
effect--it is ultimately only the researcher (relying on the underlying
science/engineering) who can conclude that causality actually exists.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33q.htm (2 of 3) [5/7/2002 3:14:43 PM]

1.3.3.26. Scatter Plot

Appearance The most popular rendition of a scatter plot is

1. some plot character (e.g., X) at the data points, and
2. no line connecting data points.
Other scatter plot format variants include
1. an optional plot character (e.g, X) at the data points, but
2. a solid line connecting data points.
In both cases, the resulting plot is referred to as a scatter plot, although
the former (discrete and disconnected) is the author's personal
preference since nothing makes it onto the screen except the data--there
are no interpolative artifacts to bias the interpretation.

Related Run Sequence Plot

Techniques Box Plot
Block Plot

Case Study The scatter plot is demonstrated in the load cell calibration data case
study.

Software Scatter plots are a fundamental technique that should be available in any
general purpose statistical software program, including Dataplot. Scatter
plots are also available in most graphics and spreadsheet programs as
well.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33q.htm (3 of 3) [5/7/2002 3:14:43 PM]

1.3.3.26.1. Scatter Plot: No Relationship

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.26. Scatter Plot

1.3.3.26.1. Scatter Plot: No Relationship

Scatter Plot
with No
Relationship

Discussion Note in the plot above how for a given value of X (say X = 0.5), the
corresponding values of Y range all over the place from Y = -2 to Y = +2.
The same is true for other values of X. This lack of predictablility in
determining Y from a given value of X, and the associated amorphous,
non-structured appearance of the scatter plot leads to the summary
conclusion: no relationship.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33q1.htm [5/7/2002 3:14:43 PM]

1.3.3.26.2. Scatter Plot: Strong Linear (positive correlation) Relationship

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.26. Scatter Plot

1.3.3.26.2. Scatter Plot: Strong Linear

(positive correlation)
Relationship
Scatter Plot
Showing
Strong
Positive
Linear
Correlation

Discussion Note in the plot above how a straight line comfortably fits through the
data; hence a linear relationship exists. The scatter about the line is quite
small, so there is a strong linear relationship. The slope of the line is
positive (small values of X correspond to small values of Y; large values
of X correspond to large values of Y), so there is a positive co-relation
(that is, a positive correlation) between X and Y.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33q2.htm [5/7/2002 3:14:44 PM]

1.3.3.26.3. Scatter Plot: Strong Linear (negative correlation) Relationship

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.26. Scatter Plot

1.3.3.26.3. Scatter Plot: Strong Linear

(negative correlation)
Relationship
Scatter Plot
Showing a
Strong
Negative
Correlation

Discussion Note in the plot above how a straight line comfortably fits through the
data; hence there is a linear relationship. The scatter about the line is
quite small, so there is a strong linear relationship. The slope of the line
is negative (small values of X correspond to large values of Y; large
values of X correspond to small values of Y), so there is a negative
co-relation (that is, a negative correlation) between X and Y.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33q3.htm [5/7/2002 3:14:44 PM]

1.3.3.26.4. Scatter Plot: Exact Linear (positive correlation) Relationship

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.26. Scatter Plot

1.3.3.26.4. Scatter Plot: Exact Linear

(positive correlation)
Relationship
Scatter Plot
Showing an
Exact
Linear
Relationship

Discussion Note in the plot above how a straight line comfortably fits through the
data; hence there is a linear relationship. The scatter about the line is
zero--there is perfect predictability between X and Y), so there is an
exact linear relationship. The slope of the line is positive (small values
of X correspond to small values of Y; large values of X correspond to
large values of Y), so there is a positive co-relation (that is, a positive
correlation) between X and Y.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33q4.htm (1 of 2) [5/7/2002 3:14:44 PM]

1.3.3.26.4. Scatter Plot: Exact Linear (positive correlation) Relationship

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33q4.htm (2 of 2) [5/7/2002 3:14:44 PM]

1.3.3.26.5. Scatter Plot: Quadratic Relationship

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.26. Scatter Plot

1.3.3.26.5. Scatter Plot: Quadratic

Relationship
Scatter Plot
Showing
Quadratic
Relationship

Discussion Note in the plot above how no imaginable simple straight line could
ever adequately describe the relationship between X and Y--a curved (or
curvilinear, or non-linear) function is needed. The simplest such
curvilinear function is a quadratic model

for some A, B, and C. Many other curvilinear functions are possible, but
the data analysis principle of parsimony suggests that we try fitting a
quadratic function first.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33q5.htm (1 of 2) [5/7/2002 3:14:45 PM]

1.3.3.26.5. Scatter Plot: Quadratic Relationship

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33q5.htm (2 of 2) [5/7/2002 3:14:45 PM]

1.3.3.26.6. Scatter Plot: Exponential Relationship

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.26. Scatter Plot

1.3.3.26.6. Scatter Plot: Exponential

Relationship
Scatter Plot
Showing
Exponential
Relationship

Discussion Note that a simple straight line is grossly inadequate in describing the
relationship between X and Y. A quadratic model would prove lacking,
especially for large values of X. In this example, the large values of X
correspond to nearly constant values of Y, and so a non-linear function
beyond the quadratic is needed. Among the many other non-linear
functions available, one of the simpler ones is the exponential model

for some A, B, and C. In this case, an exponential function would, in

fact, fit well, and so one is led to the summary conclusion of an
exponential relationship.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33q6.htm (1 of 2) [5/7/2002 3:14:46 PM]

1.3.3.26.6. Scatter Plot: Exponential Relationship

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33q6.htm (2 of 2) [5/7/2002 3:14:46 PM]

1.3.3.26.7. Scatter Plot: Sinusoidal Relationship (damped)

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.26. Scatter Plot

1.3.3.26.7. Scatter Plot: Sinusoidal

Relationship (damped)
Scatter Plot
Showing a
Sinusoidal
Relationship

Discussion The complex relationship between X and Y appears to be basically

oscillatory, and so one is naturally drawn to the trigonometric sinusoidal
model:

Closer inspection of the scatter plot reveals that the amount of swing
(the amplitude in the model) does not appear to be constant but rather
is decreasing (damping) as X gets large. We thus would be led to the
conclusion: damped sinusoidal relationship, with the simplest
corresponding model being

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33q7.htm (1 of 2) [5/7/2002 3:14:47 PM]

1.3.3.26.7. Scatter Plot: Sinusoidal Relationship (damped)

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33q7.htm (2 of 2) [5/7/2002 3:14:47 PM]

1.3.3.26.8. Scatter Plot: Variation of Y Does Not Depend on X (homoscedastic)

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.26. Scatter Plot

1.3.3.26.8. Scatter Plot: Variation of Y Does

Not Depend on X
(homoscedastic)
Scatter Plot
Showing
Homoscedastic
Variability

Discussion This scatter plot reveals a linear relationship between X and Y: for a
given value of X, the predicted value of Y will fall on a line. The plot
further reveals that the variation in Y about the predicted value is
about the same (+- 10 units), regardless of the value of X.
Statistically, this is referred to as homoscedasticity. Such
homoscedasticity is very important as it is an underlying assumption
for regression, and its violation leads to parameter estimates with
inflated variances. If the data are homoscedastic, then the usual
regression estimates can be used. If the data are not homoscedastic,
then the estimates can be improved using weighting procedures as
shown in the next example.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33q8.htm (1 of 2) [5/7/2002 3:14:53 PM]

1.3.3.26.8. Scatter Plot: Variation of Y Does Not Depend on X (homoscedastic)

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33q8.htm (2 of 2) [5/7/2002 3:14:53 PM]

1.3.3.26.9. Scatter Plot: Variation of Y Does Depend on X (heteroscedastic)

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.26. Scatter Plot

1.3.3.26.9. Scatter Plot: Variation of Y Does

Depend on X (heteroscedastic)
Scatter Plot
Showing
Heteroscedastic
Variability

Discussion This scatter plot reveals an approximate linear relationship between

X and Y, but more importantly, it reveals a statistical condition
referred to as heteroscedasticity (that is, nonconstant variation in Y
over the values of X). For a heteroscedastic data set, the variation in
Y differs depending on the value of X. In this example, small values
of X yield small scatter in Y while large values of X result in large
scatter in Y.
Heteroscedasticity complicates the analysis somewhat, but its effects
can be overcome by:
1. proper weighting of the data with noisier data being weighted
less, or by

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33q9.htm (1 of 2) [5/7/2002 3:14:53 PM]

1.3.3.26.9. Scatter Plot: Variation of Y Does Depend on X (heteroscedastic)

2. performing a Y variable transformation to achieve

homoscedasticity. The Box-Cox normality plot can help
determine a suitable transformation.

Impact of Fortunately, unweighted regression analyses on heteroscedastic data

Ignoring produce estimates of the coefficients that are unbiased. However, the
Unequal coefficients will not be as precise as they would be with proper
Variability in weighting.
the Data
Note further that if heteroscedasticity does exist, it is frequently
useful to plot and model the local variation as a
function of X, as in . This modeling has
two advantages:
1. it provides additional insight and understanding as to how the
response Y relates to X; and
2. it provides a convenient means of forming weights for a
weighted regression by simply using

The topic of non-constant variation is discussed in some detail in the

process modeling chapter.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33q9.htm (2 of 2) [5/7/2002 3:14:53 PM]

1.3.3.26.10. Scatter Plot: Outlier

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.26. Scatter Plot

1.3.3.26.10. Scatter Plot: Outlier

Scatter Plot
Showing
Outliers

Discussion The scatter plot here reveals

1. a basic linear relationship between X and Y for most of the data,
and
2. a single outlier (at X = 375).
An outlier is defined as a data point that emanates from a different
model than do the rest of the data. The data here appear to come from a
linear model with a given slope and variation except for the outlier
which appears to have been generated from some other model.
Outlier detection is important for effective modeling. Outliers should be
excluded from such model fitting. If all the data here are included in a
linear regression, then the fitted model will be poor virtually
everywhere. If the outlier is omitted from the fitting process, then the
resulting fit will be excellent almost everywhere (for all points except

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33qa.htm (1 of 2) [5/7/2002 3:14:53 PM]

1.3.3.26.10. Scatter Plot: Outlier

the outlying point).

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33qa.htm (2 of 2) [5/7/2002 3:14:53 PM]

1.3.3.26.11. Scatterplot Matrix

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.26. Scatter Plot

1.3.3.26.11. Scatterplot Matrix

Purpose: Given a set of variables X1, X2, ... , Xk, the scatterplot matrix contains
Check all the pairwise scatter plots of the variables on a single page in a
Pairwise matrix format. That is, if there are k variables, the scatterplot matrix
Relationships will have k rows and k columns and the ith row and jth column of this
Between matrix is a plot of Xi versus Xj.
Variables
Although the basic concept of the scatterplot matrix is simple, there are
numerous alternatives in the details of the plots.
1. The diagonal plot is simply a 45-degree line since we are plotting
Xi versus Xi. Although this has some usefulness in terms of
showing the univariate distribution of the variable, other
alternatives are common. Some users prefer to use the diagonal
to print the variable label. Another alternative is to plot the
univariate histogram on the diagonal. Alternatively, we could
simply leave the diagonal blank.
2. Since Xi versus Xj is equivalent to Xj versus Xi with the axes
reversed, some prefer to omit the plots below the diagonal.
3. It can be helpful to overlay some type of fitted curve on the
scatter plot. Although a linear or quadratic fit can be used, the
most common alternative is to overlay a lowess curve.
4. Due to the potentially large number of plots, it can be somewhat
tricky to provide the axes labels in a way that is both informative
and visually pleasing. One alternative that seems to work well is
to provide axis labels on alternating rows and columns. That is,
row one will have tic marks and axis labels on the left vertical
axis for the first plot only while row two will have the tic marks
and axis labels for the right vertical axis for the last plot in the
row only. This alternating pattern continues for the remaining
rows. A similar pattern is used for the columns and the horizontal
axes labels. Another alternative is to put the minimum and
maximum scale value in the diagonal plot with the variable

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33qb.htm (1 of 3) [5/7/2002 3:14:54 PM]

1.3.3.26.11. Scatterplot Matrix

name.
5. Some analysts prefer to connect the scatter plots. Others prefer to
leave a little gap between each plot.
6. Although this plot type is most commonly used for scatter plots,
the basic concept is both simple and powerful and extends easily
to other plot formats that involve pairwise plots such as the
quantile-quantile plot and the bihistogram.

Sample Plot

This sample plot was generated from pollution data collected by NIST
chemist Lloyd Currie.
There are a number of ways to view this plot. If we are primarily
interested in a particular variable, we can scan the row and column for
that variable. If we are interested in finding the strongest relationship,
we can scan all the plots and then determine which variables are
related.

Definition Given k variables, scatter plot matrices are formed by creating k rows
and k columns. Each row and column defines a single scatter plot
The individual plot for row i and column j is defined as
● Vertical axis: Variable Xi

● Horizontal axis: Variable Xj

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33qb.htm (2 of 3) [5/7/2002 3:14:54 PM]

1.3.3.26.11. Scatterplot Matrix

Questions The scatterplot matrix can provide answers to the following questions:
1. Are there pairwise relationships between the variables?
2. If there are relationships, what is the nature of these
relationships?
3. Are there outliers in the data?
4. Is there clustering by groups in the data?

Linking and The scatterplot matrix serves as the foundation for the concepts of
Brushing linking and brushing.
By linking, we mean showing how a point, or set of points, behaves in
each of the plots. This is accomplished by highlighting these points in
some fashion. For example, the highlighted points could be drawn as a
filled circle while the remaining points could be drawn as unfilled
circles. A typical application of this would be to show how an outlier
shows up in each of the individual pairwise plots. Brushing extends this
concept a bit further. In brushing, the points to be highlighted are
interactively selected by a mouse and the scatterplot matrix is
dynamically updated (ideally in real time). That is, we can select a
rectangular region of points in one plot and see how those points are
reflected in the other plots. Brushing is discussed in detail by Becker,
Cleveland, and Wilks in the paper "Dynamic Graphics for Data
Analysis" (Cleveland and McGill, 1988).

Related Star plot

Techniques Scatter plot
Conditioning plot
Locally weighted least squares

Software Scatterplot matrices are becoming increasingly common in general

purpose statistical software programs, including Dataplot. If a software
program does not generate scatterplot matrices, but it does provide
multiple plots per page and scatter plots, it should be possible to write a
macro to generate a scatterplot matrix. Brushing is available in a few of
the general purpose statistical software programs that emphasize
graphical approaches.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33qb.htm (3 of 3) [5/7/2002 3:14:54 PM]

1.3.3.26.12. Conditioning Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.26. Scatter Plot

1.3.3.26.12. Conditioning Plot

Purpose: A conditioning plot, also known as a coplot or subset plot, is a plot of
Check two variables conditional on the value of a third variable (called the
pairwise conditioning variable). The conditioning variable may be either a
relationship variable that takes on only a few discrete values or a continuous variable
between two that is divided into a limited number of subsets.
variables
conditional One limitation of the scatterplot matrix is that it cannot show interaction
on a third effects with another variable. This is the strength of the conditioning
variable plot. It is also useful for displaying scatter plots for groups in the data.
Although these groups can also be plotted on a single plot with different
plot symbols, it can often be visually easier to distinguish the groups
using the conditioning plot.
Although the basic concept of the conditioning plot matrix is simple,
there are numerous alternatives in the details of the plots.
1. It can be helpful to overlay some type of fitted curve on the
scatter plot. Although a linear or quadratic fit can be used, the
most common alternative is to overlay a lowess curve.
2. Due to the potentially large number of plots, it can be somewhat
tricky to provide the axis labels in a way that is both informative
and visually pleasing. One alternative that seems to work well is
to provide axis labels on alternating rows and columns. That is,
row one will have tic marks and axis labels on the left vertical
axis for the first plot only while row two will have the tic marks
and axis labels for the right vertical axis for the last plot in the
row only. This alternating pattern continues for the remaining
rows. A similar pattern is used for the columns and the horizontal
axis labels. Note that this approach only works if the axes limits
are fixed to common values for all of the plots.
3. Some analysts prefer to connect the scatter plots. Others prefer to
leave a little gap between each plot. Alternatively, each plot can
have its own labeling with the plots not connected.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33qc.htm (1 of 3) [5/7/2002 3:14:54 PM]

1.3.3.26.12. Conditioning Plot

4. Although this plot type is most commonly used for scatter plots,
the basic concept is both simple and powerful and extends easily
to other plot formats.

Sample Plot

In this case, temperature has six distinct values. We plot torque versus
time for each of these temperatures. This example is discussed in more
detail in the process modeling chapter.

Definition Given the variables X, Y, and Z, the conditioning plot is formed by

dividing the values of Z into k groups. There are several ways that these
groups may be formed. There may be a natural grouping of the data, the
data may be divided into several equal sized groups, the grouping may
be determined by clusters in the data, and so on. The page will be
divided into n rows and c columns where . Each row and
column defines a single scatter plot.
The individual plot for row i and column j is defined as
● Vertical axis: Variable Y

● Horizontal axis: Variable X

where only the points in the group corresponding to the ith row and jth
column are used.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33qc.htm (2 of 3) [5/7/2002 3:14:54 PM]

1.3.3.26.12. Conditioning Plot

Questions The conditioning plot can provide answers to the following questions:
1. Is there a relationship between two variables?
2. If there is a relationship, does the nature of the relationship
depend on the value of a third variable?
3. Are groups in the data similar?
4. Are there outliers in the data?

Related Scatter plot

Techniques Scatterplot matrix
Locally weighted least squares

Software Scatter plot matrices are becoming increasingly common in general

purpose statistical software programs, including Dataplot. If a software
program does not generate conditioning plots, but it does provide
multiple plots per page and scatter plots, it should be possible to write a
macro to generate a conditioning plot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33qc.htm (3 of 3) [5/7/2002 3:14:54 PM]

1.3.3.27. Spectral Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.27. Spectral Plot

Purpose: A spectral plot ( Jenkins and Watts 1968 or Bloomfield 1976) is a
Examine graphical technique for examining cyclic structure in the frequency
Cyclic domain. It is a smoothed Fourier transform of the autocovariance
Structure function.
The frequency is measured in cycles per unit time where unit time is
defined to be the distance between 2 points. A frequency of 0
corresponds to an infinite cycle while a frequency of 0.5 corresponds to
a cycle of 2 data points. Equi-spaced time series are inherently limited to
detecting frequencies between 0 and 0.5.
Trends should typically be removed from the time series before
applying the spectral plot. Trends can be detected from a run sequence
plot. Trends are typically removed by differencing the series or by
fitting a straight line (or some other polynomial curve) and applying the
spectral analysis to the residuals.
Spectral plots are often used to find a starting value for the frequency,
, in the sinusoidal model

See the beam deflection case study for an example of this.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33r.htm (1 of 3) [5/7/2002 3:14:54 PM]

1.3.3.27. Spectral Plot

Sample Plot

This spectral plot shows one dominant frequency of approximately 0.3

cycles per observation.

Definition: The spectral plot is formed by:

Variance ● Vertical axis: Smoothed variance (power)
Versus
● Horizontal axis: Frequency (cycles per observation)
Frequency
The computations for generating the smoothed variances can be
involved and are not discussed further here. The details can be found in
the Jenkins and Bloomfield references and in most texts that discuss the
frequency analysis of time series.

Questions The spectral plot can be used to answer the following questions:
1. How many cyclic components are there?
2. Is there a dominant cyclic frequency?
3. If there is a dominant cyclic frequency, what is it?

Importance The spectral plot is the primary technique for assessing the cyclic nature
Check of univariate time series in the frequency domain. It is almost always the
Cyclic second plot (after a run sequence plot) generated in a frequency domain
Behavior of analysis of a time series.
Time Series

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33r.htm (2 of 3) [5/7/2002 3:14:54 PM]

1.3.3.27. Spectral Plot

Examples 1. Random (= White Noise)

2. Strong autocorrelation and autoregressive model
3. Sinusoidal model

Related Autocorrelation Plot

Techniques Complex Demodulation Amplitude Plot
Complex Demodulation Phase Plot

Case Study The spectral plot is demonstrated in the beam deflection data case study.

Software Spectral plots are a fundamental technique in the frequency analysis of

time series. They are available in many general purpose statistical
software programs, including Dataplot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33r.htm (3 of 3) [5/7/2002 3:14:54 PM]

1.3.3.27.1. Spectral Plot: Random Data

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.27. Spectral Plot

1.3.3.27.1. Spectral Plot: Random Data

Spectral
Plot of 200
Normal
Random
Numbers

Conclusions We can make the following conclusions from the above plot.
1. There are no dominant peaks.
2. There is no identifiable pattern in the spectrum.
3. The data are random.

Discussion For random data, the spectral plot should show no dominant peaks or
distinct pattern in the spectrum. For the sample plot above, there are no
clearly dominant peaks and the peaks seem to fluctuate at random. This
type of appearance of the spectral plot indicates that there are no
significant cyclic patterns in the data.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33r1.htm (1 of 2) [5/7/2002 3:14:55 PM]

1.3.3.27.1. Spectral Plot: Random Data

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33r1.htm (2 of 2) [5/7/2002 3:14:55 PM]

1.3.3.27.2. Spectral Plot: Strong Autocorrelation and Autoregressive Model

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.27. Spectral Plot

1.3.3.27.2. Spectral Plot: Strong

Autocorrelation and
Autoregressive Model
Spectral Plot
for Random
Walk Data

Conclusions We can make the following conclusions from the above plot.
1. Strong dominant peak near zero.
2. Peak decays rapidly towards zero.
3. An autoregressive model is an appropriate model.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33r2.htm (1 of 2) [5/7/2002 3:14:55 PM]

1.3.3.27.2. Spectral Plot: Strong Autocorrelation and Autoregressive Model

Discussion This spectral plot starts with a dominant peak near zero and rapidly
decays to zero. This is the spectral plot signature of a process with
strong positive autocorrelation. Such processes are highly non-random
in that there is high association between an observation and a
succeeding observation. In short, if you know Yi you can make a
strong guess as to what Yi+1 will be.

Recommended The next step would be to determine the parameters for the
Next Step autoregressive model:

Such estimation can be done by linear regression or by fitting a

Box-Jenkins autoregressive (AR) model.

The residual standard deviation for this autoregressive model will be

much smaller than the residual standard deviation for the default
model

Then the system should be reexamined to find an explanation for the

strong autocorrelation. Is it due to the
1. phenomenon under study; or
2. drifting in the environment; or
3. contamination from the data acquisition system (DAS)?
Oftentimes the source of the problem is item (3) above where
contamination and carry-over from the data acquisition system result
because the DAS does not have time to electronically recover before
collecting the next data point. If this is the case, then consider slowing
down the sampling rate to re-achieve randomness.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33r2.htm (2 of 2) [5/7/2002 3:14:55 PM]

1.3.3.27.3. Spectral Plot: Sinusoidal Model

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.27. Spectral Plot

1.3.3.27.3. Spectral Plot: Sinusoidal Model

Spectral Plot
for Sinusoidal
Model

Conclusions We can make the following conclusions from the above plot.
1. There is a single dominant peak at approximately 0.3.
2. There is an underlying single-cycle sinusoidal model.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33r3.htm (1 of 2) [5/7/2002 3:14:55 PM]

1.3.3.27.3. Spectral Plot: Sinusoidal Model

Discussion This spectral plot shows a single dominant frequency. This indicates
that a single-cycle sinusoidal model might be appropriate.
If one were to naively assume that the data represented by the graph
could be fit by the model

and then estimate the constant by the sample mean, the analysis would
be incorrect because
● the sample mean is biased;

● the confidence interval for the mean, which is valid only for
random data, is meaningless and too small.
On the other hand, the choice of the proper model

where is the amplitude, is the frequency (between 0 and .5 cycles

per observation), and is the phase can be fit by non-linear least
squares. The beam deflection data case study demonstrates fitting this
type of model.

Recommended The recommended next steps are to:

Next Steps 1. Estimate the frequency from the spectral plot. This will be
helpful as a starting value for the subsequent non-linear fitting.
A complex demodulation phase plot can be used to fine tune the
estimate of the frequency before performing the non-linear fit.
2. Do a complex demodulation amplitude plot to obtain an initial
estimate of the amplitude and to determine if a constant
amplitude is justified.
3. Carry out a non-linear fit of the model

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33r3.htm (2 of 2) [5/7/2002 3:14:55 PM]

1.3.3.28. Standard Deviation Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.28. Standard Deviation Plot

Purpose: Standard deviation plots are used to see if the standard deviation varies
Detect between different groups of the data. The grouping is determined by the
Changes in analyst. In most cases, the data provide a specific grouping variable. For
Scale example, the groups may be the levels of a factor variable. In the sample
Between plot below, the months of the year provide the grouping.
Groups
Standard deviation plots can be used with ungrouped data to determine
if the standard deviation is changing over time. In this case, the data are
broken into an arbitrary number of equal-sized groups. For example, a
data series with 400 points can be divided into 10 groups of 40 points
each. A standard deviation plot can then be generated with these groups
to see if the standard deviation is increasing or decreasing over time.
Although the standard deviation is the most commonly used measure of
scale, the same concept applies to other measures of scale. For example,
instead of plotting the standard deviation of each group, the median
absolute deviation or the average absolute deviation might be plotted
instead. This might be done if there were significant outliers in the data
and a more robust measure of scale than the standard deviation was
desired.
Standard deviation plots are typically used in conjunction with mean
plots. The mean plot would be used to check for shifts in location while
the standard deviation plot would be used to check for shifts in scale.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33s.htm (1 of 3) [5/7/2002 3:14:55 PM]

1.3.3.28. Standard Deviation Plot

Sample Plot

This sample standard deviation plot shows

1. there is a shift in variation;
2. greatest variation is during the summer months.

Definition: Standard deviation plots are formed by:

Group ● Vertical axis: Group standard deviations
Standard
● Horizontal axis: Group identifier
Deviations
Versus A reference line is plotted at the overall standard deviation.
Group ID

Questions The standard deviation plot can be used to answer the following
questions.
1. Are there any shifts in variation?
2. What is the magnitude of the shifts in variation?
3. Is there a distinct pattern in the shifts in variation?

Importance: A common assumption in 1-factor analyses is that of equal variances.

Checking That is, the variance is the same for different levels of the factor
Assumptions variable. The standard deviation plot provides a graphical check for that
assumption. A common assumption for univariate data is that the
variance is constant. By grouping the data into equi-sized intervals, the
standard deviation plot can provide a graphical test of this assumption.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33s.htm (2 of 3) [5/7/2002 3:14:55 PM]

1.3.3.28. Standard Deviation Plot

Related Mean Plot

Techniques Dex Standard Deviation Plot

Software Most general purpose statistical software programs do not support a

standard deviation plot. However, if the statistical program can generate
the standard deviation for a group, it should be feasible to write a macro
to generate this plot. Dataplot supports a standard deviation plot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33s.htm (3 of 3) [5/7/2002 3:14:55 PM]

1.3.3.29. Star Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.29. Star Plot

Purpose: The star plot (Chambers 1983) is a method of displaying multivariate
Display data. Each star represents a single observation. Typically, star plots are
Multivariate generated in a multi-plot format with many stars on each page and each
Data star representing one observation.
Star plots are used to examine the relative values for a single data point
(e.g., point 3 is large for variables 2 and 4, small for variables 1, 3, 5,
and 6) and to locate similar points or dissimilar points.

Sample Plot The plot below contains the star plots of 16 cars. The data file actually
contains 74 cars, but we restrict the plot to what can reasonably be
shown on one page. The variable list for the sample star plot is
1 Price
2 Mileage (MPG)
3 1978 Repair Record (1 = Worst, 5 = Best)
4 1977 Repair Record (1 = Worst, 5 = Best)
5 Headroom
6 Rear Seat Room
7 Trunk Space
8 Weight
9 Length

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33t.htm (1 of 3) [5/7/2002 3:14:56 PM]

1.3.3.29. Star Plot

We can look at these plots individually or we can use them to identify

clusters of cars with similar features. For example, we can look at the
star plot of the Cadillac Seville and see that it is one of the most
expensive cars, gets below average (but not among the worst) gas
mileage, has an average repair record, and has average-to-above-average
roominess and size. We can then compare the Cadillac models (the last
three plots) with the AMC models (the first three plots). This
comparison shows distinct patterns. The AMC models tend to be
inexpensive, have below average gas mileage, and are small in both
height and weight and in roominess. The Cadillac models are expensive,
have poor gas mileage, and are large in both size and roominess.

Definition The star plot consists of a sequence of equi-angular spokes, called radii,
with each spoke representing one of the variables. The data length of a
spoke is proportional to the magnitude of the variable for the data point
relative to the maximum magnitude of the variable across all data
points. A line is drawn connecting the data values for each spoke. This
gives the plot a star-like appearance and the origin of the name of this
plot.

Questions The star plot can be used to answer the following questions:
1. What variables are dominant for a given observation?
2. Which observations are most similar, i.e., are there clusters of
observations?
3. Are there outliers?

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33t.htm (2 of 3) [5/7/2002 3:14:56 PM]

1.3.3.29. Star Plot

Weakness in Star plots are helpful for small-to-moderate-sized multivariate data sets.
Technique Their primary weakness is that their effectiveness is limited to data sets
with less than a few hundred points. After that, they tend to be
overwhelming.
Graphical techniques suited for large data sets are discussed by Scott.

Related Alternative ways to plot multivariate data are discussed in Chambers, du

Techniques Toit, and Everitt.

Software Star plots are available in some general purpose statistical software
progams, including Dataplot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33t.htm (3 of 3) [5/7/2002 3:14:56 PM]

1.3.3.30. Weibull Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.30. Weibull Plot

Purpose: The Weibull plot (Nelson 1982) is a graphical technique for
Graphical determining if a data set comes from a population tha would logically
Check To See be fit by a 2-parameter Weibull distribution (the location is assumed to
If Data Come be zero).
From a
Population The Weibull plot has special scales that are designed so that if the data
That Would do in fact follow a Weibull distribution, the points will be linear (or
Be Fit by a nearly linear). The least squares fit of this line yields estimates for the
Weibull shape and scale parameters of the Weibull distribution. Weibull
Distribution distribution (the location is assumed to be zero).

Sample Plot

This Weibull plot shows that:

1. the assumption of a Weibull distribution is reasonable;
2. the shape parameter estimate is computed to be 33.32;
3. the scale parameter estimate is computed to be 5.28; and

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33u.htm (1 of 3) [5/7/2002 3:14:56 PM]

1.3.3.30. Weibull Plot

4. there are no outliers.

Definition: The Weibull plot is formed by:

Weibull ● Vertical axis: Weibull cumulative probability expressed as a
Cumulative percentage
Probability
● Horizontal axis: LN of ordered response
Versus
LN(Ordered The vertical scale is ln-ln(1-p) where p=(i-0.3)/(n+0.4) and i is the rank
Response) of the observation. This scale is chosen in order to linearize the
resulting plot for Weibull data.

Questions The Weibull plot can be used to answer the following questions:
1. Do the data follow a 2-parameter Weibull distribution?
2. What is the best estimate of the shape parameter for the
2-parameter Weibull distribution?
3. What is the best estimate of the scale (= variation) parameter for
the 2-parameter Weibull distribution?

Importance: Many statistical analyses, particularly in the field of reliability, are

Check based on the assumption that the data follow a Weibull distribution. If
Distributional the analysis assumes the data follow a Weibull distribution, it is
Assumptions important to verify this assumption and, if verified, find good estimates
of the Weibull parameters.

Related Weibull Probability Plot

Techniques Weibull PPCC Plot
Weibull Hazard Plot

The Weibull probability plot (in conjunction with the Weibull PPCC
plot), the Weibull hazard plot, and the Weibull plot are all similar
techniques that can be used for assessing the adequacy of the Weibull
distribution as a model for the data, and additionally providing
estimation for the shape, scale, or location parameters.
The Weibull hazard plot and Weibull plot are designed to handle
censored data (which the Weibull probability plot does not).

Case Study The Weibull plot is demonstrated in the airplane glass failure data case
study.

Software Weibull plots are generally available in statistical software programs

that are designed to analyze reliability data. Dataplot supports the
Weibull plot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33u.htm (2 of 3) [5/7/2002 3:14:56 PM]

1.3.3.30. Weibull Plot

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33u.htm (3 of 3) [5/7/2002 3:14:56 PM]

1.3.3.31. Youden Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.31. Youden Plot

Purpose: Youden plots are a graphical technique for analyzing interlab data when
Interlab each lab has made two runs on the same product or one run on two
Comparisons different products.
The Youden plot is a simple but effective method for comparing both
the within-laboratory variability and the between-laboratory variability.

Sample Plot

This plot shows:

1. Not all labs are equivalent.
2. Lab 4 is biased low.
3. Lab 3 has within-lab variability problems.
4. Lab 5 has an outlying run.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3331.htm (1 of 2) [5/7/2002 3:14:56 PM]

1.3.3.31. Youden Plot

Definition: Youden plots are formed by:

Response 1 1. Vertical axis: Response variable 1 (i.e., run 1 or product 1
Versus response value)
Response 2
2. Horizontal axis: Response variable 2 (i.e., run 2 or product 2
Coded by
response value)
Lab
In addition, the plot symbol is the lab id (typically an integer from 1 to k
where k is the number of labs). Sometimes a 45-degree reference line is
drawn. Ideally, a lab generating two runs of the same product should
produce reasonably similar results. Departures from this reference line
indicate inconsistency from the lab. If two different products are being
tested, then a 45-degree line may not be appropriate. However, if the
labs are consistent, the points should lie near some fitted straight line.

Questions The Youden plot can be used to answer the following questions:
1. Are all labs equivalent?
2. What labs have between-lab problems (reproducibility)?
3. What labs have within-lab problems (repeatability)?
4. What labs are outliers?

Importance In interlaboratory studies or in comparing two runs from the same lab, it
is useful to know if consistent results are generated. Youden plots
should be a routine plot for analyzing this type of data.

DEX Youden The dex Youden plot is a specialized Youden plot used in the design of
Plot experiments. In particular, it is useful for full and fractional designs.

Related Scatter Plot

Techniques

Software The Youden plot is essentially a scatter plot, so it should be feasible to

write a macro for a Youden plot in any general purpose statistical
program that supports scatter plots. Dataplot supports a Youden plot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3331.htm (2 of 2) [5/7/2002 3:14:56 PM]

1.3.3.31.1. DEX Youden Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic
1.3.3.31. Youden Plot

1.3.3.31.1. DEX Youden Plot

DEX Youden The dex (Design of Experiments) Youden plot is a specialized Youden
Plot: plot used in the analysis of full and fractional experiment designs. In
Introduction particular, it is used in support of a Yates analysis. These designs may
have a low level, coded as "-1" or "-", and a high level, coded as "+1"
or "+", for each factor. In addition, there can optionally be one or more
center points. Center points are at the midpoint between the low and
high levels for each factor and are coded as "0".
The Yates analysis and the the dex Youden plot only use the "-1" and
"+1" points. The Yates analysis is used to estimate factor effects. The
dex Youden plot can be used to help determine the approriate model to
use from the Yates analysis.

Construction The following are the primary steps in the construction of the dex
of DEX Youden plot.
Youden Plot
1. For a given factor or interaction term, compute the mean of the
response variable for the low level of the factor and for the high
level of the factor. Any center points are omitted from the
computation.
2. Plot the point where the y-coordinate is the mean for the high
level of the factor and the x-coordinate is the mean for the low
level of the factor. The character used for the plot point should
identify the factor or interaction term (e.g., "1" for factor 1, "13"
for the interaction between factors 1 and 3).
3. Repeat steps 1 and 2 for each factor and interaction term of the
data.
The high and low values of the interaction terms are obtained by
multiplying the corresponding values of the main level factors. For
example, the interaction term X13 is obtained by multiplying the values
for X1 with the corresponding values of X3. Since the values for X1 and
X3 are either "-1" or "+1", the resulting values for X13 are also either

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33311.htm (1 of 3) [5/7/2002 3:14:57 PM]

1.3.3.31.1. DEX Youden Plot

"-1" or "+1".
In summary, the dex Youden plot is a plot of the mean of the response
variable for the high level of a factor or interaction term against the
mean of the response variable for the low level of that factor or
interaction term.
For unimportant factors and interaction terms, these mean values
should be nearly the same. For important factors and interaction terms,
these mean values should be quite different. So the interpretation of the
plot is that unimportant factors should be clustered together near the
grand mean. Points that stand apart from this cluster identify important
factors that should be included in the model.

Sample DEX The following is a dex Youden plot for the data used in the Eddy
Youden Plot current case study. The analysis in that case study demonstrated that
X1 and X2 were the most important factors.

Interpretation From the above dex Youden plot, we see that factors 1 and 2 stand out
of the Sample from the others. That is, the mean response values for the low and high
DEX Youden levels of factor 1 and factor 2 are quite different. For factor 3 and the 2
Plot and 3-term interactions, the mean response values for the low and high
levels are similar.
We would conclude from this plot that factors 1 and 2 are important
and should be included in our final model while the remaining factors
and interactions should be omitted from the final model.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33311.htm (2 of 3) [5/7/2002 3:14:57 PM]

1.3.3.31.1. DEX Youden Plot

Case Study The Eddy current case study demonstrates the use of the dex Youden
plot in the context of the analysis of a full factorial design.

Software DEX Youden plots are not typically available as built-in plots in
statistical software programs. However, it should be relatively
straightforward to write a macro to generate this plot in most general
purpose statistical software programs.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda33311.htm (3 of 3) [5/7/2002 3:14:57 PM]

1.3.3.32. 4-Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.32. 4-Plot
Purpose: The 4-plot is a collection of 4 specific EDA graphical techniques
Check whose purpose is to test the assumptions that underlie most
Underlying measurement processes. A 4-plot consists of a
Statistical 1. run sequence plot;
Assumptions
2. lag plot;
3. histogram;
4. normal probability plot.
If the 4 underlying assumptions of a typical measurement process
hold, then the above 4 plots will have a characteristic appearance (see
the normal random numbers case study below); if any of the
underlying assumptions fail to hold, then it will be revealed by an
anomalous appearance in one or more of the plots. Several commonly
encountered situations are demonstrated in the case studies below.
Although the 4-plot has an obvious use for univariate and time series
data, its usefulness extends far beyond that. Many statistical models of
the form

have the same underlying assumptions for the error term. That is, no
matter how complicated the functional fit, the assumptions on the
underlying error term are still the same. The 4-plot can and should be
routinely applied to the residuals when fitting models regardless of
whether the model is simple or complicated.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3332.htm (1 of 5) [5/7/2002 3:14:57 PM]

1.3.3.32. 4-Plot

Sample Plot:
Process Has
Fixed
Location,
Fixed
Variation,
Non-Random
(Oscillatory),
Non-Normal
U-Shaped
Distribution,
and Has 3
Outliers.

This 4-plot reveals the following:

1. the fixed location assumption is justified as shown by the run
sequence plot in the upper left corner.
2. the fixed variation assumption is justified as shown by the run
sequence plot in the upper left corner.
3. the randomness assumption is violated as shown by the
non-random (oscillatory) lag plot in the upper right corner.
4. the assumption of a common, normal distribution is violated as
shown by the histogram in the lower left corner and the normal
probability plot in the lower right corner. The distribution is
non-normal and is a U-shaped distribution.
5. there are several outliers apparent in the lag plot in the upper
right corner.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3332.htm (2 of 5) [5/7/2002 3:14:57 PM]

1.3.3.32. 4-Plot

Definition: The 4-plot consists of the following:

1. Run 1. Run sequence plot to test fixed location and variation.
Sequence
❍ Vertically: Yi
Plot;
2. Lag Plot; Horizontally: i
❍
3. Histogram; 2. Lag Plot to test randomness.
4. Normal ❍ Vertically: Yi
Probability
Plot ❍ Horizontally: Yi-1
3. Histogram to test (normal) distribution.
❍ Vertically: Counts

❍ Horizontally: Y

4. Normal probability plot to test normal distribution.

❍ Vertically: Ordered Yi

❍ Horizontally: Theoretical values from a normal N(0,1)

distribution for ordered Yi

Questions 4-plots can provide answers to many questions:

1. Is the process in-control, stable, and predictable?
2. Is the process drifting with respect to location?
3. Is the process drifting with respect to variation?
4. Are the data random?
5. Is an observation related to an adjacent observation?
6. If the data are a time series, is is white noise?
7. If the data are a time series and not white noise, is it sinusoidal,
autoregressive, etc.?
8. If the data are non-random, what is a better model?
9. Does the process follow a normal distribution?
10. If non-normal, what distribution does the process follow?
11. Is the model

valid and sufficient?

12. If the default model is insufficient, what is a better model?
13. Is the formula valid?
14. Is the sample mean a good estimator of the process location?
15. If not, what would be a better estimator?
16. Are there any outliers?

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3332.htm (3 of 5) [5/7/2002 3:14:57 PM]

1.3.3.32. 4-Plot

Importance: There are 4 assumptions that typically underlie all measurement

Testing processes; namely, that the data from the process at hand "behave
Underlying like":
Assumptions 1. random drawings;
Helps Ensure
2. from a fixed distribution;
the Validity of
the Final 3. with that distribution having a fixed location; and
Scientific and 4. with that distribution having fixed variation.
Engineering Predictability is an all-important goal in science and engineering. If
Conclusions the above 4 assumptions hold, then we have achieved probabilistic
predictability--the ability to make probability statements not only
about the process in the past, but also about the process in the future.
In short, such processes are said to be "statistically in control". If the 4
assumptions do not hold, then we have a process that is drifting (with
respect to location, variation, or distribution), is unpredictable, and is
out of control. A simple characterization of such processes by a
location estimate, a variation estimate, or a distribution "estimate"
inevitably leads to optimistic and grossly invalid engineering
conclusions.
Inasmuch as the validity of the final scientific and engineering
conclusions is inextricably linked to the validity of these same 4
underlying assumptions, it naturally follows that there is a real
necessity for all 4 assumptions to be routinely tested. The 4-plot (run
sequence plot, lag plot, histogram, and normal probability plot) is seen
as a simple, efficient, and powerful way of carrying out this routine
checking.

Interpretation: Of the 4 underlying assumptions:

Flat, 1. If the fixed location assumption holds, then the run sequence
Equi-Banded, plot will be flat and non-drifting.
Random,
2. If the fixed variation assumption holds, then the vertical spread
Bell-Shaped,
in the run sequence plot will be approximately the same over
and Linear
the entire horizontal axis.
3. If the randomness assumption holds, then the lag plot will be
structureless and random.
4. If the fixed distribution assumption holds (in particular, if the
fixed normal distribution assumption holds), then the histogram
will be bell-shaped and the normal probability plot will be
approximatelylinear.
If all 4 of the assumptions hold, then the process is "statistically in
control". In practice, many processes fall short of achieving this ideal.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3332.htm (4 of 5) [5/7/2002 3:14:57 PM]

1.3.3.32. 4-Plot

Related Run Sequence Plot

Techniques Lag Plot
Histogram
Normal Probability Plot

Autocorrelation Plot
Spectral Plot
PPCC Plot

Case Studies The 4-plot is used in most of the case studies in this chapter:
1. Normal random numbers (the ideal)
2. Uniform random numbers
3. Random walk
4. Josephson junction cryothermometry
5. Beam deflections
6. Filter transmittance
7. Standard resistor
8. Heat flow meter 1

Software It should be feasible to write a macro for the 4-plot in any general
purpose statistical software program that supports the capability for
multiple plots per page and supports the underlying plot techniques.
Dataplot supports the 4-plot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3332.htm (5 of 5) [5/7/2002 3:14:57 PM]

1.3.3.33. 6-Plot

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.3. Graphical Techniques: Alphabetic

1.3.3.33. 6-Plot
Purpose: The 6-plot is a collection of 6 specific graphical techniques whose
Graphical purpose is to assess the validity of a Y versus X fit. The fit can be a
Model linear fit, a non-linear fit, a LOWESS (locally weighted least squares)
Validation fit, a spline fit, or any other fit utilizing a single independent variable.
The 6 plots are:
1. Scatter plot of the response and predicted values versus the
independent variable;
2. Scatter plot of the residuals versus the independent variable;
3. Scatter plot of the residuals versus the predicted values;
4. Lag plot of the residuals;
5. Histogram of the residuals;
6. Normal probability plot of the residuals.

Sample Plot

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3333.htm (1 of 4) [5/7/2002 3:14:57 PM]

1.3.3.33. 6-Plot

This 6-plot, which followed a linear fit, shows that the linear model is
not adequate. It suggests that a quadratic model would be a better
model.

Definition: The 6-plot consists of the following:

6 1. Response and predicted values
Component
❍ Vertical axis: Response variable, predicted values
Plots
❍ Horizontal axis: Independent variable

2. Residuals versus independent variable

❍ Vertical axis: Residuals

❍ Horizontal axis: Independent variable

3. Residuals versus predicted values

❍ Vertical axis: Residuals

❍ Horizontal axis: Predicted values

4. Lag plot of residuals

❍ Vertical axis: RES(I)

❍ Horizontal axis: RES(I-1)

5. Histogram of residuals
❍ Vertical axis: Counts

❍ Horizontal axis: Residual values

6. Normal probability plot of residuals

❍ Vertical axis: Ordered residuals

❍ Horizontal axis: Theoretical values from a normal N(0,1)

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3333.htm (2 of 4) [5/7/2002 3:14:57 PM]

1.3.3.33. 6-Plot

distribution for ordered residuals

Questions The 6-plot can be used to answer the following questions:

1. Are the residuals approximately normally distributed with a fixed
location and scale?
2. Are there outliers?
3. Is the fit adequate?
4. Do the residuals suggest a better fit?

Importance: A model involving a response variable and a single independent variable

Validating has the form:
Model

where Y is the response variable, X is the independent variable, f is the

linear or non-linear fit function, and E is the random component. For a
good model, the error component should behave like:
1. random drawings (i.e., independent);
2. from a fixed distribution;
3. with fixed location; and
4. with fixed variation.
In addition, for fitting models it is usually further assumed that the fixed
distribution is normal and the fixed location is zero. For a good model
the fixed variation should be as small as possible. A necessary
component of fitting models is to verify these assumptions for the error
component and to assess whether the variation for the error component
is sufficiently small. The histogram, lag plot, and normal probability
plot are used to verify the fixed distribution, location, and variation
assumptions on the error component. The plot of the response variable
and the predicted values versus the independent variable is used to
assess whether the variation is sufficiently small. The plots of the
residuals versus the independent variable and the predicted values is
used to assess the independence assumption.
Assessing the validity and quality of the fit in terms of the above
assumptions is an absolutely vital part of the model-fitting process. No
fit should be considered complete without an adequate model validation
step.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3333.htm (3 of 4) [5/7/2002 3:14:57 PM]

1.3.3.33. 6-Plot

Related Linear Least Squares

Techniques Non-Linear Least Squares
Scatter Plot
Run Sequence Plot
Lag Plot
Normal Probability Plot
Histogram

Case Study The 6-plot is used in the Alaska pipeline data case study.

Software It should be feasible to write a macro for the 6-plot in any general
purpose statistical software program that supports the capability for
multiple plots per page and supports the underlying plot techniques.
Dataplot supports the 6-plot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3333.htm (4 of 4) [5/7/2002 3:14:57 PM]

1.3.4. Graphical Techniques: By Problem Category

1. Exploratory Data Analysis

1.3. EDA Techniques

1.3.4. Graphical
Techniques: By
Problem
Category
Univariate
y=c+e

Run Sequence Lag Plot: Histogram:

Plot: 1.3.3.25 1.3.3.15 1.3.3.14

Normal 4-Plot: 1.3.3.32 PPCC Plot:

Probability Plot: 1.3.3.23
1.3.3.21

Weibull Plot: Probability Plot: Box-Cox

1.3.3.30 1.3.3.22 Linearity Plot:
1.3.3.5

http://www.itl.nist.gov/div898/handbook/eda/section3/eda34.htm (1 of 4) [5/7/2002 3:14:58 PM]

1.3.4. Graphical Techniques: By Problem Category

Box-Cox Bootstrap Plot:

Normality Plot: 1.3.3.4
1.3.3.6

Time Series
y = f(t) + e

Run Sequence Spectral Plot: Autocorrelation

Plot: 1.3.3.25 1.3.3.27 Plot: 1.3.3.1

Complex Complex
Demodulation Demodulation
Amplitude Plot: Phase Plot:
1.3.3.8 1.3.3.9

1 Factor
y = f(x) + e

Scatter Plot: Box Plot: 1.3.3.7 Bihistogram:

1.3.3.26 1.3.3.2

http://www.itl.nist.gov/div898/handbook/eda/section3/eda34.htm (2 of 4) [5/7/2002 3:14:58 PM]

1.3.4. Graphical Techniques: By Problem Category

Quantile-Quantile Mean Plot: Standard

Plot: 1.3.3.24 1.3.3.20 Deviation Plot:
1.3.3.28

Multi-Factor/Comparative
y = f(xp, x1,x2,...,xk) + e

Block Plot:
1.3.3.3

Multi-Factor/Screening
y = f(x1,x2,x3,...,xk) + e

DEX Scatter DEX Mean Plot: DEX Standard

Plot: 1.3.3.11 1.3.3.12 Deviation Plot:
1.3.3.13

Contour Plot:
1.3.3.10

http://www.itl.nist.gov/div898/handbook/eda/section3/eda34.htm (3 of 4) [5/7/2002 3:14:58 PM]

1.3.4. Graphical Techniques: By Problem Category

Regression
y = f(x1,x2,x3,...,xk) + e

Scatter Plot: 6-Plot: 1.3.3.33 Linear

1.3.3.26 Correlation Plot:
1.3.3.16

Linear Intercept Linear Slope Linear Residual

Plot: 1.3.3.17 Plot: 1.3.3.18 Standard
Deviation
Plot:1.3.3.19

Interlab
(y1,y2) = f(x) + e

Youden Plot:
1.3.3.31

Multivariate
(y1,y2,...,yp)

Star Plot:
1.3.3.29

http://www.itl.nist.gov/div898/handbook/eda/section3/eda34.htm (4 of 4) [5/7/2002 3:14:58 PM]

1.3.5. Quantitative Techniques

1. Exploratory Data Analysis

1.3. EDA Techniques

1.3.5. Quantitative Techniques

Confirmatory The techniques discussed in this section are classical statistical methods
Statistics as opposed to EDA techniques. EDA and classical techniques are not
mutually exclusive and can be used in a complamentary fashion. For
example, the analysis can start with some simple graphical techniques
such as the 4-plot followed by the classical confirmatory methods
discussed herein to provide more rigorous statments about the
conclusions. If the classical methods yield different conclusions than
the graphical analysis, then some effort should be invested to explain
why. Often this is an indication that some of the assumptions of the
classical techniques are violated.
Many of the quantitative techniques fall into two broad categories:
1. Interval estimation
2. Hypothesis tests

Interval It is common in statistics to estimate a parameter from a sample of data.

Estimates The value of the parameter using all of the possible data, not just the
sample data, is called the population parameter or true value of the
parameter. An estimate of the true parameter value is made using the
sample data. This is called a point estimate or a sample estimate.
For example, the most commonly used measure of location is the mean.
The population, or true, mean is the sum of all the members of the
given population divided by the number of members in the population.
As it is typically impractical to measure every member of the
population, a random sample is drawn from the population. The sample
mean is calculated by summing the values in the sample and dividing
by the number of values in the sample. This sample mean is then used
as the point estimate of the population mean.
Interval estimates expand on point estimates by incorporating the
uncertainty of the point estimate. In the example for the mean above,
different samples from the same population will generate different
values for the sample mean. An interval estimate quantifies this
uncertainty in the sample estimate by computing lower and upper

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35.htm (1 of 4) [5/7/2002 3:14:58 PM]

1.3.5. Quantitative Techniques

values of an interval which will, with a given level of confidence (i.e.,

probability), contain the population parameter.

Hypothesis Hypothesis tests also address the uncertainty of the sample estimate.
Tests However, instead of providing an interval, a hypothesis test attempts to
refute a specific claim about a population parameter based on the
sample data. For example, the hypothesis might be one of the
following:
● the population mean is equal to 10

● the population standard deviation is equal to 5

● the means from two populations are equal

● the standard deviations from 5 populations are equal

To reject a hypothesis is to conclude that it is false. However, to accept

a hypothesis does not mean that it is true, only that we do not have
evidence to believe otherwise. Thus hypothesis tests are usually stated
in terms of both a condition that is doubted (null hypothesis) and a
condition that is believed (alternative hypothesis).
A common format for a hypothesis test is:
H0: A statement of the null hypothesis, e.g., two
population means are equal.
Ha: A statement of the alternative hypothesis, e.g., two
population means are not equal.
Test Statistic: The test statistic is based on the specific
hypothesis test.
Significance Level: The significance level, , defines the sensitivity of
the test. A value of = 0.05 means that we
inadvertently reject the null hypothesis 5% of the
time when it is in fact true. This is also called the
type I error. The choice of is somewhat
arbitrary, although in practice values of 0.1, 0.05,
and 0.01 are commonly used.
The probability of rejecting the null hypothesis
when it is in fact false is called the power of the
test and is denoted by 1 - . Its complement, the
probability of accepting the null hypothesis when
the alternative hypothesis is, in fact, true (type II
error), is called and can only be computed for a
specific alternative hypothesis.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35.htm (2 of 4) [5/7/2002 3:14:58 PM]

1.3.5. Quantitative Techniques

Critical Region: The critical region encompasses those values of

the test statistic that lead to a rejection of the null
hypothesis. Based on the distribution of the test
statistic and the significance level, a cut-off value
for the test statistic is computed. Values either
above or below or both (depending on the
direction of the test) this cut-off define the critical
region.

Practical It is important to distinguish between statistical significance and

Versus practical significance. Statistical significance simply means that we
Statistical reject the null hypothesis. The ability of the test to detect differences
Significance that lead to rejection of the null hypothesis depends on the sample size.
For example, for a particularly large sample, the test may reject the null
hypothesis that two process means are equivalent. However, in practice
the difference between the two means may be relatively small to the
point of having no real engineering significance. Similarly, if the
sample size is small, a difference that is large in engineering terms may
not lead to rejection of the null hypothesis. The analyst should not just
blindly apply the tests, but should combine engineering judgement with
statistical analysis.

Bootstrap In some cases, it is possible to mathematically derive appropriate

Uncertainty uncertainty intervals. This is particularly true for intervals based on the
Estimates assumption of a normal distribution. However, there are many cases in
which it is not possible to mathematically derive the uncertainty. In
these cases, the bootstrap provides a method for empirically
determining an appropriate interval.

Table of Some of the more common classical quantitative techniques are listed
Contents below. This list of quantitative techniques is by no means meant to be
exhaustive. Additional discussions of classical statistical techniques are
contained in the product comparisons chapter.
● Location
1. Measures of Location
2. Confidence Limits for the Mean and One Sample t-Test
3. Two Sample t-Test for Equal Means
4. One Factor Analysis of Variance
5. Multi-Factor Analysis of Variance
● Scale (or variability or spread)
1. Measures of Scale
2. Bartlett's Test

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35.htm (3 of 4) [5/7/2002 3:14:58 PM]

1.3.5. Quantitative Techniques

3. Chi-Square Test
4. F-Test
5. Levene Test
● Skewness and Kurtosis
1. Measures of Skewness and Kurtosis
● Randomness
1. Autocorrelation
2. Runs Test
● Distributional Measures
1. Anderson-Darling Test
2. Chi-Square Goodness-of-Fit Test
3. Kolmogorov-Smirnov Test
● Outliers
1. Grubbs Test
● 2-Level Factorial Designs
1. Yates Analysis

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35.htm (4 of 4) [5/7/2002 3:14:58 PM]

1.3.5.1. Measures of Location

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques

1.3.5.1. Measures of Location

Location A fundamental task in many statistical analyses is to estimate a location
parameter for the distribution; i.e., to find a typical or central value that
best describes the data.

Definition of The first step is to define what we mean by a typical value. For
Location univariate data, there are three common definitions:
1. mean - the mean is the sum of the data points divided by the
number of data points. That is,

The mean is that value that is most commonly referred to as the

average. We will use the term average as a synonym for the mean
and the term typical value to refer generically to measures of
location.
2. median - the median is the value of the point which has half the
data smaller than that point and half the data larger than that
point. That is, if X1, X2, ... ,XN is a random sample sorted from
smallest value to largest value, then the median is defined as:

3. mode - the mode is the value of the random sample that occurs
with the greatest frequency. It is not necessarily unique. The
mode is typically used in a qualitative fashion. For example, there
may be a single dominant hump in the data perhaps two or more
smaller humps in the data. This is usually evident from a
histogram of the data.
When taking samples from continuous populations, we need to be
somewhat careful in how we define the mode. That is, any

http://www.itl.nist.gov/div898/handbook/eda/section3/eda351.htm (1 of 5) [5/7/2002 3:14:59 PM]

1.3.5.1. Measures of Location

specific value may not occur more than once if the data are
continuous. What may be a more meaningful, if less exact
measure, is the midpoint of the class interval of the histogram
with the highest peak.

Why A natural question is why we have more than one measure of the typical
Different value. The following example helps to explain why these alternative
Measures definitions are useful and necessary.
This plot shows histograms for 10,000 random numbers generated from
a normal, an exponential, a Cauchy, and a lognormal distribution.

Normal The first histogram is a sample from a normal distribution. The mean is
Distribution 0.005, the median is -0.010, and the mode is -0.144 (the mode is
computed as the midpoint of the histogram interval with the highest
peak).
The normal distribution is a symmetric distribution with well-behaved
tails and a single peak at the center of the distribution. By symmetric,
we mean that the distribution can be folded about an axis so that the 2
sides coincide. That is, it behaves the same to the left and right of some
center point. For a normal distribution, the mean, median, and mode are
actually equivalent. The histogram above generates similar estimates for
the mean, median, and mode. Therefore, if a histogram or normal
probability plot indicates that your data are approximated well by a
normal distribution, then it is reasonable to use the mean as the location
estimator.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda351.htm (2 of 5) [5/7/2002 3:14:59 PM]

1.3.5.1. Measures of Location

Exponential The second histogram is a sample from an exponential distribution. The

Distribution mean is 1.001, the median is 0.684, and the mode is 0.254 (the mode is
computed as the midpoint of the histogram interval with the highest
peak).
The exponential distribution is a skewed, i. e., not symmetric,
distribution. For skewed distributions, the mean and median are not the
same. The mean will be pulled in the direction of the skewness. That is,
if the right tail is heavier than the left tail, the mean will be greater than
the median. Likewise, if the left tail is heavier than the right tail, the
mean will be less than the median.
For skewed distributions, it is not at all obvious whether the mean, the
median, or the mode is the more meaningful measure of the typical
value. In this case, all three measures are useful.

Cauchy The third histogram is a sample from a Cauchy distribution. The mean is
Distribution 3.70, the median is -0.016, and the mode is -0.362 (the mode is
computed as the midpoint of the histogram interval with the highest
peak).
For better visual comparison with the other data sets, we restricted the
histogram of the Cauchy distribution to values between -10 and 10. The
full Cauchy data set in fact has a minimum of approximately -29,000
and a maximum of approximately 89,000.
The Cauchy distribution is a symmetric distribution with heavy tails and
a single peak at the center of the distribution. The Cauchy distribution
has the interesting property that collecting more data does not provide a
more accurate estimate of the mean. That is, the sampling distribution of
the mean is equivalent to the sampling distribution of the original data.
This means that for the Cauchy distribution the mean is useless as a
measure of the typical value. For this histogram, the mean of 3.7 is well
above the vast majority of the data. This is caused by a few very
extreme values in the tail. However, the median does provide a useful
measure for the typical value.
Although the Cauchy distribution is an extreme case, it does illustrate
the importance of heavy tails in measuring the mean. Extreme values in
the tails distort the mean. However, these extreme values do not distort
the median since the median is based on ranks. In general, for data with
extreme values in the tails, the median provides a better estimate of
location than does the mean.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda351.htm (3 of 5) [5/7/2002 3:14:59 PM]

1.3.5.1. Measures of Location

Lognormal The fourth histogram is a sample from a lognormal distribution. The

Distribution mean is 1.677, the median is 0.989, and the mode is 0.680 (the mode is
computed as the midpoint of the histogram interval with the highest
peak).
The lognormal is also a skewed distribution. Therefore the mean and
median do not provide similar estimates for the location. As with the
exponential distribution, there is no obvious answer to the question of
which is the more meaningful measure of location.

Robustness There are various alternatives to the mean and median for measuring
location. These alternatives were developed to address non-normal data
since the mean is an optimal estimator if in fact your data are normal.
Tukey and Mosteller defined two types of robustness where robustness
is a lack of susceptibility to the effects of nonnormality.
1. Robustness of validity means that the confidence intervals for the
population location have a 95% chance of covering the population
location regardless of what the underlying distribution is.
2. Robustness of efficiency refers to high effectiveness in the face of
non-normal tails. That is, confidence intervals for the population
location tend to be almost as narrow as the best that could be done
if we knew the true shape of the distributuion.
The mean is an example of an estimator that is the best we can do if the
underlying distribution is normal. However, it lacks robustness of
validity. That is, confidence intervals based on the mean tend not to be
precise if the underlying distribution is in fact not normal.
The median is an example of a an estimator that tends to have
robustness of validity but not robustness of efficiency.
The alternative measures of location try to balance these two concepts of
robustness. That is, the confidence intervals for the case when the data
are normal should be almost as narrow as the confidence intervals based
on the mean. However, they should maintain their validity even if the
underlying data are not normal. In particular, these alternatives address
the problem of heavy-tailed distributions.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda351.htm (4 of 5) [5/7/2002 3:14:59 PM]

1.3.5.1. Measures of Location

Alternative A few of the more common alternative location measures are:

Measures of 1. Mid-Mean - computes a mean using the data between the 25th
Location and 75th percentiles.
2. Trimmed Mean - similar to the mid-mean except different
percentile values are used. A common choice is to trim 5% of the
points in both the lower and upper tails, i.e., calculate the mean
for data between the 5th and 95th percentiles.
3. Winsorized Mean - similar to the trimmed mean. However,
instead of trimming the points, they are set to the lowest (or
highest) value. For example, all data below the 5th percentile are
set equal to the value of the 5th percentile and all data greater
than the 95th percentile are set equal to the 95th percentile.
4. Mid-range = (smallest + largest)/2.
The first three alternative location estimators defined above have the
advantage of the median in the sense that they are not unduly affected
by extremes in the tails. However, they generate estimates that are closer
to the mean for data that are normal (or nearly so).
The mid-range, since it is based on the two most extreme points, is not
robust. Its use is typically restricted to situations in which the behavior
at the extreme points is relevant.

Case Study The uniform random numbers case study compares the performance of
several different location estimators for a particular non-normal
distribution.

Software Most general purpose statistical software programs, including Dataplot,

can compute at least some of the measures of location discussed above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda351.htm (5 of 5) [5/7/2002 3:14:59 PM]

1.3.5.2. Confidence Limits for the Mean

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques

1.3.5.2. Confidence Limits for the Mean

Purpose: Confidence limits for the mean (Snedecor and Cochran, 1989) are an interval estimate
Interval for the mean. Interval estimates are often desirable because the estimate of the mean
Estimate for varies from sample to sample. Instead of a single estimate for the mean, a confidence
Mean interval generates a lower and upper limit for the mean. The interval estimate gives an
indication of how much uncertainty there is in our estimate of the true mean. The
narrower the interval, the more precise is our estimate.
Confidence limits are expressed in terms of a confidence coefficient. Although the
choice of confidence coefficient is somewhat arbitrary, in practice 90%, 95%, and
99% intervals are often used, with 95% being the most commonly used.
As a technical note, a 95% confidence interval does not mean that there is a 95%
probability that the interval contains the true mean. The interval computed from a
given sample either contains the true mean or it does not. Instead, the level of
confidence is associated with the method of calculating the interval. The confidence
coefficient is simply the proportion of samples of a given size that may be expected to
contain the true mean. That is, for a 95% confidence interval, if many samples are
collected and the confidence interval computed, in the long run about 95% of these
intervals would contain the true mean.

Definition: Confidence limits are defined as:

Confidence
Interval
where is the sample mean, s is the sample standard deviation, N is the sample size,
is the desired significance level, and is the upper critical value of the t
distribution with N - 1 degrees of freedom. Note that the confidence coefficient is 1 -
.
From the formula, it is clear that the width of the interval is controlled by two factors:

1. As N increases, the interval gets narrower from the term.

That is, one way to obtain more precise estimates for the mean is to increase the
sample size.
2. The larger the sample standard deviation, the larger the confidence interval.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda352.htm (1 of 4) [5/7/2002 3:14:59 PM]

1.3.5.2. Confidence Limits for the Mean

This simply means that noisy data, i.e., data with a large standard deviation, are
going to generate wider intervals than data with a smaller standard deviation.

Definition: To test whether the population mean has a specific value, , against the two-sided
Hypothesis alternative that it does not have a value , the confidence interval is converted to
Test hypothesis-test form. The test is a one-sample t-test, and it is defined as:
H0:
Ha:
Test Statistic:
where , N, and are defined as above.
Significance Level: . The most commonly used value for is 0.05.
Critical Region: Reject the null hypothesis that the mean is a specified value, ,
if

or

Sample Dataplot generated the following output for a confidence interval from the
Output for ZARR13.DAT data set:
Confidence
Interval
CONFIDENCE LIMITS FOR MEAN
(2-SIDED)

NUMBER OF OBSERVATIONS = 195

MEAN = 9.261460
STANDARD DEVIATION = 0.2278881E-01
STANDARD DEVIATION OF MEAN = 0.1631940E-02

CONFIDENCE T T X SD(MEAN) LOWER UPPER

VALUE (%) VALUE LIMIT LIMIT
---------------------------------------------------------
50.000 0.676 0.110279E-02 9.26036 9.26256
75.000 1.154 0.188294E-02 9.25958 9.26334
90.000 1.653 0.269718E-02 9.25876 9.26416
95.000 1.972 0.321862E-02 9.25824 9.26468
99.000 2.601 0.424534E-02 9.25721 9.26571
99.900 3.341 0.545297E-02 9.25601 9.26691
99.990 3.973 0.648365E-02 9.25498 9.26794
99.999 4.536 0.740309E-02 9.25406 9.26886

http://www.itl.nist.gov/div898/handbook/eda/section3/eda352.htm (2 of 4) [5/7/2002 3:14:59 PM]

1.3.5.2. Confidence Limits for the Mean

Interpretation The first few lines print the sample statistics used in calculating the confidence
of the Sample interval. The table shows the confidence interval for several different significance
Output levels. The first column lists the confidence level (which is 1 - expressed as a
percent), the second column lists the t-value (i.e., ), the third column lists
the t-value times the standard error (the standard error is ), the fourth column
lists the lower confidence limit, and the fifth column lists the upper confidence limit.
For example, for a 95% confidence interval, we go to the row identified by 95.000 in
the first column and extract an interval of (9.25824, 9.26468) from the last two
columns.
Output from other statistical software may look somewhat different from the above
output.

Sample Dataplot generated the following output for a one-sample t-test from the
Output for t ZARR13.DAT data set:
Test
T TEST
(1-SAMPLE)
MU0 = 5.000000
NULL HYPOTHESIS UNDER TEST--MEAN MU = 5.000000

SAMPLE:
NUMBER OF OBSERVATIONS = 195
MEAN = 9.261460
STANDARD DEVIATION = 0.2278881E-01
STANDARD DEVIATION OF MEAN = 0.1631940E-02

TEST:
MEAN-MU0 = 4.261460
T TEST STATISTIC VALUE = 2611.284
DEGREES OF FREEDOM = 194.0000
T TEST STATISTIC CDF VALUE = 1.000000

ALTERNATIVE- ALTERNATIVE-
ALTERNATIVE- HYPOTHESIS HYPOTHESIS
HYPOTHESIS ACCEPTANCE INTERVAL CONCLUSION
MU <> 5.000000 (0,0.025) (0.975,1) ACCEPT
MU < 5.000000 (0,0.05) REJECT
MU > 5.000000 (0.95,1) ACCEPT

http://www.itl.nist.gov/div898/handbook/eda/section3/eda352.htm (3 of 4) [5/7/2002 3:14:59 PM]

1.3.5.2. Confidence Limits for the Mean

Interpretation We are testing the hypothesis that the population mean is 5. The output is divided into
of Sample three sections.
Output 1. The first section prints the sample statistics used in the computation of the t-test.
2. The second section prints the t-test statistic value, the degrees of freedom, and
the cumulative distribution function (cdf) value of the t-test statistic. The t-test
statistic cdf value is an alternative way of expressing the critical value. This cdf
value is compared to the acceptance intervals printed in section three. For an
upper one-tailed test, the alternative hypothesis acceptance interval is (1 - ,1),
the alternative hypothesis acceptance interval for a lower one-tailed test is (0,
), and the alternative hypothesis acceptance interval for a two-tailed test is (1 -
/2,1) or (0, /2). Note that accepting the alternative hypothesis is equivalent to
rejecting the null hypothesis.
3. The third section prints the conclusions for a 95% test since this is the most
common case. Results are given in terms of the alternative hypothesis for the
two-tailed test and for the one-tailed test in both directions. The alternative
hypothesis acceptance interval column is stated in terms of the cdf value printed
in section two. The last column specifies whether the alternative hypothesis is
accepted or rejected. For a different significance level, the appropriate
conclusion can be drawn from the t-test statistic cdf value printed in section
two. For example, for a significance level of 0.10, the corresponding alternative
hypothesis acceptance intervals are (0,0.05) and (0.95,1), (0, 0.10), and (0.90,1).
Output from other statistical software may look somewhat different from the above
output.

Questions Confidence limits for the mean can be used to answer the following questions:
1. What is a reasonable estimate for the mean?
2. How much variability is there in the estimate of the mean?
3. Does a given target value fall within the confidence limits?

Related Two-Sample T-Test

Techniques
Confidence intervals for other location estimators such as the median or mid-mean
tend to be mathematically difficult or intractable. For these cases, confidence intervals
can be obtained using the bootstrap.

Case Study Heat flow meter data.

Software Confidence limits for the mean and one-sample t-tests are available in just about all
general purpose statistical software programs, including Dataplot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda352.htm (4 of 4) [5/7/2002 3:14:59 PM]

1.3.5.3. Two-Sample t-Test for Equal Means

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques

1.3.5.3. Two-Sample t-Test for Equal Means

Purpose: The two-sample t-test (Snedecor and Cochran, 1989) is used to determine if two
Test if two population means are equal. A common application of this is to test if a new
population process or treatment is superior to a current process or treatment.
means are
equal There are several variations on this test.
1. The data may either be paired or not paired. By paired, we mean that there
is a one-to-one correspondence between the values in the two samples. That
is, if X1, X2, ..., Xn and Y1, Y2, ... , Yn are the two samples, then Xi
corresponds to Yi. For paired samples, the difference Xi - Yi is usually
calculated. For unpaired samples, the sample sizes for the two samples may
or may not be equal. The formulas for paired data are somewhat simpler
than the formulas for unpaired data.
2. The variances of the two samples may be assumed to be equal or unequal.
Equal variances yields somewhat simpler formulas, although with
computers this is no longer a significant issue.
3. The null hypothesis might be that the two population means are not equal (
). If so, this must be converted to the form that the difference
between the two population means is equal to some constant (
). This form might be preferred if you only want to adopt a
new process or treatment if it exceeds the current treatment by some
threshold value.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda353.htm (1 of 5) [5/7/2002 3:15:00 PM]

1.3.5.3. Two-Sample t-Test for Equal Means

Definition The two sample t test for unpaired data is defined as:
H0:
Ha:
Test
Statistic:

where N1 and N2 are the sample sizes, and are the sample
means, and and are the sample variances.

If equal variances are assumed, then the formula reduces to:

where

Significance .
Level:
Critical Reject the null hypothesis that the two means are equal if
Region:
or

where is the critical value of the t distribution with

degrees of freedom where

If equal variances are assumed, then

http://www.itl.nist.gov/div898/handbook/eda/section3/eda353.htm (2 of 5) [5/7/2002 3:15:00 PM]

1.3.5.3. Two-Sample t-Test for Equal Means

Sample Dataplot generated the following output for the t test from the AUTO83B.DAT
Output data set:

T TEST
(2-SAMPLE)
NULL HYPOTHESIS UNDER TEST--POPULATION MEANS MU1 = MU2

SAMPLE 1:
NUMBER OF OBSERVATIONS = 249
MEAN = 20.14458
STANDARD DEVIATION = 6.414700
STANDARD DEVIATION OF MEAN = 0.4065151

SAMPLE 2:
NUMBER OF OBSERVATIONS = 79
MEAN = 30.48101
STANDARD DEVIATION = 6.107710
STANDARD DEVIATION OF MEAN = 0.6871710

IF ASSUME SIGMA1 = SIGMA2:

POOLED STANDARD DEVIATION = 6.342600
DIFFERENCE (DEL) IN MEANS = -10.33643
STANDARD DEVIATION OF DEL = 0.8190135
T TEST STATISTIC VALUE = -12.62059
DEGREES OF FREEDOM = 326.0000
T TEST STATISTIC CDF VALUE = 0.000000

IF NOT ASSUME SIGMA1 = SIGMA2:

STANDARD DEVIATION SAMPLE 1 = 6.414700
STANDARD DEVIATION SAMPLE 2 = 6.107710
BARTLETT CDF VALUE = 0.402799
DIFFERENCE (DEL) IN MEANS = -10.33643
STANDARD DEVIATION OF DEL = 0.7984100
T TEST STATISTIC VALUE = -12.94627
EQUIVALENT DEG. OF FREEDOM = 136.8750
T TEST STATISTIC CDF VALUE = 0.000000

ALTERNATIVE- ALTERNATIVE-
ALTERNATIVE- HYPOTHESIS HYPOTHESIS
HYPOTHESIS ACCEPTANCE INTERVAL CONCLUSION
MU1 <> MU2 (0,0.025) (0.975,1) ACCEPT
MU1 < MU2 (0,0.05) ACCEPT
MU1 > MU2 (0.95,1) REJECT

http://www.itl.nist.gov/div898/handbook/eda/section3/eda353.htm (3 of 5) [5/7/2002 3:15:00 PM]

1.3.5.3. Two-Sample t-Test for Equal Means

Interpretation We are testing the hypothesis that the population mean is equal for the two
of Sample samples. The output is divided into five sections.
Output 1. The first section prints the sample statistics for sample one used in the
computation of the t-test.
2. The second section prints the sample statistics for sample two used in the
computation of the t-test.
3. The third section prints the pooled standard deviation, the difference in the
means, the t-test statistic value, the degrees of freedom, and the cumulative
distribution function (cdf) value of the t-test statistic under the assumption
that the standard deviations are equal. The t-test statistic cdf value is an
alternative way of expressing the critical value. This cdf value is compared
to the acceptance intervals printed in section five. For an upper one-tailed
test, the acceptance interval is (0,1 - ), the acceptance interval for a
two-tailed test is ( /2, 1 - /2), and the acceptance interval for a lower
one-tailed test is ( ,1).
4. The fourth section prints the pooled standard deviation, the difference in
the means, the t-test statistic value, the degrees of freedom, and the
cumulative distribution function (cdf) value of the t-test statistic under the
assumption that the standard deviations are not equal. The t-test statistic cdf
value is an alternative way of expressing the critical value. cdf value is
compared to the acceptance intervals printed in section five. For an upper
one-tailed test, the alternative hypothesis acceptance interval is (1 - ,1),
the alternative hypothesis acceptance interval for a lower one-tailed test is
(0, ), and the alternative hypothesis acceptance interval for a two-tailed
test is (1 - /2,1) or (0, /2). Note that accepting the alternative hypothesis
is equivalent to rejecting the null hypothesis.
5. The fifth section prints the conclusions for a 95% test under the assumption
that the standard deviations are not equal since a 95% test is the most
common case. Results are given in terms of the alternative hypothesis for
the two-tailed test and for the one-tailed test in both directions. The
alternative hypothesis acceptance interval column is stated in terms of the
cdf value printed in section four. The last column specifies whether the
alternative hypothesis is accepted or rejected. For a different significance
level, the appropriate conclusion can be drawn from the t-test statistic cdf
value printed in section four. For example, for a significance level of 0.10,
the corresponding alternative hypothesis acceptance intervals are (0,0.05)
and (0.95,1), (0, 0.10), and (0.90,1).
Output from other statistical software may look somewhat different from the
above output.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda353.htm (4 of 5) [5/7/2002 3:15:00 PM]

1.3.5.3. Two-Sample t-Test for Equal Means

Questions Two-sample t-tests can be used to answer the following questions:

1. Is process 1 equivalent to process 2?
2. Is the new process better than the current process?
3. Is the new process better than the current process by at least some
pre-determined threshold amount?

Related Confidence Limits for the Mean

Techniques Analysis of Variance

Case Study Ceramic strength data.

Software Two-sample t-tests are available in just about all general purpose statistical
software programs, including Dataplot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda353.htm (5 of 5) [5/7/2002 3:15:00 PM]

1.3.5.3.1. Data Used for Two-Sample t-Test

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques
1.3.5.3. Two-Sample t-Test for Equal Means

1.3.5.3.1. Data Used for Two-Sample t-Test

Data Used The following is the data used for the two-sample t-test example. The
for first column is miles per gallon for U.S. cars and the second column is
Two-Sample miles per gallon for Japanese cars. For the t-test example, rows with the
t-Test second column equal to -999 were deleted.
Example
18 24
15 27
18 27
16 25
17 31
15 35
14 24
14 19
14 28
15 23
15 27
14 20
15 22
14 18
22 20
18 31
21 32
21 31
10 32
10 24
11 26
9 29
28 24
25 24
19 33
16 33
17 32
19 28

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3531.htm (1 of 6) [5/7/2002 3:15:01 PM]

1.3.5.3.1. Data Used for Two-Sample t-Test

18 19
14 32
14 34
14 26
14 30
12 22
13 22
13 33
18 39
22 36
19 28
18 27
23 21
26 24
25 30
20 34
21 32
13 38
14 37
15 30
14 31
17 37
11 32
13 47
12 41
13 45
15 34
13 33
13 24
14 32
22 39
28 35
13 32
14 37
13 38
14 34
15 34
12 32
13 33
13 32
14 25
13 24
12 37
13 31
18 36
16 36

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3531.htm (2 of 6) [5/7/2002 3:15:01 PM]

1.3.5.3.1. Data Used for Two-Sample t-Test

18 34
18 38
23 32
11 38
12 32
13 -999
12 -999
18 -999
21 -999
19 -999
21 -999
15 -999
16 -999
15 -999
11 -999
20 -999
21 -999
19 -999
15 -999
26 -999
25 -999
16 -999
16 -999
18 -999
16 -999
13 -999
14 -999
14 -999
14 -999
28 -999
19 -999
18 -999
15 -999
15 -999
16 -999
15 -999
16 -999
14 -999
17 -999
16 -999
15 -999
18 -999
21 -999
20 -999
13 -999
23 -999

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3531.htm (3 of 6) [5/7/2002 3:15:01 PM]

1.3.5.3.1. Data Used for Two-Sample t-Test

20 -999
23 -999
18 -999
19 -999
25 -999
26 -999
18 -999
16 -999
16 -999
15 -999
22 -999
22 -999
24 -999
23 -999
29 -999
25 -999
20 -999
18 -999
19 -999
18 -999
27 -999
13 -999
17 -999
13 -999
13 -999
13 -999
30 -999
26 -999
18 -999
17 -999
16 -999
15 -999
18 -999
21 -999
19 -999
19 -999
16 -999
16 -999
16 -999
16 -999
25 -999
26 -999
31 -999
34 -999
36 -999
20 -999

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3531.htm (4 of 6) [5/7/2002 3:15:01 PM]

1.3.5.3.1. Data Used for Two-Sample t-Test

19 -999
20 -999
19 -999
21 -999
20 -999
25 -999
21 -999
19 -999
21 -999
21 -999
19 -999
18 -999
19 -999
18 -999
18 -999
18 -999
30 -999
31 -999
23 -999
24 -999
22 -999
20 -999
22 -999
20 -999
21 -999
17 -999
18 -999
17 -999
18 -999
17 -999
16 -999
19 -999
19 -999
36 -999
27 -999
23 -999
24 -999
34 -999
35 -999
28 -999
29 -999
27 -999
34 -999
32 -999
28 -999
26 -999

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3531.htm (5 of 6) [5/7/2002 3:15:01 PM]

1.3.5.3.1. Data Used for Two-Sample t-Test

24 -999
19 -999
28 -999
24 -999
27 -999
27 -999
26 -999
24 -999
30 -999
39 -999
35 -999
34 -999
30 -999
22 -999
27 -999
20 -999
18 -999
28 -999
27 -999
34 -999
31 -999
29 -999
27 -999
24 -999
23 -999
38 -999
36 -999
25 -999
38 -999
26 -999
22 -999
36 -999
27 -999
27 -999
32 -999
28 -999
31 -999

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3531.htm (6 of 6) [5/7/2002 3:15:01 PM]

1.3.5.4. One-Factor ANOVA

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques

1.3.5.4. One-Factor ANOVA

Purpose: One factor analysis of variance (Snedecor and Cochran, 1989) is a
Test for special case of analysis of variance (ANOVA), for one factor of interest,
Equal and a generalization of the two-sample t-test. The two-sample t-test is
Means used to decide whether two groups (levels) of a factor have the same
Across mean. One-way analysis of variance generalizes this to levels where k,
Groups the number of levels, is greater than or equal to 2.
For example, data collected on, say, five instruments have one factor
(instruments) at five levels. The ANOVA tests whether instruments
have a significant effect on the results.

Definition The Product and Process Comparisons chapter (chapter 7) contains a

more extensive discussion of 1-factor ANOVA, including the details for
the mathematical computations of one-way analysis of variance.
The model for the analysis of variance can be stated in two
mathematically equivalent ways. In the following discussion, each level
of each factor is called a cell. For the one-way case, a cell and a level
are equivalent since there is only one factor. In the following, the
subscript i refers to the level and the subscript j refers to the observation
within a level. For example, Y23 refers to the third observation in the
second level.
The first model is

This model decomposes the response into a mean for each cell and an
error term. The analysis of variance provides estimates for each cell
mean. These estimated cell means are the predicted values of the model
and the differences between the response variable and the estimated cell
means are the residuals. That is

The second model is

This model decomposes the response into an overall (grand) mean, the
effect of the ith factor level, and an error term. The analysis of variance
provides estimates of the grand mean and the effect of the ith factor
level. The predicted values and the residuals of the model are

The distinction between these models is that the second model divides

http://www.itl.nist.gov/div898/handbook/eda/section3/eda354.htm (1 of 4) [5/7/2002 3:15:01 PM]

1.3.5.4. One-Factor ANOVA
the cell mean into an overall mean and the effect of the ith factor level.
This second model makes the factor effect more explicit, so we will
emphasize this approach.

Model Note that the ANOVA model assumes that the error term, Eij, should
Validation follow the assumptions for a univariate measurement process. That is,
after performing an analysis of variance, the model should be validated
by analyzing the residuals.

Sample Dataplot generated the following output for the one-way analysis of variance from the GEAR.DAT data set.
Output

NUMBER OF OBSERVATIONS = 100

NUMBER OF FACTORS = 1
NUMBER OF LEVELS FOR FACTOR 1 = 10
BALANCED CASE
RESIDUAL STANDARD DEVIATION = 0.59385783970E-02
RESIDUAL DEGREES OF FREEDOM = 90
REPLICATION CASE
REPLICATION STANDARD DEVIATION = 0.59385774657E-02
REPLICATION DEGREES OF FREEDOM = 90
NUMBER OF DISTINCT CELLS = 10

*****************
* ANOVA TABLE *
*****************

SOURCE DF SUM OF SQUARES MEAN SQUARE F STATISTIC F CDF SIG

-------------------------------------------------------------------------------
TOTAL (CORRECTED) 99 0.003903 0.000039
-------------------------------------------------------------------------------
FACTOR 1 9 0.000729 0.000081 2.2969 97.734% *
-------------------------------------------------------------------------------
RESIDUAL 90 0.003174 0.000035

RESIDUAL STANDARD DEVIATION = 0.00593857840

RESIDUAL DEGREES OF FREEDOM = 90
REPLICATION STANDARD DEVIATION = 0.00593857747
REPLICATION DEGREES OF FREEDOM = 90

****************
* ESTIMATION *
****************

GRAND MEAN = 0.99764001369E+00

GRAND STANDARD DEVIATION = 0.62789078802E-02

LEVEL-ID NI MEAN EFFECT SD(EFFECT)

--------------------------------------------------------------------
FACTOR 1-- 1.00000 10. 0.99800 0.00036 0.00178
-- 2.00000 10. 0.99910 0.00146 0.00178
-- 3.00000 10. 0.99540 -0.00224 0.00178
-- 4.00000 10. 0.99820 0.00056 0.00178
-- 5.00000 10. 0.99190 -0.00574 0.00178
-- 6.00000 10. 0.99880 0.00116 0.00178
-- 7.00000 10. 1.00150 0.00386 0.00178

http://www.itl.nist.gov/div898/handbook/eda/section3/eda354.htm (2 of 4) [5/7/2002 3:15:01 PM]

1.3.5.4. One-Factor ANOVA
-- 8.00000 10. 1.00040 0.00276 0.00178
-- 9.00000 10. 0.99830 0.00066 0.00178
-- 10.00000 10. 0.99480 -0.00284 0.00178

MODEL RESIDUAL STANDARD DEVIATION

-------------------------------------------------------
CONSTANT ONLY-- 0.0062789079
CONSTANT & FACTOR 1 ONLY-- 0.0059385784

Interpretation The output is divided into three sections.

of Sample 1. The first section prints the number of observations (100), the
Output number of factors (10), and the number of levels for each factor
(10 levels for factor 1). It also prints some overall summary
statistics. In particular, the residual standard deviation is 0.0059.
The smaller the residual standard deviation, the more we have
accounted for the variance in the data.
2. The second section prints an ANOVA table. The ANOVA table
decomposes the variance into the following component sum of
squares:
❍ Total sum of squares. The degrees of freedom for this
entry is the number of observations minus one.
❍ Sum of squares for the factor. The degrees of freedom for
this entry is the number of levels minus one. The mean
square is the sum of squares divided by the number of
degrees of freedom.
❍ Residual sum of squares. The degrees of freedom is the
total degrees of freedom minus the factor degrees of
freedom. The mean square is the sum of squares divided
by the number of degrees of freedom.
That is, it summarizes how much of the variance in the data
(total sum of squares) is accounted for by the factor effect (factor
sum of squares) and how much is random error (residual sum of
squares). Ideally, we would like most of the variance to be
explained by the factor effect. The ANOVA table provides a
formal F test for the factor effect. The F-statistic is the mean
square for the factor divided by the mean square for the error.
This statistic follows an F distribution with (k-1) and (N-k)
degrees of freedom. If the F CDF column for the factor effect is
greater than 95%, then the factor is significant at the 5% level.
3. The third section prints an estimation section. It prints an overall
mean and overall standard deviation. Then for each level of each
factor, it prints the number of observations, the mean for the
observations of each cell ( in the above terminology), the
factor effect ( in the above terminology), and the standard
deviation of the factor effect. Finally, it prints the residual
standard deviation for the various possible models. For the
one-way ANOVA, the two models are the constant model, i.e.,

and the model with a factor effect

http://www.itl.nist.gov/div898/handbook/eda/section3/eda354.htm (3 of 4) [5/7/2002 3:15:01 PM]

1.3.5.4. One-Factor ANOVA

For these data, including the factor effect reduces the residual
standard deviation from 0.00623 to 0.0059. That is, although the
factor is statistically significant, it has minimal improvement
over a simple constant model. This is because the factor is just
barely significant.
Output from other statistical software may look somewhat different
from the above output.
In addition to the quantitative ANOVA output, it is recommended that
any analysis of variance be complemented with model validation. At a
minimum, this should include
1. A run sequence plot of the residuals.
2. A normal probability plot of the residuals.
3. A scatter plot of the predicted values against the residuals.

Question The analysis of variance can be used to answer the following question
● Are means the same across groups in the data?

Importance The analysis of uncertainty depends on whether the factor significantly

affects the outcome.

Related Two-sample t-test

Techniques Multi-factor analysis of variance
Regression
Box plot

Software Most general purpose statistical software programs, including Dataplot,

can generate an analysis of variance.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda354.htm (4 of 4) [5/7/2002 3:15:01 PM]

1.3.5.5. Multi-factor Analysis of Variance

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques

1.3.5.5. Multi-factor Analysis of Variance

Purpose: The analysis of variance (ANOVA) (Neter, Wasserman, and Kunter,
Detect 1990) is used to detect significant factors in a multi-factor model. In the
significant multi-factor model, there is a response (dependent) variable and one or
factors more factor (independent) variables. This is a common model in
designed experiments where the experimenter sets the values for each of
the factor variables and then measures the response variable.
Each factor can take on a certain number of values. These are referred to
as the levels of a factor. The number of levels can vary betweeen
factors. For designed experiments, the number of levels for a given
factor tends to be small. Each factor and level combination is a cell.
Balanced designs are those in which the cells have an equal number of
observations and unbalanced designs are those in which the number of
observations varies among cells. It is customary to use balanced designs
in designed experiments.

Definition The Product and Process Comparisons chapter (chapter 7) contains a

more extensive discussion of 2-factor ANOVA, including the details for
the mathematical computations.
The model for the analysis of variance can be stated in two
mathematically equivalent ways. We explain the model for a two-way
ANOVA (the concepts are the same for additional factors). In the
following discussion, each combination of factors and levels is called a
cell. In the following, the subscript i refers to the level of factor 1, j
refers to the level of factor 2, and the subscript k refers to the kth
observation within the (i,j)th cell. For example, Y235 refers to the fifth
observation in the second level of factor 1 and the third level of factor 2.
The first model is

This model decomposes the response into a mean for each cell and an
error term. The analysis of variance provides estimates for each cell
mean. These cell means are the predicted values of the model and the
differences between the response variable and the estimated cell means
are the residuals. That is

The second model is

This model decomposes the response into an overall (grand) mean,

factor effects ( and represent the effects of the ith level of the first

http://www.itl.nist.gov/div898/handbook/eda/section3/eda355.htm (1 of 5) [5/7/2002 3:15:02 PM]

1.3.5.5. Multi-factor Analysis of Variance
factor and the jth level of the second factor, respectively), and an error
term. The analysis of variance provides estimates of the grand mean and
the factor effects. The predicted values and the residuals of the model
are

The distinction between these models is that the second model divides
the cell mean into an overall mean and factor effects. This second model
makes the factor effect more explicit, so we will emphasize this
approach.

Model Note that the ANOVA model assumes that the error term, Eijk, should
Validation follow the assumptions for a univariate measurement process. That is,
after performing an analysis of variance, the model should be validated
by analyzing the residuals.

Sample Dataplot generated the following ANOVA output for the JAHANMI2.DAT data set:
Output

**********************************
**********************************
** 4-WAY ANALYSIS OF VARIANCE **
**********************************
**********************************

NUMBER OF OBSERVATIONS = 480

NUMBER OF FACTORS = 4
NUMBER OF LEVELS FOR FACTOR 1 = 2
NUMBER OF LEVELS FOR FACTOR 2 = 2
NUMBER OF LEVELS FOR FACTOR 3 = 2
NUMBER OF LEVELS FOR FACTOR 4 = 2
BALANCED CASE
RESIDUAL STANDARD DEVIATION = 0.63057727814E+02
RESIDUAL DEGREES OF FREEDOM = 475
REPLICATION CASE
REPLICATION STANDARD DEVIATION = 0.61890106201E+02
REPLICATION DEGREES OF FREEDOM = 464
NUMBER OF DISTINCT CELLS = 16

*****************
* ANOVA TABLE *
*****************

SOURCE DF SUM OF SQUARES MEAN SQUARE F STATISTIC F CDF SIG

-------------------------------------------------------------------------------
TOTAL (CORRECTED) 479 2668446.000000 5570.868652
-------------------------------------------------------------------------------
FACTOR 1 1 26672.726562 26672.726562 6.7080 99.011% **
FACTOR 2 1 11524.053711 11524.053711 2.8982 91.067%
FACTOR 3 1 14380.633789 14380.633789 3.6166 94.219%
FACTOR 4 1 727143.125000 727143.125000 182.8703 100.000% **
-------------------------------------------------------------------------------
RESIDUAL 475 1888731.500000 3976.276855

RESIDUAL STANDARD DEVIATION = 63.05772781

http://www.itl.nist.gov/div898/handbook/eda/section3/eda355.htm (2 of 5) [5/7/2002 3:15:02 PM]

1.3.5.5. Multi-factor Analysis of Variance
RESIDUAL DEGREES OF FREEDOM = 475
REPLICATION STANDARD DEVIATION = 61.89010620
REPLICATION DEGREES OF FREEDOM = 464
LACK OF FIT F RATIO = 2.6447 = THE 99.7269% POINT OF THE
F DISTRIBUTION WITH 11 AND 464 DEGREES OF FREEDOM

****************
* ESTIMATION *
****************

GRAND MEAN = 0.65007739258E+03

GRAND STANDARD DEVIATION = 0.74638252258E+02

LEVEL-ID NI MEAN EFFECT SD(EFFECT)

--------------------------------------------------------------------
FACTOR 1-- -1.00000 240. 657.53168 7.45428 2.87818
-- 1.00000 240. 642.62286 -7.45453 2.87818
FACTOR 2-- -1.00000 240. 645.17755 -4.89984 2.87818
-- 1.00000 240. 654.97723 4.89984 2.87818
FACTOR 3-- -1.00000 240. 655.55084 5.47345 2.87818
-- 1.00000 240. 644.60376 -5.47363 2.87818
FACTOR 4-- 1.00000 240. 688.99890 38.92151 2.87818
-- 2.00000 240. 611.15594 -38.92145 2.87818

MODEL RESIDUAL STANDARD DEVIATION

-------------------------------------------------------
CONSTANT ONLY-- 74.6382522583
CONSTANT & FACTOR 1 ONLY-- 74.3419036865
CONSTANT & FACTOR 2 ONLY-- 74.5548019409
CONSTANT & FACTOR 3 ONLY-- 74.5147094727
CONSTANT & FACTOR 4 ONLY-- 63.7284545898
CONSTANT & ALL 4 FACTORS -- 63.0577278137

Interpretation The output is divided into three sections.

of Sample
Output 1. The first section prints the number of observations (480), the
number of factors (4), and the number of levels for each factor (2
levels for each factor). It also prints some overall summary
statistics. In particular, the residual standard deviation is 63.058.
The smaller the residual standard deviation, the more we have
accounted for the variance in the data.
2. The second section prints an ANOVA table. The ANOVA table
decomposes the variance into the following component sum of
squares:
❍ Total sum of squares. The degrees of freedom for this
entry is the number of observations minus one.
❍ Sum of squares for each of the factors. The degrees of
freedom for these entries are the number of levels for the
factor minus one. The mean square is the sum of squares
divided by the number of degrees of freedom.
❍ Residual sum of squares. The degrees of freedom is the
total degrees of freedom minus the sum of the factor
degrees of freedom. The mean square is the sum of
squares divided by the number of degrees of freedom.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda355.htm (3 of 5) [5/7/2002 3:15:02 PM]

1.3.5.5. Multi-factor Analysis of Variance
That is, it summarizes how much of the variance in the data
(total sum of squares) is accounted for by the factor effects
(factor sum of squares) and how much is random error (residual
sum of squares). Ideally, we would like most of the variance to
be explained by the factor effects. The ANOVA table provides a
formal F test for the factor effects. The F-statistic is the mean
square for the factor divided by the mean square for the error.
This statistic follows an F distribution with (k-1) and (N-k)
degrees of freedom where k is the number of levels for the given
factor. If the F CDF column for the factor effect is greater than
95%, then the factor is significant at the 5% level. Here, we see
that the size of the effect of factor 4 dominates the size of the
other effects. The F test shows that factors one and four are
significant at the 1% level while factors two and three are not
significant at the 5% level.
3. The third section is an estimation section. It prints an overall
mean and overall standard deviation. Then for each level of each
factor, it prints the number of observations, the mean for the
observations of each cell ( in the above terminology), the
factor effects ( and in the above terminology), and the
standard deviation of the factor effect. Finally, it prints the
residual standard deviation for the various possible models. For
the four-way ANOVA here, it prints the constant model

a model with each factor individually, and the model with all
four factors included.
For these data, we see that including factor 4 has a significant
impact on the residual standard deviation (63.73 when only the
factor 4 effect is included compared to 63.058 when all four
factors are included).
Output from other statistical software may look somewhat different
from the above output.
In addition to the quantitative ANOVA output, it is recommended that
any analysis of variance be complemented with model validation. At a
minimum, this should include
1. A run sequence plot of the residuals.
2. A normal probability plot of the residuals.
3. A scatter plot of the predicted values against the residuals.

Questions The analysis of variance can be used to answer the following

questions:
1. Do any of the factors have a significant effect?
2. Which is the most important factor?
3. Can we account for most of the variability in the data?

Related One-factor analysis of variance

Techniques Two-sample t-test
Box plot
Block plot
Dex mean plot

Case Study The quantitative ANOVA approach can be contrasted with the more
graphical EDA approach in the ceramic strength case study.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda355.htm (4 of 5) [5/7/2002 3:15:02 PM]

1.3.5.5. Multi-factor Analysis of Variance

Software Most general purpose statistical software programs, including Dataplot,

can perform multi-factor analysis of variance.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda355.htm (5 of 5) [5/7/2002 3:15:02 PM]

1.3.5.6. Measures of Scale

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques

1.3.5.6. Measures of Scale

Scale, A fundamental task in many statistical analyses is to characterize the
Variability, or spread, or variability, of a data set. Measures of scale are simply
Spread attempts to estimate this variability.
When assessing the variability of a data set, there are two key
components:
1. How spread out are the data values near the center?
2. How spread out are the tails?
Different numerical summaries will give different weight to these two
elements. The choice of scale estimator is often driven by which of
these components you want to emphasize.
The histogram is an effective graphical technique for showing both of
these components of the spread.

Definitions of For univariate data, there are several common numerical measures of
Variability the spread:
1. variance - the variance is defined as

where is the mean of the data.

The variance is roughly the arithmetic average of the squared
distance from the mean. Squaring the distance from the mean
has the effect of giving greater weight to values that are further
from the mean. For example, a point 2 units from the mean
adds 4 to the above sum while a point 10 units from the mean
adds 100 to the sum. Although the variance is intended to be an
overall measure of spread, it can be greatly affected by the tail
behavior.
2. standard deviation - the standard deviation is the square root of
the variance. That is,

http://www.itl.nist.gov/div898/handbook/eda/section3/eda356.htm (1 of 6) [5/7/2002 3:15:02 PM]

1.3.5.6. Measures of Scale

The standard deviation restores the units of the spread to the

original data units (the variance squares the units).
3. range - the range is the largest value minus the smallest value in
a data set. Note that this measure is based only on the lowest
and highest extreme values in the sample. The spread near the
center of the data is not captured at all.
4. average absolute deviation - the average absolute deviation
(AAD) is defined as

where is the mean of the data and |Y| is the absolute value of
Y. This measure does not square the distance from the mean, so
it is less affected by extreme observations than are the variance
and standard deviation.
5. median absolute deviation - the median absolute deviation
(MAD) is defined as

where is the median of the data and |Y| is the absolute value
of Y. This is a variation of the average absolute deviation that is
even less affected by extremes in the tail because the data in the
tails have less influence on the calculation of the median than
they do on the mean.
6. interquartile range - this is the value of the 75th percentile
minus the value of the 25th percentile. This measure of scale
attempts to measure the variability of points near the center.
In summary, the variance, standard deviation, average absolute
deviation, and median absolute deviation measure both aspects of the
variability; that is, the variability near the center and the variability in
the tails. They differ in that the average absolute deviation and median
absolute deviation do not give undue weight to the tail behavior. On
the other hand, the range only uses the two most extreme points and
the interquartile range only uses the middle portion of the data.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda356.htm (2 of 6) [5/7/2002 3:15:02 PM]

1.3.5.6. Measures of Scale

Why Different The following example helps to clarify why these alternative
Measures? defintions of spread are useful and necessary.
This plot shows histograms for 10,000 random numbers generated
from a normal, a double exponential, a Cauchy, and a Tukey-Lambda
distribution.

Normal The first histogram is a sample from a normal distribution. The

Distribution standard deviation is 0.997, the median absolute deviation is 0.681,
and the range is 7.87.
The normal distribution is a symmetric distribution with well-behaved
tails and a single peak at the center of the distribution. By symmetric,
we mean that the distribution can be folded about an axis so that the
two sides coincide. That is, it behaves the same to the left and right of
some center point. In this case, the median absolute deviation is a bit
less than the standard deviation due to the downweighting of the tails.
The range of a little less than 8 indicates the extreme values fall
within about 4 standard deviations of the mean. If a histogram or
normal probability plot indicates that your data are approximated well
by a normal distribution, then it is reasonable to use the standard
deviation as the spread estimator.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda356.htm (3 of 6) [5/7/2002 3:15:02 PM]

1.3.5.6. Measures of Scale

Double The second histogram is a sample from a double exponential

Exponential distribution. The standard deviation is 1.417, the median absolute
Distribution deviation is 0.706, and the range is 17.556.
Comparing the double exponential and the normal histograms shows
that the double exponential has a stronger peak at the center, decays
more rapidly near the center, and has much longer tails. Due to the
longer tails, the standard deviation tends to be inflated compared to
the normal. On the other hand, the median absolute deviation is only
slightly larger than it is for the normal data. The longer tails are
clearly reflected in the value of the range, which shows that the
extremes fall about 12 standard deviations from the mean compared to
about 4 for the normal data.

Cauchy The third histogram is a sample from a Cauchy distribution. The

Distribution standard deviation is 998.389, the median absolute deviation is 1.16,
and the range is 118,953.6.
The Cauchy distribution is a symmetric distribution with heavy tails
and a single peak at the center of the distribution. The Cauchy
distribution has the interesting property that collecting more data does
not provide a more accurate estimate for the mean or standard
deviation. That is, the sampling distribution of the means and standard
deviation are equivalent to the sampling distribution of the original
data. That means that for the Cauchy distribution the standard
deviation is useless as a measure of the spread. From the histogram, it
is clear that just about all the data are between about -5 and 5.
However, a few very extreme values cause both the standard deviation
and range to be extremely large. However, the median absolute
deviation is only slightly larger than it is for the normal distribution.
In this case, the median absolute deviation is clearly the better
measure of spread.
Although the Cauchy distribution is an extreme case, it does illustrate
the importance of heavy tails in measuring the spread. Extreme values
in the tails can distort the standard deviation. However, these extreme
values do not distort the median absolute deviation since the median
absolute deviation is based on ranks. In general, for data with extreme
values in the tails, the median absolute deviation or interquartile range
can provide a more stable estimate of spread than the standard
deviation.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda356.htm (4 of 6) [5/7/2002 3:15:02 PM]

1.3.5.6. Measures of Scale

Tukey-Lambda The fourth histogram is a sample from a Tukey lambda distribution

Distribution with shape parameter = 1.2. The standard deviation is 0.49, the
median absolute deviation is 0.427, and the range is 1.666.
The Tukey lambda distribution has a range limited to .
That is, it has truncated tails. In this case the standard deviation and
median absolute deviation have closer values than for the other three
examples which have significant tails.

Robustness Tukey and Mosteller defined two types of robustness where

robustness is a lack of susceptibility to the effects of nonnormality.
1. Robustness of validity means that the confidence intervals for a
measure of the population spread (e.g., the standard deviation)
have a 95% chance of covering the true value (i.e., the
population value) of that measure of spread regardless of the
underlying distribution.
2. Robustness of efficiency refers to high effectiveness in the face
of non-normal tails. That is, confidence intervals for the
measure of spread tend to be almost as narrow as the best that
could be done if we knew the true shape of the distribution.
The standard deviation is an example of an estimator that is the best
we can do if the underlying distribution is normal. However, it lacks
robustness of validity. That is, confidence intervals based on the
standard deviation tend to lack precision if the underlying distribution
is in fact not normal.
The median absolute deviation and the interquartile range are
estimates of scale that have robustness of validity. However, they are
not particularly strong for robustness of efficiency.
If histograms and probability plots indicate that your data are in fact
reasonably approximated by a normal distribution, then it makes sense
to use the standard deviation as the estimate of scale. However, if your
data are not normal, and in particular if there are long tails, then using
an alternative measure such as the median absolute deviation, average
absolute deviation, or interquartile range makes sense. The range is
used in some applications, such as quality control, for its simplicity. In
addition, comparing the range to the standard deviation gives an
indication of the spread of the data in the tails.
Since the range is determined by the two most extreme points in the
data set, we should be cautious about its use for large values of N.
Tukey and Mosteller give a scale estimator that has both robustness of

http://www.itl.nist.gov/div898/handbook/eda/section3/eda356.htm (5 of 6) [5/7/2002 3:15:02 PM]

1.3.5.6. Measures of Scale

validity and robustness of efficiency. However, it is more complicated

and we do not give the formula here.

Software Most general purpose statistical software programs, including

Dataplot, can generate at least some of the measures of scale
discusssed above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda356.htm (6 of 6) [5/7/2002 3:15:02 PM]

1.3.5.7. Bartlett's Test

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques

1.3.5.7. Bartlett's Test

Purpose: Bartlett's test ( Snedecor and Cochran, 1983) is used to test if k samples have equal
Test for variances. Equal variances across samples is called homogeneity of variances. Some
Homogeneity statistical tests, for example the analysis of variance, assume that variances are equal
of Variances across groups or samples. The Bartlett test can be used to verify that assumption.
Bartlett's test is sensitive to departures from normality. That is, if your samples come
from non-normal distributions, then Bartlett's test may simply be testing for
non-normality. The Levene test is an alternative to the Bartlett test that is less sensitive to
departures from normality.

Definition The Bartlett test is defined as:

H0:
Ha: for at least one pair (i,j).
Test The Bartlett test statistic is designed to test for equality of variances across
Statistic: groups against the alternative that variances are unequal for at least two
groups.

In the above, si2 is the variance of the ith group, N is the total sample size,
Ni is the sample size of the ith group, k is the number of groups, and sp2 is
the pooled variance. The pooled variance is a weighted average of the
group variances and is defined as:

Significance
Level:

http://www.itl.nist.gov/div898/handbook/eda/section3/eda357.htm (1 of 3) [5/7/2002 3:15:03 PM]

1.3.5.7. Bartlett's Test

Critical The variances are judged to be unequal if,

Region:

where is the upper critical value of the chi-square distribution

with k - 1 degrees of freedom and a significance level of .
In the above formulas for the critical regions, the Handbook follows the
convention that is the upper critical value from the chi-square
distribution and is the lower critical value from the chi-square
distribution. Note that this is the opposite of some texts and software
programs. In particular, Dataplot uses the opposite convention.
An alternate definition (Dixon and Massey, 1969) is based on an approximation to the F
distribution. This definition is given in the Product and Process Comparisons chapter
(chapter 7).

Sample Dataplot generated the following output for Bartlett's test using the GEAR.DAT
Output data set:

BARTLETT TEST
(STANDARD DEFINITION)
NULL HYPOTHESIS UNDER TEST--ALL SIGMA(I) ARE EQUAL

TEST:
DEGREES OF FREEDOM = 9.000000

TEST STATISTIC VALUE = 20.78580

CUTOFF: 95% PERCENT POINT = 16.91898
CUTOFF: 99% PERCENT POINT = 21.66600

CHI-SQUARE CDF VALUE = 0.986364

NULL NULL HYPOTHESIS NULL HYPOTHESIS

HYPOTHESIS ACCEPTANCE INTERVAL CONCLUSION
ALL SIGMA EQUAL (0.000,0.950) REJECT

http://www.itl.nist.gov/div898/handbook/eda/section3/eda357.htm (2 of 3) [5/7/2002 3:15:03 PM]

1.3.5.7. Bartlett's Test

Interpretation We are testing the hypothesis that the group variances are all equal.
of Sample The output is divided into two sections.
Output 1. The first section prints the value of the Bartlett test statistic, the
degrees of freedom (k-1), the upper critical value of the
chi-square distribution corresponding to significance levels of
0.05 (the 95% percent point) and 0.01 (the 99% percent point).
We reject the null hypothesis at that significance level if the
value of the Bartlett test statistic is greater than the
corresponding critical value.
2. The second section prints the conclusion for a 95% test.
Output from other statistical software may look somewhat different
from the above output.

Question Bartlett's test can be used to answer the following question:

● Is the assumption of equal variances valid?

Importance Bartlett's test is useful whenever the assumption of equal variances is

made. In particular, this assumption is made for the frequently used
one-way analysis of variance. In this case, Bartlett's or Levene's test
should be applied to verify the assumption.

Related Standard Deviation Plot

Techniques Box Plot
Levene Test
Chi-Square Test
Analysis of Variance

Case Study Heat flow meter data

Software The Bartlett test is available in many general purpose statistical

software programs, including Dataplot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda357.htm (3 of 3) [5/7/2002 3:15:03 PM]

1.3.5.8. Chi-Square Test for the Standard Deviation

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques

1.3.5.8. Chi-Square Test for the Standard Deviation

Purpose: A chi-square test ( Snedecor and Cochran, 1983) can be used to test if the standard deviation of
Test if a population is equal to a specified value. This test can be either a two-sided test or a one-sided
standard test. The two-sided version tests against the alternative that the true standard deviation is either
deviation is less than or greater than the specified value. The one-sided version only tests in one direction.
equal to a The choice of a two-sided or one-sided test is determined by the problem. For example, if we
specified are testing a new process, we may only be concerned if its variability is greater than the
value variability of the current process.

Definition The chi-square hypothesis test is defined as:

H0:
Ha: for a lower one-tailed test

for an upper one-tailed test

for a two-tailed test

Test Statistic: T=

where N is the sample size and is the sample standard deviation. The key
element of this formula is the ratio which compares the ratio of the
sample standard deviation to the target standard deviation. The more this
ratio deviates from 1, the more likely we are to reject the null hypothesis.
Significance Level: .
Critical Region: Reject the null hypothesis that the standard deviation is a specified value,
, if

for an upper one-tailed alternative

for a lower one-tailed alternative

for a two-tailed test

or

where is the critical value of the chi-square distribution with N -

1 degrees of freedom.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda358.htm (1 of 3) [5/7/2002 3:15:03 PM]

1.3.5.8. Chi-Square Test for the Standard Deviation

In the above formulas for the critical regions, the Handbook follows the
convention that is the upper critical value from the chi-square
distribution and is the lower critical value from the chi-square
distribution. Note that this is the opposite of some texts and software
programs. In particular, Dataplot uses the opposite convention.

The formula for the hypothesis test can easily be converted to form an interval estimate for the
standard deviation:

Sample Dataplot generated the following output for a chi-square test from the GEAR.DAT data set:
Output
CHI-SQUARED TEST
SIGMA0 = 0.1000000
NULL HYPOTHESIS UNDER TEST--STANDARD DEVIATION SIGMA = .1000000

SAMPLE:
NUMBER OF OBSERVATIONS = 100
MEAN = 0.9976400
STANDARD DEVIATION S = 0.6278908E-02

TEST:
S/SIGMA0 = 0.6278908E-01
CHI-SQUARED STATISTIC = 0.3903044
DEGREES OF FREEDOM = 99.00000
CHI-SQUARED CDF VALUE = 0.000000

ALTERNATIVE- ALTERNATIVE-
ALTERNATIVE- HYPOTHESIS HYPOTHESIS
HYPOTHESIS ACCEPTANCE INTERVAL CONCLUSION
SIGMA <> .1000000 (0,0.025), (0.975,1) ACCEPT
SIGMA < .1000000 (0,0.05) ACCEPT
SIGMA > .1000000 (0.95,1) REJECT

Interpretation We are testing the hypothesis that the population standard deviation is 0.1. The output is
of Sample divided into three sections.
Output 1. The first section prints the sample statistics used in the computation of the chi-square
test.
2. The second section prints the chi-square test statistic value, the degrees of freedom, and
the cumulative distribution function (cdf) value of the chi-square test statistic. The
chi-square test statistic cdf value is an alternative way of expressing the critical value.
This cdf value is compared to the acceptance intervals printed in section three. For an
upper one-tailed test, the alternative hypothesis acceptance interval is (1 - ,1), the
alternative hypothesis acceptance interval for a lower one-tailed test is (0, ), and the
alternative hypothesis acceptance interval for a two-tailed test is (1 - /2,1) or (0, /2).
Note that accepting the alternative hypothesis is equivalent to rejecting the null
hypothesis.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda358.htm (2 of 3) [5/7/2002 3:15:03 PM]

1.3.5.8. Chi-Square Test for the Standard Deviation

3. The third section prints the conclusions for a 95% test since this is the most common
case. Results are given in terms of the alternative hypothesis for the two-tailed test and
for the one-tailed test in both directions. The alternative hypothesis acceptance interval
column is stated in terms of the cdf value printed in section two. The last column
specifies whether the alternative hypothesis is accepted or rejected. For a different
significance level, the appropriate conclusion can be drawn from the chi-square test
statistic cdf value printed in section two. For example, for a significance level of 0.10,
the corresponding alternative hypothesis acceptance intervals are (0,0.05) and (0.95,1),
(0, 0.10), and (0.90,1).
Output from other statistical software may look somewhat different from the above output.

Questions The chi-square test can be used to answer the following questions:
1. Is the standard deviation equal to some pre-determined threshold value?
2. Is the standard deviation greater than some pre-determined threshold value?
3. Is the standard deviation less than some pre-determined threshold value?

Related F Test
Techniques Bartlett Test
Levene Test

Software The chi-square test for the standard deviation is available in many general purpose statistical
software programs, including Dataplot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda358.htm (3 of 3) [5/7/2002 3:15:03 PM]

1.3.5.8.1. Data Used for Chi-Square Test for the Standard Deviation

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques
1.3.5.8. Chi-Square Test for the Standard Deviation

1.3.5.8.1. Data Used for Chi-Square Test for

the Standard Deviation
Data Used The following are the data used for the chi-square test for the standard
for deviation example. The first column is gear diameter and the second
Chi-Square column is batch number. Only the first column is used for this example.
Test for the
Standard
1.006 1.000
Deviation
0.996 1.000
Example
0.998 1.000
1.000 1.000
0.992 1.000
0.993 1.000
1.002 1.000
0.999 1.000
0.994 1.000
1.000 1.000
0.998 2.000
1.006 2.000
1.000 2.000
1.002 2.000
0.997 2.000
0.998 2.000
0.996 2.000
1.000 2.000
1.006 2.000
0.988 2.000
0.991 3.000
0.987 3.000
0.997 3.000
0.999 3.000
0.995 3.000
0.994 3.000
1.000 3.000

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3581.htm (1 of 3) [5/7/2002 3:15:04 PM]

1.3.5.8.1. Data Used for Chi-Square Test for the Standard Deviation

0.999 3.000
0.996 3.000
0.996 3.000
1.005 4.000
1.002 4.000
0.994 4.000
1.000 4.000
0.995 4.000
0.994 4.000
0.998 4.000
0.996 4.000
1.002 4.000
0.996 4.000
0.998 5.000
0.998 5.000
0.982 5.000
0.990 5.000
1.002 5.000
0.984 5.000
0.996 5.000
0.993 5.000
0.980 5.000
0.996 5.000
1.009 6.000
1.013 6.000
1.009 6.000
0.997 6.000
0.988 6.000
1.002 6.000
0.995 6.000
0.998 6.000
0.981 6.000
0.996 6.000
0.990 7.000
1.004 7.000
0.996 7.000
1.001 7.000
0.998 7.000
1.000 7.000
1.018 7.000
1.010 7.000
0.996 7.000
1.002 7.000
0.998 8.000
1.000 8.000
1.006 8.000

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3581.htm (2 of 3) [5/7/2002 3:15:04 PM]

1.3.5.8.1. Data Used for Chi-Square Test for the Standard Deviation

1.000 8.000
1.002 8.000
0.996 8.000
0.998 8.000
0.996 8.000
1.002 8.000
1.006 8.000
1.002 9.000
0.998 9.000
0.996 9.000
0.995 9.000
0.996 9.000
1.004 9.000
1.004 9.000
0.998 9.000
0.999 9.000
0.991 9.000
0.991 10.000
0.995 10.000
0.984 10.000
0.994 10.000
0.997 10.000
0.997 10.000
0.991 10.000
0.998 10.000
1.004 10.000
0.997 10.000

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3581.htm (3 of 3) [5/7/2002 3:15:04 PM]

1.3.5.9. F-Test for Equality of Two Standard Deviations

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques

1.3.5.9. F-Test for Equality of Two Standard

Deviations
Purpose: An F-test ( Snedecor and Cochran, 1983) is used to test if the standard deviations of two
Test if populations are equal. This test can be a two-tailed test or a one-tailed test. The
standard two-tailed version tests against the alternative that the standard deviations are not equal.
deviations The one-tailed version only tests in one direction, that is the standard deviation from the
from two first population is either greater than or less than (but not both) the second population
populations standard deviation . The choice is determined by the problem. For example, if we are
are equal testing a new process, we may only be interested in knowing if the new process is less
variable than the old process.

Definition The F hypothesis test is defined as:

H0:
Ha: for a lower one tailed test

for an upper one tailed test

for a two tailed test

Test F=
Statistic:
where and are the sample variances. The more this ratio deviates
from 1, the stronger the evidence for unequal population variances.
Significance
Level:

http://www.itl.nist.gov/div898/handbook/eda/section3/eda359.htm (1 of 3) [5/7/2002 3:15:04 PM]

1.3.5.9. F-Test for Equality of Two Standard Deviations

Critical The hypothesis that the two standard deviations are equal is rejected if
Region:
for an upper one-tailed test

for a lower one-tailed test

for a two-tailed test

or

where is the critical value of the F distribution with and

degrees of freedom and a significance level of .
In the above formulas for the critical regions, the Handbook follows the
convention that is the upper critical value from the F distribution and
is the lower critical value from the F distribution. Note that this is
the opposite of the designation used by some texts and software programs.
In particular, Dataplot uses the opposite convention.

Sample Dataplot generated the following output for an F-test from the JAHANMI2.DAT data
Output set:

F TEST
NULL HYPOTHESIS UNDER TEST--SIGMA1 = SIGMA2
ALTERNATIVE HYPOTHESIS UNDER TEST--SIGMA1 NOT EQUAL SIGMA2

SAMPLE 1:
NUMBER OF OBSERVATIONS = 240
MEAN = 688.9987
STANDARD DEVIATION = 65.54909

SAMPLE 2:
NUMBER OF OBSERVATIONS = 240
MEAN = 611.1559
STANDARD DEVIATION = 61.85425

TEST:
STANDARD DEV. (NUMERATOR) = 65.54909
STANDARD DEV. (DENOMINATOR) = 61.85425
F TEST STATISTIC VALUE = 1.123037
DEG. OF FREEDOM (NUMER.) = 239.0000
DEG. OF FREEDOM (DENOM.) = 239.0000
F TEST STATISTIC CDF VALUE = 0.814808

NULL NULL HYPOTHESIS NULL HYPOTHESIS

HYPOTHESIS ACCEPTANCE INTERVAL CONCLUSION
SIGMA1 = SIGMA2 (0.000,0.950) ACCEPT

http://www.itl.nist.gov/div898/handbook/eda/section3/eda359.htm (2 of 3) [5/7/2002 3:15:04 PM]

1.3.5.9. F-Test for Equality of Two Standard Deviations

Interpretation We are testing the hypothesis that the standard deviations for sample one and sample
of Sample two are equal. The output is divided into four sections.
Output 1. The first section prints the sample statistics for sample one used in the
computation of the F-test.
2. The second section prints the sample statistics for sample two used in the
computation of the F-test.
3. The third section prints the numerator and denominator standard deviations, the
F-test statistic value, the degrees of freedom, and the cumulative distribution
function (cdf) value of the F-test statistic. The F-test statistic cdf value is an
alternative way of expressing the critical value. This cdf value is compared to the
acceptance interval printed in section four. The acceptance interval for a
two-tailed test is (0,1 - ).
4. The fourth section prints the conclusions for a 95% test since this is the most
common case. Results are printed for an upper one-tailed test. The acceptance
interval column is stated in terms of the cdf value printed in section three. The
last column specifies whether the null hypothesis is accepted or rejected. For a
different significance level, the appropriate conclusion can be drawn from the
F-test statistic cdf value printed in section four. For example, for a significance
level of 0.10, the corresponding acceptance interval become (0.000,0.9000).
Output from other statistical software may look somewhat different from the above
output.

Questions The F-test can be used to answer the following questions:

1. Do two samples come from populations with equal standard deviations?
2. Does a new process, treatment, or test reduce the variability of the current
process?

Related Quantile-Quantile Plot

Techniques Bihistogram
Chi-Square Test
Bartlett's Test
Levene Test

Case Study Ceramic strength data.

Software The F-test for equality of two standard deviations is available in many general purpose
statistical software programs, including Dataplot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda359.htm (3 of 3) [5/7/2002 3:15:04 PM]

1.3.5.10. Levene Test for Equality of Variances

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques

1.3.5.10. Levene Test for Equality of

Variances
Purpose: Levene's test ( Levene 1960) is used to test if k samples have equal
Test for variances. Equal variances across samples is called homogeneity of
Homogeneity variance. Some statistical tests, for example the analysis of variance,
of Variances assume that variances are equal across groups or samples. The Levene test
can be used to verify that assumption.
Levene's test is an alternative to the Bartlett test. The Levene test is less
sensitive than the Bartlett test to departures from normality. If you have
strong evidence that your data do in fact come from a normal, or nearly
normal, distribution, then Bartlett's test has better performance.

Definition The Levene test is defined as:

H0:
Ha: for at least one pair (i,j).
Test Given a variable Y with sample of size N divided into k
Statistic: subgroups, where Ni is the sample size of the ith subgroup,
the Levene test statistic is defined as:

where Zij can have one of the following three definitions:

1.

where is the mean of the ith subgroup.

2.

where is the median of the ith subgroup.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35a.htm (1 of 4) [5/7/2002 3:15:05 PM]

1.3.5.10. Levene Test for Equality of Variances
3.

where is the 10% trimmed mean of the ith

subgroup.
are the group means of the Zij and is the overall
mean of the Zij.

The three choices for defining Zij determine the robustness

and power of Levene's test. By robustness, we mean the
ability of the test to not falsely detect unequal variances
when the underlying data are not normally distributed and
the variables are in fact equal. By power, we mean the
ability of the test to detect unequal variances when the
variances are in fact unequal.
Levene's original paper only proposed using the mean.
Brown and Forsythe (1974)) extended Levene's test to use
either the median or the trimmed mean in addition to the
mean. They performed Monte Carlo studies that indicated
that using the trimmed mean performed best when the
underlying data followed a Cauchy distribution (i.e.,
heavy-tailed) and the median performed best when the
underlying data followed a (i.e., skewed) distribution.
Using the mean provided the best power for symmetric,
moderate-tailed, distributions.
Although the optimal choice depends on the underlying
distribution, the definition based on the median is
recommended as the choice that provides good robustness
against many types of non-normal data while retaining
good power. If you have knowledge of the underlying
distribution of the data, this may indicate using one of the
other choices.
Significance
Level:

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35a.htm (2 of 4) [5/7/2002 3:15:05 PM]

1.3.5.10. Levene Test for Equality of Variances

Critical The Levene test rejects the hypothesis that the variances are
Region: equal if

where is the upper critical value of the F

distribution with k - 1 and N - k degrees of freedom at a
significance level of .
In the above formulas for the critical regions, the Handbook
follows the convention that is the upper critical value
from the F distribution and is the lower critical
value. Note that this is the opposite of some texts and
software programs. In particular, Dataplot uses the opposite
convention.

Sample Dataplot generated the following output for Levene's test using the
Output GEAR.DAT data set:

LEVENE F-TEST FOR SHIFT IN VARIATION

(ASSUMPTION: NORMALITY)

1. STATISTICS
NUMBER OF OBSERVATIONS = 100
NUMBER OF GROUPS = 10
LEVENE F TEST STATISTIC = 1.705910

2. FOR LEVENE TEST STATISTIC

0 % POINT = 0.
50 % POINT = 0.9339308
75 % POINT = 1.296365
90 % POINT = 1.702053
95 % POINT = 1.985595
99 % POINT = 2.610880
99.9 % POINT = 3.478882

90.09152 % Point: 1.705910

3. CONCLUSION (AT THE 5% LEVEL):

THERE IS NO SHIFT IN VARIATION.
THUS: HOMOGENEOUS WITH RESPECT TO VARIATION.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35a.htm (3 of 4) [5/7/2002 3:15:05 PM]

1.3.5.10. Levene Test for Equality of Variances

Interpretation We are testing the hypothesis that the group variances are equal. The
of Sample output is divided into three sections.
Output 1. The first section prints the number of observations (N), the number
of groups (k), and the value of the Levene test statistic.
2. The second section prints the upper critical value of the F
distribution corresponding to various significance levels. The value
in the first column, the confidence level of the test, is equivalent to
100(1- ). We reject the null hypothesis at that significance level if
the value of the Levene F test statistic printed in section one is
greater than the critical value printed in the last column.
3. The third section prints the conclusion for a 95% test. For a
different significance level, the appropriate conclusion can be drawn
from the table printed in section two. For example, for = 0.10, we
look at the row for 90% confidence and compare the critical value
1.702 to the Levene test statistic 1.7059. Since the test statistic is
greater than the critical value, we reject the null hypothesis at the
= 0.10 level.
Output from other statistical software may look somewhat different from
the above output.

Question Levene's test can be used to answer the following question:

● Is the assumption of equal variances valid?

Related Standard Deviation Plot

Techniques Box Plot
Bartlett Test
Chi-Square Test
Analysis of Variance

Software The Levene test is available in some general purpose statistical software
programs, including Dataplot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35a.htm (4 of 4) [5/7/2002 3:15:05 PM]

1.3.5.11. Measures of Skewness and Kurtosis

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques

1.3.5.11. Measures of Skewness and

Kurtosis
Skewness A fundamental task in many statistical analyses is to characterize the
and Kurtosis location and variability of a data set. A further characterization of the
data includes skewness and kurtosis.
Skewness is a measure of symmetry, or more precisely, the lack of
symmetry. A distribution, or data set, is symmetric if it looks the same
to the left and right of the center point.
Kurtosis is a measure of whether the data are peaked or flat relative to a
normal distribution. That is, data sets with high kurtosis tend to have a
distinct peak near the mean, decline rather rapidly, and have heavy tails.
Data sets with low kurtosis tend to have a flat top near the mean rather
than a sharp peak. A uniform distribution would be the extreme case.
The histogram is an effective graphical technique for showing both the
skewness and kurtosis of data set.

Definition of For univariate data Y1, Y2, ..., YN, the formula for skewness is:
Skewness

where is the mean, is the standard deviation, and N is the number of

data points. The skewness for a normal distribution is zero, and any
symmetric data should have a skewness near zero. Negative values for
the skewness indicate data that are skewed left and positive values for
the skewness indicate data that are skewed right. By skewed left, we
mean that the left tail is heavier than the right tail. Similarly, skewed
right means that the right tail is heavier than the left tail. Some
measurements have a lower bound and are skewed right. For example,
in reliability studies, failure times cannot be negative.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35b.htm (1 of 4) [5/7/2002 3:15:06 PM]

1.3.5.11. Measures of Skewness and Kurtosis

Definition of For univariate data Y1, Y2, ..., YN, the formula for kurtosis is:
Kurtosis

where is the mean, is the standard deviation, and N is the number of

data points.
The kurtosis for a standard normal distribution is three. For this reason,
excess kurtosis is defined as

so that the standard normal distribution has a kurtosis of zero. Positive

kurtosis indicates a "peaked" distribution and negative kurtosis indicates
a "flat" distribution.

Examples The following example shows histograms for 10,000 random numbers
generated from a normal, a double exponential, a Cauchy, and a Weibull
distribution.

Normal The first histogram is a sample from a normal distribution. The normal
Distribution distribution is a symmetric distribution with well-behaved tails. This is
indicated by the skewness of 0.03. The kurtosis of 2.96 is near the
expected value of 3. The histogram verifies the symmetry.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35b.htm (2 of 4) [5/7/2002 3:15:06 PM]

1.3.5.11. Measures of Skewness and Kurtosis

Double The second histogram is a sample from a double exponential

Exponential distribution. The double exponential is a symmetric distribution.
Distribution Compared to the normal, it has a stronger peak, more rapid decay, and
heavier tails. That is, we would expect a skewness near zero and a
kurtosis higher than 3. The skewness is 0.06 and the kurtosis is 5.9.

Cauchy The third histogram is a sample from a Cauchy distribution.

Distribution
For better visual comparison with the other data sets, we restricted the
histogram of the Cauchy distribution to values between -10 and 10. The
full data set for the Cauchy data in fact has a minimum of approximately
-29,000 and a maximum of approximately 89,000.
The Cauchy distribution is a symmetric distribution with heavy tails and
a single peak at the center of the distribution. Since it is symmetric, we
would expect a skewness near zero. Due to the heavier tails, we might
expect the kurtosis to be larger than for a normal distribution. In fact the
skewness is 69.99 and the kurtosis is 6,693. These extremely high
values can be explained by the heavy tails. Just as the mean and
standard deviation can be distorted by extreme values in the tails, so too
can the skewness and kurtosis measures.

Weibull The fourth histogram is a sample from a Weibull distribution with shape
Distribution parameter 1.5. The Weibull distribution is a skewed distribution with the
amount of skewness depending on the value of the shape parameter. The
degree of decay as we move away from the center also depends on the
value of the shape parameter. For this data set, the skewness is 1.08 and
the kurtosis is 4.46, which indicates moderate skewness and kurtosis.

Dealing Many classical statistical tests and intervals depend on normality

with assumptions. Significant skewness and kurtosis clearly indicate that data
Skewness are not normal. If a data set exhibits significant skewness or kurtosis (as
and Kurtosis indicated by a histogram or the numerical measures), what can we do
about it?
One approach is to apply some type of transformation to try to make the
data normal, or more nearly normal. The Box-Cox transformation is a
useful technique for trying to normalize a data set. In particular, taking
the log or square root of a data set is often useful for data that exhibit
moderate right skewness.
Another approach is to use techniques based on distributions other than
the normal. For example, in reliability studies, the exponential, Weibull,
and lognormal distributions are typically used as a basis for modeling
rather than using the normal distribution. The probability plot

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35b.htm (3 of 4) [5/7/2002 3:15:06 PM]

1.3.5.11. Measures of Skewness and Kurtosis

correlation coefficient plot and the probability plot are useful tools for
determining a good distributional model for the data.

Software The skewness and kurtosis coefficients are available in most general
purpose statistical software programs, including Dataplot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35b.htm (4 of 4) [5/7/2002 3:15:06 PM]

1.3.5.12. Autocorrelation

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques

1.3.5.12. Autocorrelation
Purpose: The autocorrelation ( Box and Jenkins, 1976) function can be used for
Detect the following two purposes:
Non-Randomness, 1. To detect non-randomness in data.
Time Series
Modeling 2. To identify an appropriate time series model if the data are not
random.

Definition Given measurements, Y1, Y2, ..., YN at time X1, X2, ..., XN, the lag k
autocorrelation function is defined as

Although the time variable, X, is not used in the formula for

autocorrelation, the assumption is that the observations are equi-spaced.
Autocorrelation is a correlation coefficient. However, instead of
correlation between two different variables, the correlation is between
two values of the same variable at times Xi and Xi+k.

When the autocorrelation is used to detect non-randomness, it is

usually only the first (lag 1) autocorrelation that is of interest. When the
autocorrelation is used to identify an appropriate time series model, the
autocorrelations are usually plotted for many lags.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35c.htm (1 of 4) [5/7/2002 3:15:13 PM]

1.3.5.12. Autocorrelation

Sample Output Dataplot generated the following autocorrelation output using the
LEW.DAT data set:

THE LAG-ONE AUTOCORRELATION COEFFICIENT OF THE

200 OBSERVATIONS = -0.3073048E+00

THE COMPUTED VALUE OF THE CONSTANT A = -0.30730480E+00

lag autocorrelation
0. 1.00
1. -0.31
2. -0.74
3. 0.77
4. 0.21
5. -0.90
6. 0.38
7. 0.63
8. -0.77
9. -0.12
10. 0.82
11. -0.40
12. -0.55
13. 0.73
14. 0.07
15. -0.76
16. 0.40
17. 0.48
18. -0.70
19. -0.03
20. 0.70
21. -0.41
22. -0.43
23. 0.67
24. 0.00
25. -0.66
26. 0.42
27. 0.39
28. -0.65
29. 0.03
30. 0.63
31. -0.42
32. -0.36
33. 0.64
34. -0.05
35. -0.60
36. 0.43
37. 0.32
38. -0.64
39. 0.08
40. 0.58

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35c.htm (2 of 4) [5/7/2002 3:15:13 PM]

1.3.5.12. Autocorrelation

41. -0.45
42. -0.28
43. 0.62
44. -0.10
45. -0.55
46. 0.45
47. 0.25
48. -0.61
49. 0.14

Questions The autocorrelation function can be used to answer the following

questions
1. Was this sample data set generated from a random process?
2. Would a non-linear or time series model be a more appropriate
model for these data than a simple constant plus error model?

Importance Randomness is one of the key assumptions in determining if a

univariate statistical process is in control. If the assumptions of
constant location and scale, randomness, and fixed distribution are
reasonable, then the univariate process can be modeled as:

where Ei is an error term.

If the randomness assumption is not valid, then a different model needs

to be used. This will typically be either a time series model or a
non-linear model (with time as the independent variable).

Related Autocorrelation Plot

Techniques Run Sequence Plot
Lag Plot
Runs Test

Case Study The heat flow meter data demonstrate the use of autocorrelation in
determining if the data are from a random process.
The beam deflection data demonstrate the use of autocorrelation in
developing a non-linear sinusoidal model.

Software The autocorrelation capability is available in most general purpose

statistical software programs, including Dataplot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35c.htm (3 of 4) [5/7/2002 3:15:13 PM]

1.3.5.12. Autocorrelation

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35c.htm (4 of 4) [5/7/2002 3:15:13 PM]

1.3.5.13. Runs Test for Detecting Non-randomness

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques

1.3.5.13. Runs Test for Detecting

Non-randomness
Purpose: The runs test ( Bradley, 1968) can be used to decide if a data set is from a
Detect random process.
Non-Randomness
A run is defined as a series of increasing values or a series of decreasing
values. The number of increasing, or decreasing, values is the length of the
run. In a random data set, the probability that the (I+1)th value is larger or
smaller than the Ith value follows a binomial distribution, which forms the
basis of the runs test.

Typical Analysis The first step in the runs test is to compute the sequential differences (Yi -
and Test Yi-1). Positive values indicate an increasing value and negative values
Statistics indicate a decreasing value. A runs test should include information such as
the output shown below from Dataplot for the LEW.DAT data set. The
output shows a table of:
1. runs of length exactly I for I = 1, 2, ..., 10
2. number of runs of length I
3. expected number of runs of length I
4. standard deviation of the number of runs of length I
5. a z-score where the z-score is defined to be

where is the sample mean and s is the sample standard deviation.

The z-score column is compared to a standard normal table. That is, at the
5% significance level, a z-score with an absolute value greater than 1.96
indicates non-randomness.
There are several alternative formulations of the runs test in the literature. For
example, a series of coin tosses would record a series of heads and tails. A

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35d.htm (1 of 5) [5/7/2002 3:15:13 PM]

1.3.5.13. Runs Test for Detecting Non-randomness

run of length r is r consecutive heads or r consecutive tails. To use the

Dataplot RUNS command, you could code a sequence of the N = 10 coin
tosses HHHHTTHTHH as
1234323234
that is, a heads is coded as an increasing value and a tails is coded as a
decreasing value.
Another alternative is to code values above the median as positive and values
below the median as negative. There are other formulations as well. All of
them can be converted to the Dataplot formulation. Just remember that it
ultimately reduces to 2 choices. To use the Dataplot runs test, simply code
one choice as an increasing value and the other as a decreasing value as in the
heads/tails example above. If you are using other statistical software, you
need to check the conventions used by that program.

Sample Output Dataplot generated the following runs test output using the LEW.DAT data
set:

RUNS UP

STATISTIC = NUMBER OF RUNS UP

OF LENGTH EXACTLY I

I STAT EXP(STAT) SD(STAT) Z

1 18.0 41.7083 6.4900 -3.65

2 40.0 18.2167 3.3444 6.51
3 2.0 5.2125 2.0355 -1.58
4 0.0 1.1302 1.0286 -1.10
5 0.0 0.1986 0.4424 -0.45
6 0.0 0.0294 0.1714 -0.17
7 0.0 0.0038 0.0615 -0.06
8 0.0 0.0004 0.0207 -0.02
9 0.0 0.0000 0.0066 -0.01
10 0.0 0.0000 0.0020 0.00

STATISTIC = NUMBER OF RUNS UP

OF LENGTH I OR MORE

I STAT EXP(STAT) SD(STAT) Z

1 60.0 66.5000 4.1972 -1.55

2 42.0 24.7917 2.8083 6.13

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35d.htm (2 of 5) [5/7/2002 3:15:13 PM]

1.3.5.13. Runs Test for Detecting Non-randomness

3 2.0 6.5750 2.1639 -2.11

4 0.0 1.3625 1.1186 -1.22
5 0.0 0.2323 0.4777 -0.49
6 0.0 0.0337 0.1833 -0.18
7 0.0 0.0043 0.0652 -0.07
8 0.0 0.0005 0.0218 -0.02
9 0.0 0.0000 0.0069 -0.01
10 0.0 0.0000 0.0021 0.00

RUNS DOWN

STATISTIC = NUMBER OF RUNS DOWN

OF LENGTH EXACTLY I

I STAT EXP(STAT) SD(STAT) Z

1 25.0 41.7083 6.4900 -2.57

2 35.0 18.2167 3.3444 5.02
3 0.0 5.2125 2.0355 -2.56
4 0.0 1.1302 1.0286 -1.10
5 0.0 0.1986 0.4424 -0.45
6 0.0 0.0294 0.1714 -0.17
7 0.0 0.0038 0.0615 -0.06
8 0.0 0.0004 0.0207 -0.02
9 0.0 0.0000 0.0066 -0.01
10 0.0 0.0000 0.0020 0.00

STATISTIC = NUMBER OF RUNS DOWN

OF LENGTH I OR MORE

I STAT EXP(STAT) SD(STAT) Z

1 60.0 66.5000 4.1972 -1.55

2 35.0 24.7917 2.8083 3.63
3 0.0 6.5750 2.1639 -3.04
4 0.0 1.3625 1.1186 -1.22
5 0.0 0.2323 0.4777 -0.49
6 0.0 0.0337 0.1833 -0.18
7 0.0 0.0043 0.0652 -0.07
8 0.0 0.0005 0.0218 -0.02
9 0.0 0.0000 0.0069 -0.01
10 0.0 0.0000 0.0021 0.00

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35d.htm (3 of 5) [5/7/2002 3:15:13 PM]

1.3.5.13. Runs Test for Detecting Non-randomness

RUNS TOTAL = RUNS UP + RUNS DOWN

STATISTIC = NUMBER OF RUNS TOTAL

OF LENGTH EXACTLY I

I STAT EXP(STAT) SD(STAT) Z

1 43.0 83.4167 9.1783 -4.40

2 75.0 36.4333 4.7298 8.15
3 2.0 10.4250 2.8786 -2.93
4 0.0 2.2603 1.4547 -1.55
5 0.0 0.3973 0.6257 -0.63
6 0.0 0.0589 0.2424 -0.24
7 0.0 0.0076 0.0869 -0.09
8 0.0 0.0009 0.0293 -0.03
9 0.0 0.0001 0.0093 -0.01
10 0.0 0.0000 0.0028 0.00

STATISTIC = NUMBER OF RUNS TOTAL

OF LENGTH I OR MORE

I STAT EXP(STAT) SD(STAT) Z

1 120.0 133.0000 5.9358 -2.19

2 77.0 49.5833 3.9716 6.90
3 2.0 13.1500 3.0602 -3.64
4 0.0 2.7250 1.5820 -1.72
5 0.0 0.4647 0.6756 -0.69
6 0.0 0.0674 0.2592 -0.26
7 0.0 0.0085 0.0923 -0.09
8 0.0 0.0010 0.0309 -0.03
9 0.0 0.0001 0.0098 -0.01
10 0.0 0.0000 0.0030 0.00

LENGTH OF THE LONGEST RUN UP = 3

LENGTH OF THE LONGEST RUN DOWN = 2
LENGTH OF THE LONGEST RUN UP OR DOWN = 3

NUMBER OF POSITIVE DIFFERENCES = 104

NUMBER OF NEGATIVE DIFFERENCES = 95
NUMBER OF ZERO DIFFERENCES = 0

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35d.htm (4 of 5) [5/7/2002 3:15:13 PM]

1.3.5.13. Runs Test for Detecting Non-randomness

Interpretation of Scanning the last column labeled "Z", we note that most of the z-scores for
Sample Output run lengths 1, 2, and 3 have an absolute value greater than 1.96. This is strong
evidence that these data are in fact not random.
Output from other statistical software may look somewhat different from the
above output.

Question The runs test can be used to answer the following question:
● Were these sample data generated from a random process?

Importance Randomness is one of the key assumptions in determining if a univariate

statistical process is in control. If the assumptions of constant location and
scale, randomness, and fixed distribution are reasonable, then the univariate
process can be modeled as:

where Ei is an error term.

If the randomness assumption is not valid, then a different model needs to be

used. This will typically be either a times series model or a non-linear model
(with time as the independent variable).

Related Autocorrelation
Techniques Run Sequence Plot
Lag Plot

Case Study Heat flow meter data

Software Most general purpose statistical software programs, including Dataplot,

support a runs test.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35d.htm (5 of 5) [5/7/2002 3:15:13 PM]

1.3.5.14. Anderson-Darling Test

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques

1.3.5.14. Anderson-Darling Test

Purpose: The Anderson-Darling test (Stephens, 1974) is used to test if a sample of data
Test for came from a population with a specific distribution. It is a modification of the
Distributional Kolmogorov-Smirnov (K-S) test and gives more weight to the tails than does
Adequacy the K-S test. The K-S test is distribution free in the sense that the critical values
do not depend on the specific distribution being tested. The Anderson-Darling
test makes use of the specific distribution in calculating critical values. This
has the advantage of allowing a more sensitive test and the disadvantage that
critical values must be calculated for each distribution. Currently, tables of
critical values are available for the normal, lognormal, exponential, Weibull,
extreme value type I, and logistic distributions. We do not provide the tables of
critical values in this Handbook (see Stephens 1974, 1976, 1977, and 1979)
since this test is usually applied with a statistical software program that will
print the relevant critical values.
The Anderson-Darling test is an alternative to the chi-square and
Kolmogorov-Smirnov goodness-of-fit tests.

Definition The Anderson-Darling test is defined as:

H0: The data follow a specified distribution.
Ha: The data do not follow the specified distribution
Test The Anderson-Darling test statistic is defined as
Statistic:
where

F is the cumulative distribution function of the specified

distribution. Note that the Yi are the ordered data.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35e.htm (1 of 5) [5/7/2002 3:15:14 PM]

1.3.5.14. Anderson-Darling Test

Significance
Level:
Critical The critical values for the Anderson-Darling test are dependent
Region: on the specific distribution that is being tested. Tabulated values
and formulas have been published (Stephens, 1974, 1976, 1977,
1979) for a few specific distributions (normal, lognormal,
exponential, Weibull, logistic, extreme value type 1). The test is
a one-sided test and the hypothesis that the distribution is of a
specific form is rejected if the test statistic, A, is greater than the
critical value.
Note that for a given distribution, the Anderson-Darling statistic
may be multiplied by a constant (which usually depends on the
sample size, n). These constants are given in the various papers
by Stephens. In the sample output below, this is the "adjusted
Anderson-Darling" statistic. This is what should be compared
against the critical values. Also, be aware that different constants
(and therefore critical values) have been published. You just
need to be aware of what constant was used for a given set of
critical values (the needed constant is typically given with the
critical values).

Sample Dataplot generated the following output for the Anderson-Darling test. 1,000
Output random numbers were generated for a normal, double exponential, Cauchy,
and lognormal distribution. In all four cases, the Anderson-Darling test was
applied to test for a normal distribution. When the data were generated using a
normal distribution, the test statistic was small and the hypothesis was
accepted. When the data were generated using the double exponential, Cauchy,
and lognormal distributions, the statistics were significant, and the hypothesis
of an underlying normal distribution was rejected at significance levels of 0.10,
0.05, and 0.01.
The normal random numbers were stored in the variable Y1, the double
exponential random numbers were stored in the variable Y2, the Cauchy
random numbers were stored in the variable Y3, and the lognormal random
numbers were stored in the variable Y4.

***************************************
** anderson darling normal test y1 **
***************************************

ANDERSON-DARLING 1-SAMPLE TEST

THAT THE DATA CAME FROM A NORMAL DISTRIBUTION

1. STATISTICS:
NUMBER OF OBSERVATIONS = 1000
MEAN = 0.4359940E-02

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35e.htm (2 of 5) [5/7/2002 3:15:14 PM]

1.3.5.14. Anderson-Darling Test

STANDARD DEVIATION = 1.001816

ANDERSON-DARLING TEST STATISTIC VALUE = 0.2565918

ADJUSTED TEST STATISTIC VALUE = 0.2576117

2. CRITICAL VALUES:
90 % POINT = 0.6560000
95 % POINT = 0.7870000
97.5 % POINT = 0.9180000
99 % POINT = 1.092000

3. CONCLUSION (AT THE 5% LEVEL):

THE DATA DO COME FROM A NORMAL DISTRIBUTION.

***************************************
** anderson darling normal test y2 **
***************************************

ANDERSON-DARLING 1-SAMPLE TEST

THAT THE DATA CAME FROM A NORMAL DISTRIBUTION

1. STATISTICS:
NUMBER OF OBSERVATIONS = 1000
MEAN = 0.2034888E-01
STANDARD DEVIATION = 1.321627

ANDERSON-DARLING TEST STATISTIC VALUE = 5.826050

ADJUSTED TEST STATISTIC VALUE = 5.849208

2. CRITICAL VALUES:
90 % POINT = 0.6560000
95 % POINT = 0.7870000
97.5 % POINT = 0.9180000
99 % POINT = 1.092000

3. CONCLUSION (AT THE 5% LEVEL):

THE DATA DO NOT COME FROM A NORMAL DISTRIBUTION.

***************************************
** anderson darling normal test y3 **
***************************************

ANDERSON-DARLING 1-SAMPLE TEST

THAT THE DATA CAME FROM A NORMAL DISTRIBUTION

1. STATISTICS:
NUMBER OF OBSERVATIONS = 1000
MEAN = 1.503854
STANDARD DEVIATION = 35.13059

ANDERSON-DARLING TEST STATISTIC VALUE = 287.6429

ADJUSTED TEST STATISTIC VALUE = 288.7863

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35e.htm (3 of 5) [5/7/2002 3:15:14 PM]

1.3.5.14. Anderson-Darling Test

2. CRITICAL VALUES:
90 % POINT = 0.6560000
95 % POINT = 0.7870000
97.5 % POINT = 0.9180000
99 % POINT = 1.092000

3. CONCLUSION (AT THE 5% LEVEL):

THE DATA DO NOT COME FROM A NORMAL DISTRIBUTION.

***************************************
** anderson darling normal test y4 **
***************************************

ANDERSON-DARLING 1-SAMPLE TEST

THAT THE DATA CAME FROM A NORMAL DISTRIBUTION

1. STATISTICS:
NUMBER OF OBSERVATIONS = 1000
MEAN = 1.518372
STANDARD DEVIATION = 1.719969

ANDERSON-DARLING TEST STATISTIC VALUE = 83.06335

ADJUSTED TEST STATISTIC VALUE = 83.39352

2. CRITICAL VALUES:
90 % POINT = 0.6560000
95 % POINT = 0.7870000
97.5 % POINT = 0.9180000
99 % POINT = 1.092000

3. CONCLUSION (AT THE 5% LEVEL):

THE DATA DO NOT COME FROM A NORMAL DISTRIBUTION.

Interpretation The output is divided into three sections.

of the Sample
1. The first section prints the number of observations and estimates for the
Output
location and scale parameters.
2. The second section prints the upper critical value for the
Anderson-Darling test statistic distribution corresponding to various
significance levels. The value in the first column, the confidence level of
the test, is equivalent to 100(1- ). We reject the null hypothesis at that
significance level if the value of the Anderson-Darling test statistic
printed in section one is greater than the critical value printed in the last
column.
3. The third section prints the conclusion for a 95% test. For a different
significance level, the appropriate conclusion can be drawn from the
table printed in section two. For example, for = 0.10, we look at the
row for 90% confidence and compare the critical value 1.062 to the

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35e.htm (4 of 5) [5/7/2002 3:15:14 PM]

1.3.5.14. Anderson-Darling Test

Anderson-Darling test statistic (for the normal data) 0.256. Since the test
statistic is less than the critical value, we do not reject the null
hypothesis at the = 0.10 level.
As we would hope, the Anderson-Darling test accepts the hypothesis of
normality for the normal random numbers and rejects it for the 3 non-normal
cases.
The output from other statistical software programs may differ somewhat from
the output above.

Questions The Anderson-Darling test can be used to answer the following questions:
● Are the data from a normal distribution?

● Are the data from a log-normal distribution?

● Are the data from a Weibull distribution?

● Are the data from an exponential distribution?

● Are the data from a logistic distribution?

Importance Many statistical tests and procedures are based on specific distributional
assumptions. The assumption of normality is particularly common in classical
statistical tests. Much reliability modeling is based on the assumption that the
data follow a Weibull distribution.
There are many non-parametric and robust techniques that do not make strong
distributional assumptions. However, techniques based on specific
distributional assumptions are in general more powerful than non-parametric
and robust techniques. Therefore, if the distributional assumptions can be
validated, they are generally preferred.

Related Chi-Square goodness-of-fit Test

Techniques Kolmogorov-Smirnov Test
Shapiro-Wilk Normality Test
Probability Plot
Probability Plot Correlation Coefficient Plot

Case Study Airplane glass failure time data.

Software The Anderson-Darling goodness-of-fit test is available in some general purpose

statistical software programs, including Dataplot.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35e.htm (5 of 5) [5/7/2002 3:15:14 PM]

1.3.5.15. Chi-Square Goodness-of-Fit Test

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques

1.3.5.15. Chi-Square Goodness-of-Fit Test

Purpose: The chi-square test (Snedecor and Cochran, 1989) is used to test if a sample of data came
Test for from a population with a specific distribution.
distributional
adequacy An attractive feature of the chi-square goodness-of-fit test is that it can be applied to any
univariate distribution for which you can calculate the cumulative distribution function.
The chi-square goodness-of-fit test is applied to binned data (i.e., data put into classes).
This is actually not a restriction since for non-binned data you can simply calculate a
histogram or frequency table before generating the chi-square test. However, the value of
the chi-square test statistic are dependent on how the data is binned. Another
disadvantage of the chi-square test is that it requires a sufficient sample size in order for
the chi-square approximation to be valid.
The chi-square test is an alternative to the Anderson-Darling and Kolmogorov-Smirnov
goodness-of-fit tests. The chi-square goodness-of-fit test can be applied to discrete
distributions such as the binomial and the Poisson. The Kolmogorov-Smirnov and
Anderson-Darling tests are restricted to continuous distributions.
Additional discussion of the chi-square goodness-of-fit test is contained in the product
and process comparisons chapter (chapter 7).

Definition The chi-square test is defined for the hypothesis:

H0: The data follow a specified distribution.
Ha: The data do not follow the specified distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35f.htm (1 of 6) [5/7/2002 3:15:14 PM]

1.3.5.15. Chi-Square Goodness-of-Fit Test

Test Statistic: For the chi-square goodness-of-fit computation, the data are divided
into k bins and the test statistic is defined as

where is the observed frequency for bin i and is the expected

frequency for bin i. The expected frequency is calculated by

where F is the cumulative Distribution function for the distribution

being tested, Yu is the upper limit for class i, Yl is the lower limit for
class i, and N is the sample size.
This test is sensitive to the choice of bins. There is no optimal choice
for the bin width (since the optimal bin width depends on the
distribution). Most reasonable choices should produce similar, but
not identical, results. Dataplot uses 0.3*s, where s is the sample
standard deviation, for the class width. The lower and upper bins are
at the sample mean plus and minus 6.0*s, respectively. For the
chi-square approximation to be valid, the expected frequency should
be at least 5. This test is not valid for small samples, and if some of
the counts are less than five, you may need to combine some bins in
the tails.
Significance Level: .
Critical Region: The test statistic follows, approximately, a chi-square distribution
with (k - c) degrees of freedom where k is the number of non-empty
cells and c = the number of estimated parameters (including location
and scale parameters and shape parameters) for the distribution + 1.
For example, for a 3-parameter Weibull distribution, c = 4.
Therefore, the hypothesis that the data are from a population with
the specified distribution is rejected if

where is the chi-square percent point function with k - c

degrees of freedom and a significance level of .
In the above formulas for the critical regions, the Handbook follows
the convention that is the upper critical value from the
chi-square distribution and is the lower critical value from the
chi-square distribution. Note that this is the opposite of what is used
in some texts and software programs. In particular, Dataplot uses the
opposite convention.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35f.htm (2 of 6) [5/7/2002 3:15:14 PM]

1.3.5.15. Chi-Square Goodness-of-Fit Test

Sample Dataplot generated the following output for the chi-square test where 1,000 random
Output numbers were generated for the normal, double exponential, t with 3 degrees of freedom,
and lognormal distributions. In all cases, the chi-square test was applied to test for a
normal distribution. The test statistics show the characteristics of the test; when the data
are from a normal distribution, the test statistic is small and the hypothesis is accepted;
when the data are from the double exponential, t, and lognormal distributions, the
statistics are significant and the hypothesis of an underlying normal distribution is
rejected at significance levels of 0.10, 0.05, and 0.01.
The normal random numbers were stored in the variable Y1, the double exponential
random numbers were stored in the variable Y2, the t random numbers were stored in the
variable Y3, and the lognormal random numbers were stored in the variable Y4.

*************************************************
** normal chi-square goodness of fit test y1 **
*************************************************

CHI-SQUARED GOODNESS-OF-FIT TEST

NULL HYPOTHESIS H0: DISTRIBUTION FITS THE DATA

ALTERNATE HYPOTHESIS HA: DISTRIBUTION DOES NOT FIT THE DATA
DISTRIBUTION: NORMAL

SAMPLE:
NUMBER OF OBSERVATIONS = 1000
NUMBER OF NON-EMPTY CELLS = 24
NUMBER OF PARAMETERS USED = 0

TEST:
CHI-SQUARED TEST STATISTIC = 17.52155
DEGREES OF FREEDOM = 23
CHI-SQUARED CDF VALUE = 0.217101

ALPHA LEVEL CUTOFF CONCLUSION

10% 32.00690 ACCEPT H0
5% 35.17246 ACCEPT H0
1% 41.63840 ACCEPT H0

CELL NUMBER, BIN MIDPOINT, OBSERVED FREQUENCY,

AND EXPECTED FREQUENCY
WRITTEN TO FILE DPST1F.DAT

*************************************************
** normal chi-square goodness of fit test y2 **
*************************************************

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35f.htm (3 of 6) [5/7/2002 3:15:14 PM]

1.3.5.15. Chi-Square Goodness-of-Fit Test

CHI-SQUARED GOODNESS-OF-FIT TEST

NULL HYPOTHESIS H0: DISTRIBUTION FITS THE DATA

ALTERNATE HYPOTHESIS HA: DISTRIBUTION DOES NOT FIT THE DATA
DISTRIBUTION: NORMAL

SAMPLE:
NUMBER OF OBSERVATIONS = 1000
NUMBER OF NON-EMPTY CELLS = 26
NUMBER OF PARAMETERS USED = 0

TEST:
CHI-SQUARED TEST STATISTIC = 2030.784
DEGREES OF FREEDOM = 25
CHI-SQUARED CDF VALUE = 1.000000

ALPHA LEVEL CUTOFF CONCLUSION

10% 34.38158 REJECT H0
5% 37.65248 REJECT H0
1% 44.31411 REJECT H0

CELL NUMBER, BIN MIDPOINT, OBSERVED FREQUENCY,

AND EXPECTED FREQUENCY
WRITTEN TO FILE DPST1F.DAT

*************************************************
** normal chi-square goodness of fit test y3 **
*************************************************

CHI-SQUARED GOODNESS-OF-FIT TEST

NULL HYPOTHESIS H0: DISTRIBUTION FITS THE DATA

ALTERNATE HYPOTHESIS HA: DISTRIBUTION DOES NOT FIT THE DATA
DISTRIBUTION: NORMAL

SAMPLE:
NUMBER OF OBSERVATIONS = 1000
NUMBER OF NON-EMPTY CELLS = 25
NUMBER OF PARAMETERS USED = 0

TEST:
CHI-SQUARED TEST STATISTIC = 103165.4
DEGREES OF FREEDOM = 24
CHI-SQUARED CDF VALUE = 1.000000

ALPHA LEVEL CUTOFF CONCLUSION

10% 33.19624 REJECT H0
5% 36.41503 REJECT H0

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35f.htm (4 of 6) [5/7/2002 3:15:14 PM]

1.3.5.15. Chi-Square Goodness-of-Fit Test

1% 42.97982 REJECT H0

CELL NUMBER, BIN MIDPOINT, OBSERVED FREQUENCY,

AND EXPECTED FREQUENCY
WRITTEN TO FILE DPST1F.DAT

*************************************************
** normal chi-square goodness of fit test y4 **
*************************************************

CHI-SQUARED GOODNESS-OF-FIT TEST

NULL HYPOTHESIS H0: DISTRIBUTION FITS THE DATA

ALTERNATE HYPOTHESIS HA: DISTRIBUTION DOES NOT FIT THE DATA
DISTRIBUTION: NORMAL

SAMPLE:
NUMBER OF OBSERVATIONS = 1000
NUMBER OF NON-EMPTY CELLS = 10
NUMBER OF PARAMETERS USED = 0

TEST:
CHI-SQUARED TEST STATISTIC = 1162098.
DEGREES OF FREEDOM = 9
CHI-SQUARED CDF VALUE = 1.000000

ALPHA LEVEL CUTOFF CONCLUSION

10% 14.68366 REJECT H0
5% 16.91898 REJECT H0
1% 21.66600 REJECT H0

CELL NUMBER, BIN MIDPOINT, OBSERVED FREQUENCY,

AND EXPECTED FREQUENCY
WRITTEN TO FILE DPST1F.DAT

As we would hope, the chi-square test does not reject the normality hypothesis for the
normal distribution data set and rejects it for the three non-normal cases.

Questions The chi-square test can be used to answer the following types of questions:
● Are the data from a normal distribution?

● Are the data from a log-normal distribution?

● Are the data from a Weibull distribution?

● Are the data from an exponential distribution?

● Are the data from a logistic distribution?

● Are the data from a binomial distribution?

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35f.htm (5 of 6) [5/7/2002 3:15:14 PM]

1.3.5.15. Chi-Square Goodness-of-Fit Test

Importance Many statistical tests and procedures are based on specific distributional assumptions.
The assumption of normality is particularly common in classical statistical tests. Much
reliability modeling is based on the assumption that the distribution of the data follows a
Weibull distribution.
There are many non-parametric and robust techniques that are not based on strong
distributional assumptions. By non-parametric, we mean a technique, such as the sign
test, that is not based on a specific distributional assumption. By robust, we mean a
statistical technique that performs well under a wide range of distributional assumptions.
However, techniques based on specific distributional assumptions are in general more
powerful than these non-parametric and robust techniques. By power, we mean the ability
to detect a difference when that difference actually exists. Therefore, if the distributional
assumption can be confirmed, the parametric techniques are generally preferred.
If you are using a technique that makes a normality (or some other type of distributional)
assumption, it is important to confirm that this assumption is in fact justified. If it is, the
more powerful parametric techniques can be used. If the distributional assumption is not
justified, a non-parametric or robust technique may be required.

Related Anderson-Darling Goodness-of-Fit Test

Techniques Kolmogorov-Smirnov Test
Shapiro-Wilk Normality Test
Probability Plots
Probability Plot Correlation Coefficient Plot

Case Study Airplane glass failure times data.

Software Some general purpose statistical software programs, including Dataplot, provide a
chi-square goodness-of-fit test for at least some of the common distributions.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35f.htm (6 of 6) [5/7/2002 3:15:14 PM]

1.3.5.16. Kolmogorov-Smirnov Goodness-of-Fit Test

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques

1.3.5.16. Kolmogorov-Smirnov Goodness-of-Fit

Test
Purpose: The Kolmogorov-Smirnov test (Chakravart, Laha, and Roy, 1967) is used to
Test for decide if a sample comes from a population with a specific distribution.
Distributional
Adequacy The Kolmogorov-Smirnov (K-S) test is based on the empirical distribution
function (ECDF). Given N ordered data points Y1, Y2, ..., YN, the ECDF is
defined as

where n(i) is the number of points less than Yi and the Yi are ordered from
smallest to largest value. This is a step function that increases by 1/N at the value
of each ordered data point.
The graph below is a plot of the empirical distribution function with a normal
cumulative distribution function for 100 normal random numbers. The K-S test is
based on the maximum distance between these two curves.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35g.htm (1 of 6) [5/7/2002 3:15:15 PM]

1.3.5.16. Kolmogorov-Smirnov Goodness-of-Fit Test

Characteristics An attractive feature of this test is that the distribution of the K-S test statistic
and itself does not depend on the underlying cumulative distribution function being
Limitations of tested. Another advantage is that it is an exact test (the chi-square goodness-of-fit
the K-S Test test depends on an adequate sample size for the approximations to be valid).
Despite these advantages, the K-S test has several important limitations:
1. It only applies to continuous distributions.
2. It tends to be more sensitive near the center of the distribution than at the
tails.
3. Perhaps the most serious limitation is that the distribution must be fully
specified. That is, if location, scale, and shape parameters are estimated
from the data, the critical region of the K-S test is no longer valid. It
typically must be determined by simulation.
Due to limitations 2 and 3 above, many analysts prefer to use the
Anderson-Darling goodness-of-fit test. However, the Anderson-Darling test is
only available for a few specific distributions.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35g.htm (2 of 6) [5/7/2002 3:15:15 PM]

1.3.5.16. Kolmogorov-Smirnov Goodness-of-Fit Test

Definition The Kolmogorov-Smirnov test is defined by:

H0: The data follow a specified distribution
Ha: The data do not follow the specified distribution
Test Statistic: The Kolmogorov-Smirnov test statistic is defined as

where F is the theoretical cumulative distribution of the

distribution being tested which must be a continuous
distribution (i.e., no discrete distributions such as the
binomial or Poisson), and it must be fully specified (i.e., the
location, scale, and shape parameters cannot be estimated
from the data).
Significance Level: .
Critical Values: The hypothesis regarding the distributional form is rejected if
the test statistic, D, is greater than the critical value obtained
from a table. There are several variations of these tables in
the literature that use somewhat different scalings for the
K-S test statistic and critical regions. These alternative
formulations should be equivalent, but it is necessary to
ensure that the test statistic is calculated in a way that is
consistent with how the critical values were tabulated.
We do not provide the K-S tables in the Handbook since
software programs that perform a K-S test will provide the
relevant critical values.

Sample Output Dataplot generated the following output for the Kolmogorov-Smirnov test where
1,000 random numbers were generated for a normal, double exponential, t with 3
degrees of freedom, and lognormal distributions. In all cases, the
Kolmogorov-Smirnov test was applied to test for a normal distribution. The
Kolmogorov-Smirnov test accepts the normality hypothesis for the case of normal
data and rejects it for the double exponential, t, and lognormal data with the
exception of the double exponential data being significant at the 0.01 significance
level.
The normal random numbers were stored in the variable Y1, the double
exponential random numbers were stored in the variable Y2, the t random
numbers were stored in the variable Y3, and the lognormal random numbers were
stored in the variable Y4.

*********************************************************
** normal Kolmogorov-Smirnov goodness of fit test y1 **
*********************************************************

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35g.htm (3 of 6) [5/7/2002 3:15:15 PM]

1.3.5.16. Kolmogorov-Smirnov Goodness-of-Fit Test

KOLMOGOROV-SMIRNOV GOODNESS-OF-FIT TEST

NULL HYPOTHESIS H0: DISTRIBUTION FITS THE DATA

ALTERNATE HYPOTHESIS HA: DISTRIBUTION DOES NOT FIT THE DATA
DISTRIBUTION: NORMAL
NUMBER OF OBSERVATIONS = 1000

TEST:
KOLMOGOROV-SMIRNOV TEST STATISTIC = 0.2414924E-01

ALPHA LEVEL CUTOFF CONCLUSION

10% 0.03858 ACCEPT H0
5% 0.04301 ACCEPT H0
1% 0.05155 ACCEPT H0

*********************************************************
** normal Kolmogorov-Smirnov goodness of fit test y2 **
*********************************************************

KOLMOGOROV-SMIRNOV GOODNESS-OF-FIT TEST

NULL HYPOTHESIS H0: DISTRIBUTION FITS THE DATA

ALTERNATE HYPOTHESIS HA: DISTRIBUTION DOES NOT FIT THE DATA
DISTRIBUTION: NORMAL
NUMBER OF OBSERVATIONS = 1000

TEST:
KOLMOGOROV-SMIRNOV TEST STATISTIC = 0.5140864E-01

ALPHA LEVEL CUTOFF CONCLUSION

10% 0.03858 REJECT H0
5% 0.04301 REJECT H0
1% 0.05155 ACCEPT H0

*********************************************************
** normal Kolmogorov-Smirnov goodness of fit test y3 **
*********************************************************

KOLMOGOROV-SMIRNOV GOODNESS-OF-FIT TEST

NULL HYPOTHESIS H0: DISTRIBUTION FITS THE DATA

ALTERNATE HYPOTHESIS HA: DISTRIBUTION DOES NOT FIT THE DATA
DISTRIBUTION: NORMAL
NUMBER OF OBSERVATIONS = 1000

TEST:
KOLMOGOROV-SMIRNOV TEST STATISTIC = 0.6119353E-01

ALPHA LEVEL CUTOFF CONCLUSION

10% 0.03858 REJECT H0
5% 0.04301 REJECT H0
1% 0.05155 REJECT H0

*********************************************************
** normal Kolmogorov-Smirnov goodness of fit test y4 **

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35g.htm (4 of 6) [5/7/2002 3:15:15 PM]

1.3.5.16. Kolmogorov-Smirnov Goodness-of-Fit Test

*********************************************************

KOLMOGOROV-SMIRNOV GOODNESS-OF-FIT TEST

NULL HYPOTHESIS H0: DISTRIBUTION FITS THE DATA

ALTERNATE HYPOTHESIS HA: DISTRIBUTION DOES NOT FIT THE DATA
DISTRIBUTION: NORMAL
NUMBER OF OBSERVATIONS = 1000

TEST:
KOLMOGOROV-SMIRNOV TEST STATISTIC = 0.5354889

ALPHA LEVEL CUTOFF CONCLUSION

10% 0.03858 REJECT H0
5% 0.04301 REJECT H0
1% 0.05155 REJECT H0

Questions The Kolmogorov-Smirnov test can be used to answer the following types of
questions:
● Are the data from a normal distribution?

● Are the data from a log-normal distribution?

● Are the data from a Weibull distribution?

● Are the data from an exponential distribution?

● Are the data from a logistic distribution?

Importance Many statistical tests and procedures are based on specific distributional
assumptions. The assumption of normality is particularly common in classical
statistical tests. Much reliability modeling is based on the assumption that the
data follow a Weibull distribution.
There are many non-parametric and robust techniques that are not based on strong
distributional assumptions. By non-parametric, we mean a technique, such as the
sign test, that is not based on a specific distributional assumption. By robust, we
mean a statistical technique that performs well under a wide range of
distributional assumptions. However, techniques based on specific distributional
assumptions are in general more powerful than these non-parametric and robust
techniques. By power, we mean the ability to detect a difference when that
difference actually exists. Therefore, if the distributional assumptions can be
confirmed, the parametric techniques are generally preferred.
If you are using a technique that makes a normality (or some other type of
distributional) assumption, it is important to confirm that this assumption is in
fact justified. If it is, the more powerful parametric techniques can be used. If the
distributional assumption is not justified, using a non-parametric or robust
technique may be required.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35g.htm (5 of 6) [5/7/2002 3:15:15 PM]

1.3.5.16. Kolmogorov-Smirnov Goodness-of-Fit Test

Related Anderson-Darling goodness-of-fit Test

Techniques Chi-Square goodness-of-fit Test
Shapiro-Wilk Normality Test
Probability Plots
Probability Plot Correlation Coefficient Plot

Case Study Airplane glass failure times data

Software Some general purpose statistical software programs, including Dataplot, support
the Kolmogorov-Smirnov goodness-of-fit test, at least for some of the more
common distributions.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35g.htm (6 of 6) [5/7/2002 3:15:15 PM]

1.3.5.17. Grubbs' Test for Outliers

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques

1.3.5.17. Grubbs' Test for Outliers

Purpose: Grubbs' test (Grubbs 1969 and Stefansky 1972) is used to detect
Detection of outliers in a univariate data set. It is based on the assumption of
Outliers normality. That is, you should first verify that your data can be
reasonably approximated by a normal distribution before applying the
Grubbs' test.
Grubbs' test detects one outlier at a time. This outlier is expunged from
the dataset and the test is iterated until no outliers are detected.
However, multiple iterations change the probabilities of detection, and
the test should not be used for sample sizes of six or less since it
frequently tags most of the points as outliers.
Grubbs' test is also known as the maximum normed residual test.

Definition Grubbs' test is defined for the hypothesis:

H0: There are no outliers in the data set
Ha: There is at least one outlier in the data set
Test The Grubbs' test statistic is defined as:
Statistic:

where and are the sample mean and standard

deviation. The Grubbs test statistic is the largest absolute
deviation from the sample mean in units of the sample
standard deviation.
Significance .
Level:

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35h.htm (1 of 4) [5/7/2002 3:15:15 PM]

1.3.5.17. Grubbs' Test for Outliers

Critical The hypothesis of no outliers is rejected if

Region:

where is the critical value of the

t-distribution with (N-2) degrees of freedom and a
significance level of /(2N).
In the above formulas for the critical regions, the
Handbook follows the convention that is the upper
critical value from the t-distribution and is the
lower critical value from the t-distribution. Note that this
is the opposite of what is used in some texts and software
programs. In particular, Dataplot uses the opposite
convention.

Sample Dataplot generated the following output for the ZARR13.DAT data set
Output showing that Grubbs' test finds no outliers in the dataset:

*********************
** grubbs test y **
*********************

GRUBBS TEST FOR OUTLIERS

(ASSUMPTION: NORMALITY)

1. STATISTICS:
NUMBER OF OBSERVATIONS = 195
MINIMUM = 9.196848
MEAN = 9.261460
MAXIMUM = 9.327973
STANDARD DEVIATION = 0.2278881E-01

GRUBBS TEST STATISTIC = 2.918673

2. PERCENT POINTS OF THE REFERENCE DISTRIBUTION

FOR GRUBBS TEST STATISTIC
0 % POINT = 0.
50 % POINT = 2.984294
75 % POINT = 3.181226
90 % POINT = 3.424672
95 % POINT = 3.597898
99 % POINT = 3.970215

37.59665 % POINT: 2.918673

3. CONCLUSION (AT THE 5% LEVEL):

THERE ARE NO OUTLIERS.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35h.htm (2 of 4) [5/7/2002 3:15:15 PM]

1.3.5.17. Grubbs' Test for Outliers

Interpretation The output is divided into three sections.

of Sample 1. The first section prints the sample statistics used in the
Output computation of the Grubbs' test and the value of the Grubbs' test
statistic.
2. The second section prints the upper critical value for the Grubbs'
test statistic distribution corresponding to various significance
levels. The value in the first column, the confidence level of the
test, is equivalent to 100(1- ). We reject the null hypothesis at
that significance level if the value of the Grubbs' test statistic
printed in section one is greater than the critical value printed in
the last column.
3. The third section prints the conclusion for a 95% test. For a
different significance level, the appropriate conclusion can be
drawn from the table printed in section two. For example, for
= 0.10, we look at the row for 90% confidence and compare the
critical value 3.42 to the Grubbs' test statistic 2.92. Since the test
statistic is less than the critical value, we accept the null
hypothesis at the = 0.10 level.
Output from other statistical software may look somewhat different
from the above output.

Questions Grubbs' test can be used to answer the following questions:

1. Does the data set contain any outliers?
2. How many outliers does it contain?

Importance Many statistical techniques are sensitive to the presence of outliers. For
example, simple calculations of the mean and standard deviation may
be distorted by a single grossly inaccurate data point.
Checking for outliers should be a routine part of any data analysis.
Potential outliers should be examined to see if they are possibly
erroneous. If the data point is in error, it should be corrected if possible
and deleted if it is not possible. If there is no reason to believe that the
outlying point is in error, it should not be deleted without careful
consideration. However, the use of more robust techniques may be
warranted. Robust techniques will often downweight the effect of
outlying points without deleting them.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35h.htm (3 of 4) [5/7/2002 3:15:15 PM]

1.3.5.17. Grubbs' Test for Outliers

Related Several graphical techniques can, and should, be used to detect

Techniques outliers. A simple run sequence plot, a box plot, or a histogram should
show any obviously outlying points.
Run Sequence Plot
Histogram
Box Plot
Normal Probability Plot
Lag Plot

Case Study Heat flow meter data.

Software Some general purpose statistical software programs, including

Dataplot, support the Grubbs' test.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35h.htm (4 of 4) [5/7/2002 3:15:15 PM]

1.3.5.18. Yates Analysis

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques

1.3.5.18. Yates Analysis

Purpose: Full factorial and fractional factorial designs are common in designed experiments for
Estimate engineering and scientific applications.
Factor
Effects in In these designs, each factor is assigned two levels. These are typically called the low
a 2-Level and high levels. For computational purposes, the factors are scaled so that the low
Factorial level is assigned a value of -1 and the high level is assigned a value of +1. These are
Design also commonly referred to as "-" and "+".
A full factorial design contains all possible combinations of low/high levels for all the
factors. A fractional factorial design contains a carefully chosen subset of these
combinations. The criterion for choosing the subsets is discussed in detail in the
process improvement chapter.
The Yates analysis exploits the special structure of these designs to generate least
squares estimates for factor effects for all factors and all relevant interactions.
The mathematical details of the Yates analysis are given in chapter 10 of Box, Hunter,
and Hunter (1978).
The Yates analysis is typically complemented by a number of graphical techniques
such as the dex mean plot and the dex contour plot ("dex" represents "design of
experiments"). This is demonstrated in the Eddy current case study.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35i.htm (1 of 5) [5/7/2002 3:15:16 PM]

1.3.5.18. Yates Analysis

Yates Before performing a Yates analysis, the data should be arranged in "Yates order". That
Order is, given k factors, the kth column consists of 2k-1 minus signs (i.e., the low level of the
factor) followed by 2k-1 plus signs (i.e., the high level of the factor). For example, for
a full factorial design with three factors, the design matrix is

- - -
+ - -
- + -
+ + -
- - +
+ - +
- + +
+ + +

Determining the Yates order for fractional factorial designs requires knowledge of the
confounding structure of the fractional factorial design.

Yates
A Yates analysis generates the following output.
Output
1. A factor identifier (from Yates order). The specific identifier will vary
depending on the program used to generate the Yates analysis. Dataplot, for
example, uses the following for a 3-factor model.
1 = factor 1
2 = factor 2
3 = factor 3
12 = interaction of factor 1 and factor 2
13 = interaction of factor 1 and factor 3
23 = interaction of factor 2 and factor 3
123 =interaction of factors 1, 2, and 3
2. Least squares estimated factor effects ordered from largest in magnitude (most
significant) to smallest in magnitude (least significant).
That is, we obtain a ranked list of important factors.
3. A t-value for the individual factor effect estimates. The t-value is computed as

where e is the estimated factor effect and is the standard deviation of the
estimated factor effect.
4. The residual standard deviation that results from the model with the single term
only. That is, the residual standard deviation from the model
response = constant + 0.5 (Xi)

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35i.htm (2 of 5) [5/7/2002 3:15:16 PM]

1.3.5.18. Yates Analysis

where Xi is the estimate of the ith factor or interaction effect.

5. The cumulative residual standard deviation that results from the model using the
current term plus all terms preceding that term. That is,
response = constant + 0.5 (all effect estimates down to and including the
effect of interest)
This consists of a monotonically decreasing set of residual standard deviations
(indicating a better fit as the number of terms in the model increases). The first
cumulative residual standard deviation is for the model
response = constant
where the constant is the overall mean of the response variable. The last
cumulative residual standard deviation is for the model
response = constant + 0.5*(all factor and interaction estimates)
This last model will have a residual standard deviation of zero.

Sample Dataplot generated the following Yates analysis output for the Eddy current data set:
Output

(NOTE--DATA MUST BE IN STANDARD ORDER)

NUMBER OF OBSERVATIONS = 8
NUMBER OF FACTORS = 3
NO REPLICATION CASE

PSEUDO-REPLICATION STAND. DEV. = 0.20152531564E+00

PSEUDO-DEGREES OF FREEDOM = 1
(THE PSEUDO-REP. STAND. DEV. ASSUMES ALL
3, 4, 5, ...-TERM INTERACTIONS ARE NOT REAL,
BUT MANIFESTATIONS OF RANDOM ERROR)

STANDARD DEVIATION OF A COEF. = 0.14249992371E+00

(BASED ON PSEUDO-REP. ST. DEV.)

GRAND MEAN = 0.26587500572E+01

GRAND STANDARD DEVIATION = 0.17410624027E+01

99% CONFIDENCE LIMITS (+-) = 0.90710897446E+01

95% CONFIDENCE LIMITS (+-) = 0.18106349707E+01
99.5% POINT OF T DISTRIBUTION = 0.63656803131E+02
97.5% POINT OF T DISTRIBUTION = 0.12706216812E+02

IDENTIFIER EFFECT RESSD: RESSD:T VALUE

MEAN + MEAN +
TERM CUM TERMS
----------------------------------------------------------

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35i.htm (3 of 5) [5/7/2002 3:15:16 PM]

1.3.5.18. Yates Analysis

MEAN 2.65875 1.74106 1.74106

1 3.10250 21.8* 0.57272 0.57272
2 -0.86750 -6.1 1.81264 0.30429
23 0.29750 2.1 1.87270 0.26737
13 0.24750 1.7 1.87513 0.23341
3 0.21250 1.5 1.87656 0.19121
123 0.14250 1.0 1.87876 0.18031
12 0.12750 0.9 1.87912 0.00000

Interpretation In summary, the Yates analysis provides us with the following ranked
of Sample list of important factors along with the estimated effect estimate.
Output 1. X1: effect estimate = 3.1025 ohms
2. X2: effect estimate = -0.8675 ohms
3. X2*X3: effect estimate = 0.2975 ohms
4. X1*X3: effect estimate = 0.2475 ohms
5. X3: effect estimate = 0.2125 ohms
6. X1*X2*X3: effect estimate = 0.1425 ohms
7. X1*X2: effect estimate = 0.1275 ohms

Model From the above Yates output, we can define the potential models from
Selection and the Yates analysis. An important component of a Yates analysis is
Validation selecting the best model from the available potential models.
Once a tentative model has been selected, the error term should follow
the assumptions for a univariate measurement process. That is, the
model should be validated by analyzing the residuals.

Graphical Some analysts may prefer a more graphical presentation of the Yates
Presentation results. In particular, the following plots may be useful:
1. Ordered data plot
2. Ordered absolute effects plot
3. Cumulative residual standard deviation plot

Questions The Yates analysis can be used to answer the following questions:
1. What is the ranked list of factors?
2. What is the goodness-of-fit (as measured by the residual
standard deviation) for the various models?

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35i.htm (4 of 5) [5/7/2002 3:15:16 PM]

1.3.5.18. Yates Analysis

Related Multi-factor analysis of variance

Techniques Dex mean plot
Block plot
Dex contour plot

Case Study The Yates analysis is demonstrated in the Eddy current case study.

Software Many general purpose statistical software programs, including

Dataplot, can perform a Yates analysis.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35i.htm (5 of 5) [5/7/2002 3:15:16 PM]

1.3.5.18.1. Defining Models and Prediction Equations

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques
1.3.5.18. Yates Analysis

1.3.5.18.1. Defining Models and Prediction

Equations
Parameter In most cases of least squares fitting, the model coefficients for previously added terms
Estimates change depending on what was successively added. For example, the X1 coefficient
Don't might change depending on whether or not an X2 term was included in the model. This
Change as is not the case when the design is orthogonal, as is a 23 full factorial design. For
Additional orthogonal designs, the estimates for the previously included terms do not change as
Terms additional terms are added. This means the ranked list of effect estimates
Added simultaneously serves as the least squares coefficient estimates for progressively more
complicated models.

Yates For convenience, we list the sample Yates output for the Eddy current data set here.
Table

(NOTE--DATA MUST BE IN STANDARD ORDER)

NUMBER OF OBSERVATIONS = 8
NUMBER OF FACTORS = 3
NO REPLICATION CASE

PSEUDO-REPLICATION STAND. DEV. = 0.20152531564E+00

PSEUDO-DEGREES OF FREEDOM = 1
(THE PSEUDO-REP. STAND. DEV. ASSUMES ALL
3, 4, 5, ...-TERM INTERACTIONS ARE NOT REAL,
BUT MANIFESTATIONS OF RANDOM ERROR)

STANDARD DEVIATION OF A COEF. = 0.14249992371E+00

(BASED ON PSEUDO-REP. ST. DEV.)

GRAND MEAN = 0.26587500572E+01

GRAND STANDARD DEVIATION = 0.17410624027E+01

99% CONFIDENCE LIMITS (+-) = 0.90710897446E+01

95% CONFIDENCE LIMITS (+-) = 0.18106349707E+01
99.5% POINT OF T DISTRIBUTION = 0.63656803131E+02

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35i1.htm (1 of 3) [5/7/2002 3:15:16 PM]

1.3.5.18.1. Defining Models and Prediction Equations

97.5% POINT OF T DISTRIBUTION = 0.12706216812E+02

IDENTIFIER EFFECT RESSD: T VALUE

RESSD:
MEAN + MEAN +
TERM CUM TERMS
----------------------------------------------------------
MEAN 2.65875 1.74106 1.74106
1 3.10250 21.8* 0.57272 0.57272
2 -0.86750 -6.1 1.81264 0.30429
23 0.29750 2.1 1.87270 0.26737
13 0.24750 1.7 1.87513 0.23341
3 0.21250 1.5 1.87656 0.19121
123 0.14250 1.0 1.87876 0.18031
12 0.12750 0.9 1.87912 0.00000

The last column of the Yates table gives the residual standard deviation for 8 possible
models, each with one more term than the previous model.

Potential For this example, we can summarize the possible prediction equations using the second
Models and last columns of the Yates table:
●

has a residual standard deviation of 1.74106 ohms. Note that this is the default
model. That is, if no factors are important, the model is simply the overall mean.
●

has a residual standard deviation of 0.57272 ohms. (Here, X1 is either a +1 or -1,

and similarly for the other factors and interactions (products).)
●

has a residual standard deviation of 0.30429 ohms.

●

has a residual standard deviation of 0.26737 ohms.

●

has a residual standard deviation of 0.23341 ohms

●

has a residual standard deviation of 0.19121 ohms.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35i1.htm (2 of 3) [5/7/2002 3:15:16 PM]

1.3.5.18.1. Defining Models and Prediction Equations
●

has a residual standard deviation of 0.18031 ohms.

●

has a residual standard deviation of 0.0 ohms. Note that the model with all
possible terms included will have a zero residual standard deviation. This will
always occur with an unreplicated two-level factorial design.

Model The above step lists all the potential models. From this list, we want to select the most
Selection appropriate model. This requires balancing the following two goals.
1. We want the model to include all important factors.
2. We want the model to be parsimonious. That is, the model should be as simple as
possible.
Note that the residual standard deviation alone is insufficient for determining the most
appropriate model as it will always be decreased by adding additional factors. The next
section describes a number of approaches for determining which factors (and
interactions) to include in the model.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35i1.htm (3 of 3) [5/7/2002 3:15:16 PM]

1.3.5.18.2. Important Factors

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.5. Quantitative Techniques
1.3.5.18. Yates Analysis

1.3.5.18.2. Important Factors

Identify The Yates analysis generates a large number of potential models. From this list, we want to select
Important the most appropriate model. This requires balancing the following two goals.
Factors 1. We want the model to include all important factors.
2. We want the model to be parsimonious. That is, the model should be as simple as possible.
In short, we want our model to include all the important factors and interactions and to omit the
unimportant factors and interactions.
Seven criteria are utilized to define important factors. These seven criteria are not all equally
important, nor will they yield identical subsets, in which case a consensus subset or a weighted
consensus subset must be extracted. In practice, some of these criteria may not apply in all
situations.
These criteria will be examined in the context of the Eddy current data set. The Yates Analysis
page gave the sample Yates output for these data and the Defining Models and Predictions page
listed the potential models from the Yates analysis.
In practice, not all of these criteria will be used with every analysis (and some analysts may have
additional criteria). These critierion are given as useful guidelines. Mosts analysts will focus on
those criteria that they find most useful.

Criteria for The seven criteria that we can use in determining whether to keep a factor in the model can be
Including summarized as follows.
Terms in the 1. Effects: Engineering Significance
Model
2. Effects: Order of Magnitude
3. Effects: Statistical Significance
4. Effects: Probability Plots
5. Averages: Youden Plot
6. Residual Standard Deviation: Engineering Significance
7. Residual Standard Deviation: Statistical Significance
The first four criteria focus on effect estimates with three numeric criteria and one graphical
criteria. The fifth criteria focuses on averages. The last two criteria focus on the residual standard
deviation of the model. We discuss each of these seven criteria in detail in the following sections.
The last section summarizes the conclusions based on all of the criteria.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35i2.htm (1 of 7) [5/7/2002 3:15:17 PM]

1.3.5.18.2. Important Factors

Effects: The minimum engineering significant difference is defined as

Engineering
Significance
where is the absolute value of the parameter estimate (i.e., the effect) and is the minimum
engineering significant difference.
That is, declare a factor as "important" if the effect is greater than some a priori declared
engineering difference. This implies that the engineering staff have in fact stated what a minimum
effect will be. Oftentimes this is not the case. In the absence of an a priori difference, a good
rough rule for the minimum engineering significant is to keep only those factors whose effect
is greater than, say, 10% of the current production average. In this case, let's say that the average
detector has a sensitivity of 2.5 ohms. This would suggest that we would declare all factors whose
effect is greater than 10% of 2.5 ohms = 0.25 ohm to be significant (from an engineering point of
view).
Based on this minimum engineering significant difference criterion, we conclude that we should
keep two terms: X1 and X2.

Effects: The order of magnitude criterion is defined as

Order of
Magnitude
That is, exclude any factor that is less than 10% of the maximum effect size. We may or may not
keep the other factors. This criterion is neither engineering nor statistical, but it does offer some
additional numerical insight. For the current example, the largest effect is from X1 (3.10250
ohms), and so 10% of that is 0.31 ohms, which suggests keeping all factors whose effects exceed
0.31 ohms.
Based on the order-of-magnitude criterion, we thus conclude that we should keep two terms: X1
and X2. A third term, X2*X3 (.29750), is just slightly under the cutoff level, so we may consider
keeping it based on the other criterion.

Effects: Statistical significance is defined as

Statistical
Significance

That is, declare a factor as important if its effect is more than 2 standard deviations away from 0
(0, by definition, meaning "no effect").
The "2" comes from normal theory (more specifically, a value of 1.96 yields a 95% confidence
interval). More precise values would come from t-distribution theory.
The difficulty with this is that in order to invoke this criterion we need the standard deviation, ,
of an observation. This is problematic because
1. the engineer may not know ;
2. the experiment might not have replication, and so a model-free estimate of is not
obtainable;
3. obtaining an estimate of by assuming the sometimes- employed assumption of ignoring
3-term interactions and higher may be incorrect from an engineering point of view.
For the Eddy current example:
1. the engineer did not know ;
2. the design (a 23 full factorial) did not have replication;

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35i2.htm (2 of 7) [5/7/2002 3:15:17 PM]

1.3.5.18.2. Important Factors

3. ignoring 3-term interactions and higher interactions leads to an estimate of based on

omitting only a single term: the X1*X2*X3 interaction.
For the current example, if one assumes that the 3-term interaction is nil and hence represents a
single drawing from a population centered at zero, then an estimate of the standard deviation of
an effect is simply the estimate of the 3-factor interaction (0.1425). In the Dataplot output for our
example, this is the effect estimate for the X1*X2*X3 interaction term (the EFFECT column for
the row labeled "123"). Two standard deviations is thus 0.2850. For this example, the rule is thus
to keep all > 0.2850.

This results in keeping three terms: X1 (3.10250), X2 (-.86750), and X1*X2 (.29750).

Effects: Probability plots can be used in the following manner.

Probability 1. Normal Probability Plot: Keep a factor as "important" if it is well off the line through zero
Plots on a normal probability plot of the effect estimates.
2. Half-Normal Probability Plot: Keep a factor as "important" if it is well off the line near
zero on a half-normal probability plot of the absolute value of effect estimates.
Both of these methods are based on the fact that the least squares estimates of effects for these
2-level orthogonal designs are simply the difference of averages and so the central limit theorem,
loosely applied, suggests that (if no factor were important) the effect estimates should have
approximately a normal distribution with mean zero and the absolute value of the estimates
should have a half-normal distribution.
Since the half-normal probability plot is only concerned with effect magnitudes as opposed to
signed effects (which are subject to the vagaries of how the initial factor codings +1 and -1 were
assigned), the half-normal probability plot is preferred by some over the normal probability plot.

Normal The following half-normal plot shows the normal probability plot of the effect estimates and the
Probablity half-normal probability plot of the absolute value of the estimates for the Eddy current data.
Plot of
Effects and
Half-Normal
Probability
Plot of
Effects

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35i2.htm (3 of 7) [5/7/2002 3:15:17 PM]

1.3.5.18.2. Important Factors

For the example at hand, both probability plots clearly show two factors displaced off the line,
and from the third plot (with factor tags included), we see that those two factors are factor 1 and
factor 2. All of the remaining five effects are behaving like random drawings from a normal
distribution centered at zero, and so are deemed to be statistically non-significant. In conclusion,
this rule keeps two factors: X1 (3.10250) and X2 (-.86750).

Effects: A Youden plot can be used in the following way. Keep a factor as "important" if it is displaced
Youden Plot away from the central-tendancy "bunch" in a Youden plot of high and low averages. By
definition, a factor is important when its average response for the low (-1) setting is significantly
different from its average response for the high (+1) setting. Conversely, if the low and high
averages are about the same, then what difference does it make which setting to use and so why
would such a factor be considered important? This fact in combination with the intrinsic benefits
of the Youden plot for comparing pairs of items leads to the technique of generating a Youden
plot of the low and high averages.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35i2.htm (4 of 7) [5/7/2002 3:15:17 PM]

1.3.5.18.2. Important Factors

Youden Plot The following is the Youden plot of the effect estimatess for the Eddy current data.
of Effect
Estimatess

For the example at hand, the Youden plot clearly shows a cluster of points near the grand average
(2.65875) with two displaced points above (factor 1) and below (factor 2). Based on the Youden
plot, we conclude to keep two factors: X1 (3.10250) and X2 (-.86750).

Residual This criterion is defined as

Standard
Deviation: Residual Standard Deviation > Cutoff
Engineering That is, declare a factor as "important" if the cumulative model that includes the factor (and all
Significance larger factors) has a residual standard deviation smaller than an a priori engineering-specified
minimum residual standard deviation.
This criterion is different from the others in that it is model focused. In practice, this criterion
states that starting with the largest effect, we cumulatively keep adding terms to the model and
monitor how the residual standard deviation for each progressively more complicated model
becomes smaller. At some point, the cumulative model will become complicated enough and
comprehensive enough that the resulting residual standard deviation will drop below the
pre-specified engineering cutoff for the residual standard deviation. At that point, we stop adding
terms and declare all of the model-included terms to be "important" and everything not in the
model to be "unimportant".
This approach implies that the engineer has considered what a minimum residual standard
deviation should be. In effect, this relates to what the engineer can tolerate for the magnitude of
the typical residual (= difference between the raw data and the predicted value from the model).

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35i2.htm (5 of 7) [5/7/2002 3:15:17 PM]

1.3.5.18.2. Important Factors

In other words, how good does the engineer want the prediction equation to be. Unfortunately,
this engineering specification has not always been formulated and so this criterion can become
moot.
In the absence of a prior specified cutoff, a good rough rule for the minimum engineering residual
standard deviation is to keep adding terms until the residual standard deviation just dips below,
say, 5% of the current production average. For the Eddy current data, let's say that the average
detector has a sensitivity of 2.5 ohms. Then this would suggest that we would keep adding terms
to the model until the residual standard deviation falls below 5% of 2.5 ohms = 0.125 ohms.
Based on the minimum residual standard deviation criteria, and by scanning the far right column
of the Yates table, we would conclude to keep the following terms:
1. X1 (with a cumulative residual standard deviation = 0.57272)
2. X2 (with a cumulative residual standard deviation = 0.30429)
3. X2*X3 (with a cumulative residual standard deviation = 0.26737)
4. X1*X3 (with a cumulative residual standard deviation = 0.23341)
5. X3 (with a cumulative residual standard deviation = 0.19121)
6. X1*X2*X3 (with a cumulative residual standard deviation = 0.18031)
7. X1*X2 (with a cumulative residual standard deviation = 0.00000)
Note that we must include all terms in order to drive the residual standard deviation below 0.125.
Again, the 5% rule is a rough-and-ready rule that has no basis in engineering or statistics, but is
simply a "numerics". Ideally, the engineer has a better cutoff for the residual standard deviation
that is based on how well he/she wants the equation to peform in practice. If such a number were
available, then for this criterion and data set we would select something less than the entire
collection of terms.

Residual This criterion is defined as

Standard Residual Standard Deviation >
Deviation:
where is the standard deviation of an observation under replicated conditions.
Statistical
Significance That is, declare a term as "important" until the cumulative model that includes the term has a
residual standard deviation smaller than . In essence, we are allowing that we cannot demand a
model fit any better than what we would obtain if we had replicated data; that is, we cannot
demand that the residual standard deviation from any fitted model be any smaller than the
(theoretical or actual) replication standard deviation. We can drive the fitted standard deviation
down (by adding terms) until it achieves a value close to , but to attempt to drive it down further
means that we are, in effect, trying to fit noise.
In practice, this criterion may be difficult to apply because
1. the engineer may not know ;
2. the experiment might not have replication, and so a model-free estimate of is not
obtainable.
For the current case study:
1. the engineer did not know ;
2. the design (a 23 full factorial) did not have replication. The most common way of having
replication in such designs is to have replicated center points at the center of the cube
((X1,X2,X3) = (0,0,0)).
Thus for this current case, this criteria could not be used to yield a subset of "important" factors.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35i2.htm (6 of 7) [5/7/2002 3:15:17 PM]

1.3.5.18.2. Important Factors

Conclusions In summary, the seven criteria for specifying "important" factors yielded the following for the
Eddy current data:
1. Effects, Engineering Significance: X1, X2
2. Effects, Numerically Significant: X1, X2
3. Effects, Statistically Significant: X1, X2, X2*X3
4. Effects, Probability Plots: X1, X2
5. Averages, Youden Plot: X1, X2
6. Residual SD, Engineering Significance: all 7 terms
7. Residual SD, Statistical Significance: not applicable
Such conflicting results are common. Arguably, the three most important criteria (listed in order
of most important) are:
4. Effects, Probability Plots: X1, X2
1. Effects, Engineering Significance: X1, X2
3. Residual SD, Engineering Significance: all 7 terms
Scanning all of the above, we thus declare the following consensus for the Eddy current data:
1. Important Factors: X1 and X2
2. Parsimonious Prediction Equation:

(with a residual standard deviation of .30429 ohms)

Note that this is the initial model selection. We still need to perform model validation with a
residual analysis.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda35i2.htm (7 of 7) [5/7/2002 3:15:17 PM]

1.3.6. Probability Distributions

1. Exploratory Data Analysis

1.3. EDA Techniques

1.3.6. Probability Distributions

Probability Probability distributions are a fundamental concept in statistics. They
Distributions are used both on a theoretical level and a practical level.
Some practical uses of probability distributions are:
● To calculate confidence intervals for parameters and to calculate
critical regions for hypothesis tests.
● For univariate data, it is often useful to determine a reasonable
distributional model for the data.
● Statistical intervals and hypothesis tests are often based on
specific distributional assumptions. Before computing an
interval or test based on a distributional assumption, we need to
verify that the assumption is justified for the given data set. In
this case, the distribution does not need to be the best-fitting
distribution for the data, but an adequate enough model so that
the statistical technique yields valid conclusions.
● Simulation studies with random numbers generated from using a
specific probability distribution are often needed.

Table of 1. What is a probability distribution?

Contents 2. Related probability functions
3. Families of distributions
4. Location and scale parameters
5. Estimating the parameters of a distribution
6. A gallery of common distributions
7. Tables for probability distributions

http://www.itl.nist.gov/div898/handbook/eda/section3/eda36.htm [5/7/2002 3:15:17 PM]

1.3.6.1. What is a Probability Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions

1.3.6.1. What is a Probability Distribution

Discrete The mathematical definition of a discrete probability function, p(x), is a
Distributions function that satisfies the following properties.
1. The probability that x can take a specific value is p(x). That is

2. p(x) is non-negative for all real x.

3. The sum of p(x) over all possible values of x is 1, that is

where j represents all possible values that x can have and pj is the
probability at xj.

One consequence of properties 2 and 3 is that 0 <= p(x) <= 1.

What does this actually mean? A discrete probability function is a
function that can take a discrete number of values (not necessarily
finite). This is most often the non-negative integers or some subset of
the non-negative integers. There is no mathematical restriction that
discrete probability functions only be defined at integers, but in practice
this is usually what makes sense. For example, if you toss a coin 6
times, you can get 2 heads or 3 heads but not 2 1/2 heads. Each of the
discrete values has a certain probability of occurrence that is between
zero and one. That is, a discrete function that allows negative values or
values greater than one is not a probability function. The condition that
the probabilities sum to one means that at least one of the values has to
occur.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda361.htm (1 of 2) [5/7/2002 3:15:18 PM]

1.3.6.1. What is a Probability Distribution

Continuous The mathematical definition of a continuous probability function, f(x),

Distributions is a function that satisfies the following properties.
1. The probability that x is between two points a and b is

2. It is non-negative for all real x.

3. The integral of the probability function is one, that is

What does this actually mean? Since continuous probability functions

are defined for an infinite number of points over a continuous interval,
the probability at a single point is always zero. Probabilities are
measured over intervals, not single points. That is, the area under the
curve between two distinct points defines the probability for that
interval. This means that the height of the probability function can in
fact be greater than one. The property that the integral must equal one is
equivalent to the property for discrete distributions that the sum of all
the probabilities must equal one.

Probability Discrete probability functions are referred to as probability mass

Mass functions and continuous probability functions are referred to as
Functions probability density functions. The term probability functions covers
Versus both discrete and continuous distributions. When we are referring to
Probability probability functions in generic terms, we may use the term probability
Density density functions to mean both discrete and continuous probability
Functions functions.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda361.htm (2 of 2) [5/7/2002 3:15:18 PM]

1.3.6.2. Related Distributions

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions

1.3.6.2. Related Distributions

Probability distributions are typically defined in terms of the probability
density function. However, there are a number of probability functions
used in applications.

Probability For a continuous function, the probability density function (pdf) is the
Density probability that the variate has the value x. Since for continuous
Function distributions the probability at a single point is zero, this is often
expressed in terms of an integral between two points.

For a discrete distribution, the pdf is the probability that the variate takes
the value x.

The following is the plot of the normal probability density function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda362.htm (1 of 8) [5/7/2002 3:15:19 PM]

1.3.6.2. Related Distributions

Cumulative The cumulative distribution function (cdf) is the probability that the
Distribution variable takes a value less than or equal to x. That is
Function

For a continuous distribution, this can be expressed mathematically as

For a discrete distribution, the cdf can be expressed as

The following is the plot of the normal cumulative distribution function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda362.htm (2 of 8) [5/7/2002 3:15:19 PM]

1.3.6.2. Related Distributions

The horizontal axis is the allowable domain for the given probability
function. Since the vertical axis is a probability, it must fall between
zero and one. It increases from zero to one as we go from left to right on
the horizontal axis.

Percent The percent point function (ppf) is the inverse of the cumulative
Point distribution function. For this reason, the percent point function is also
Function commonly referred to as the inverse distribution function. That is, for a
distribution function we calculate the probability that the variable is less
than or equal to x for a given x. For the percent point function, we start
with the probability and compute the corresponding x for the cumulative
distribution. Mathematically, this can be expressed as

or alternatively

The following is the plot of the normal percent point function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda362.htm (3 of 8) [5/7/2002 3:15:19 PM]

1.3.6.2. Related Distributions

Since the horizontal axis is a probability, it goes from zero to one. The
vertical axis goes from the smallest to the largest value of the
cumulative distribution function.

Hazard The hazard function is the ratio of the probability density function to the
Function survival function, S(x).

The following is the plot of the normal distribution hazard function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda362.htm (4 of 8) [5/7/2002 3:15:19 PM]

1.3.6.2. Related Distributions

Hazard plots are most commonly used in reliability applications. Note

that Johnson, Kotz, and Balakrishnan refer to this as the conditional
failure density function rather than the hazard function.

Cumulative The cumulative hazard function is the integral of the hazard function. It
Hazard can be interpreted as the probability of failure at time x given survival
Function until time x.

This can alternatively be expressed as

The following is the plot of the normal cumulative hazard function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda362.htm (5 of 8) [5/7/2002 3:15:19 PM]

1.3.6.2. Related Distributions

Cumulative hazard plots are most commonly used in reliability

applications. Note that Johnson, Kotz, and Balakrishnan refer to this as
the hazard function rather than the cumulative hazard function.

Survival Survival functions are most often used in reliability and related fields.
Function The survival function is the probability that the variate takes a value
greater than x.

The following is the plot of the normal distribution survival function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda362.htm (6 of 8) [5/7/2002 3:15:19 PM]

1.3.6.2. Related Distributions

For a survival function, the y value on the graph starts at 1 and

monotonically decreases to zero. The survival function should be
compared to the cumulative distribution function.

Inverse Just as the percent point function is the inverse of the cumulative
Survival distribution function, the survival function also has an inverse function.
Function The inverse survival function can be defined in terms of the percent
point function.

The following is the plot of the normal distribution inverse survival

function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda362.htm (7 of 8) [5/7/2002 3:15:19 PM]

1.3.6.2. Related Distributions

As with the percent point function, the horizontal axis is a probability.

Therefore the horizontal axis goes from 0 to 1 regardless of the
particular distribution. The appearance is similar to the percent point
function. However, instead of going from the smallest to the largest
value on the vertical axis, it goes from the largest to the smallest value.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda362.htm (8 of 8) [5/7/2002 3:15:19 PM]

1.3.6.3. Families of Distributions

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions

1.3.6.3. Families of Distributions

Shape Many probability distributions are not a single distribution, but are in
Parameters fact a family of distributions. This is due to the distribution having one
or more shape parameters.
Shape parameters allow a distribution to take on a variety of shapes,
depending on the value of the shape parameter. These distributions are
particularly useful in modeling applications since they are flexible
enough to model a variety of data sets.

Example: The Weibull distribution is an example of a distribution that has a shape

Weibull
parameter. The following graph plots the Weibull pdf with the following
Distribution
values for the shape parameter: 0.5, 1.0, 2.0, and 5.0.

The shapes above include an exponential distribution, a right-skewed

distribution, and a relatively symmetric distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda363.htm (1 of 2) [5/7/2002 3:15:19 PM]

1.3.6.3. Families of Distributions

The Weibull distribution has a relatively simple distributional form.

However, the shape parameter allows the Weibull to assume a wide
variety of shapes. This combination of simplicity and flexibility in the
shape of the Weibull distribution has made it an effective distributional
model in reliability applications. This ability to model a wide variety of
distributional shapes using a relatively simple distributional form is
possible with many other distributional families as well.

PPCC Plots The PPCC plot is an effective graphical tool for selecting the member of
a distributional family with a single shape parameter that best fits a
given set of data.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda363.htm (2 of 2) [5/7/2002 3:15:19 PM]

1.3.6.4. Location and Scale Parameters

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions

1.3.6.4. Location and Scale Parameters

Normal A probability distribution is characterized by location and scale
PDF parameters. Location and scale parameters are typically used in
modeling applications.
For example, the following graph is the probability density function for
the standard normal distribution, which has the location parameter equal
to zero and scale parameter equal to one.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda364.htm (1 of 5) [5/7/2002 3:15:20 PM]

1.3.6.4. Location and Scale Parameters

Location The next plot shows the probability density function for a normal
Parameter distribution with a location parameter of 10 and a scale parameter of 1.

The effect of the location parameter is to translate the graph, relative to

the standard normal distribution, 10 units to the right on the horizontal
axis. A location parameter of -10 would have shifted the graph 10 units
to the left on the horizontal axis.
That is, a location parameter simply shifts the graph left or right on the
horizontal axis.

Scale
Parameter

The next plot has a scale parameter of 3 (and a location parameter of

zero). The effect of the scale parameter is to stretch out the graph. The
maximum y value is approximately 0.13 as opposed 0.4 in the previous
graphs. The y value, i.e., the vertical axis value, approaches zero at
about (+/-) 9 as opposed to (+/-) 3 with the first graph.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda364.htm (2 of 5) [5/7/2002 3:15:20 PM]

1.3.6.4. Location and Scale Parameters

In contrast, the next graph has a scale parameter of 1/3 (=0.333). The
effect of this scale parameter is to squeeze the pdf. That is, the
maximum y value is approximately 1.2 as opposed to 0.4 and the y
value is near zero at (+/-) 1 as opposed to (+/-) 3.

The effect of a scale parameter greater than one is to stretch the pdf. The
greater the magnitude, the greater the stretching. The effect of a scale
parameter less than one is to compress the pdf. The compressing
approaches a spike as the scale parameter goes to zero. A scale

http://www.itl.nist.gov/div898/handbook/eda/section3/eda364.htm (3 of 5) [5/7/2002 3:15:20 PM]

1.3.6.4. Location and Scale Parameters

parameter of 1 leaves the pdf unchanged (if the scale parameter is 1 to

begin with) and non-positive scale parameters are not allowed.

Location The following graph shows the effect of both a location and a scale
and Scale parameter. The plot has been shifted right 10 units and stretched by a
Together factor of 3.

Standard The standard form of any distribution is the form that has location
Form parameter zero and scale parameter one.
It is common in statistical software packages to only compute the
standard form of the distribution. There are formulas for converting
from the standard form to the form with other location and scale
parameters. These formulas are independent of the particular probability
distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda364.htm (4 of 5) [5/7/2002 3:15:20 PM]

1.3.6.4. Location and Scale Parameters

Formulas The following are the formulas for computing various probability
for Location functions based on the standard form of the distribution. The parameter
and Scale a refers to the location parameter and the parameter b refers to the scale
Based on parameter. Shape parameters are not included.
the Standard
Cumulative Distribution Function F(x;a,b) = F((x-a)/b;0,1)
Form
Probability Density Function f(x;a,b) = (1/b)f((x-a)/b;0,1)
Percent Point Function G( ;a,b) = a + bG( ;0,1)
Hazard Function h(x;a,b) = (1/b)h((x-a)/b;0,1)
Cumulative Hazard Function H(x;a,b) = H((x-a)/b;0,1)
Survival Function S(x;a,b) = S((x-a)/b;0,1)
Inverse Survival Function Z( ;a,b) = a + bZ( ;0,1)
Random Numbers Y(a,b) = a + bY(0,1)

Relationship For the normal distribution, the location and scale parameters
to Mean and correspond to the mean and standard deviation, respectively. However,
Standard this is not necessarily true for other distributions. In fact, it is not true
Deviation for most distributions.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda364.htm (5 of 5) [5/7/2002 3:15:20 PM]

1.3.6.5. Estimating the Parameters of a Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions

1.3.6.5. Estimating the Parameters of a

Distribution
Model a One common application of probability distributions is modeling
univariate univariate data with a specific probability distribution. This involves the
data set with following two steps:
a 1. Determination of the "best-fitting" distribution.
probability
2. Estimation of the parameters (shape, location, and scale
distribution
parameters) for that distribution.

Various There are various methods, both numerical and graphical, for estimating
Methods the parameters of a probability distribution.
1. Method of moments
2. Maximum likelihood
3. Least squares
4. PPCC and probability plots

http://www.itl.nist.gov/div898/handbook/eda/section3/eda365.htm [5/7/2002 3:15:20 PM]

1.3.6.5.1. Method of Moments

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.5. Estimating the Parameters of a Distribution

1.3.6.5.1. Method of Moments

Method of The method of moments equates sample moments to parameter
Moments estimates. When moment methods are available, they have the
advantage of simplicity. The disadvantage is that they are often not
available and they do not have the desirable optimality properties of
maximum likelihood and least squares estimators.
The primary use of moment estimates is as starting values for the more
precise maximum likelihood and least squares estimates.

Software Most general purpose statistical software does not include explicit
method of moments parameter estimation commands. However, when
utilized, the method of moment formulas tend to be straightforward and
can be easily implemented in most statistical software programs.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3651.htm [5/7/2002 3:15:20 PM]

1.3.6.5.2. Maximum Likelihood

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.5. Estimating the Parameters of a Distribution

1.3.6.5.2. Maximum Likelihood

Maximum Maximum likelihood estimation begins with the mathematical
Likelihood expression known as a likelihood function of the sample data. Loosely
speaking, the likelihood of a set of data is the probability of obtaining
that particular set of data given the chosen probability model. This
expression contains the unknown parameters. Those values of the
parameter that maximize the sample likelihood are known as the
maximum likelihood estimates.
The reliability chapter contains some examples of the likelihood
functions for a few of the commonly used distributions in reliability
analysis.

Advantages The advantages of this method are:

● Maximum likelihood provides a consistent approach to
parameter estimation problems. This means that maximum
likelihood estimates can be developed for a large variety of
estimation situations. For example, they can be applied in
reliability analysis to censored data under various censoring
models.
● Maximum likelihood methods have desirable mathematical and
optimality properties. Specifically,
1. They become minimum variance unbiased estimators as
the sample size increases. By unbiased, we mean that if
we take (a very large number of) random samples with
replacement from a population, the average value of the
parameter estimates will be theoretically exactly equal to
the population value. By minimum variance, we mean
that the estimator has the smallest variance, and thus the
narrowest confidence interval, of all estimators of that
type.
2. They have approximate normal distributions and
approximate sample variances that can be used to

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3652.htm (1 of 3) [5/7/2002 3:15:21 PM]

1.3.6.5.2. Maximum Likelihood

generate confidence bounds and hypothesis tests for the

parameters.
● Several popular statistical software packages provide excellent
algorithms for maximum likelihood estimates for many of the
commonly used distributions. This helps mitigate the
computational complexity of maximum likelihood estimation.

Disadvantages The disadvantages of this method are:

● The likelihood equations need to be specifically worked out for
a given distribution and estimation problem. The mathematics is
often non-trivial, particularly if confidence intervals for the
parameters are desired.
● The numerical estimation is usually non-trivial. Except for a
few cases where the maximum likelihood formulas are in fact
simple, it is generally best to rely on high quality statistical
software to obtain maximum likelihood estimates. Fortunately,
high quality maximum likelihood software is becoming
increasingly common.
● Maximum likelihood estimates can be heavily biased for small
samples. The optimality properties may not apply for small
samples.
● Maximum likelihood can be sensitive to the choice of starting
values.

Software Most general purpose statistical software programs support maximum

likelihood estimation (MLE) in some form. MLE estimation can be
supported in two ways.
1. A software program may provide a generic function
minimization (or equivalently, maximization) capability. This is
also referred to as function optimization. Maximum likelihood
estimation is essentially a function optimization problem.
This type of capability is particularly common in mathematical
software programs.
2. A software program may provide MLE computations for a
specific problem. For example, it may generate ML estimates
for the parameters of a Weibull distribution.
Statistical software programs will often provide ML estimates
for many specific problems even when they do not support
general function optimization.
The advantage of function minimization software is that it can be
applied to many different MLE problems. The drawback is that you
have to specify the maximum likelihood equations to the software. As

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3652.htm (2 of 3) [5/7/2002 3:15:21 PM]

1.3.6.5.2. Maximum Likelihood

the functions can be non-trivial, there is potential for error in entering

the equations.
The advantage of the specific MLE procedures is that greater
efficiency and better numerical stability can often be obtained by
taking advantage of the properties of the specific estimation problem.
The specific methods often return explicit confidence intervals. In
addition, you do not have to know or specify the likelihood equations
to the software. The disadvantage is that each MLE problem must be
specifically coded.
Dataplot supports MLE for a limited number of distributions.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3652.htm (3 of 3) [5/7/2002 3:15:21 PM]

1.3.6.5.3. Least Squares

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.5. Estimating the Parameters of a Distribution

1.3.6.5.3. Least Squares

Least Squares Non-linear least squares provides an alternative to maximum
likelihood.

Advantages The advantages of this method are:

● Non-linear least squares software may be available in many
statistical software packages that do not support maximum
likelihood estimates.
● It can be applied more generally than maximum likelihood.
That is, if your software provides non-linear fitting and it has
the ability to specify the probability function you are interested
in, then you can generate least squares estimates for that
distribution. This will allow you to obtain reasonable estimates
for distributions even if the software does not provide
maximum likelihood estimates.

Disadvantages The disadvantages of this method are:

● It is not readily applicable to censored data.

● It is generally considered to have less desirable optimality

properties than maximum likelihood.
● It can be quite sensitive to the choice of starting values.

Software Non-linear least squares fitting is available in many general purpose

statistical software programs. The macro developed for Dataplot can
be adapted to many software programs that provide least squares
estimation.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3653.htm [5/7/2002 3:15:21 PM]

1.3.6.5.4. PPCC and Probability Plots

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.5. Estimating the Parameters of a Distribution

1.3.6.5.4. PPCC and Probability Plots

PPCC and The PPCC plot can be used to estimate the shape parameter of a
Probability distribution with a single shape parameter. After finding the best value
Plots of the shape parameter, the probability plot can be used to estimate the
location and scale parameters of a probability distribution.

Advantages The advantages of this method are:

● It is based on two well-understood concepts.

1. The linearity (i.e., straightness) of the probability plot is a

good measure of the adequacy of the distributional fit.
2. The correlation coefficient between the points on the
probability plot is a good measure of the linearity of the
probability plot.
● It is an easy technique to implement for a wide variety of
distributions with a single shape parameter. The basic
requirement is to be able to compute the percent point function,
which is needed in the computation of both the probability plot
and the PPCC plot.
● The PPCC plot provides insight into the sensitivity of the shape
parameter. That is, if the PPCC plot is relatively flat in the
neighborhood of the optimal value of the shape parameter, this
is a strong indication that the fitted model will not be sensitive
to small deviations, or even large deviations in some cases, in
the value of the shape parameter.
● The maximum correlation value provides a method for
comparing across distributions as well as identifying the best
value of the shape parameter for a given distribution. For
example, we could use the PPCC and probability fits for the
Weibull, lognormal, and possibly several other distributions.
Comparing the maximum correlation coefficient achieved for
each distribution can help in selecting which is the best
distribution to use.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3654.htm (1 of 2) [5/7/2002 3:15:21 PM]

1.3.6.5.4. PPCC and Probability Plots

Disadvantages The disadvantages of this method are:

● It is limited to distributions with a single shape parameter.

● PPCC plots are not widely available in statistical software

packages other than Dataplot (Dataplot provides PPCC plots for
40+ distributions). Probability plots are generally available.
However, many statistical software packages only provide them
for a limited number of distributions.
● Significance levels for the correlation coefficient (i.e., if the
maximum correlation value is above a given value, then the
distribution provides an adequate fit for the data with a given
confidence level) have only been worked out for a limited
number of distributions.

Case Study The airplane glass failure time case study demonstrates the use of the
PPCC and probability plots in finding the best distributional model
and the parameter estimation of the distributional model.

Other For reliability applications, the hazard plot and the Weibull plot are
Graphical alternative graphical methods that are commonly used to estimate
Methods parameters.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3654.htm (2 of 2) [5/7/2002 3:15:21 PM]

1.3.6.6. Gallery of Distributions

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions

1.3.6.6. Gallery of Distributions

Gallery of Detailed information on a few of the most common distributions is
Common available below. There are a large number of distributions used in
Distributions statistical applications. It is beyond the scope of this Handbook to
discuss more than a few of these. Two excellent sources for additional
detailed information on a large array of distributions are Johnson,
Kotz, and Balakrishnan and Evans, Hastings, and Peacock. Equations
for the probability functions are given for the standard form of the
distribution. Formulas exist for defining the functions with location
and scale parameters in terms of the standard form of the distribution.

The sections on parameter estimation are restricted to the method of

moments and maximum likelihood. This is because the least squares
and PPCC and probability plot estimation procedures are generic. The
maximum likelihood equations are not listed if they involve solving
simultaneous equations. This is because these methods require
sophisticated computer software to solve. Except where the maximum
likelihood estimates are trivial, you should depend on a statistical
software program to compute them. References are given for those
who are interested.
Be aware that different sources may give formulas that are different
from those shown here. In some cases, these are simply
mathematically equivalent formulations. In other cases, a different
parameterization may be used.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366.htm (1 of 3) [5/7/2002 3:15:22 PM]

1.3.6.6. Gallery of Distributions

Continuous
Distributions

Normal Distribution Uniform Cauchy Distribution

Distribution

t Distribution F Distribution Chi-Square

Distribution

Exponential Weibull Distribution Lognormal

Distribution Distribution

Fatigue Life Gamma Distribution Double Exponential

Distribution Distribution

Power Normal Power Lognormal Tukey-Lambda

Distribution Distribution Distribution

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366.htm (2 of 3) [5/7/2002 3:15:22 PM]

1.3.6.6. Gallery of Distributions

Extreme Value Type

I Distribution

Discrete
Distributions

Binomial Poisson Distribution

Distribution

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366.htm (3 of 3) [5/7/2002 3:15:22 PM]

1.3.6.6.1. Normal Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.6. Gallery of Distributions

1.3.6.6.1. Normal Distribution

Probability The general formula for the probability density function of the normal
Density distribution is
Function

where is the location parameter and is the scale parameter. The case
where = 0 and = 1 is called the standard normal distribution. The
equation for the standard normal distribution is

Since the general form of probability functions can be expressed in

terms of the standard distribution, all subsequent formulas in this section
are given for the standard form of the function.
The following is the plot of the standard normal probability density
function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3661.htm (1 of 7) [5/7/2002 3:15:23 PM]

1.3.6.6.1. Normal Distribution

Cumulative The formula for the cumulative distribution function of the normal
Distribution distribution does not exist in a simple closed formula. It is computed
Function numerically.
The following is the plot of the normal cumulative distribution function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3661.htm (2 of 7) [5/7/2002 3:15:23 PM]

1.3.6.6.1. Normal Distribution

Percent The formula for the percent point function of the normal distribution
Point does not exist in a simple closed formula. It is computed numerically.
Function
The following is the plot of the normal percent point function.

Hazard The formula for the hazard function of the normal distribution is
Function

where is the cumulative distribution function of the standard normal

distribution and is the probability density function of the standard
normal distribution.

The following is the plot of the normal hazard function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3661.htm (3 of 7) [5/7/2002 3:15:23 PM]

1.3.6.6.1. Normal Distribution

Cumulative The normal cumulative hazard function can be computed from the
Hazard normal cumulative distribution function.
Function
The following is the plot of the normal cumulative hazard function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3661.htm (4 of 7) [5/7/2002 3:15:23 PM]

1.3.6.6.1. Normal Distribution

Survival The normal survival function can be computed from the normal
Function cumulative distribution function.
The following is the plot of the normal survival function.

Inverse The normal inverse survival function can be computed from the normal
Survival percent point function.
Function
The following is the plot of the normal inverse survival function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3661.htm (5 of 7) [5/7/2002 3:15:23 PM]

1.3.6.6.1. Normal Distribution

Common Mean The location parameter .

Statistics Median The location parameter .
Mode The location parameter .
Range Infinity in both directions.
Standard Deviation The scale parameter .
Coefficient of
Variation
Coefficient of 0
Skewness
Coefficient of 3
Kurtosis

Parameter The location and scale parameters of the normal distribution can be
Estimation estimated with the sample mean and sample standard deviation,
respectively.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3661.htm (6 of 7) [5/7/2002 3:15:23 PM]

1.3.6.6.1. Normal Distribution

Comments For both theoretical and practical reasons, the normal distribution is
probably the most important distribution in statistics. For example,
● Many classical statistical tests are based on the assumption that
the data follow a normal distribution. This assumption should be
tested before applying these tests.
● In modeling applications, such as linear and non-linear regression,
the error term is often assumed to follow a normal distribution
with fixed location and scale.
● The normal distribution is used to find significance levels in many
hypothesis tests and confidence intervals.

Theroretical The normal distribution is widely used. Part of the appeal is that it is
Justification well behaved and mathematically tractable. However, the central limit
- Central theorem provides a theoretical basis for why it has wide applicability.
Limit
Theorem The central limit theorem basically states that as the sample size (N)
becomes large, the following occur:
1. The sampling distribution of the mean becomes approximately
normal regardless of the distribution of the original variable.
2. The sampling distribution of the mean is centered at the
population mean, , of the original variable. In addition, the
standard deviation of the sampling distribution of the mean
approaches .

Software Most general purpose statistical software programs, including Dataplot,

support at least some of the probability functions for the normal
distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3661.htm (7 of 7) [5/7/2002 3:15:23 PM]

1.3.6.6.2. Uniform Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.6. Gallery of Distributions

1.3.6.6.2. Uniform Distribution

Probability The general formula for the probability density function of the uniform
Density distribution is
Function

where A is the location parameter and (B - A) is the scale parameter. The case
where A = 0 and B = 1 is called the standard uniform distribution. The
equation for the standard uniform distribution is

Since the general form of probability functions can be expressed in terms of

the standard distribution, all subsequent formulas in this section are given for
the standard form of the function.
The following is the plot of the uniform probability density function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3662.htm (1 of 7) [5/7/2002 3:15:24 PM]

1.3.6.6.2. Uniform Distribution

Cumulative The formula for the cumulative distribution function of the uniform
Distribution distribution is
Function

The following is the plot of the uniform cumulative distribution function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3662.htm (2 of 7) [5/7/2002 3:15:24 PM]

1.3.6.6.2. Uniform Distribution

Percent The formula for the percent point function of the uniform distribution is
Point
Function
The following is the plot of the uniform percent point function.

Hazard The formula for the hazard function of the uniform distribution is
Function

The following is the plot of the uniform hazard function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3662.htm (3 of 7) [5/7/2002 3:15:24 PM]

1.3.6.6.2. Uniform Distribution

Cumulative The formula for the cumulative hazard function of the uniform distribution is
Hazard
Function
The following is the plot of the uniform cumulative hazard function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3662.htm (4 of 7) [5/7/2002 3:15:24 PM]

1.3.6.6.2. Uniform Distribution

Survival The uniform survival function can be computed from the uniform cumulative
Function distribution function.
The following is the plot of the uniform survival function.

Inverse The uniform inverse survival function can be computed from the uniform
Survival percent point function.
Function
The following is the plot of the uniform inverse survival function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3662.htm (5 of 7) [5/7/2002 3:15:24 PM]

1.3.6.6.2. Uniform Distribution

Common Mean (A + B)/2

Statistics Median (A + B)/2
Range B-A
Standard Deviation

Coefficient of
Variation

Coefficient of 0
Skewness
Coefficient of Kurtosis 9/5

Parameter The method of moments estimators for A and B are

Estimation

The maximum likelihood estimators for A and B are

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3662.htm (6 of 7) [5/7/2002 3:15:24 PM]

1.3.6.6.2. Uniform Distribution

Comments The uniform distribution defines equal probability over a given range for a
continuous distribution. For this reason, it is important as a reference
distribution.
One of the most important applications of the uniform distribution is in the
generation of random numbers. That is, almost all random number generators
generate random numbers on the (0,1) interval. For other distributions, some
transformation is applied to the uniform random numbers.

Software Most general purpose statistical software programs, including Dataplot,

support at least some of the probability functions for the uniform distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3662.htm (7 of 7) [5/7/2002 3:15:24 PM]

1.3.6.6.3. Cauchy Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.6. Gallery of Distributions

1.3.6.6.3. Cauchy Distribution

Probability The general formula for the probability density function of the Cauchy
Density distribution is
Function

where t is the location parameter and s is the scale parameter. The case
where t = 0 and s = 1 is called the standard Cauchy distribution. The
equation for the standard Cauchy distribution reduces to

Since the general form of probability functions can be expressed in

terms of the standard distribution, all subsequent formulas in this section
are given for the standard form of the function.
The following is the plot of the standard Cauchy probability density
function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3663.htm (1 of 7) [5/7/2002 3:15:25 PM]

1.3.6.6.3. Cauchy Distribution

Cumulative The formula for the cumulative distribution function for the Cauchy
Distribution distribution is
Function

The following is the plot of the Cauchy cumulative distribution function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3663.htm (2 of 7) [5/7/2002 3:15:25 PM]

1.3.6.6.3. Cauchy Distribution

Percent The formula for the percent point function of the Cauchy distribution is
Point
Function
The following is the plot of the Cauchy percent point function.

Hazard The Cauchy hazard function can be computed from the Cauchy
Function probability density and cumulative distribution functions.
The following is the plot of the Cauchy hazard function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3663.htm (3 of 7) [5/7/2002 3:15:25 PM]

1.3.6.6.3. Cauchy Distribution

Cumulative The Cauchy cumulative hazard function can be computed from the
Hazard Cauchy cumulative distribution function.
Function
The following is the plot of the Cauchy cumulative hazard function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3663.htm (4 of 7) [5/7/2002 3:15:25 PM]

1.3.6.6.3. Cauchy Distribution

Survival The Cauchy survival function can be computed from the Cauchy
Function cumulative distribution function.
The following is the plot of the Cauchy survival function.

Inverse The Cauchy inverse survival function can be computed from the Cauchy
Survival percent point function.
Function
The following is the plot of the Cauchy inverse survival function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3663.htm (5 of 7) [5/7/2002 3:15:25 PM]

1.3.6.6.3. Cauchy Distribution

Common Mean The mean is undefined.

Statistics Median The location parameter t.
Mode The location parameter t.
Range Infinity in both directions.
Standard Deviation The standard deviation is undefined.
Coefficient of The coefficient of variation is undefined.
Variation
Coefficient of The coefficient of skewness is undefined.
Skewness
Coefficient of The coefficient of kurtosis is undefined.
Kurtosis

Parameter The likelihood functions for the Cauchy maximum likelihood estimates
Estimation are given in chapter 16 of Johnson, Kotz, and Balakrishnan. These
equations typically must be solved numerically on a computer.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3663.htm (6 of 7) [5/7/2002 3:15:25 PM]

1.3.6.6.3. Cauchy Distribution

Comments The Cauchy distribution is important as an example of a pathological

case. Cauchy distributions look similar to a normal distribution.
However, they have much heavier tails. When studying hypothesis tests
that assume normality, seeing how the tests perform on data from a
Cauchy distribution is a good indicator of how sensitive the tests are to
heavy-tail departures from normality. Likewise, it is a good check for
robust techniques that are designed to work well under a wide variety of
distributional assumptions.
The mean and standard deviation of the Cauchy distribution are
undefined. The practical meaning of this is that collecting 1,000 data
points gives no more accurate an estimate of the mean and standard
deviation than does a single point.

Software Many general purpose statistical software programs, including Dataplot,

support at least some of the probability functions for the Cauchy
distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3663.htm (7 of 7) [5/7/2002 3:15:25 PM]

1.3.6.6.4. t Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.6. Gallery of Distributions

1.3.6.6.4. t Distribution
Probability The formula for the probability density function of the t distribution is
Density
Function

where is the beta function and is a positive integer shape parameter.

The formula for the beta function is

In a testing context, the t distribution is treated as a "standardized

distribution" (i.e., no location or scale parameters). However, in a
distributional modeling context (as with other probability distributions),
the t distribution itself can be transformed with a location parameter, ,
and a scale parameter, .

The following is the plot of the t probability density function for 4

different values of the shape parameter.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3664.htm (1 of 4) [5/7/2002 3:15:26 PM]

1.3.6.6.4. t Distribution

These plots all have a similar shape. The difference is in the heaviness
of the tails. In fact, the t distribution with equal to 1 is a Cauchy
distribution. The t distribution approaches a normal distribution as
becomes large. The approximation is quite good for values of > 30.

Cumulative The formula for the cumulative distribution function of the t distribution
Distribution is complicated and is not included here. It is given in the Evans,
Function Hastings, and Peacock book.

The following are the plots of the t cumulative distribution function with
the same values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3664.htm (2 of 4) [5/7/2002 3:15:26 PM]

1.3.6.6.4. t Distribution

Percent The formula for the percent point function of the t distribution does not
Point exist in a simple closed form. It is computed numerically.
Function
The following are the plots of the t percent point function with the same
values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3664.htm (3 of 4) [5/7/2002 3:15:26 PM]

1.3.6.6.4. t Distribution

Other Since the t distribution is typically used to develop hypothesis tests and
Probability confidence intervals and rarely for modeling applications, we omit the
Functions formulas and plots for the hazard, cumulative hazard, survival, and
inverse survival probability functions.

Common Mean 0 (It is undefined for equal to 1.)

Statistics Median 0
Mode 0
Range Infinity in both directions.
Standard Deviation

It is undefined for equal to 1 or 2.

Coefficient of Undefined
Variation
Coefficient of 0. It is undefined for less than or equal to 3.
Skewness However, the t distribution is symmetric in all
cases.
Coefficient of
Kurtosis

It is undefined for less than or equal to 4.

Parameter Since the t distribution is typically used to develop hypothesis tests and
Estimation confidence intervals and rarely for modeling applications, we omit any
discussion of parameter estimation.

Comments The t distribution is used in many cases for the critical regions for
hypothesis tests and in determining confidence intervals. The most
common example is testing if data are consistent with the assumed
process mean.

Software Most general purpose statistical software programs, including Dataplot,

support at least some of the probability functions for the t distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3664.htm (4 of 4) [5/7/2002 3:15:26 PM]

1.3.6.6.5. F Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.6. Gallery of Distributions

1.3.6.6.5. F Distribution
Probability The F distribution is the ratio of two chi-square distributions with
Density degrees of freedom and , respectively. The formula for the
Function probability density function of the F distribution is

where and are the shape parameters and is the gamma function.
The formula for the gamma function is

In a testing context, the F distribution is treated as a "standardized

distribution" (i.e., no location or scale parameters). However, in a
distributional modeling context (as with other probability distributions),
the F distribution itself can be transformed with a location parameter, ,
and a scale parameter, .
The following is the plot of the F probability density function for 4
different values of the shape parameters.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3665.htm (1 of 4) [5/7/2002 3:15:26 PM]

1.3.6.6.5. F Distribution

Cumulative The formula for the Cumulative distribution function of the F

Distribution distribution is
Function

where k = / ( + *x) and Ik is the incomplete beta function. The

formula for the incomplete beta function is

where B is the beta function

The following is the plot of the F cumulative distribution function with

the same values of and as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3665.htm (2 of 4) [5/7/2002 3:15:26 PM]

1.3.6.6.5. F Distribution

Percent The formula for the percent point function of the F distribution does not
Point exist in a simple closed form. It is computed numerically.
Function
The following is the plot of the F percent point function with the same
values of and as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3665.htm (3 of 4) [5/7/2002 3:15:26 PM]

1.3.6.6.5. F Distribution

Other Since the F distribution is typically used to develop hypothesis tests and
Probability confidence intervals and rarely for modeling applications, we omit the
Functions formulas and plots for the hazard, cumulative hazard, survival, and
inverse survival probability functions.

Common The formulas below are for the case where the location parameter is
Statistics zero and the scale parameter is one.
Mean

Mode

Range 0 to positive infinity

Standard Deviation

Coefficient of
Variation

Coefficient of
Skewness

Parameter Since the F distribution is typically used to develop hypothesis tests and
Estimation confidence intervals and rarely for modeling applications, we omit any
discussion of parameter estimation.

Comments The F distribution is used in many cases for the critical regions for
hypothesis tests and in determining confidence intervals. Two common
examples are the analysis of variance and the F test to determine if the
variances of two populations are equal.

Software Most general purpose statistical software programs, including Dataplot,

support at least some of the probability functions for the F distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3665.htm (4 of 4) [5/7/2002 3:15:26 PM]

1.3.6.6.6. Chi-Square Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.6. Gallery of Distributions

1.3.6.6.6. Chi-Square Distribution

Probability The chi-square distribution results when independent variables with
Density standard normal distributions are squared and summed. The formula for
Function the probability density function of the chi-square distribution is

where is the shape parameter and is the gamma function. The

formula for the gamma function is

In a testing context, the chi-square distribution is treated as a

"standardized distribution" (i.e., no location or scale parameters).
However, in a distributional modeling context (as with other probability
distributions), the chi-square distribution itself can be transformed with
a location parameter, , and a scale parameter, .

The following is the plot of the chi-square probability density function

for 4 different values of the shape parameter.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3666.htm (1 of 4) [5/7/2002 3:15:27 PM]

1.3.6.6.6. Chi-Square Distribution

Cumulative The formula for the cumulative distribution function of the chi-square
Distribution distribution is
Function

where is the gamma function defined above and is the incomplete

gamma function. The formula for the incomplete gamma function is

The following is the plot of the chi-square cumulative distribution

function with the same values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3666.htm (2 of 4) [5/7/2002 3:15:27 PM]

1.3.6.6.6. Chi-Square Distribution

Percent The formula for the percent point function of the chi-square distribution
Point does not exist in a simple closed form. It is computed numerically.
Function
The following is the plot of the chi-square percent point function with
the same values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3666.htm (3 of 4) [5/7/2002 3:15:27 PM]

1.3.6.6.6. Chi-Square Distribution

Other Since the chi-square distribution is typically used to develop hypothesis

Probability tests and confidence intervals and rarely for modeling applications, we
Functions omit the formulas and plots for the hazard, cumulative hazard, survival,
and inverse survival probability functions.

Common Mean
Statistics Median approximately - 2/3 for large
Mode
Range 0 to positive infinity
Standard Deviation
Coefficient of
Variation

Coefficient of
Skewness

Coefficient of
Kurtosis

Parameter Since the chi-square distribution is typically used to develop hypothesis

Estimation tests and confidence intervals and rarely for modeling applications, we
omit any discussion of parameter estimation.

Comments The chi-square distribution is used in many cases for the critical regions
for hypothesis tests and in determining confidence intervals. Two
common examples are the chi-square test for independence in an RxC
contingency table and the chi-square test to determine if the standard
deviation of a population is equal to a pre-specified value.

Software Most general purpose statistical software programs, including Dataplot,

support at least some of the probability functions for the chi-square
distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3666.htm (4 of 4) [5/7/2002 3:15:27 PM]

1.3.6.6.7. Exponential Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.6. Gallery of Distributions

1.3.6.6.7. Exponential Distribution

Probability The general formula for the probability density function of the
Density exponential distribution is
Function

where is the location parameter and is the scale parameter (the

scale parameter is often referred to as which equals ). The case
where = 0 and = 1 is called the standard exponential distribution.
The equation for the standard exponential distribution is

The general form of probability functions can be expressed in terms of

the standard distribution. Subsequent formulas in this section are given
for the 1-parameter (i.e., with scale parameter) form of the function.
The following is the plot of the exponential probability density function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3667.htm (1 of 7) [5/7/2002 3:15:28 PM]

1.3.6.6.7. Exponential Distribution

Cumulative The formula for the cumulative distribution function of the exponential
Distribution distribution is
Function

The following is the plot of the exponential cumulative distribution

function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3667.htm (2 of 7) [5/7/2002 3:15:28 PM]

1.3.6.6.7. Exponential Distribution

Percent The formula for the percent point function of the exponential
Point distribution is
Function

The following is the plot of the exponential percent point function.

Hazard The formula for the hazard function of the exponential distribution is
Function

The following is the plot of the exponential hazard function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3667.htm (3 of 7) [5/7/2002 3:15:28 PM]

1.3.6.6.7. Exponential Distribution

Cumulative The formula for the cumulative hazard function of the exponential
Hazard distribution is
Function

The following is the plot of the exponential cumulative hazard function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3667.htm (4 of 7) [5/7/2002 3:15:28 PM]

1.3.6.6.7. Exponential Distribution

Survival The formula for the survival function of the exponential distribution is
Function

The following is the plot of the exponential survival function.

Inverse The formula for the inverse survival function of the exponential
Survival distribution is
Function

The following is the plot of the exponential inverse survival function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3667.htm (5 of 7) [5/7/2002 3:15:28 PM]

1.3.6.6.7. Exponential Distribution

Common Mean
Statistics Median
Mode Zero
Range Zero to plus infinity
Standard Deviation
Coefficient of 1
Variation
Coefficient of 2
Skewness
Coefficient of 9
Kurtosis

Parameter For the full sample case, the maximum likelihood estimator of the scale
Estimation parameter is the sample mean. Maximum likelihood estimation for the
exponential distribution is discussed in the chapter on reliability
(Chapter 8). It is also discussed in chapter 19 of Johnson, Kotz, and
Balakrishnan.

Comments The exponential distribution is primarily used in reliability applications.

The exponential distribution is used to model data with a constant
failure rate (indicated by the hazard plot which is simply equal to a
constant).

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3667.htm (6 of 7) [5/7/2002 3:15:28 PM]

1.3.6.6.7. Exponential Distribution

Software Most general purpose statistical software programs, including Dataplot,

support at least some of the probability functions for the exponential
distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3667.htm (7 of 7) [5/7/2002 3:15:28 PM]

1.3.6.6.8. Weibull Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.6. Gallery of Distributions

1.3.6.6.8. Weibull Distribution

Probability The formula for the probability density function of the general Weibull distribution
Density is
Function

where is the shape parameter, is the location parameter and is the scale
parameter. The case where = 0 and = 1 is called the standard Weibull
distribution. The case where = 0 is called the 2-parameter Weibull distribution.
The equation for the standard Weibull distribution reduces to

Since the general form of probability functions can be expressed in terms of the
standard distribution, all subsequent formulas in this section are given for the
standard form of the function.
The following is the plot of the Weibull probability density function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3668.htm (1 of 7) [5/7/2002 3:15:29 PM]

1.3.6.6.8. Weibull Distribution

Cumulative The formula for the cumulative distribution function of the Weibull distribution is
Distribution
Function

The following is the plot of the Weibull cumulative distribution function with the
same values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3668.htm (2 of 7) [5/7/2002 3:15:29 PM]

1.3.6.6.8. Weibull Distribution

Percent The formula for the percent point function of the Weibull distribution is
Point
Function

The following is the plot of the Weibull percent point function with the same
values of as the pdf plots above.

Hazard The formula for the hazard function of the Weibull distribution is
Function

The following is the plot of the Weibull hazard function with the same values of
as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3668.htm (3 of 7) [5/7/2002 3:15:29 PM]

1.3.6.6.8. Weibull Distribution

Cumulative The formula for the cumulative hazard function of the Weibull distribution is
Hazard
Function
The following is the plot of the Weibull cumulative hazard function with the same
values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3668.htm (4 of 7) [5/7/2002 3:15:29 PM]

1.3.6.6.8. Weibull Distribution

Survival The formula for the survival function of the Weibull distribution is
Function

The following is the plot of the Weibull survival function with the same values of
as the pdf plots above.

Inverse The formula for the inverse survival function of the Weibull distribution is
Survival
Function

The following is the plot of the Weibull inverse survival function with the same
values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3668.htm (5 of 7) [5/7/2002 3:15:29 PM]

1.3.6.6.8. Weibull Distribution

Common The formulas below are with the location parameter equal to zero and the scale
Statistics parameter equal to one.
Mean

where is the gamma function

Median
Mode

Range Zero to positive infinity.

Standard Deviation

Coefficient of Variation

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3668.htm (6 of 7) [5/7/2002 3:15:29 PM]

1.3.6.6.8. Weibull Distribution

Parameter Maximum likelihood estimation for the Weibull distribution is discussed in the
Estimation Reliability chapter (Chapter 8). It is also discussed in Chapter 21 of Johnson, Kotz,
and Balakrishnan.

Comments The Weibull distribution is used extensively in reliability applications to model

failure times.

Software Most general purpose statistical software programs, including Dataplot, support at
least some of the probability functions for the Weibull distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3668.htm (7 of 7) [5/7/2002 3:15:29 PM]

1.3.6.6.9. Lognormal Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.6. Gallery of Distributions

1.3.6.6.9. Lognormal Distribution

Probability A variable X is lognormally distributed if Y = LN(X) is normally
Density distributed with "LN" denoting the natural logarithm. The general
Function formula for the probability density function of the lognormal
distribution is

where is the shape parameter, is the location parameter and m is the

scale parameter. The case where = 0 and m = 1 is called the standard
lognormal distribution. The case where equals zero is called the
2-parameter lognormal distribution.
The equation for the standard lognormal distribution is

Since the general form of probability functions can be expressed in

terms of the standard distribution, all subsequent formulas in this section
are given for the standard form of the function.
The following is the plot of the lognormal probability density function
for four values of .

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3669.htm (1 of 8) [5/7/2002 3:15:31 PM]

1.3.6.6.9. Lognormal Distribution

There are several common parameterizations of the lognormal

distribution. The form given here is from Evans, Hastings, and Peacock.

Cumulative The formula for the cumulative distribution function of the lognormal
Distribution distribution is
Function

where is the cumulative distribution function of the normal

distribution.

The following is the plot of the lognormal cumulative distribution

function with the same values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3669.htm (2 of 8) [5/7/2002 3:15:31 PM]

1.3.6.6.9. Lognormal Distribution

Percent The formula for the percent point function of the lognormal distribution
Point is
Function

where is the percent point function of the normal distribution.

The following is the plot of the lognormal percent point function with
the same values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3669.htm (3 of 8) [5/7/2002 3:15:31 PM]

1.3.6.6.9. Lognormal Distribution

Hazard The formula for the hazard function of the lognormal distribution is
Function

where is the probability density function of the normal distribution

and is the cumulative distribution function of the normal distribution.
The following is the plot of the lognormal hazard function with the same
values of as the pdf plots above.

Cumulative The formula for the cumulative hazard function of the lognormal
Hazard distribution is
Function

where is the cumulative distribution function of the normal

distribution.

The following is the plot of the lognormal cumulative hazard function

with the same values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3669.htm (4 of 8) [5/7/2002 3:15:31 PM]

1.3.6.6.9. Lognormal Distribution

Survival The formula for the survival function of the lognormal distribution is
Function

where is the cumulative distribution function of the normal

distribution.

The following is the plot of the lognormal survival function with the
same values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3669.htm (5 of 8) [5/7/2002 3:15:31 PM]

1.3.6.6.9. Lognormal Distribution

Inverse The formula for the inverse survival function of the lognormal
Survival distribution is
Function

where is the percent point function of the normal distribution.

The following is the plot of the lognormal inverse survival function with
the same values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3669.htm (6 of 8) [5/7/2002 3:15:31 PM]

1.3.6.6.9. Lognormal Distribution

Common The formulas below are with the location parameter equal to zero and
Statistics the scale parameter equal to one.
Mean
Median Scale parameter m (= 1 if scale parameter not
specified).
Mode

Range Zero to positive infinity

Standard Deviation

Coefficient of
Skewness
Coefficient of
Kurtosis
Coefficient of
Variation

Parameter The maximum likelihood estimates for the scale parameter, m, and the
Estimation shape parameter, , are

and

where

If the location parameter is known, it can be subtracted from the original

data points before computing the maximum likelihood estimates of the
shape and scale parameters.

Comments The lognormal distribution is used extensively in reliability applications

to model failure times. The lognormal and Weibull distributions are
probably the most commonly used distributions in reliability
applications.

Software Most general purpose statistical software programs, including Dataplot,

support at least some of the probability functions for the lognormal
distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3669.htm (7 of 8) [5/7/2002 3:15:31 PM]

1.3.6.6.9. Lognormal Distribution

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3669.htm (8 of 8) [5/7/2002 3:15:31 PM]

1.3.6.6.10. Fatigue Life Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.6. Gallery of Distributions

1.3.6.6.10. Fatigue Life Distribution

Probability The fatigue life distribution is also commonly known as the Birnbaum-Saunders
Density distribution. There are several alternative formulations of the fatigue life
Function distribution in the literature.
The general formula for the probability density function of the fatigue life
distribution is

where is the shape parameter, is the location parameter, is the scale

parameter, is the probability density function of the standard normal
distribution, and is the cumulative distribution function of the standard normal
distribution. The case where = 0 and = 1 is called the standard fatigue life
distribution. The equation for the standard fatigue life distribution reduces to

Since the general form of probability functions can be expressed in terms of the
standard distribution, all subsequent formulas in this section are given for the
standard form of the function.
The following is the plot of the fatigue life probability density function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366a.htm (1 of 7) [5/7/2002 3:15:32 PM]

1.3.6.6.10. Fatigue Life Distribution

Cumulative The formula for the cumulative distribution function of the fatigue life
Distribution distribution is
Function

where is the cumulative distribution function of the standard normal

distribution. The following is the plot of the fatigue life cumulative distribution
function with the same values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366a.htm (2 of 7) [5/7/2002 3:15:32 PM]

1.3.6.6.10. Fatigue Life Distribution

Percent The formula for the percent point function of the fatigue life distribution is
Point
Function

where is the percent point function of the standard normal distribution. The
following is the plot of the fatigue life percent point function with the same
values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366a.htm (3 of 7) [5/7/2002 3:15:32 PM]

1.3.6.6.10. Fatigue Life Distribution

Hazard The fatigue life hazard function can be computed from the fatigue life probability
Function density and cumulative distribution functions.
The following is the plot of the fatigue life hazard function with the same values
of as the pdf plots above.

Cumulative The fatigue life cumulative hazard function can be computed from the fatigue life
Hazard cumulative distribution function.
Function
The following is the plot of the fatigue cumulative hazard function with the same
values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366a.htm (4 of 7) [5/7/2002 3:15:32 PM]

1.3.6.6.10. Fatigue Life Distribution

Survival The fatigue life survival function can be computed from the fatigue life
Function cumulative distribution function.
The following is the plot of the fatigue survival function with the same values of
as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366a.htm (5 of 7) [5/7/2002 3:15:32 PM]

1.3.6.6.10. Fatigue Life Distribution

Inverse The fatigue life inverse survival function can be computed from the fatigue life
Survival percent point function.
Function
The following is the plot of the gamma inverse survival function with the same
values of as the pdf plots above.

Common The formulas below are with the location parameter equal to zero and the scale
Statistics parameter equal to one.
Mean

Range Zero to positive infinity.

Standard Deviation

Coefficient of Variation

Parameter Maximum likelihood estimation for the fatigue life distribution is discussed in the
Estimation Reliability chapter.

Comments The fatigue life distribution is used extensively in reliability applications to model
failure times.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366a.htm (6 of 7) [5/7/2002 3:15:32 PM]

1.3.6.6.10. Fatigue Life Distribution

Software Some general purpose statistical software programs, including Dataplot, support
at least some of the probability functions for the fatigue life distribution. Support
for this distribution is likely to be available for statistical programs that
emphasize reliability applications.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366a.htm (7 of 7) [5/7/2002 3:15:32 PM]

1.3.6.6.11. Gamma Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.6. Gallery of Distributions

1.3.6.6.11. Gamma Distribution

Probability The general formula for the probability density function of the gamma
Density distribution is
Function

where is the shape parameter, is the location parameter, is the

scale parameter, and is the gamma function which has the formula

The case where = 0 and = 1 is called the standard gamma

distribution. The equation for the standard gamma distribution reduces
to

Since the general form of probability functions can be expressed in

terms of the standard distribution, all subsequent formulas in this section
are given for the standard form of the function.
The following is the plot of the gamma probability density function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366b.htm (1 of 7) [5/7/2002 3:15:33 PM]

1.3.6.6.11. Gamma Distribution

Cumulative The formula for the cumulative distribution function of the gamma
Distribution distribution is
Function

where is the gamma function defined above and is the

incomplete gamma function. The incomplete gamma function has the
formula

The following is the plot of the gamma cumulative distribution function

with the same values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366b.htm (2 of 7) [5/7/2002 3:15:33 PM]

1.3.6.6.11. Gamma Distribution

Percent The formula for the percent point function of the gamma distribution
Point does not exist in a simple closed form. It is computed numerically.
Function
The following is the plot of the gamma percent point function with the
same values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366b.htm (3 of 7) [5/7/2002 3:15:33 PM]

1.3.6.6.11. Gamma Distribution

Hazard The formula for the hazard function of the gamma distribution is
Function

The following is the plot of the gamma hazard function with the same
values of as the pdf plots above.

Cumulative The formula for the cumulative hazard function of the gamma
Hazard distribution is
Function

where is the gamma function defined above and is the

incomplete gamma function defined above.
The following is the plot of the gamma cumulative hazard function with
the same values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366b.htm (4 of 7) [5/7/2002 3:15:33 PM]

1.3.6.6.11. Gamma Distribution

Survival The formula for the survival function of the gamma distribution is
Function

where is the gamma function defined above and is the

incomplete gamma function defined above.
The following is the plot of the gamma survival function with the same
values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366b.htm (5 of 7) [5/7/2002 3:15:33 PM]

1.3.6.6.11. Gamma Distribution

Inverse The gamma inverse survival function does not exist in simple closed
Survival form. It is computed numberically.
Function
The following is the plot of the gamma inverse survival function with
the same values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366b.htm (6 of 7) [5/7/2002 3:15:33 PM]

1.3.6.6.11. Gamma Distribution

Common The formulas below are with the location parameter equal to zero and
Statistics the scale parameter equal to one.
Mean
Mode
Range Zero to positive infinity.
Standard Deviation
Skewness

Kurtosis

Coefficient of
Variation

Parameter The method of moments estimators of the gamma distribution are

Estimation

where and s are the sample mean and standard deviation, respectively.
The equations for the maximum likelihood estimation of the shape and
scale parameters are given in Chapter 18 of Evans, Hastings, and
Peacock and Chapter 17 of Johnson, Kotz, and Balakrishnan. These
equations need to be solved numerically; this is typically accomplished
by using statistical software packages.

Software Some general purpose statistical software programs, including Dataplot,

support at least some of the probability functions for the gamma
distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366b.htm (7 of 7) [5/7/2002 3:15:33 PM]

1.3.6.6.12. Double Exponential Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.6. Gallery of Distributions

1.3.6.6.12. Double Exponential Distribution

Probability The general formula for the probability density function of the double
Density exponential distribution is
Function

where is the location parameter and is the scale parameter. The

case where = 0 and = 1 is called the standard double exponential
distribution. The equation for the standard double exponential
distribution is

Since the general form of probability functions can be expressed in

terms of the standard distribution, all subsequent formulas in this section
are given for the standard form of the function.
The following is the plot of the double exponential probability density
function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366c.htm (1 of 7) [5/7/2002 3:15:34 PM]

1.3.6.6.12. Double Exponential Distribution

Cumulative The formula for the cumulative distribution function of the double
Distribution exponential distribution is
Function

The following is the plot of the double exponential cumulative

distribution function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366c.htm (2 of 7) [5/7/2002 3:15:34 PM]

1.3.6.6.12. Double Exponential Distribution

Percent The formula for the percent point function of the double exponential
Point distribution is
Function

The following is the plot of the double exponential percent point

function.

Hazard The formula for the hazard function of the double exponential
Function distribution is

The following is the plot of the double exponential hazard function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366c.htm (3 of 7) [5/7/2002 3:15:34 PM]

1.3.6.6.12. Double Exponential Distribution

Cumulative The formula for the cumulative hazard function of the double
Hazard exponential distribution is
Function

The following is the plot of the double exponential cumulative hazard

function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366c.htm (4 of 7) [5/7/2002 3:15:34 PM]

1.3.6.6.12. Double Exponential Distribution

Survival The double exponential survival function can be computed from the
Function cumulative distribution function of the double exponential distribution.
The following is the plot of the double exponential survival function.

Inverse The formula for the inverse survival function of the double exponential
Survival distribution is
Function

The following is the plot of the double exponential inverse survival

function.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366c.htm (5 of 7) [5/7/2002 3:15:34 PM]

1.3.6.6.12. Double Exponential Distribution

Common Mean
Statistics Median
Mode
Range Negative infinity to positive infinity
Standard Deviation
Skewness 0
Kurtosis 6
Coefficient of
Variation

Parameter The maximum likelihood estimators of the location and scale parameters
Estimation of the double exponential distribution are

where is the sample median.

Software Some general purpose statistical software programs, including Dataplot,

support at least some of the probability functions for the double
exponential distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366c.htm (6 of 7) [5/7/2002 3:15:34 PM]

1.3.6.6.12. Double Exponential Distribution

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366c.htm (7 of 7) [5/7/2002 3:15:34 PM]

1.3.6.6.13. Power Normal Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.6. Gallery of Distributions

1.3.6.6.13. Power Normal Distribution

Probability The formula for the probability density function of the standard form of
Density the power normal distribution is
Function

where p is the shape parameter (also referred to as the power parameter),

is the cumulative distribution function of the standard normal
distribution, and is the probability density function of the standard
normal distribution.

As with other probability distributions, the power normal distribution

can be transformed with a location parameter, , and a scale parameter,
. We omit the equation for the general form of the power normal
distribution. Since the general form of probability functions can be
expressed in terms of the standard distribution, all subsequent formulas
in this section are given for the standard form of the function.
The following is the plot of the power normal probability density
function with four values of p.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366d.htm (1 of 7) [5/7/2002 3:15:35 PM]

1.3.6.6.13. Power Normal Distribution

Cumulative The formula for the cumulative distribution function of the power
Distribution normal distribution is
Function

where is the cumulative distribution function of the standard normal

distribution.
The following is the plot of the power normal cumulative distribution
function with the same values of p as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366d.htm (2 of 7) [5/7/2002 3:15:35 PM]

1.3.6.6.13. Power Normal Distribution

Percent The formula for the percent point function of the power normal
Point distribution is
Function

where is the percent point function of the standard normal

distribution.
The following is the plot of the power normal percent point function
with the same values of p as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366d.htm (3 of 7) [5/7/2002 3:15:35 PM]

1.3.6.6.13. Power Normal Distribution

Hazard The formula for the hazard function of the power normal distribution is
Function

The following is the plot of the power normal hazard function with the
same values of p as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366d.htm (4 of 7) [5/7/2002 3:15:35 PM]

1.3.6.6.13. Power Normal Distribution

Cumulative The formula for the cumulative hazard function of the power normal
Hazard distribution is
Function

The following is the plot of the power normal cumulative hazard

function with the same values of p as the pdf plots above.

Survival The formula for the survival function of the power normal distribution is
Function

The following is the plot of the power normal survival function with the
same values of p as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366d.htm (5 of 7) [5/7/2002 3:15:35 PM]

1.3.6.6.13. Power Normal Distribution

Inverse The formula for the inverse survival function of the power normal
Survival distribution is
Function

The following is the plot of the power normal inverse survival function
with the same values of p as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366d.htm (6 of 7) [5/7/2002 3:15:35 PM]

1.3.6.6.13. Power Normal Distribution

Common The statistics for the power normal distribution are complicated and
Statistics require tables. Nelson discusses the mean, median, mode, and standard
deviation of the power normal distribution and provides references to
the appropriate tables.

Software Most general purpose statistical software programs do not support the
probability functions for the power normal distribution. Dataplot does
support them.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366d.htm (7 of 7) [5/7/2002 3:15:35 PM]

1.3.6.6.14. Power Lognormal Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.6. Gallery of Distributions

1.3.6.6.14. Power Lognormal Distribution

Probability The formula for the probability density function of the standard form of the power
Density lognormal distribution is
Function

where p (also referred to as the power parameter) and are the shape parameters,
is the cumulative distribution function of the standard normal distribution, and
is the probability density function of the standard normal distribution.

As with other probability distributions, the power lognormal distribution can be

transformed with a location parameter, , and a scale parameter, B. We omit the
equation for the general form of the power lognormal distribution. Since the
general form of probability functions can be expressed in terms of the standard
distribution, all subsequent formulas in this section are given for the standard form
of the function.
The following is the plot of the power lognormal probability density function with
four values of p and set to 1.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366e.htm (1 of 6) [5/7/2002 3:15:42 PM]

1.3.6.6.14. Power Lognormal Distribution

Cumulative The formula for the cumulative distribution function of the power lognormal
Distribution distribution is
Function

where is the cumulative distribution function of the standard normal distribution.

The following is the plot of the power lognormal cumulative distribution function
with the same values of p as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366e.htm (2 of 6) [5/7/2002 3:15:42 PM]

1.3.6.6.14. Power Lognormal Distribution

Percent The formula for the percent point function of the power lognormal distribution is
Point
Function

where is the percent point function of the standard normal distribution.

The following is the plot of the power lognormal percent point function with the
same values of p as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366e.htm (3 of 6) [5/7/2002 3:15:42 PM]

1.3.6.6.14. Power Lognormal Distribution

Hazard The formula for the hazard function of the power lognormal distribution is
Function

where is the cumulative distribution function of the standard normal distribution,

and is the probability density function of the standard normal distribution.
Note that this is simply a multiple (p) of the lognormal hazard function.

The following is the plot of the power lognormal hazard function with the same
values of p as the pdf plots above.

Cumulative The formula for the cumulative hazard function of the power lognormal
Hazard distribution is
Function

The following is the plot of the power lognormal cumulative hazard function with
the same values of p as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366e.htm (4 of 6) [5/7/2002 3:15:42 PM]

1.3.6.6.14. Power Lognormal Distribution

Survival The formula for the survival function of the power lognormal distribution is
Function

The following is the plot of the power lognormal survival function with the same
values of p as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366e.htm (5 of 6) [5/7/2002 3:15:42 PM]

1.3.6.6.14. Power Lognormal Distribution

Inverse The formula for the inverse survival function of the power lognormal distribution is
Survival
Function

The following is the plot of the power lognormal inverse survival function with the
same values of p as the pdf plots above.

Common The statistics for the power lognormal distribution are complicated and require
Statistics tables. Nelson discusses the mean, median, mode, and standard deviation of the
power lognormal distribution and provides references to the appropriate tables.

Parameter Nelson discusses maximum likelihood estimation for the power lognormal
Estimation distribution. These estimates need to be performed with computer software.
Software for maximum likelihood estimation of the parameters of the power
lognormal distribution is not as readily available as for other reliability
distributions such as the exponential, Weibull, and lognormal.

Software Most general purpose statistical software programs do not support the probability
functions for the power lognormal distribution. Dataplot does support them.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366e.htm (6 of 6) [5/7/2002 3:15:42 PM]

1.3.6.6.15. Tukey-Lambda Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.6. Gallery of Distributions

1.3.6.6.15. Tukey-Lambda Distribution

Probability The Tukey-Lambda density function does not have a simple, closed
Density form. It is computed numerically.
Function
The Tukey-Lambda distribution has the shape parameter . As with
other probability distributions, the Tukey-Lambda distribution can be
transformed with a location parameter, , and a scale parameter, .
Since the general form of probability functions can be expressed in
terms of the standard distribution, all subsequent formulas in this section
are given for the standard form of the function.
The following is the plot of the Tukey-Lambda probability density
function for four values of .

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366f.htm (1 of 4) [5/7/2002 3:15:42 PM]

1.3.6.6.15. Tukey-Lambda Distribution

Cumulative The Tukey-Lambda distribution does not have a simple, closed form. It
Distribution is computed numerically.
Function
The following is the plot of the Tukey-Lambda cumulative distribution
function with the same values of as the pdf plots above.

Percent The formula for the percent point function of the standard form of the
Point Tukey-Lambda distribution is
Function

The following is the plot of the Tukey-Lambda percent point function

with the same values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366f.htm (2 of 4) [5/7/2002 3:15:42 PM]

1.3.6.6.15. Tukey-Lambda Distribution

Other The Tukey-Lambda distribution is typically used to identify an

Probability appropriate distribution (see the comments below) and not used in
Functions statistical models directly. For this reason, we omit the formulas, and
plots for the hazard, cumulative hazard, survival, and inverse survival
functions. We also omit the common statistics and parameter estimation
sections.

Comments The Tukey-Lambda distribution is actually a family of distributions that

can approximate a number of common distributions. For example,
= -1 approximately Cauchy
= 0 exactly logistic
= 0.14 approximately normal
= 0.5 U-shaped
= 1 exactly uniform (from -1 to +1)
The most common use of this distribution is to generate a
Tukey-Lambda PPCC plot of a data set. Based on the ppcc plot, an
appropriate model for the data is suggested. For example, if the
maximum correlation occurs for a value of at or near 0.14, then the
data can be modeled with a normal distribution. Values of less than
this imply a heavy-tailed distribution (with -1 approximating a Cauchy).
That is, as the optimal value of goes from 0.14 to -1, increasingly
heavy tails are implied. Similarly, as the optimal value of becomes
greater than 0.14, shorter tails are implied.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366f.htm (3 of 4) [5/7/2002 3:15:42 PM]

1.3.6.6.15. Tukey-Lambda Distribution

As the Tukey-Lambda distribution is a symmetric distribution, the use

of the Tukey-Lambda PPCC plot to determine a reasonable distribution
to model the data only applies to symmetric distributuins. A histogram
of the data should provide evidence as to whether the data can be
reasonably modeled with a symmetric distribution.

Software Most general purpose statistical software programs do not support the
probability functions for the Tukey-Lambda distribution. Dataplot does
support them.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366f.htm (4 of 4) [5/7/2002 3:15:42 PM]

1.3.6.6.16. Extreme Value Type I Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.6. Gallery of Distributions

1.3.6.6.16. Extreme Value Type I

Distribution
Probability The extreme value type I distribution has two forms. One is based on the
Density smallest extreme and the other is based on the largest extreme. We call
Function these the minimum and maximum cases, respectively. Formulas and
plots for both cases are given. The extreme value type I distribution is
also referred to as the Gumbel distribution.
The general formula for the probability density function of the Gumbel
(minimum) distribution is

where is the location parameter and is the scale parameter. The

case where = 0 and = 1 is called the standard Gumbel
distribution. The equation for the standard Gumbel distribution
(minimum) reduces to

The following is the plot of the Gumbel probability density function for
the minimum case.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366g.htm (1 of 12) [5/7/2002 3:15:44 PM]

1.3.6.6.16. Extreme Value Type I Distribution

The general formula for the probability density function of the Gumbel
(maximum) distribution is

where is the location parameter and is the scale parameter. The

case where = 0 and = 1 is called the standard Gumbel
distribution. The equation for the standard Gumbel distribution
(maximum) reduces to

The following is the plot of the Gumbel probability density function for
the maximum case.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366g.htm (2 of 12) [5/7/2002 3:15:44 PM]

1.3.6.6.16. Extreme Value Type I Distribution

Since the general form of probability functions can be expressed in

terms of the standard distribution, all subsequent formulas in this section
are given for the standard form of the function.

Cumulative The formula for the cumulative distribution function of the Gumbel
Distribution distribution (minimum) is
Function

The following is the plot of the Gumbel cumulative distribution function

for the minimum case.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366g.htm (3 of 12) [5/7/2002 3:15:44 PM]

1.3.6.6.16. Extreme Value Type I Distribution

The formula for the cumulative distribution function of the Gumbel

distribution (maximum) is

The following is the plot of the Gumbel cumulative distribution function

for the maximum case.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366g.htm (4 of 12) [5/7/2002 3:15:44 PM]

1.3.6.6.16. Extreme Value Type I Distribution

Percent The formula for the percent point function of the Gumbel distribution
Point (minimum) is
Function

The following is the plot of the Gumbel percent point function for the
minimum case.

The formula for the percent point function of the Gumbel distribution
(maximum) is

The following is the plot of the Gumbel percent point function for the
maximum case.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366g.htm (5 of 12) [5/7/2002 3:15:44 PM]

1.3.6.6.16. Extreme Value Type I Distribution

Hazard The formula for the hazard function of the Gumbel distribution
Function (minimum) is

The following is the plot of the Gumbel hazard function for the
minimum case.

The formula for the hazard function of the Gumbel distribution

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366g.htm (6 of 12) [5/7/2002 3:15:44 PM]

1.3.6.6.16. Extreme Value Type I Distribution

(maximum) is

The following is the plot of the Gumbel hazard function for the
maximum case.

Cumulative The formula for the cumulative hazard function of the Gumbel
Hazard distribution (minimum) is
Function

The following is the plot of the Gumbel cumulative hazard function for
the minimum case.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366g.htm (7 of 12) [5/7/2002 3:15:44 PM]

1.3.6.6.16. Extreme Value Type I Distribution

The formula for the cumulative hazard function of the Gumbel

distribution (maximum) is

The following is the plot of the Gumbel cumulative hazard function for
the maximum case.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366g.htm (8 of 12) [5/7/2002 3:15:44 PM]

1.3.6.6.16. Extreme Value Type I Distribution

Survival The formula for the survival function of the Gumbel distribution
Function (minimum) is

The following is the plot of the Gumbel survival function for the
minimum case.

The formula for the survival function of the Gumbel distribution

(maximum) is

The following is the plot of the Gumbel survival function for the
maximum case.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366g.htm (9 of 12) [5/7/2002 3:15:44 PM]

1.3.6.6.16. Extreme Value Type I Distribution

Inverse The formula for the inverse survival function of the Gumbel distribution
Survival (minimum) is
Function

The following is the plot of the Gumbel inverse survival function for the
minimum case.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366g.htm (10 of 12) [5/7/2002 3:15:44 PM]

1.3.6.6.16. Extreme Value Type I Distribution

The formula for the inverse survival function of the Gumbel distribution
(maximum) is

The following is the plot of the Gumbel inverse survival function for the
maximum case.

Common The formulas below are for the maximum order statistic case.
Statistics Mean

The constant 0.5772 is Euler's number.

Median
Mode
Range Negative infinity to positive infinity.
Standard Deviation

Skewness 1.13955
Kurtosis 5.4
Coefficient of
Variation

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366g.htm (11 of 12) [5/7/2002 3:15:44 PM]

1.3.6.6.16. Extreme Value Type I Distribution

Parameter The method of moments estimators of the Gumbel (maximum)

Estimation distribution are

where and s are the sample mean and standard deviation,

respectively.
The equations for the maximum likelihood estimation of the shape and
scale parameters are discussed in Chapter 15 of Evans, Hastings, and
Peacock and Chapter 22 of Johnson, Kotz, and Balakrishnan. These
equations need to be solved numerically and this is typically
accomplished by using statistical software packages.

Software Some general purpose statistical software programs, including Dataplot,

support at least some of the probability functions for the extreme value
type I distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366g.htm (12 of 12) [5/7/2002 3:15:44 PM]

1.3.6.6.17. Binomial Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.6. Gallery of Distributions

1.3.6.6.17. Binomial Distribution

Probability The binomial distribution is used when there are exactly two mutually
Mass exclusive outcomes of a trial. These outcomes are appropriately labeled
Function "success" and "failure". The binomial distribution is used to obtain the
probability of observing x successes in N trials, with the probability of success
on a single trial denoted by p. The binomial distribution assumes that p is fixed
for all trials.
The formula for the binomial probability mass function is

where

The following is the plot of the binomial probability density function for four
values of p and n = 100.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366h.htm (1 of 4) [5/7/2002 3:15:45 PM]

1.3.6.6.17. Binomial Distribution

Cumulative The formula for the binomial cumulative probability function is

Distribution
Function

The following is the plot of the binomial cumulative distribution function with
the same values of p as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366h.htm (2 of 4) [5/7/2002 3:15:45 PM]

1.3.6.6.17. Binomial Distribution

Percent The binomial percent point function does not exist in simple closed form. It is
Point computed numerically. Note that because this is a discrete distribution that is
Function only defined for integer values of x, the percent point function is not smooth in
the way the percent point function typically is for a continuous distribution.
The following is the plot of the binomial percent point function with the same
values of p as the pdf plots above.

Common Mean
Statistics Mode
Range 0 to N
Standard Deviation

Coefficient of
Variation

Coefficient of
Skewness

Coefficient of Kurtosis

Comments The binomial distribution is probably the most commonly used discrete
distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366h.htm (3 of 4) [5/7/2002 3:15:45 PM]

1.3.6.6.17. Binomial Distribution

Parameter The maximum likelihood estimator of p (n is fixed) is

Estimation

Software Most general purpose statistical software programs, including Dataplot, support
at least some of the probability functions for the binomial distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366h.htm (4 of 4) [5/7/2002 3:15:45 PM]

1.3.6.6.18. Poisson Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.6. Gallery of Distributions

1.3.6.6.18. Poisson Distribution

Probability The Poisson distribution is used to model the number of events
Mass occurring within a given time interval.
Function
The formula for the Poisson probability mass function is

is the shape parameter which indicates the average number of events

in the given time interval.
The following is the plot of the Poisson probability density function for
four values of .

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366i.htm (1 of 4) [5/7/2002 3:15:45 PM]

1.3.6.6.18. Poisson Distribution

Cumulative The formula for the Poisson cumulative probability function is

Distribution
Function

The following is the plot of the Poisson cumulative distribution function

with the same values of as the pdf plots above.

Percent The Poisson percent point function does not exist in simple closed form.
Point It is computed numerically. Note that because this is a discrete
Function distribution that is only defined for integer values of x, the percent point
function is not smooth in the way the percent point function typically is
for a continuous distribution.
The following is the plot of the Poisson percent point function with the
same values of as the pdf plots above.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366i.htm (2 of 4) [5/7/2002 3:15:45 PM]

1.3.6.6.18. Poisson Distribution

Common Mean
Statistics Mode For non-integer , it is the largest integer less
than . For integer , x = and x = - 1 are
both the mode.
Range 0 to positive infinity
Standard Deviation
Coefficient of
Variation
Coefficient of
Skewness
Coefficient of
Kurtosis

Parameter The maximum likelihood estimator of is

Estimation

where is the sample mean.

Software Most general purpose statistical software programs, including Dataplot,

support at least some of the probability functions for the Poisson
distribution.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366i.htm (3 of 4) [5/7/2002 3:15:45 PM]

1.3.6.6.18. Poisson Distribution

http://www.itl.nist.gov/div898/handbook/eda/section3/eda366i.htm (4 of 4) [5/7/2002 3:15:45 PM]

1.3.6.7. Tables for Probability Distributions

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions

1.3.6.7. Tables for Probability Distributions

Tables Several commonly used tables for probability distributions can be
referenced below.
The values from these tables can also be obtained from most general
purpose statistical software programs. Most introductory statistics
textbooks (e.g., Snedecor and Cochran) contain more extensive tables
than are included here. These tables are included for convenience.
1. Cumulative distribution function for the standard normal
distribution
2. Upper critical values of Student's t-distribution with degrees of
freedom
3. Upper critical values of the F-distribution with and degrees
of freedom
4. Upper critical values of the chi-square distribution with degrees
of freedom
5. Critical values of t* distribution for testing the output of a linear
calibration line at 3 points
6. Upper critical values of the normal PPCC distribution

http://www.itl.nist.gov/div898/handbook/eda/section3/eda367.htm [5/7/2002 3:15:46 PM]

1.3.6.7.1. Cumulative Distribution Function of the Standard Normal Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.7. Tables for Probability Distributions

1.3.6.7.1. Cumulative Distribution Function

of the Standard Normal
Distribution
How to Use The table below contains the area under the standard normal curve from
This Table 0 to z. This can be used to compute the cumulative distribution function
values for the standard normal distribution.

The table utilizes the symmetry of the normal distribution, so what in

fact is given is

where a is the value of interest. This is demonstrated in the graph below

for a = 0.5. The shaded area of the curve represents the probability that x
is between 0 and a.

This can be clarified by a few simple examples.

1. What is the probability that x is less than or equal to 1.53? Look
for 1.5 in the X column, go right to the 0.03 column to find the
value 0.43699. Now add 0.5 (for the probability less than zero) to
obtain the final result of 0.93699.
2. What is the probability that x is less than or equal to -1.53? For
negative values, use the relationship

From the first example, this gives 1 - 0.93699 = 0.06301.

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3671.htm (1 of 3) [5/7/2002 3:15:46 PM]

1.3.6.7.1. Cumulative Distribution Function of the Standard Normal Distribution
3. What is the probability that x is between -1 and 0.5? Look up the
values for 0.5 (0.5 + 0.19146 = 0.69146) and -1 (1 - (0.5 +
0.34134) = 0.15866). Then subtract the results (0.69146 -
0.15866) to obtain the result 0.5328.
To use this table with a non-standard normal distribution (either the
location parameter is not 0 or the scale parameter is not 1), standardize
your value by subtracting the mean and dividing the result by the
standard deviation. Then look up the value for this standardized value.
A few particularly important numbers derived from the table below,
specifically numbers that are commonly used in significance tests, are
summarized in the following table:
p 0.001 0.005 0.010 0.025 0.050 0.100
Zp -3.090 -2.576 -2.326 -1.960 -1.645 -1.282
p 0.999 0.995 0.990 0.975 0.950 0.900
Zp +3.090 +2.576 +2.326 +1.960 +1.645 +1.282

These are critical values for the normal distribution.

Area under the Normal Curve from 0 to X

X 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09

0.0 0.00000 0.00399 0.00798 0.01197 0.01595 0.01994 0.02392 0.02790 0.03188
0.03586
0.1 0.03983 0.04380 0.04776 0.05172 0.05567 0.05962 0.06356 0.06749 0.07142
0.07535
0.2 0.07926 0.08317 0.08706 0.09095 0.09483 0.09871 0.10257 0.10642 0.11026
0.11409
0.3 0.11791 0.12172 0.12552 0.12930 0.13307 0.13683 0.14058 0.14431 0.14803
0.15173
0.4 0.15542 0.15910 0.16276 0.16640 0.17003 0.17364 0.17724 0.18082 0.18439
0.18793
0.5 0.19146 0.19497 0.19847 0.20194 0.20540 0.20884 0.21226 0.21566 0.21904
0.22240
0.6 0.22575 0.22907 0.23237 0.23565 0.23891 0.24215 0.24537 0.24857 0.25175
0.25490
0.7 0.25804 0.26115 0.26424 0.26730 0.27035 0.27337 0.27637 0.27935 0.28230
0.28524
0.8 0.28814 0.29103 0.29389 0.29673 0.29955 0.30234 0.30511 0.30785 0.31057
0.31327
0.9 0.31594 0.31859 0.32121 0.32381 0.32639 0.32894 0.33147 0.33398 0.33646
0.33891
1.0 0.34134 0.34375 0.34614 0.34849 0.35083 0.35314 0.35543 0.35769 0.35993
0.36214
1.1 0.36433 0.36650 0.36864 0.37076 0.37286 0.37493 0.37698 0.37900 0.38100
0.38298
1.2 0.38493 0.38686 0.38877 0.39065 0.39251 0.39435 0.39617 0.39796 0.39973
0.40147
1.3 0.40320 0.40490 0.40658 0.40824 0.40988 0.41149 0.41308 0.41466 0.41621
0.41774
1.4 0.41924 0.42073 0.42220 0.42364 0.42507 0.42647 0.42785 0.42922 0.43056
0.43189

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3671.htm (2 of 3) [5/7/2002 3:15:46 PM]

1.3.6.7.1. Cumulative Distribution Function of the Standard Normal Distribution
1.5 0.43319 0.43448 0.43574 0.43699 0.43822 0.43943 0.44062 0.44179 0.44295
0.44408
1.6 0.44520 0.44630 0.44738 0.44845 0.44950 0.45053 0.45154 0.45254 0.45352
0.45449
1.7 0.45543 0.45637 0.45728 0.45818 0.45907 0.45994 0.46080 0.46164 0.46246
0.46327
1.8 0.46407 0.46485 0.46562 0.46638 0.46712 0.46784 0.46856 0.46926 0.46995
0.47062
1.9 0.47128 0.47193 0.47257 0.47320 0.47381 0.47441 0.47500 0.47558 0.47615
0.47670
2.0 0.47725 0.47778 0.47831 0.47882 0.47932 0.47982 0.48030 0.48077 0.48124
0.48169
2.1 0.48214 0.48257 0.48300 0.48341 0.48382 0.48422 0.48461 0.48500 0.48537
0.48574
2.2 0.48610 0.48645 0.48679 0.48713 0.48745 0.48778 0.48809 0.48840 0.48870
0.48899
2.3 0.48928 0.48956 0.48983 0.49010 0.49036 0.49061 0.49086 0.49111 0.49134
0.49158
2.4 0.49180 0.49202 0.49224 0.49245 0.49266 0.49286 0.49305 0.49324 0.49343
0.49361
2.5 0.49379 0.49396 0.49413 0.49430 0.49446 0.49461 0.49477 0.49492 0.49506
0.49520
2.6 0.49534 0.49547 0.49560 0.49573 0.49585 0.49598 0.49609 0.49621 0.49632
0.49643
2.7 0.49653 0.49664 0.49674 0.49683 0.49693 0.49702 0.49711 0.49720 0.49728
0.49736
2.8 0.49744 0.49752 0.49760 0.49767 0.49774 0.49781 0.49788 0.49795 0.49801
0.49807
2.9 0.49813 0.49819 0.49825 0.49831 0.49836 0.49841 0.49846 0.49851 0.49856
0.49861
3.0 0.49865 0.49869 0.49874 0.49878 0.49882 0.49886 0.49889 0.49893 0.49896
0.49900
3.1 0.49903 0.49906 0.49910 0.49913 0.49916 0.49918 0.49921 0.49924 0.49926
0.49929
3.2 0.49931 0.49934 0.49936 0.49938 0.49940 0.49942 0.49944 0.49946 0.49948
0.49950
3.3 0.49952 0.49953 0.49955 0.49957 0.49958 0.49960 0.49961 0.49962 0.49964
0.49965
3.4 0.49966 0.49968 0.49969 0.49970 0.49971 0.49972 0.49973 0.49974 0.49975
0.49976
3.5 0.49977 0.49978 0.49978 0.49979 0.49980 0.49981 0.49981 0.49982 0.49983
0.49983
3.6 0.49984 0.49985 0.49985 0.49986 0.49986 0.49987 0.49987 0.49988 0.49988
0.49989
3.7 0.49989 0.49990 0.49990 0.49990 0.49991 0.49991 0.49992 0.49992 0.49992
0.49992
3.8 0.49993 0.49993 0.49993 0.49994 0.49994 0.49994 0.49994 0.49995 0.49995
0.49995
3.9 0.49995 0.49995 0.49996 0.49996 0.49996 0.49996 0.49996 0.49996 0.49997
0.49997
4.0 0.49997 0.49997 0.49997 0.49997 0.49997 0.49997 0.49998 0.49998 0.49998
0.49998

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3671.htm (3 of 3) [5/7/2002 3:15:46 PM]

1.3.6.7.2. Upper Critical Values of the Student's-t Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.7. Tables for Probability Distributions

1.3.6.7.2. Upper Critical Values of the Student's-t

Distribution
How to This table contains the upper critical values of the Student's t-distribution. The upper critical
Use This values are computed using the percent point function. Due to the symmetry of the
Table t-distribution, this table can be used for both 1-sided (lower and upper) and 2-sided tests
using the appropriate value of .
The significance level, , is demonstrated with the graph below which plots a t distribution
with 10 degrees of freedom. The most commonly used significance level is = 0.05. For a
two-sided test, we compute the percent point function at /2 (0.025). If the absolute value of
the test statistic is greater than the upper critical value (0.025), then we reject the null
hypothesis. Due to the symmetry of the t-distribution, we only tabulate the upper critical
values in the table below.

Given a specified value for :

1. For a two-sided test, find the column corresponding to /2 and reject the null
hypothesis if the absolute value of the test statistic is greater than the value of in
the table below.
2. For an upper one-sided test, find the column corresponding to and reject the null

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3672.htm (1 of 5) [5/7/2002 3:15:46 PM]

1.3.6.7.2. Upper Critical Values of the Student's-t Distribution

hypothesis if the test statistic is greater than the tabled value.

3. For an lower one-sided test, find the column corresponding to and reject the null
hypothesis if the test statistic is less than the negative of the tabled value.

Upper critical values of Student's t distribution with degrees of freedom

Probability of exceeding the critical value

0.10 0.05 0.025 0.01 0.001

1 3.078 6.314 12.706 31.821 318.313

2 1.886 2.920 4.303 6.965 22.327
3 1.638 2.353 3.182 4.541 10.215
4 1.533 2.132 2.776 3.747 7.173
5 1.476 2.015 2.571 3.365 5.893
6 1.440 1.943 2.447 3.143 5.208
7 1.415 1.895 2.365 2.998 4.782
8 1.397 1.860 2.306 2.896 4.499
9 1.383 1.833 2.262 2.821 4.296
10 1.372 1.812 2.228 2.764 4.143
11 1.363 1.796 2.201 2.718 4.024
12 1.356 1.782 2.179 2.681 3.929
13 1.350 1.771 2.160 2.650 3.852
14 1.345 1.761 2.145 2.624 3.787
15 1.341 1.753 2.131 2.602 3.733
16 1.337 1.746 2.120 2.583 3.686
17 1.333 1.740 2.110 2.567 3.646
18 1.330 1.734 2.101 2.552 3.610
19 1.328 1.729 2.093 2.539 3.579
20 1.325 1.725 2.086 2.528 3.552
21 1.323 1.721 2.080 2.518 3.527
22 1.321 1.717 2.074 2.508 3.505
23 1.319 1.714 2.069 2.500 3.485
24 1.318 1.711 2.064 2.492 3.467
25 1.316 1.708 2.060 2.485 3.450
26 1.315 1.706 2.056 2.479 3.435

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3672.htm (2 of 5) [5/7/2002 3:15:46 PM]

1.3.6.7.2. Upper Critical Values of the Student's-t Distribution

27 1.314 1.703 2.052 2.473 3.421

28 1.313 1.701 2.048 2.467 3.408
29 1.311 1.699 2.045 2.462 3.396
30 1.310 1.697 2.042 2.457 3.385
31 1.309 1.696 2.040 2.453 3.375
32 1.309 1.694 2.037 2.449 3.365
33 1.308 1.692 2.035 2.445 3.356
34 1.307 1.691 2.032 2.441 3.348
35 1.306 1.690 2.030 2.438 3.340
36 1.306 1.688 2.028 2.434 3.333
37 1.305 1.687 2.026 2.431 3.326
38 1.304 1.686 2.024 2.429 3.319
39 1.304 1.685 2.023 2.426 3.313
40 1.303 1.684 2.021 2.423 3.307
41 1.303 1.683 2.020 2.421 3.301
42 1.302 1.682 2.018 2.418 3.296
43 1.302 1.681 2.017 2.416 3.291
44 1.301 1.680 2.015 2.414 3.286
45 1.301 1.679 2.014 2.412 3.281
46 1.300 1.679 2.013 2.410 3.277
47 1.300 1.678 2.012 2.408 3.273
48 1.299 1.677 2.011 2.407 3.269
49 1.299 1.677 2.010 2.405 3.265
50 1.299 1.676 2.009 2.403 3.261
51 1.298 1.675 2.008 2.402 3.258
52 1.298 1.675 2.007 2.400 3.255
53 1.298 1.674 2.006 2.399 3.251
54 1.297 1.674 2.005 2.397 3.248
55 1.297 1.673 2.004 2.396 3.245
56 1.297 1.673 2.003 2.395 3.242
57 1.297 1.672 2.002 2.394 3.239
58 1.296 1.672 2.002 2.392 3.237
59 1.296 1.671 2.001 2.391 3.234
60 1.296 1.671 2.000 2.390 3.232
61 1.296 1.670 2.000 2.389 3.229
62 1.295 1.670 1.999 2.388 3.227
63 1.295 1.669 1.998 2.387 3.225
64 1.295 1.669 1.998 2.386 3.223
65 1.295 1.669 1.997 2.385 3.220

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3672.htm (3 of 5) [5/7/2002 3:15:46 PM]

1.3.6.7.2. Upper Critical Values of the Student's-t Distribution

66 1.295 1.668 1.997 2.384 3.218

67 1.294 1.668 1.996 2.383 3.216
68 1.294 1.668 1.995 2.382 3.214
69 1.294 1.667 1.995 2.382 3.213
70 1.294 1.667 1.994 2.381 3.211
71 1.294 1.667 1.994 2.380 3.209
72 1.293 1.666 1.993 2.379 3.207
73 1.293 1.666 1.993 2.379 3.206
74 1.293 1.666 1.993 2.378 3.204
75 1.293 1.665 1.992 2.377 3.202
76 1.293 1.665 1.992 2.376 3.201
77 1.293 1.665 1.991 2.376 3.199
78 1.292 1.665 1.991 2.375 3.198
79 1.292 1.664 1.990 2.374 3.197
80 1.292 1.664 1.990 2.374 3.195
81 1.292 1.664 1.990 2.373 3.194
82 1.292 1.664 1.989 2.373 3.193
83 1.292 1.663 1.989 2.372 3.191
84 1.292 1.663 1.989 2.372 3.190
85 1.292 1.663 1.988 2.371 3.189
86 1.291 1.663 1.988 2.370 3.188
87 1.291 1.663 1.988 2.370 3.187
88 1.291 1.662 1.987 2.369 3.185
89 1.291 1.662 1.987 2.369 3.184
90 1.291 1.662 1.987 2.368 3.183
91 1.291 1.662 1.986 2.368 3.182
92 1.291 1.662 1.986 2.368 3.181
93 1.291 1.661 1.986 2.367 3.180
94 1.291 1.661 1.986 2.367 3.179
95 1.291 1.661 1.985 2.366 3.178
96 1.290 1.661 1.985 2.366 3.177
97 1.290 1.661 1.985 2.365 3.176
98 1.290 1.661 1.984 2.365 3.175
99 1.290 1.660 1.984 2.365 3.175
100 1.290 1.660 1.984 2.364 3.174
1.282 1.645 1.960 2.326 3.090

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3672.htm (4 of 5) [5/7/2002 3:15:46 PM]

1.3.6.7.2. Upper Critical Values of the Student's-t Distribution

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3672.htm (5 of 5) [5/7/2002 3:15:46 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.7. Tables for Probability Distributions

1.3.6.7.3. Upper Critical Values of the F

Distribution
How to Use This table contains the upper critical values of the F distribution. This
This Table table is used for one-sided F tests at the = 0.05, 0.10, and 0.01 levels.
More specifically, a test statistic is computed with and degrees of
freedom, and the result is compared to this table. For a one-sided test,
the null hypothesis is rejected when the test statistic is greater than the
tabled value. This is demonstrated with the graph of an F distribution
with = 10 and = 10. The shaded area of the graph indicates the
rejection region at the significance level. Since this is a one-sided test,
we have probability in the upper tail of exceeding the critical value
and zero in the lower tail. Because the F distribution is asymmetric, a
two-sided test requires a set of of tables (not included here) that contain
the rejection regions for both the lower and upper tails.

Contents The following tables for from 1 to 100 are included:

1. One sided, 5% significance level, = 1 - 10
2. One sided, 5% significance level, = 11 - 20
3. One sided, 10% significance level, = 1 - 10
4. One sided, 10% significance level, = 11 - 20
5. One sided, 1% significance level, = 1 - 10
6. One sided, 1% significance level, = 11 - 20

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (1 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

Upper critical values of the F distribution

for numerator degrees of freedom and denominator degrees of freedom

5% significance level

\ 1 2 3 4 5 6 7 8
9 10

1 161.448 199.500 215.707 224.583 230.162 233.986 236.768

238.882 240.543 241.882
2 18.513 19.000 19.164 19.247 19.296 19.330 19.353
19.371 19.385 19.396
3 10.128 9.552 9.277 9.117 9.013 8.941 8.887
8.845 8.812 8.786
4 7.709 6.944 6.591 6.388 6.256 6.163 6.094
6.041 5.999 5.964
5 6.608 5.786 5.409 5.192 5.050 4.950 4.876
4.818 4.772 4.735
6 5.987 5.143 4.757 4.534 4.387 4.284 4.207
4.147 4.099 4.060
7 5.591 4.737 4.347 4.120 3.972 3.866 3.787
3.726 3.677 3.637
8 5.318 4.459 4.066 3.838 3.687 3.581 3.500
3.438 3.388 3.347
9 5.117 4.256 3.863 3.633 3.482 3.374 3.293
3.230 3.179 3.137
10 4.965 4.103 3.708 3.478 3.326 3.217 3.135
3.072 3.020 2.978
11 4.844 3.982 3.587 3.357 3.204 3.095 3.012
2.948 2.896 2.854
12 4.747 3.885 3.490 3.259 3.106 2.996 2.913
2.849 2.796 2.753
13 4.667 3.806 3.411 3.179 3.025 2.915 2.832
2.767 2.714 2.671
14 4.600 3.739 3.344 3.112 2.958 2.848 2.764
2.699 2.646 2.602
15 4.543 3.682 3.287 3.056 2.901 2.790 2.707
2.641 2.588 2.544

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (2 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

16 4.494 3.634 3.239 3.007 2.852 2.741 2.657

2.591 2.538 2.494
17 4.451 3.592 3.197 2.965 2.810 2.699 2.614
2.548 2.494 2.450
18 4.414 3.555 3.160 2.928 2.773 2.661 2.577
2.510 2.456 2.412
19 4.381 3.522 3.127 2.895 2.740 2.628 2.544
2.477 2.423 2.378
20 4.351 3.493 3.098 2.866 2.711 2.599 2.514
2.447 2.393 2.348
21 4.325 3.467 3.072 2.840 2.685 2.573 2.488
2.420 2.366 2.321
22 4.301 3.443 3.049 2.817 2.661 2.549 2.464
2.397 2.342 2.297
23 4.279 3.422 3.028 2.796 2.640 2.528 2.442
2.375 2.320 2.275
24 4.260 3.403 3.009 2.776 2.621 2.508 2.423
2.355 2.300 2.255
25 4.242 3.385 2.991 2.759 2.603 2.490 2.405
2.337 2.282 2.236
26 4.225 3.369 2.975 2.743 2.587 2.474 2.388
2.321 2.265 2.220
27 4.210 3.354 2.960 2.728 2.572 2.459 2.373
2.305 2.250 2.204
28 4.196 3.340 2.947 2.714 2.558 2.445 2.359
2.291 2.236 2.190
29 4.183 3.328 2.934 2.701 2.545 2.432 2.346
2.278 2.223 2.177
30 4.171 3.316 2.922 2.690 2.534 2.421 2.334
2.266 2.211 2.165
31 4.160 3.305 2.911 2.679 2.523 2.409 2.323
2.255 2.199 2.153
32 4.149 3.295 2.901 2.668 2.512 2.399 2.313
2.244 2.189 2.142
33 4.139 3.285 2.892 2.659 2.503 2.389 2.303
2.235 2.179 2.133
34 4.130 3.276 2.883 2.650 2.494 2.380 2.294
2.225 2.170 2.123
35 4.121 3.267 2.874 2.641 2.485 2.372 2.285
2.217 2.161 2.114
36 4.113 3.259 2.866 2.634 2.477 2.364 2.277
2.209 2.153 2.106
37 4.105 3.252 2.859 2.626 2.470 2.356 2.270
2.201 2.145 2.098
38 4.098 3.245 2.852 2.619 2.463 2.349 2.262
2.194 2.138 2.091

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (3 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

39 4.091 3.238 2.845 2.612 2.456 2.342 2.255

2.187 2.131 2.084
40 4.085 3.232 2.839 2.606 2.449 2.336 2.249
2.180 2.124 2.077
41 4.079 3.226 2.833 2.600 2.443 2.330 2.243
2.174 2.118 2.071
42 4.073 3.220 2.827 2.594 2.438 2.324 2.237
2.168 2.112 2.065
43 4.067 3.214 2.822 2.589 2.432 2.318 2.232
2.163 2.106 2.059
44 4.062 3.209 2.816 2.584 2.427 2.313 2.226
2.157 2.101 2.054
45 4.057 3.204 2.812 2.579 2.422 2.308 2.221
2.152 2.096 2.049
46 4.052 3.200 2.807 2.574 2.417 2.304 2.216
2.147 2.091 2.044
47 4.047 3.195 2.802 2.570 2.413 2.299 2.212
2.143 2.086 2.039
48 4.043 3.191 2.798 2.565 2.409 2.295 2.207
2.138 2.082 2.035
49 4.038 3.187 2.794 2.561 2.404 2.290 2.203
2.134 2.077 2.030
50 4.034 3.183 2.790 2.557 2.400 2.286 2.199
2.130 2.073 2.026
51 4.030 3.179 2.786 2.553 2.397 2.283 2.195
2.126 2.069 2.022
52 4.027 3.175 2.783 2.550 2.393 2.279 2.192
2.122 2.066 2.018
53 4.023 3.172 2.779 2.546 2.389 2.275 2.188
2.119 2.062 2.015
54 4.020 3.168 2.776 2.543 2.386 2.272 2.185
2.115 2.059 2.011
55 4.016 3.165 2.773 2.540 2.383 2.269 2.181
2.112 2.055 2.008
56 4.013 3.162 2.769 2.537 2.380 2.266 2.178
2.109 2.052 2.005
57 4.010 3.159 2.766 2.534 2.377 2.263 2.175
2.106 2.049 2.001
58 4.007 3.156 2.764 2.531 2.374 2.260 2.172
2.103 2.046 1.998
59 4.004 3.153 2.761 2.528 2.371 2.257 2.169
2.100 2.043 1.995
60 4.001 3.150 2.758 2.525 2.368 2.254 2.167
2.097 2.040 1.993
61 3.998 3.148 2.755 2.523 2.366 2.251 2.164
2.094 2.037 1.990

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (4 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

62 3.996 3.145 2.753 2.520 2.363 2.249 2.161

2.092 2.035 1.987
63 3.993 3.143 2.751 2.518 2.361 2.246 2.159
2.089 2.032 1.985
64 3.991 3.140 2.748 2.515 2.358 2.244 2.156
2.087 2.030 1.982
65 3.989 3.138 2.746 2.513 2.356 2.242 2.154
2.084 2.027 1.980
66 3.986 3.136 2.744 2.511 2.354 2.239 2.152
2.082 2.025 1.977
67 3.984 3.134 2.742 2.509 2.352 2.237 2.150
2.080 2.023 1.975
68 3.982 3.132 2.740 2.507 2.350 2.235 2.148
2.078 2.021 1.973
69 3.980 3.130 2.737 2.505 2.348 2.233 2.145
2.076 2.019 1.971
70 3.978 3.128 2.736 2.503 2.346 2.231 2.143
2.074 2.017 1.969
71 3.976 3.126 2.734 2.501 2.344 2.229 2.142
2.072 2.015 1.967
72 3.974 3.124 2.732 2.499 2.342 2.227 2.140
2.070 2.013 1.965
73 3.972 3.122 2.730 2.497 2.340 2.226 2.138
2.068 2.011 1.963
74 3.970 3.120 2.728 2.495 2.338 2.224 2.136
2.066 2.009 1.961
75 3.968 3.119 2.727 2.494 2.337 2.222 2.134
2.064 2.007 1.959
76 3.967 3.117 2.725 2.492 2.335 2.220 2.133
2.063 2.006 1.958
77 3.965 3.115 2.723 2.490 2.333 2.219 2.131
2.061 2.004 1.956
78 3.963 3.114 2.722 2.489 2.332 2.217 2.129
2.059 2.002 1.954
79 3.962 3.112 2.720 2.487 2.330 2.216 2.128
2.058 2.001 1.953
80 3.960 3.111 2.719 2.486 2.329 2.214 2.126
2.056 1.999 1.951
81 3.959 3.109 2.717 2.484 2.327 2.213 2.125
2.055 1.998 1.950
82 3.957 3.108 2.716 2.483 2.326 2.211 2.123
2.053 1.996 1.948
83 3.956 3.107 2.715 2.482 2.324 2.210 2.122
2.052 1.995 1.947
84 3.955 3.105 2.713 2.480 2.323 2.209 2.121
2.051 1.993 1.945

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (5 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

85 3.953 3.104 2.712 2.479 2.322 2.207 2.119

2.049 1.992 1.944
86 3.952 3.103 2.711 2.478 2.321 2.206 2.118
2.048 1.991 1.943
87 3.951 3.101 2.709 2.476 2.319 2.205 2.117
2.047 1.989 1.941
88 3.949 3.100 2.708 2.475 2.318 2.203 2.115
2.045 1.988 1.940
89 3.948 3.099 2.707 2.474 2.317 2.202 2.114
2.044 1.987 1.939
90 3.947 3.098 2.706 2.473 2.316 2.201 2.113
2.043 1.986 1.938
91 3.946 3.097 2.705 2.472 2.315 2.200 2.112
2.042 1.984 1.936
92 3.945 3.095 2.704 2.471 2.313 2.199 2.111
2.041 1.983 1.935
93 3.943 3.094 2.703 2.470 2.312 2.198 2.110
2.040 1.982 1.934
94 3.942 3.093 2.701 2.469 2.311 2.197 2.109
2.038 1.981 1.933
95 3.941 3.092 2.700 2.467 2.310 2.196 2.108
2.037 1.980 1.932
96 3.940 3.091 2.699 2.466 2.309 2.195 2.106
2.036 1.979 1.931
97 3.939 3.090 2.698 2.465 2.308 2.194 2.105
2.035 1.978 1.930
98 3.938 3.089 2.697 2.465 2.307 2.193 2.104
2.034 1.977 1.929
99 3.937 3.088 2.696 2.464 2.306 2.192 2.103
2.033 1.976 1.928
100 3.936 3.087 2.696 2.463 2.305 2.191 2.103
2.032 1.975 1.927

\ 11 12 13 14 15 16 17 18
19 20

1 242.983 243.906 244.690 245.364 245.950 246.464 246.918

247.323 247.686 248.013
2 19.405 19.413 19.419 19.424 19.429 19.433 19.437
19.440 19.443 19.446
3 8.763 8.745 8.729 8.715 8.703 8.692 8.683
8.675 8.667 8.660

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (6 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

4 5.936 5.912 5.891 5.873 5.858 5.844 5.832

5.821 5.811 5.803
5 4.704 4.678 4.655 4.636 4.619 4.604 4.590
4.579 4.568 4.558
6 4.027 4.000 3.976 3.956 3.938 3.922 3.908
3.896 3.884 3.874
7 3.603 3.575 3.550 3.529 3.511 3.494 3.480
3.467 3.455 3.445
8 3.313 3.284 3.259 3.237 3.218 3.202 3.187
3.173 3.161 3.150
9 3.102 3.073 3.048 3.025 3.006 2.989 2.974
2.960 2.948 2.936
10 2.943 2.913 2.887 2.865 2.845 2.828 2.812
2.798 2.785 2.774
11 2.818 2.788 2.761 2.739 2.719 2.701 2.685
2.671 2.658 2.646
12 2.717 2.687 2.660 2.637 2.617 2.599 2.583
2.568 2.555 2.544
13 2.635 2.604 2.577 2.554 2.533 2.515 2.499
2.484 2.471 2.459
14 2.565 2.534 2.507 2.484 2.463 2.445 2.428
2.413 2.400 2.388
15 2.507 2.475 2.448 2.424 2.403 2.385 2.368
2.353 2.340 2.328
16 2.456 2.425 2.397 2.373 2.352 2.333 2.317
2.302 2.288 2.276
17 2.413 2.381 2.353 2.329 2.308 2.289 2.272
2.257 2.243 2.230
18 2.374 2.342 2.314 2.290 2.269 2.250 2.233
2.217 2.203 2.191
19 2.340 2.308 2.280 2.256 2.234 2.215 2.198
2.182 2.168 2.155
20 2.310 2.278 2.250 2.225 2.203 2.184 2.167
2.151 2.137 2.124
21 2.283 2.250 2.222 2.197 2.176 2.156 2.139
2.123 2.109 2.096
22 2.259 2.226 2.198 2.173 2.151 2.131 2.114
2.098 2.084 2.071
23 2.236 2.204 2.175 2.150 2.128 2.109 2.091
2.075 2.061 2.048
24 2.216 2.183 2.155 2.130 2.108 2.088 2.070
2.054 2.040 2.027
25 2.198 2.165 2.136 2.111 2.089 2.069 2.051
2.035 2.021 2.007
26 2.181 2.148 2.119 2.094 2.072 2.052 2.034
2.018 2.003 1.990

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (7 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

27 2.166 2.132 2.103 2.078 2.056 2.036 2.018

2.002 1.987 1.974
28 2.151 2.118 2.089 2.064 2.041 2.021 2.003
1.987 1.972 1.959
29 2.138 2.104 2.075 2.050 2.027 2.007 1.989
1.973 1.958 1.945
30 2.126 2.092 2.063 2.037 2.015 1.995 1.976
1.960 1.945 1.932
31 2.114 2.080 2.051 2.026 2.003 1.983 1.965
1.948 1.933 1.920
32 2.103 2.070 2.040 2.015 1.992 1.972 1.953
1.937 1.922 1.908
33 2.093 2.060 2.030 2.004 1.982 1.961 1.943
1.926 1.911 1.898
34 2.084 2.050 2.021 1.995 1.972 1.952 1.933
1.917 1.902 1.888
35 2.075 2.041 2.012 1.986 1.963 1.942 1.924
1.907 1.892 1.878
36 2.067 2.033 2.003 1.977 1.954 1.934 1.915
1.899 1.883 1.870
37 2.059 2.025 1.995 1.969 1.946 1.926 1.907
1.890 1.875 1.861
38 2.051 2.017 1.988 1.962 1.939 1.918 1.899
1.883 1.867 1.853
39 2.044 2.010 1.981 1.954 1.931 1.911 1.892
1.875 1.860 1.846
40 2.038 2.003 1.974 1.948 1.924 1.904 1.885
1.868 1.853 1.839
41 2.031 1.997 1.967 1.941 1.918 1.897 1.879
1.862 1.846 1.832
42 2.025 1.991 1.961 1.935 1.912 1.891 1.872
1.855 1.840 1.826
43 2.020 1.985 1.955 1.929 1.906 1.885 1.866
1.849 1.834 1.820
44 2.014 1.980 1.950 1.924 1.900 1.879 1.861
1.844 1.828 1.814
45 2.009 1.974 1.945 1.918 1.895 1.874 1.855
1.838 1.823 1.808
46 2.004 1.969 1.940 1.913 1.890 1.869 1.850
1.833 1.817 1.803
47 1.999 1.965 1.935 1.908 1.885 1.864 1.845
1.828 1.812 1.798
48 1.995 1.960 1.930 1.904 1.880 1.859 1.840
1.823 1.807 1.793
49 1.990 1.956 1.926 1.899 1.876 1.855 1.836
1.819 1.803 1.789

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (8 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

50 1.986 1.952 1.921 1.895 1.871 1.850 1.831

1.814 1.798 1.784
51 1.982 1.947 1.917 1.891 1.867 1.846 1.827
1.810 1.794 1.780
52 1.978 1.944 1.913 1.887 1.863 1.842 1.823
1.806 1.790 1.776
53 1.975 1.940 1.910 1.883 1.859 1.838 1.819
1.802 1.786 1.772
54 1.971 1.936 1.906 1.879 1.856 1.835 1.816
1.798 1.782 1.768
55 1.968 1.933 1.903 1.876 1.852 1.831 1.812
1.795 1.779 1.764
56 1.964 1.930 1.899 1.873 1.849 1.828 1.809
1.791 1.775 1.761
57 1.961 1.926 1.896 1.869 1.846 1.824 1.805
1.788 1.772 1.757
58 1.958 1.923 1.893 1.866 1.842 1.821 1.802
1.785 1.769 1.754
59 1.955 1.920 1.890 1.863 1.839 1.818 1.799
1.781 1.766 1.751
60 1.952 1.917 1.887 1.860 1.836 1.815 1.796
1.778 1.763 1.748
61 1.949 1.915 1.884 1.857 1.834 1.812 1.793
1.776 1.760 1.745
62 1.947 1.912 1.882 1.855 1.831 1.809 1.790
1.773 1.757 1.742
63 1.944 1.909 1.879 1.852 1.828 1.807 1.787
1.770 1.754 1.739
64 1.942 1.907 1.876 1.849 1.826 1.804 1.785
1.767 1.751 1.737
65 1.939 1.904 1.874 1.847 1.823 1.802 1.782
1.765 1.749 1.734
66 1.937 1.902 1.871 1.845 1.821 1.799 1.780
1.762 1.746 1.732
67 1.935 1.900 1.869 1.842 1.818 1.797 1.777
1.760 1.744 1.729
68 1.932 1.897 1.867 1.840 1.816 1.795 1.775
1.758 1.742 1.727
69 1.930 1.895 1.865 1.838 1.814 1.792 1.773
1.755 1.739 1.725
70 1.928 1.893 1.863 1.836 1.812 1.790 1.771
1.753 1.737 1.722
71 1.926 1.891 1.861 1.834 1.810 1.788 1.769
1.751 1.735 1.720
72 1.924 1.889 1.859 1.832 1.808 1.786 1.767
1.749 1.733 1.718

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (9 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

73 1.922 1.887 1.857 1.830 1.806 1.784 1.765

1.747 1.731 1.716
74 1.921 1.885 1.855 1.828 1.804 1.782 1.763
1.745 1.729 1.714
75 1.919 1.884 1.853 1.826 1.802 1.780 1.761
1.743 1.727 1.712
76 1.917 1.882 1.851 1.824 1.800 1.778 1.759
1.741 1.725 1.710
77 1.915 1.880 1.849 1.822 1.798 1.777 1.757
1.739 1.723 1.708
78 1.914 1.878 1.848 1.821 1.797 1.775 1.755
1.738 1.721 1.707
79 1.912 1.877 1.846 1.819 1.795 1.773 1.754
1.736 1.720 1.705
80 1.910 1.875 1.845 1.817 1.793 1.772 1.752
1.734 1.718 1.703
81 1.909 1.874 1.843 1.816 1.792 1.770 1.750
1.733 1.716 1.702
82 1.907 1.872 1.841 1.814 1.790 1.768 1.749
1.731 1.715 1.700
83 1.906 1.871 1.840 1.813 1.789 1.767 1.747
1.729 1.713 1.698
84 1.905 1.869 1.838 1.811 1.787 1.765 1.746
1.728 1.712 1.697
85 1.903 1.868 1.837 1.810 1.786 1.764 1.744
1.726 1.710 1.695
86 1.902 1.867 1.836 1.808 1.784 1.762 1.743
1.725 1.709 1.694
87 1.900 1.865 1.834 1.807 1.783 1.761 1.741
1.724 1.707 1.692
88 1.899 1.864 1.833 1.806 1.782 1.760 1.740
1.722 1.706 1.691
89 1.898 1.863 1.832 1.804 1.780 1.758 1.739
1.721 1.705 1.690
90 1.897 1.861 1.830 1.803 1.779 1.757 1.737
1.720 1.703 1.688
91 1.895 1.860 1.829 1.802 1.778 1.756 1.736
1.718 1.702 1.687
92 1.894 1.859 1.828 1.801 1.776 1.755 1.735
1.717 1.701 1.686
93 1.893 1.858 1.827 1.800 1.775 1.753 1.734
1.716 1.699 1.684
94 1.892 1.857 1.826 1.798 1.774 1.752 1.733
1.715 1.698 1.683
95 1.891 1.856 1.825 1.797 1.773 1.751 1.731
1.713 1.697 1.682

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (10 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

96 1.890 1.854 1.823 1.796 1.772 1.750 1.730

1.712 1.696 1.681
97 1.889 1.853 1.822 1.795 1.771 1.749 1.729
1.711 1.695 1.680
98 1.888 1.852 1.821 1.794 1.770 1.748 1.728
1.710 1.694 1.679
99 1.887 1.851 1.820 1.793 1.769 1.747 1.727
1.709 1.693 1.678
100 1.886 1.850 1.819 1.792 1.768 1.746 1.726
1.708 1.691 1.676

Upper critical values of the F distribution

for numerator degrees of freedom and denominator degrees of freedom

10% significance level

\ 1 2 3 4 5 6 7 8
9 10

1 39.863 49.500 53.593 55.833 57.240 58.204 58.906

59.439 59.858 60.195
2 8.526 9.000 9.162 9.243 9.293 9.326 9.349
9.367 9.381 9.392
3 5.538 5.462 5.391 5.343 5.309 5.285 5.266
5.252 5.240 5.230
4 4.545 4.325 4.191 4.107 4.051 4.010 3.979
3.955 3.936 3.920
5 4.060 3.780 3.619 3.520 3.453 3.405 3.368
3.339 3.316 3.297
6 3.776 3.463 3.289 3.181 3.108 3.055 3.014
2.983 2.958 2.937
7 3.589 3.257 3.074 2.961 2.883 2.827 2.785
2.752 2.725 2.703
8 3.458 3.113 2.924 2.806 2.726 2.668 2.624
2.589 2.561 2.538
9 3.360 3.006 2.813 2.693 2.611 2.551 2.505
2.469 2.440 2.416

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (11 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

10 3.285 2.924 2.728 2.605 2.522 2.461 2.414

2.377 2.347 2.323
11 3.225 2.860 2.660 2.536 2.451 2.389 2.342
2.304 2.274 2.248
12 3.177 2.807 2.606 2.480 2.394 2.331 2.283
2.245 2.214 2.188
13 3.136 2.763 2.560 2.434 2.347 2.283 2.234
2.195 2.164 2.138
14 3.102 2.726 2.522 2.395 2.307 2.243 2.193
2.154 2.122 2.095
15 3.073 2.695 2.490 2.361 2.273 2.208 2.158
2.119 2.086 2.059
16 3.048 2.668 2.462 2.333 2.244 2.178 2.128
2.088 2.055 2.028
17 3.026 2.645 2.437 2.308 2.218 2.152 2.102
2.061 2.028 2.001
18 3.007 2.624 2.416 2.286 2.196 2.130 2.079
2.038 2.005 1.977
19 2.990 2.606 2.397 2.266 2.176 2.109 2.058
2.017 1.984 1.956
20 2.975 2.589 2.380 2.249 2.158 2.091 2.040
1.999 1.965 1.937
21 2.961 2.575 2.365 2.233 2.142 2.075 2.023
1.982 1.948 1.920
22 2.949 2.561 2.351 2.219 2.128 2.060 2.008
1.967 1.933 1.904
23 2.937 2.549 2.339 2.207 2.115 2.047 1.995
1.953 1.919 1.890
24 2.927 2.538 2.327 2.195 2.103 2.035 1.983
1.941 1.906 1.877
25 2.918 2.528 2.317 2.184 2.092 2.024 1.971
1.929 1.895 1.866
26 2.909 2.519 2.307 2.174 2.082 2.014 1.961
1.919 1.884 1.855
27 2.901 2.511 2.299 2.165 2.073 2.005 1.952
1.909 1.874 1.845
28 2.894 2.503 2.291 2.157 2.064 1.996 1.943
1.900 1.865 1.836
29 2.887 2.495 2.283 2.149 2.057 1.988 1.935
1.892 1.857 1.827
30 2.881 2.489 2.276 2.142 2.049 1.980 1.927
1.884 1.849 1.819
31 2.875 2.482 2.270 2.136 2.042 1.973 1.920
1.877 1.842 1.812
32 2.869 2.477 2.263 2.129 2.036 1.967 1.913
1.870 1.835 1.805

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (12 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

33 2.864 2.471 2.258 2.123 2.030 1.961 1.907

1.864 1.828 1.799
34 2.859 2.466 2.252 2.118 2.024 1.955 1.901
1.858 1.822 1.793
35 2.855 2.461 2.247 2.113 2.019 1.950 1.896
1.852 1.817 1.787
36 2.850 2.456 2.243 2.108 2.014 1.945 1.891
1.847 1.811 1.781
37 2.846 2.452 2.238 2.103 2.009 1.940 1.886
1.842 1.806 1.776
38 2.842 2.448 2.234 2.099 2.005 1.935 1.881
1.838 1.802 1.772
39 2.839 2.444 2.230 2.095 2.001 1.931 1.877
1.833 1.797 1.767
40 2.835 2.440 2.226 2.091 1.997 1.927 1.873
1.829 1.793 1.763
41 2.832 2.437 2.222 2.087 1.993 1.923 1.869
1.825 1.789 1.759
42 2.829 2.434 2.219 2.084 1.989 1.919 1.865
1.821 1.785 1.755
43 2.826 2.430 2.216 2.080 1.986 1.916 1.861
1.817 1.781 1.751
44 2.823 2.427 2.213 2.077 1.983 1.913 1.858
1.814 1.778 1.747
45 2.820 2.425 2.210 2.074 1.980 1.909 1.855
1.811 1.774 1.744
46 2.818 2.422 2.207 2.071 1.977 1.906 1.852
1.808 1.771 1.741
47 2.815 2.419 2.204 2.068 1.974 1.903 1.849
1.805 1.768 1.738
48 2.813 2.417 2.202 2.066 1.971 1.901 1.846
1.802 1.765 1.735
49 2.811 2.414 2.199 2.063 1.968 1.898 1.843
1.799 1.763 1.732
50 2.809 2.412 2.197 2.061 1.966 1.895 1.840
1.796 1.760 1.729
51 2.807 2.410 2.194 2.058 1.964 1.893 1.838
1.794 1.757 1.727
52 2.805 2.408 2.192 2.056 1.961 1.891 1.836
1.791 1.755 1.724
53 2.803 2.406 2.190 2.054 1.959 1.888 1.833
1.789 1.752 1.722
54 2.801 2.404 2.188 2.052 1.957 1.886 1.831
1.787 1.750 1.719
55 2.799 2.402 2.186 2.050 1.955 1.884 1.829
1.785 1.748 1.717

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (13 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

56 2.797 2.400 2.184 2.048 1.953 1.882 1.827

1.782 1.746 1.715
57 2.796 2.398 2.182 2.046 1.951 1.880 1.825
1.780 1.744 1.713
58 2.794 2.396 2.181 2.044 1.949 1.878 1.823
1.779 1.742 1.711
59 2.793 2.395 2.179 2.043 1.947 1.876 1.821
1.777 1.740 1.709
60 2.791 2.393 2.177 2.041 1.946 1.875 1.819
1.775 1.738 1.707
61 2.790 2.392 2.176 2.039 1.944 1.873 1.818
1.773 1.736 1.705
62 2.788 2.390 2.174 2.038 1.942 1.871 1.816
1.771 1.735 1.703
63 2.787 2.389 2.173 2.036 1.941 1.870 1.814
1.770 1.733 1.702
64 2.786 2.387 2.171 2.035 1.939 1.868 1.813
1.768 1.731 1.700
65 2.784 2.386 2.170 2.033 1.938 1.867 1.811
1.767 1.730 1.699
66 2.783 2.385 2.169 2.032 1.937 1.865 1.810
1.765 1.728 1.697
67 2.782 2.384 2.167 2.031 1.935 1.864 1.808
1.764 1.727 1.696
68 2.781 2.382 2.166 2.029 1.934 1.863 1.807
1.762 1.725 1.694
69 2.780 2.381 2.165 2.028 1.933 1.861 1.806
1.761 1.724 1.693
70 2.779 2.380 2.164 2.027 1.931 1.860 1.804
1.760 1.723 1.691
71 2.778 2.379 2.163 2.026 1.930 1.859 1.803
1.758 1.721 1.690
72 2.777 2.378 2.161 2.025 1.929 1.858 1.802
1.757 1.720 1.689
73 2.776 2.377 2.160 2.024 1.928 1.856 1.801
1.756 1.719 1.687
74 2.775 2.376 2.159 2.022 1.927 1.855 1.800
1.755 1.718 1.686
75 2.774 2.375 2.158 2.021 1.926 1.854 1.798
1.754 1.716 1.685
76 2.773 2.374 2.157 2.020 1.925 1.853 1.797
1.752 1.715 1.684
77 2.772 2.373 2.156 2.019 1.924 1.852 1.796
1.751 1.714 1.683
78 2.771 2.372 2.155 2.018 1.923 1.851 1.795
1.750 1.713 1.682

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (14 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

79 2.770 2.371 2.154 2.017 1.922 1.850 1.794

1.749 1.712 1.681
80 2.769 2.370 2.154 2.016 1.921 1.849 1.793
1.748 1.711 1.680
81 2.769 2.369 2.153 2.016 1.920 1.848 1.792
1.747 1.710 1.679
82 2.768 2.368 2.152 2.015 1.919 1.847 1.791
1.746 1.709 1.678
83 2.767 2.368 2.151 2.014 1.918 1.846 1.790
1.745 1.708 1.677
84 2.766 2.367 2.150 2.013 1.917 1.845 1.790
1.744 1.707 1.676
85 2.765 2.366 2.149 2.012 1.916 1.845 1.789
1.744 1.706 1.675
86 2.765 2.365 2.149 2.011 1.915 1.844 1.788
1.743 1.705 1.674
87 2.764 2.365 2.148 2.011 1.915 1.843 1.787
1.742 1.705 1.673
88 2.763 2.364 2.147 2.010 1.914 1.842 1.786
1.741 1.704 1.672
89 2.763 2.363 2.146 2.009 1.913 1.841 1.785
1.740 1.703 1.671
90 2.762 2.363 2.146 2.008 1.912 1.841 1.785
1.739 1.702 1.670
91 2.761 2.362 2.145 2.008 1.912 1.840 1.784
1.739 1.701 1.670
92 2.761 2.361 2.144 2.007 1.911 1.839 1.783
1.738 1.701 1.669
93 2.760 2.361 2.144 2.006 1.910 1.838 1.782
1.737 1.700 1.668
94 2.760 2.360 2.143 2.006 1.910 1.838 1.782
1.736 1.699 1.667
95 2.759 2.359 2.142 2.005 1.909 1.837 1.781
1.736 1.698 1.667
96 2.759 2.359 2.142 2.004 1.908 1.836 1.780
1.735 1.698 1.666
97 2.758 2.358 2.141 2.004 1.908 1.836 1.780
1.734 1.697 1.665
98 2.757 2.358 2.141 2.003 1.907 1.835 1.779
1.734 1.696 1.665
99 2.757 2.357 2.140 2.003 1.906 1.835 1.778
1.733 1.696 1.664
100 2.756 2.356 2.139 2.002 1.906 1.834 1.778
1.732 1.695 1.663

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (15 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

\ 11 12 13 14 15 16 17 18
19 20

1 60.473 60.705 60.903 61.073 61.220 61.350 61.464

61.566 61.658 61.740
2 9.401 9.408 9.415 9.420 9.425 9.429 9.433
9.436 9.439 9.441
3 5.222 5.216 5.210 5.205 5.200 5.196 5.193
5.190 5.187 5.184
4 3.907 3.896 3.886 3.878 3.870 3.864 3.858
3.853 3.849 3.844
5 3.282 3.268 3.257 3.247 3.238 3.230 3.223
3.217 3.212 3.207
6 2.920 2.905 2.892 2.881 2.871 2.863 2.855
2.848 2.842 2.836
7 2.684 2.668 2.654 2.643 2.632 2.623 2.615
2.607 2.601 2.595
8 2.519 2.502 2.488 2.475 2.464 2.455 2.446
2.438 2.431 2.425
9 2.396 2.379 2.364 2.351 2.340 2.329 2.320
2.312 2.305 2.298
10 2.302 2.284 2.269 2.255 2.244 2.233 2.224
2.215 2.208 2.201
11 2.227 2.209 2.193 2.179 2.167 2.156 2.147
2.138 2.130 2.123
12 2.166 2.147 2.131 2.117 2.105 2.094 2.084
2.075 2.067 2.060
13 2.116 2.097 2.080 2.066 2.053 2.042 2.032
2.023 2.014 2.007
14 2.073 2.054 2.037 2.022 2.010 1.998 1.988
1.978 1.970 1.962
15 2.037 2.017 2.000 1.985 1.972 1.961 1.950
1.941 1.932 1.924
16 2.005 1.985 1.968 1.953 1.940 1.928 1.917
1.908 1.899 1.891
17 1.978 1.958 1.940 1.925 1.912 1.900 1.889
1.879 1.870 1.862
18 1.954 1.933 1.916 1.900 1.887 1.875 1.864
1.854 1.845 1.837
19 1.932 1.912 1.894 1.878 1.865 1.852 1.841
1.831 1.822 1.814
20 1.913 1.892 1.875 1.859 1.845 1.833 1.821
1.811 1.802 1.794

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (16 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

21 1.896 1.875 1.857 1.841 1.827 1.815 1.803

1.793 1.784 1.776
22 1.880 1.859 1.841 1.825 1.811 1.798 1.787
1.777 1.768 1.759
23 1.866 1.845 1.827 1.811 1.796 1.784 1.772
1.762 1.753 1.744
24 1.853 1.832 1.814 1.797 1.783 1.770 1.759
1.748 1.739 1.730
25 1.841 1.820 1.802 1.785 1.771 1.758 1.746
1.736 1.726 1.718
26 1.830 1.809 1.790 1.774 1.760 1.747 1.735
1.724 1.715 1.706
27 1.820 1.799 1.780 1.764 1.749 1.736 1.724
1.714 1.704 1.695
28 1.811 1.790 1.771 1.754 1.740 1.726 1.715
1.704 1.694 1.685
29 1.802 1.781 1.762 1.745 1.731 1.717 1.705
1.695 1.685 1.676
30 1.794 1.773 1.754 1.737 1.722 1.709 1.697
1.686 1.676 1.667
31 1.787 1.765 1.746 1.729 1.714 1.701 1.689
1.678 1.668 1.659
32 1.780 1.758 1.739 1.722 1.707 1.694 1.682
1.671 1.661 1.652
33 1.773 1.751 1.732 1.715 1.700 1.687 1.675
1.664 1.654 1.645
34 1.767 1.745 1.726 1.709 1.694 1.680 1.668
1.657 1.647 1.638
35 1.761 1.739 1.720 1.703 1.688 1.674 1.662
1.651 1.641 1.632
36 1.756 1.734 1.715 1.697 1.682 1.669 1.656
1.645 1.635 1.626
37 1.751 1.729 1.709 1.692 1.677 1.663 1.651
1.640 1.630 1.620
38 1.746 1.724 1.704 1.687 1.672 1.658 1.646
1.635 1.624 1.615
39 1.741 1.719 1.700 1.682 1.667 1.653 1.641
1.630 1.619 1.610
40 1.737 1.715 1.695 1.678 1.662 1.649 1.636
1.625 1.615 1.605
41 1.733 1.710 1.691 1.673 1.658 1.644 1.632
1.620 1.610 1.601
42 1.729 1.706 1.687 1.669 1.654 1.640 1.628
1.616 1.606 1.596
43 1.725 1.703 1.683 1.665 1.650 1.636 1.624
1.612 1.602 1.592

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (17 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

44 1.721 1.699 1.679 1.662 1.646 1.632 1.620

1.608 1.598 1.588
45 1.718 1.695 1.676 1.658 1.643 1.629 1.616
1.605 1.594 1.585
46 1.715 1.692 1.672 1.655 1.639 1.625 1.613
1.601 1.591 1.581
47 1.712 1.689 1.669 1.652 1.636 1.622 1.609
1.598 1.587 1.578
48 1.709 1.686 1.666 1.648 1.633 1.619 1.606
1.594 1.584 1.574
49 1.706 1.683 1.663 1.645 1.630 1.616 1.603
1.591 1.581 1.571
50 1.703 1.680 1.660 1.643 1.627 1.613 1.600
1.588 1.578 1.568
51 1.700 1.677 1.658 1.640 1.624 1.610 1.597
1.586 1.575 1.565
52 1.698 1.675 1.655 1.637 1.621 1.607 1.594
1.583 1.572 1.562
53 1.695 1.672 1.652 1.635 1.619 1.605 1.592
1.580 1.570 1.560
54 1.693 1.670 1.650 1.632 1.616 1.602 1.589
1.578 1.567 1.557
55 1.691 1.668 1.648 1.630 1.614 1.600 1.587
1.575 1.564 1.555
56 1.688 1.666 1.645 1.628 1.612 1.597 1.585
1.573 1.562 1.552
57 1.686 1.663 1.643 1.625 1.610 1.595 1.582
1.571 1.560 1.550
58 1.684 1.661 1.641 1.623 1.607 1.593 1.580
1.568 1.558 1.548
59 1.682 1.659 1.639 1.621 1.605 1.591 1.578
1.566 1.555 1.546
60 1.680 1.657 1.637 1.619 1.603 1.589 1.576
1.564 1.553 1.543
61 1.679 1.656 1.635 1.617 1.601 1.587 1.574
1.562 1.551 1.541
62 1.677 1.654 1.634 1.616 1.600 1.585 1.572
1.560 1.549 1.540
63 1.675 1.652 1.632 1.614 1.598 1.583 1.570
1.558 1.548 1.538
64 1.673 1.650 1.630 1.612 1.596 1.582 1.569
1.557 1.546 1.536
65 1.672 1.649 1.628 1.610 1.594 1.580 1.567
1.555 1.544 1.534
66 1.670 1.647 1.627 1.609 1.593 1.578 1.565
1.553 1.542 1.532

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (18 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

67 1.669 1.646 1.625 1.607 1.591 1.577 1.564

1.552 1.541 1.531
68 1.667 1.644 1.624 1.606 1.590 1.575 1.562
1.550 1.539 1.529
69 1.666 1.643 1.622 1.604 1.588 1.574 1.560
1.548 1.538 1.527
70 1.665 1.641 1.621 1.603 1.587 1.572 1.559
1.547 1.536 1.526
71 1.663 1.640 1.619 1.601 1.585 1.571 1.557
1.545 1.535 1.524
72 1.662 1.639 1.618 1.600 1.584 1.569 1.556
1.544 1.533 1.523
73 1.661 1.637 1.617 1.599 1.583 1.568 1.555
1.543 1.532 1.522
74 1.659 1.636 1.616 1.597 1.581 1.567 1.553
1.541 1.530 1.520
75 1.658 1.635 1.614 1.596 1.580 1.565 1.552
1.540 1.529 1.519
76 1.657 1.634 1.613 1.595 1.579 1.564 1.551
1.539 1.528 1.518
77 1.656 1.632 1.612 1.594 1.578 1.563 1.550
1.538 1.527 1.516
78 1.655 1.631 1.611 1.593 1.576 1.562 1.548
1.536 1.525 1.515
79 1.654 1.630 1.610 1.592 1.575 1.561 1.547
1.535 1.524 1.514
80 1.653 1.629 1.609 1.590 1.574 1.559 1.546
1.534 1.523 1.513
81 1.652 1.628 1.608 1.589 1.573 1.558 1.545
1.533 1.522 1.512
82 1.651 1.627 1.607 1.588 1.572 1.557 1.544
1.532 1.521 1.511
83 1.650 1.626 1.606 1.587 1.571 1.556 1.543
1.531 1.520 1.509
84 1.649 1.625 1.605 1.586 1.570 1.555 1.542
1.530 1.519 1.508
85 1.648 1.624 1.604 1.585 1.569 1.554 1.541
1.529 1.518 1.507
86 1.647 1.623 1.603 1.584 1.568 1.553 1.540
1.528 1.517 1.506
87 1.646 1.622 1.602 1.583 1.567 1.552 1.539
1.527 1.516 1.505
88 1.645 1.622 1.601 1.583 1.566 1.551 1.538
1.526 1.515 1.504
89 1.644 1.621 1.600 1.582 1.565 1.550 1.537
1.525 1.514 1.503

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (19 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

90 1.643 1.620 1.599 1.581 1.564 1.550 1.536

1.524 1.513 1.503
91 1.643 1.619 1.598 1.580 1.564 1.549 1.535
1.523 1.512 1.502
92 1.642 1.618 1.598 1.579 1.563 1.548 1.534
1.522 1.511 1.501
93 1.641 1.617 1.597 1.578 1.562 1.547 1.534
1.521 1.510 1.500
94 1.640 1.617 1.596 1.578 1.561 1.546 1.533
1.521 1.509 1.499
95 1.640 1.616 1.595 1.577 1.560 1.545 1.532
1.520 1.509 1.498
96 1.639 1.615 1.594 1.576 1.560 1.545 1.531
1.519 1.508 1.497
97 1.638 1.614 1.594 1.575 1.559 1.544 1.530
1.518 1.507 1.497
98 1.637 1.614 1.593 1.575 1.558 1.543 1.530
1.517 1.506 1.496
99 1.637 1.613 1.592 1.574 1.557 1.542 1.529
1.517 1.505 1.495
100 1.636 1.612 1.592 1.573 1.557 1.542 1.528
1.516 1.505 1.494

Upper critical values of the F distribution

for numerator degrees of freedom and denominator degrees of freedom

1% significance level

\ 1 2 3 4 5 6 7 8
9 10

1 4052.19 4999.52 5403.34 5624.62 5763.65 5858.97 5928.33

5981.10 6022.50 6055.85
2 98.502 99.000 99.166 99.249 99.300 99.333 99.356
99.374 99.388 99.399
3 34.116 30.816 29.457 28.710 28.237 27.911 27.672

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (20 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

27.489 27.345 27.229

4 21.198 18.000 16.694 15.977 15.522 15.207 14.976
14.799 14.659 14.546
5 16.258 13.274 12.060 11.392 10.967 10.672 10.456
10.289 10.158 10.051
6 13.745 10.925 9.780 9.148 8.746 8.466 8.260
8.102 7.976 7.874
7 12.246 9.547 8.451 7.847 7.460 7.191 6.993
6.840 6.719 6.620
8 11.259 8.649 7.591 7.006 6.632 6.371 6.178
6.029 5.911 5.814
9 10.561 8.022 6.992 6.422 6.057 5.802 5.613
5.467 5.351 5.257
10 10.044 7.559 6.552 5.994 5.636 5.386 5.200
5.057 4.942 4.849
11 9.646 7.206 6.217 5.668 5.316 5.069 4.886
4.744 4.632 4.539
12 9.330 6.927 5.953 5.412 5.064 4.821 4.640
4.499 4.388 4.296
13 9.074 6.701 5.739 5.205 4.862 4.620 4.441
4.302 4.191 4.100
14 8.862 6.515 5.564 5.035 4.695 4.456 4.278
4.140 4.030 3.939
15 8.683 6.359 5.417 4.893 4.556 4.318 4.142
4.004 3.895 3.805
16 8.531 6.226 5.292 4.773 4.437 4.202 4.026
3.890 3.780 3.691
17 8.400 6.112 5.185 4.669 4.336 4.102 3.927
3.791 3.682 3.593
18 8.285 6.013 5.092 4.579 4.248 4.015 3.841
3.705 3.597 3.508
19 8.185 5.926 5.010 4.500 4.171 3.939 3.765
3.631 3.523 3.434
20 8.096 5.849 4.938 4.431 4.103 3.871 3.699
3.564 3.457 3.368
21 8.017 5.780 4.874 4.369 4.042 3.812 3.640
3.506 3.398 3.310
22 7.945 5.719 4.817 4.313 3.988 3.758 3.587
3.453 3.346 3.258
23 7.881 5.664 4.765 4.264 3.939 3.710 3.539
3.406 3.299 3.211
24 7.823 5.614 4.718 4.218 3.895 3.667 3.496
3.363 3.256 3.168
25 7.770 5.568 4.675 4.177 3.855 3.627 3.457
3.324 3.217 3.129
26 7.721 5.526 4.637 4.140 3.818 3.591 3.421

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (21 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

3.288 3.182 3.094

27 7.677 5.488 4.601 4.106 3.785 3.558 3.388
3.256 3.149 3.062
28 7.636 5.453 4.568 4.074 3.754 3.528 3.358
3.226 3.120 3.032
29 7.598 5.420 4.538 4.045 3.725 3.499 3.330
3.198 3.092 3.005
30 7.562 5.390 4.510 4.018 3.699 3.473 3.305
3.173 3.067 2.979
31 7.530 5.362 4.484 3.993 3.675 3.449 3.281
3.149 3.043 2.955
32 7.499 5.336 4.459 3.969 3.652 3.427 3.258
3.127 3.021 2.934
33 7.471 5.312 4.437 3.948 3.630 3.406 3.238
3.106 3.000 2.913
34 7.444 5.289 4.416 3.927 3.611 3.386 3.218
3.087 2.981 2.894
35 7.419 5.268 4.396 3.908 3.592 3.368 3.200
3.069 2.963 2.876
36 7.396 5.248 4.377 3.890 3.574 3.351 3.183
3.052 2.946 2.859
37 7.373 5.229 4.360 3.873 3.558 3.334 3.167
3.036 2.930 2.843
38 7.353 5.211 4.343 3.858 3.542 3.319 3.152
3.021 2.915 2.828
39 7.333 5.194 4.327 3.843 3.528 3.305 3.137
3.006 2.901 2.814
40 7.314 5.179 4.313 3.828 3.514 3.291 3.124
2.993 2.888 2.801
41 7.296 5.163 4.299 3.815 3.501 3.278 3.111
2.980 2.875 2.788
42 7.280 5.149 4.285 3.802 3.488 3.266 3.099
2.968 2.863 2.776
43 7.264 5.136 4.273 3.790 3.476 3.254 3.087
2.957 2.851 2.764
44 7.248 5.123 4.261 3.778 3.465 3.243 3.076
2.946 2.840 2.754
45 7.234 5.110 4.249 3.767 3.454 3.232 3.066
2.935 2.830 2.743
46 7.220 5.099 4.238 3.757 3.444 3.222 3.056
2.925 2.820 2.733
47 7.207 5.087 4.228 3.747 3.434 3.213 3.046
2.916 2.811 2.724
48 7.194 5.077 4.218 3.737 3.425 3.204 3.037
2.907 2.802 2.715
49 7.182 5.066 4.208 3.728 3.416 3.195 3.028

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (22 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

2.898 2.793 2.706

50 7.171 5.057 4.199 3.720 3.408 3.186 3.020
2.890 2.785 2.698
51 7.159 5.047 4.191 3.711 3.400 3.178 3.012
2.882 2.777 2.690
52 7.149 5.038 4.182 3.703 3.392 3.171 3.005
2.874 2.769 2.683
53 7.139 5.030 4.174 3.695 3.384 3.163 2.997
2.867 2.762 2.675
54 7.129 5.021 4.167 3.688 3.377 3.156 2.990
2.860 2.755 2.668
55 7.119 5.013 4.159 3.681 3.370 3.149 2.983
2.853 2.748 2.662
56 7.110 5.006 4.152 3.674 3.363 3.143 2.977
2.847 2.742 2.655
57 7.102 4.998 4.145 3.667 3.357 3.136 2.971
2.841 2.736 2.649
58 7.093 4.991 4.138 3.661 3.351 3.130 2.965
2.835 2.730 2.643
59 7.085 4.984 4.132 3.655 3.345 3.124 2.959
2.829 2.724 2.637
60 7.077 4.977 4.126 3.649 3.339 3.119 2.953
2.823 2.718 2.632
61 7.070 4.971 4.120 3.643 3.333 3.113 2.948
2.818 2.713 2.626
62 7.062 4.965 4.114 3.638 3.328 3.108 2.942
2.813 2.708 2.621
63 7.055 4.959 4.109 3.632 3.323 3.103 2.937
2.808 2.703 2.616
64 7.048 4.953 4.103 3.627 3.318 3.098 2.932
2.803 2.698 2.611
65 7.042 4.947 4.098 3.622 3.313 3.093 2.928
2.798 2.693 2.607
66 7.035 4.942 4.093 3.618 3.308 3.088 2.923
2.793 2.689 2.602
67 7.029 4.937 4.088 3.613 3.304 3.084 2.919
2.789 2.684 2.598
68 7.023 4.932 4.083 3.608 3.299 3.080 2.914
2.785 2.680 2.593
69 7.017 4.927 4.079 3.604 3.295 3.075 2.910
2.781 2.676 2.589
70 7.011 4.922 4.074 3.600 3.291 3.071 2.906
2.777 2.672 2.585
71 7.006 4.917 4.070 3.596 3.287 3.067 2.902
2.773 2.668 2.581
72 7.001 4.913 4.066 3.591 3.283 3.063 2.898

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (23 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

2.769 2.664 2.578

73 6.995 4.908 4.062 3.588 3.279 3.060 2.895
2.765 2.660 2.574
74 6.990 4.904 4.058 3.584 3.275 3.056 2.891
2.762 2.657 2.570
75 6.985 4.900 4.054 3.580 3.272 3.052 2.887
2.758 2.653 2.567
76 6.981 4.896 4.050 3.577 3.268 3.049 2.884
2.755 2.650 2.563
77 6.976 4.892 4.047 3.573 3.265 3.046 2.881
2.751 2.647 2.560
78 6.971 4.888 4.043 3.570 3.261 3.042 2.877
2.748 2.644 2.557
79 6.967 4.884 4.040 3.566 3.258 3.039 2.874
2.745 2.640 2.554
80 6.963 4.881 4.036 3.563 3.255 3.036 2.871
2.742 2.637 2.551
81 6.958 4.877 4.033 3.560 3.252 3.033 2.868
2.739 2.634 2.548
82 6.954 4.874 4.030 3.557 3.249 3.030 2.865
2.736 2.632 2.545
83 6.950 4.870 4.027 3.554 3.246 3.027 2.863
2.733 2.629 2.542
84 6.947 4.867 4.024 3.551 3.243 3.025 2.860
2.731 2.626 2.539
85 6.943 4.864 4.021 3.548 3.240 3.022 2.857
2.728 2.623 2.537
86 6.939 4.861 4.018 3.545 3.238 3.019 2.854
2.725 2.621 2.534
87 6.935 4.858 4.015 3.543 3.235 3.017 2.852
2.723 2.618 2.532
88 6.932 4.855 4.012 3.540 3.233 3.014 2.849
2.720 2.616 2.529
89 6.928 4.852 4.010 3.538 3.230 3.012 2.847
2.718 2.613 2.527
90 6.925 4.849 4.007 3.535 3.228 3.009 2.845
2.715 2.611 2.524
91 6.922 4.846 4.004 3.533 3.225 3.007 2.842
2.713 2.609 2.522
92 6.919 4.844 4.002 3.530 3.223 3.004 2.840
2.711 2.606 2.520
93 6.915 4.841 3.999 3.528 3.221 3.002 2.838
2.709 2.604 2.518
94 6.912 4.838 3.997 3.525 3.218 3.000 2.835
2.706 2.602 2.515
95 6.909 4.836 3.995 3.523 3.216 2.998 2.833

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (24 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

2.704 2.600 2.513

96 6.906 4.833 3.992 3.521 3.214 2.996 2.831
2.702 2.598 2.511
97 6.904 4.831 3.990 3.519 3.212 2.994 2.829
2.700 2.596 2.509
98 6.901 4.829 3.988 3.517 3.210 2.992 2.827
2.698 2.594 2.507
99 6.898 4.826 3.986 3.515 3.208 2.990 2.825
2.696 2.592 2.505
100 6.895 4.824 3.984 3.513 3.206 2.988 2.823
2.694 2.590 2.503

\ 11 12 13 14 15 16 17 18
19 20

1. 6083.35 6106.35 6125.86 6142.70 6157.28 6170.12 6181.42

6191.52 6200.58 6208.74
2. 99.408 99.416 99.422 99.428 99.432 99.437 99.440
99.444 99.447 99.449
3. 27.133 27.052 26.983 26.924 26.872 26.827 26.787
26.751 26.719 26.690
4. 14.452 14.374 14.307 14.249 14.198 14.154 14.115
14.080 14.048 14.020
5. 9.963 9.888 9.825 9.770 9.722 9.680 9.643
9.610 9.580 9.553
6. 7.790 7.718 7.657 7.605 7.559 7.519 7.483
7.451 7.422 7.396
7. 6.538 6.469 6.410 6.359 6.314 6.275 6.240
6.209 6.181 6.155
8. 5.734 5.667 5.609 5.559 5.515 5.477 5.442
5.412 5.384 5.359
9. 5.178 5.111 5.055 5.005 4.962 4.924 4.890
4.860 4.833 4.808
10. 4.772 4.706 4.650 4.601 4.558 4.520 4.487
4.457 4.430 4.405
11. 4.462 4.397 4.342 4.293 4.251 4.213 4.180
4.150 4.123 4.099
12. 4.220 4.155 4.100 4.052 4.010 3.972 3.939
3.909 3.883 3.858
13. 4.025 3.960 3.905 3.857 3.815 3.778 3.745
3.716 3.689 3.665
14. 3.864 3.800 3.745 3.698 3.656 3.619 3.586

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (25 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

3.556 3.529 3.505

15. 3.730 3.666 3.612 3.564 3.522 3.485 3.452
3.423 3.396 3.372
16. 3.616 3.553 3.498 3.451 3.409 3.372 3.339
3.310 3.283 3.259
17. 3.519 3.455 3.401 3.353 3.312 3.275 3.242
3.212 3.186 3.162
18. 3.434 3.371 3.316 3.269 3.227 3.190 3.158
3.128 3.101 3.077
19. 3.360 3.297 3.242 3.195 3.153 3.116 3.084
3.054 3.027 3.003
20. 3.294 3.231 3.177 3.130 3.088 3.051 3.018
2.989 2.962 2.938
21. 3.236 3.173 3.119 3.072 3.030 2.993 2.960
2.931 2.904 2.880
22. 3.184 3.121 3.067 3.019 2.978 2.941 2.908
2.879 2.852 2.827
23. 3.137 3.074 3.020 2.973 2.931 2.894 2.861
2.832 2.805 2.781
24. 3.094 3.032 2.977 2.930 2.889 2.852 2.819
2.789 2.762 2.738
25. 3.056 2.993 2.939 2.892 2.850 2.813 2.780
2.751 2.724 2.699
26. 3.021 2.958 2.904 2.857 2.815 2.778 2.745
2.715 2.688 2.664
27. 2.988 2.926 2.871 2.824 2.783 2.746 2.713
2.683 2.656 2.632
28. 2.959 2.896 2.842 2.795 2.753 2.716 2.683
2.653 2.626 2.602
29. 2.931 2.868 2.814 2.767 2.726 2.689 2.656
2.626 2.599 2.574
30. 2.906 2.843 2.789 2.742 2.700 2.663 2.630
2.600 2.573 2.549
31. 2.882 2.820 2.765 2.718 2.677 2.640 2.606
2.577 2.550 2.525
32. 2.860 2.798 2.744 2.696 2.655 2.618 2.584
2.555 2.527 2.503
33. 2.840 2.777 2.723 2.676 2.634 2.597 2.564
2.534 2.507 2.482
34. 2.821 2.758 2.704 2.657 2.615 2.578 2.545
2.515 2.488 2.463
35. 2.803 2.740 2.686 2.639 2.597 2.560 2.527
2.497 2.470 2.445
36. 2.786 2.723 2.669 2.622 2.580 2.543 2.510
2.480 2.453 2.428
37. 2.770 2.707 2.653 2.606 2.564 2.527 2.494

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (26 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

2.464 2.437 2.412

38. 2.755 2.692 2.638 2.591 2.549 2.512 2.479
2.449 2.421 2.397
39. 2.741 2.678 2.624 2.577 2.535 2.498 2.465
2.434 2.407 2.382
40. 2.727 2.665 2.611 2.563 2.522 2.484 2.451
2.421 2.394 2.369
41. 2.715 2.652 2.598 2.551 2.509 2.472 2.438
2.408 2.381 2.356
42. 2.703 2.640 2.586 2.539 2.497 2.460 2.426
2.396 2.369 2.344
43. 2.691 2.629 2.575 2.527 2.485 2.448 2.415
2.385 2.357 2.332
44. 2.680 2.618 2.564 2.516 2.475 2.437 2.404
2.374 2.346 2.321
45. 2.670 2.608 2.553 2.506 2.464 2.427 2.393
2.363 2.336 2.311
46. 2.660 2.598 2.544 2.496 2.454 2.417 2.384
2.353 2.326 2.301
47. 2.651 2.588 2.534 2.487 2.445 2.408 2.374
2.344 2.316 2.291
48. 2.642 2.579 2.525 2.478 2.436 2.399 2.365
2.335 2.307 2.282
49. 2.633 2.571 2.517 2.469 2.427 2.390 2.356
2.326 2.299 2.274
50. 2.625 2.562 2.508 2.461 2.419 2.382 2.348
2.318 2.290 2.265
51. 2.617 2.555 2.500 2.453 2.411 2.374 2.340
2.310 2.282 2.257
52. 2.610 2.547 2.493 2.445 2.403 2.366 2.333
2.302 2.275 2.250
53. 2.602 2.540 2.486 2.438 2.396 2.359 2.325
2.295 2.267 2.242
54. 2.595 2.533 2.479 2.431 2.389 2.352 2.318
2.288 2.260 2.235
55. 2.589 2.526 2.472 2.424 2.382 2.345 2.311
2.281 2.253 2.228
56. 2.582 2.520 2.465 2.418 2.376 2.339 2.305
2.275 2.247 2.222
57. 2.576 2.513 2.459 2.412 2.370 2.332 2.299
2.268 2.241 2.215
58. 2.570 2.507 2.453 2.406 2.364 2.326 2.293
2.262 2.235 2.209
59. 2.564 2.502 2.447 2.400 2.358 2.320 2.287
2.256 2.229 2.203
60. 2.559 2.496 2.442 2.394 2.352 2.315 2.281

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (27 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

2.251 2.223 2.198

61. 2.553 2.491 2.436 2.389 2.347 2.309 2.276
2.245 2.218 2.192
62. 2.548 2.486 2.431 2.384 2.342 2.304 2.270
2.240 2.212 2.187
63. 2.543 2.481 2.426 2.379 2.337 2.299 2.265
2.235 2.207 2.182
64. 2.538 2.476 2.421 2.374 2.332 2.294 2.260
2.230 2.202 2.177
65. 2.534 2.471 2.417 2.369 2.327 2.289 2.256
2.225 2.198 2.172
66. 2.529 2.466 2.412 2.365 2.322 2.285 2.251
2.221 2.193 2.168
67. 2.525 2.462 2.408 2.360 2.318 2.280 2.247
2.216 2.188 2.163
68. 2.520 2.458 2.403 2.356 2.314 2.276 2.242
2.212 2.184 2.159
69. 2.516 2.454 2.399 2.352 2.310 2.272 2.238
2.208 2.180 2.155
70. 2.512 2.450 2.395 2.348 2.306 2.268 2.234
2.204 2.176 2.150
71. 2.508 2.446 2.391 2.344 2.302 2.264 2.230
2.200 2.172 2.146
72. 2.504 2.442 2.388 2.340 2.298 2.260 2.226
2.196 2.168 2.143
73. 2.501 2.438 2.384 2.336 2.294 2.256 2.223
2.192 2.164 2.139
74. 2.497 2.435 2.380 2.333 2.290 2.253 2.219
2.188 2.161 2.135
75. 2.494 2.431 2.377 2.329 2.287 2.249 2.215
2.185 2.157 2.132
76. 2.490 2.428 2.373 2.326 2.284 2.246 2.212
2.181 2.154 2.128
77. 2.487 2.424 2.370 2.322 2.280 2.243 2.209
2.178 2.150 2.125
78. 2.484 2.421 2.367 2.319 2.277 2.239 2.206
2.175 2.147 2.122
79. 2.481 2.418 2.364 2.316 2.274 2.236 2.202
2.172 2.144 2.118
80. 2.478 2.415 2.361 2.313 2.271 2.233 2.199
2.169 2.141 2.115
81. 2.475 2.412 2.358 2.310 2.268 2.230 2.196
2.166 2.138 2.112
82. 2.472 2.409 2.355 2.307 2.265 2.227 2.193
2.163 2.135 2.109
83. 2.469 2.406 2.352 2.304 2.262 2.224 2.191

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (28 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.3. Upper Critical Values of the F Distribution

2.160 2.132 2.106

84. 2.466 2.404 2.349 2.302 2.259 2.222 2.188
2.157 2.129 2.104
85. 2.464 2.401 2.347 2.299 2.257 2.219 2.185
2.154 2.126 2.101
86. 2.461 2.398 2.344 2.296 2.254 2.216 2.182
2.152 2.124 2.098
87. 2.459 2.396 2.342 2.294 2.252 2.214 2.180
2.149 2.121 2.096
88. 2.456 2.393 2.339 2.291 2.249 2.211 2.177
2.147 2.119 2.093
89. 2.454 2.391 2.337 2.289 2.247 2.209 2.175
2.144 2.116 2.091
90. 2.451 2.389 2.334 2.286 2.244 2.206 2.172
2.142 2.114 2.088
91. 2.449 2.386 2.332 2.284 2.242 2.204 2.170
2.139 2.111 2.086
92. 2.447 2.384 2.330 2.282 2.240 2.202 2.168
2.137 2.109 2.083
93. 2.444 2.382 2.327 2.280 2.237 2.200 2.166
2.135 2.107 2.081
94. 2.442 2.380 2.325 2.277 2.235 2.197 2.163
2.133 2.105 2.079
95. 2.440 2.378 2.323 2.275 2.233 2.195 2.161
2.130 2.102 2.077
96. 2.438 2.375 2.321 2.273 2.231 2.193 2.159
2.128 2.100 2.075
97. 2.436 2.373 2.319 2.271 2.229 2.191 2.157
2.126 2.098 2.073
98. 2.434 2.371 2.317 2.269 2.227 2.189 2.155
2.124 2.096 2.071
99. 2.432 2.369 2.315 2.267 2.225 2.187 2.153
2.122 2.094 2.069
100. 2.430 2.368 2.313 2.265 2.223 2.185 2.151
2.120 2.092 2.067

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm (29 of 29) [5/7/2002 3:15:48 PM]

1.3.6.7.4. Critical Values of the Chi-Square Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.7. Tables for Probability Distributions

1.3.6.7.4. Critical Values of the Chi-Square

Distribution
How to Use This table contains the critical values of the chi-square distribution.
This Table Because of the lack of symmetry of the chi-square distribution, separate
tables are provided for the upper and lower tails of the distribution.
A test statistic with degrees of freedom is computed from the data. For
upper one-sided tests, the test statistic is compared with a value from the
table of upper critical values. For two-sided tests, the test statistic is
compared with values from both the table for the upper critical value
and the table for the lower critical value.
The significance level, , is demonstrated with the graph below which
shows a chi-square distribution with 3 degrees of freedom for a
two-sided test at significance level = 0.05. If the test statistic is
greater than the upper critical value or less than the lower critical value,
we reject the null hypothesis. Specific instructions are given below.

Given a specified value for :

1. For a two-sided test, find the column corresponding to /2 in the
table for upper critical values and reject the null hypothesis if the
test statistic is greater than the tabled value. Similarly, find the

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3674.htm (1 of 7) [5/7/2002 3:15:49 PM]

1.3.6.7.4. Critical Values of the Chi-Square Distribution

column corresponding to 1 - /2 in the table for lower critical

values and reject the null hypothesis if the test statistic is less than
the tabled value.
2. For an upper one-sided test, find the column corresponding to
in the upper critical values table and reject the null hypothesis if
the test statistic is greater than the tabled value.
3. For a lower one-sided test, find the column corresponding to 1 -
in the lower critical values table and reject the null hypothesis
if the computed test statistic is less than the tabled value.

Upper critical values of chi-square distribution with degrees of freedom

Probability of exceeding the critical value

0.10 0.05 0.025 0.01 0.001

1 2.706 3.841 5.024 6.635 10.828

2 4.605 5.991 7.378 9.210 13.816
3 6.251 7.815 9.348 11.345 16.266
4 7.779 9.488 11.143 13.277 18.467
5 9.236 11.070 12.833 15.086 20.515
6 10.645 12.592 14.449 16.812 22.458
7 12.017 14.067 16.013 18.475 24.322
8 13.362 15.507 17.535 20.090 26.125
9 14.684 16.919 19.023 21.666 27.877
10 15.987 18.307 20.483 23.209 29.588
11 17.275 19.675 21.920 24.725 31.264
12 18.549 21.026 23.337 26.217 32.910
13 19.812 22.362 24.736 27.688 34.528
14 21.064 23.685 26.119 29.141 36.123
15 22.307 24.996 27.488 30.578 37.697
16 23.542 26.296 28.845 32.000 39.252
17 24.769 27.587 30.191 33.409 40.790
18 25.989 28.869 31.526 34.805 42.312
19 27.204 30.144 32.852 36.191 43.820
20 28.412 31.410 34.170 37.566 45.315
21 29.615 32.671 35.479 38.932 46.797

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3674.htm (2 of 7) [5/7/2002 3:15:49 PM]

1.3.6.7.4. Critical Values of the Chi-Square Distribution

22 30.813 33.924 36.781 40.289 48.268

23 32.007 35.172 38.076 41.638 49.728
24 33.196 36.415 39.364 42.980 51.179
25 34.382 37.652 40.646 44.314 52.620
26 35.563 38.885 41.923 45.642 54.052
27 36.741 40.113 43.195 46.963 55.476
28 37.916 41.337 44.461 48.278 56.892
29 39.087 42.557 45.722 49.588 58.301
30 40.256 43.773 46.979 50.892 59.703
31 41.422 44.985 48.232 52.191 61.098
32 42.585 46.194 49.480 53.486 62.487
33 43.745 47.400 50.725 54.776 63.870
34 44.903 48.602 51.966 56.061 65.247
35 46.059 49.802 53.203 57.342 66.619
36 47.212 50.998 54.437 58.619 67.985
37 48.363 52.192 55.668 59.893 69.347
38 49.513 53.384 56.896 61.162 70.703
39 50.660 54.572 58.120 62.428 72.055
40 51.805 55.758 59.342 63.691 73.402
41 52.949 56.942 60.561 64.950 74.745
42 54.090 58.124 61.777 66.206 76.084
43 55.230 59.304 62.990 67.459 77.419
44 56.369 60.481 64.201 68.710 78.750
45 57.505 61.656 65.410 69.957 80.077
46 58.641 62.830 66.617 71.201 81.400
47 59.774 64.001 67.821 72.443 82.720
48 60.907 65.171 69.023 73.683 84.037
49 62.038 66.339 70.222 74.919 85.351
50 63.167 67.505 71.420 76.154 86.661
51 64.295 68.669 72.616 77.386 87.968
52 65.422 69.832 73.810 78.616 89.272
53 66.548 70.993 75.002 79.843 90.573
54 67.673 72.153 76.192 81.069 91.872
55 68.796 73.311 77.380 82.292 93.168
56 69.919 74.468 78.567 83.513 94.461
57 71.040 75.624 79.752 84.733 95.751
58 72.160 76.778 80.936 85.950 97.039
59 73.279 77.931 82.117 87.166 98.324
60 74.397 79.082 83.298 88.379 99.607
61 75.514 80.232 84.476 89.591 100.888
62 76.630 81.381 85.654 90.802 102.166

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3674.htm (3 of 7) [5/7/2002 3:15:49 PM]

1.3.6.7.4. Critical Values of the Chi-Square Distribution

63 77.745 82.529 86.830 92.010 103.442

64 78.860 83.675 88.004 93.217 104.716
65 79.973 84.821 89.177 94.422 105.988
66 81.085 85.965 90.349 95.626 107.258
67 82.197 87.108 91.519 96.828 108.526
68 83.308 88.250 92.689 98.028 109.791
69 84.418 89.391 93.856 99.228 111.055
70 85.527 90.531 95.023 100.425 112.317
71 86.635 91.670 96.189 101.621 113.577
72 87.743 92.808 97.353 102.816 114.835
73 88.850 93.945 98.516 104.010 116.092
74 89.956 95.081 99.678 105.202 117.346
75 91.061 96.217 100.839 106.393 118.599
76 92.166 97.351 101.999 107.583 119.850
77 93.270 98.484 103.158 108.771 121.100
78 94.374 99.617 104.316 109.958 122.348
79 95.476 100.749 105.473 111.144 123.594
80 96.578 101.879 106.629 112.329 124.839
81 97.680 103.010 107.783 113.512 126.083
82 98.780 104.139 108.937 114.695 127.324
83 99.880 105.267 110.090 115.876 128.565
84 100.980 106.395 111.242 117.057 129.804
85 102.079 107.522 112.393 118.236 131.041
86 103.177 108.648 113.544 119.414 132.277
87 104.275 109.773 114.693 120.591 133.512
88 105.372 110.898 115.841 121.767 134.746
89 106.469 112.022 116.989 122.942 135.978
90 107.565 113.145 118.136 124.116 137.208
91 108.661 114.268 119.282 125.289 138.438
92 109.756 115.390 120.427 126.462 139.666
93 110.850 116.511 121.571 127.633 140.893
94 111.944 117.632 122.715 128.803 142.119
95 113.038 118.752 123.858 129.973 143.344
96 114.131 119.871 125.000 131.141 144.567
97 115.223 120.990 126.141 132.309 145.789
98 116.315 122.108 127.282 133.476 147.010
99 117.407 123.225 128.422 134.642 148.230
100 118.498 124.342 129.561 135.807 149.449
100 118.498 124.342 129.561 135.807 149.449

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3674.htm (4 of 7) [5/7/2002 3:15:49 PM]

1.3.6.7.4. Critical Values of the Chi-Square Distribution

Lower critical values of chi-square distribution with degrees of freedom

Probability of exceeding the critical value

0.90 0.95 0.975 0.99 0.999

1. .016 .004 .001 .000 .000

2. .211 .103 .051 .020 .002
3. .584 .352 .216 .115 .024
4. 1.064 .711 .484 .297 .091
5. 1.610 1.145 .831 .554 .210
6. 2.204 1.635 1.237 .872 .381
7. 2.833 2.167 1.690 1.239 .598
8. 3.490 2.733 2.180 1.646 .857
9. 4.168 3.325 2.700 2.088 1.152
10. 4.865 3.940 3.247 2.558 1.479
11. 5.578 4.575 3.816 3.053 1.834
12. 6.304 5.226 4.404 3.571 2.214
13. 7.042 5.892 5.009 4.107 2.617
14. 7.790 6.571 5.629 4.660 3.041
15. 8.547 7.261 6.262 5.229 3.483
16. 9.312 7.962 6.908 5.812 3.942
17. 10.085 8.672 7.564 6.408 4.416
18. 10.865 9.390 8.231 7.015 4.905
19. 11.651 10.117 8.907 7.633 5.407
20. 12.443 10.851 9.591 8.260 5.921
21. 13.240 11.591 10.283 8.897 6.447
22. 14.041 12.338 10.982 9.542 6.983
23. 14.848 13.091 11.689 10.196 7.529
24. 15.659 13.848 12.401 10.856 8.085
25. 16.473 14.611 13.120 11.524 8.649
26. 17.292 15.379 13.844 12.198 9.222
27. 18.114 16.151 14.573 12.879 9.803
28. 18.939 16.928 15.308 13.565 10.391
29. 19.768 17.708 16.047 14.256 10.986
30. 20.599 18.493 16.791 14.953 11.588

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3674.htm (5 of 7) [5/7/2002 3:15:49 PM]

1.3.6.7.4. Critical Values of the Chi-Square Distribution

31. 21.434 19.281 17.539 15.655 12.196

32. 22.271 20.072 18.291 16.362 12.811
33. 23.110 20.867 19.047 17.074 13.431
34. 23.952 21.664 19.806 17.789 14.057
35. 24.797 22.465 20.569 18.509 14.688
36. 25.643 23.269 21.336 19.233 15.324
37. 26.492 24.075 22.106 19.960 15.965
38. 27.343 24.884 22.878 20.691 16.611
39. 28.196 25.695 23.654 21.426 17.262
40. 29.051 26.509 24.433 22.164 17.916
41. 29.907 27.326 25.215 22.906 18.575
42. 30.765 28.144 25.999 23.650 19.239
43. 31.625 28.965 26.785 24.398 19.906
44. 32.487 29.787 27.575 25.148 20.576
45. 33.350 30.612 28.366 25.901 21.251
46. 34.215 31.439 29.160 26.657 21.929
47. 35.081 32.268 29.956 27.416 22.610
48. 35.949 33.098 30.755 28.177 23.295
49. 36.818 33.930 31.555 28.941 23.983
50. 37.689 34.764 32.357 29.707 24.674
51. 38.560 35.600 33.162 30.475 25.368
52. 39.433 36.437 33.968 31.246 26.065
53. 40.308 37.276 34.776 32.018 26.765
54. 41.183 38.116 35.586 32.793 27.468
55. 42.060 38.958 36.398 33.570 28.173
56. 42.937 39.801 37.212 34.350 28.881
57. 43.816 40.646 38.027 35.131 29.592
58. 44.696 41.492 38.844 35.913 30.305
59. 45.577 42.339 39.662 36.698 31.020
60. 46.459 43.188 40.482 37.485 31.738
61. 47.342 44.038 41.303 38.273 32.459
62. 48.226 44.889 42.126 39.063 33.181
63. 49.111 45.741 42.950 39.855 33.906
64. 49.996 46.595 43.776 40.649 34.633
65. 50.883 47.450 44.603 41.444 35.362
66. 51.770 48.305 45.431 42.240 36.093
67. 52.659 49.162 46.261 43.038 36.826
68. 53.548 50.020 47.092 43.838 37.561
69. 54.438 50.879 47.924 44.639 38.298
70. 55.329 51.739 48.758 45.442 39.036
71. 56.221 52.600 49.592 46.246 39.777

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3674.htm (6 of 7) [5/7/2002 3:15:49 PM]

1.3.6.7.4. Critical Values of the Chi-Square Distribution

72. 57.113 53.462 50.428 47.051 40.519

73. 58.006 54.325 51.265 47.858 41.264
74. 58.900 55.189 52.103 48.666 42.010
75. 59.795 56.054 52.942 49.475 42.757
76. 60.690 56.920 53.782 50.286 43.507
77. 61.586 57.786 54.623 51.097 44.258
78. 62.483 58.654 55.466 51.910 45.010
79. 63.380 59.522 56.309 52.725 45.764
80. 64.278 60.391 57.153 53.540 46.520
81. 65.176 61.261 57.998 54.357 47.277
82. 66.076 62.132 58.845 55.174 48.036
83. 66.976 63.004 59.692 55.993 48.796
84. 67.876 63.876 60.540 56.813 49.557
85. 68.777 64.749 61.389 57.634 50.320
86. 69.679 65.623 62.239 58.456 51.085
87. 70.581 66.498 63.089 59.279 51.850
88. 71.484 67.373 63.941 60.103 52.617
89. 72.387 68.249 64.793 60.928 53.386
90. 73.291 69.126 65.647 61.754 54.155
91. 74.196 70.003 66.501 62.581 54.926
92. 75.100 70.882 67.356 63.409 55.698
93. 76.006 71.760 68.211 64.238 56.472
94. 76.912 72.640 69.068 65.068 57.246
95. 77.818 73.520 69.925 65.898 58.022
96. 78.725 74.401 70.783 66.730 58.799
97. 79.633 75.282 71.642 67.562 59.577
98. 80.541 76.164 72.501 68.396 60.356
99. 81.449 77.046 73.361 69.230 61.137
100. 82.358 77.929 74.222 70.065 61.918

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3674.htm (7 of 7) [5/7/2002 3:15:49 PM]

1.3.6.7.5. Critical Values of the t* Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.7. Tables for Probability Distributions

1.3.6.7.5. Critical Values of the t*

Distribution
How to Use This table contains upper critical values of the t* distribution that are
This Table appropriate for determining whether or not a calibration line is in a state
of statistical control from measurements on a check standard at three
points in the calibration interval. A test statistic with degrees of
freedom is compared with the critical value. If the absolute value of the
test statistic exceeds the tabled value, the calibration of the instrument is
judged to be out of control.

Upper critical values of t* distribution at significance level 0.05

for testing the output of a linear calibration line at 3 points

1 37.544 61 2.455
2 7.582 62 2.454
3 4.826 63 2.453
4 3.941 64 2.452
5 3.518 65 2.451
6 3.274 66 2.450
7 3.115 67 2.449
8 3.004 68 2.448
9 2.923 69 2.447
10 2.860 70 2.446
11 2.811 71 2.445
12 2.770 72 2.445
13 2.737 73 2.444
14 2.709 74 2.443
15 2.685 75 2.442

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3675.htm (1 of 3) [5/7/2002 3:15:49 PM]

1.3.6.7.5. Critical Values of the t* Distribution

16 2.665 76 2.441
17 2.647 77 2.441
18 2.631 78 2.440
19 2.617 79 2.439
20 2.605 80 2.439
21 2.594 81 2.438
22 2.584 82 2.437
23 2.574 83 2.437
24 2.566 84 2.436
25 2.558 85 2.436
26 2.551 86 2.435
27 2.545 87 2.435
28 2.539 88 2.434
29 2.534 89 2.434
30 2.528 90 2.433
31 2.524 91 2.432
32 2.519 92 2.432
33 2.515 93 2.431
34 2.511 94 2.431
35 2.507 95 2.431
36 2.504 96 2.430
37 2.501 97 2.430
38 2.498 98 2.429
39 2.495 99 2.429
40 2.492 100 2.428
41 2.489 101 2.428
42 2.487 102 2.428
43 2.484 103 2.427
44 2.482 104 2.427
45 2.480 105 2.426
46 2.478 106 2.426
47 2.476 107 2.426
48 2.474 108 2.425
49 2.472 109 2.425
50 2.470 110 2.425
51 2.469 111 2.424
52 2.467 112 2.424
53 2.466 113 2.424
54 2.464 114 2.423
55 2.463 115 2.423
56 2.461 116 2.423
57 2.460 117 2.422
58 2.459 118 2.422
59 2.457 119 2.422
60 2.456 120 2.422

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3675.htm (2 of 3) [5/7/2002 3:15:49 PM]

1.3.6.7.5. Critical Values of the t* Distribution

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3675.htm (3 of 3) [5/7/2002 3:15:49 PM]

1.3.6.7.6. Critical Values of the Normal PPCC Distribution

1. Exploratory Data Analysis

1.3. EDA Techniques
1.3.6. Probability Distributions
1.3.6.7. Tables for Probability Distributions

1.3.6.7.6. Critical Values of the Normal

PPCC Distribution
How to Use This table contains the critical values of the normal probability plot
This Table correlation coefficient (PPCC) distribution that are appropriate for
determining whether or not a data set came from a population with
approximately a normal distribution. It is used in conjuction with a
normal probability plot. The test statistic is the correlation coefficient of
the points that make up a normal probability plot. This test statistic is
compared with the critical value below. If the test statistic is less than
the tabulated value, the null hypothesis that the data came from a
population with a normal distribution is rejected.
For example, suppose a set of 50 data points had a correlation
coefficient of 0.985 from the normal probability plot. At the 5%
significance level, the critical value is 0.965. Since 0.985 is greater than
0.965, we cannot reject the null hypothesis that the data came from a
population with a normal distribution.

Critical values of the normal PPCC for testing if data come from
a normal distribution

N 0.05 0.01

3 0.8687 0.8790
4 0.8234 0.8666
5 0.8240 0.8786
6 0.8351 0.8880
7 0.8474 0.8970
8 0.8590 0.9043
9 0.8689 0.9115
10 0.8765 0.9173

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3676.htm (1 of 4) [5/7/2002 3:15:49 PM]

1.3.6.7.6. Critical Values of the Normal PPCC Distribution

11 0.8838 0.9223
12 0.8918 0.9267
13 0.8974 0.9310
14 0.9029 0.9343
15 0.9080 0.9376
16 0.9121 0.9405
17 0.9160 0.9433
18 0.9196 0.9452
19 0.9230 0.9479
20 0.9256 0.9498
21 0.9285 0.9515
22 0.9308 0.9535
23 0.9334 0.9548
24 0.9356 0.9564
25 0.9370 0.9575
26 0.9393 0.9590
27 0.9413 0.9600
28 0.9428 0.9615
29 0.9441 0.9622
30 0.9462 0.9634
31 0.9476 0.9644
32 0.9490 0.9652
33 0.9505 0.9661
34 0.9521 0.9671
35 0.9530 0.9678
36 0.9540 0.9686
37 0.9551 0.9693
38 0.9555 0.9700
39 0.9568 0.9704
40 0.9576 0.9712
41 0.9589 0.9719
42 0.9593 0.9723
43 0.9609 0.9730
44 0.9611 0.9734
45 0.9620 0.9739
46 0.9629 0.9744
47 0.9637 0.9748
48 0.9640 0.9753
49 0.9643 0.9758
50 0.9654 0.9761
55 0.9683 0.9781
60 0.9706 0.9797
65 0.9723 0.9809
70 0.9742 0.9822
75 0.9758 0.9831
80 0.9771 0.9841

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3676.htm (2 of 4) [5/7/2002 3:15:49 PM]

1.3.6.7.6. Critical Values of the Normal PPCC Distribution

85 0.9784 0.9850
90 0.9797 0.9857
95 0.9804 0.9864
100 0.9814 0.9869
110 0.9830 0.9881
120 0.9841 0.9889
130 0.9854 0.9897
140 0.9865 0.9904
150 0.9871 0.9909
160 0.9879 0.9915
170 0.9887 0.9919
180 0.9891 0.9923
190 0.9897 0.9927
200 0.9903 0.9930
210 0.9907 0.9933
220 0.9910 0.9936
230 0.9914 0.9939
240 0.9917 0.9941
250 0.9921 0.9943
260 0.9924 0.9945
270 0.9926 0.9947
280 0.9929 0.9949
290 0.9931 0.9951
300 0.9933 0.9952
310 0.9936 0.9954
320 0.9937 0.9955
330 0.9939 0.9956
340 0.9941 0.9957
350 0.9942 0.9958
360 0.9944 0.9959
370 0.9945 0.9960
380 0.9947 0.9961
390 0.9948 0.9962
400 0.9949 0.9963
410 0.9950 0.9964
420 0.9951 0.9965
430 0.9953 0.9966
440 0.9954 0.9966
450 0.9954 0.9967
460 0.9955 0.9968
470 0.9956 0.9968
480 0.9957 0.9969
490 0.9958 0.9969
500 0.9959 0.9970
525 0.9961 0.9972
550 0.9963 0.9973

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3676.htm (3 of 4) [5/7/2002 3:15:49 PM]

1.3.6.7.6. Critical Values of the Normal PPCC Distribution

575 0.9964 0.9974

600 0.9965 0.9975
625 0.9967 0.9976
650 0.9968 0.9977
675 0.9969 0.9977
700 0.9970 0.9978
725 0.9971 0.9979
750 0.9972 0.9980
775 0.9973 0.9980
800 0.9974 0.9981
825 0.9975 0.9981
850 0.9975 0.9982
875 0.9976 0.9982
900 0.9977 0.9983
925 0.9977 0.9983
950 0.9978 0.9984
975 0.9978 0.9984
1000 0.9979 0.9984

http://www.itl.nist.gov/div898/handbook/eda/section3/eda3676.htm (4 of 4) [5/7/2002 3:15:49 PM]

1.4. EDA Case Studies

1. Exploratory Data Analysis

1.4. EDA Case Studies

Summary This section presents a series of case studies that demonstrate the
application of EDA methods to specific problems. In some cases, we
have focused on just one EDA technique that uncovers virtually all there
is to know about the data. For other case studies, we need several EDA
techniques, the selection of which is dictated by the outcome of the
previous step in the analaysis sequence. Note in these case studies how
the flow of the analysis is motivated by the focus on underlying
assumptions and general EDA principles.

Table of 1. Introduction
Contents for
Section 4 2. By Problem Category

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4.htm [5/7/2002 3:15:49 PM]

1.4.1. Case Studies Introduction

1. Exploratory Data Analysis

1.4. EDA Case Studies

1.4.1. Case Studies Introduction

Purpose The purpose of the first eight case studies is to show how EDA
graphics and quantitative measures and tests are applied to data from
scientific processes and to critique those data with regard to the
following assumptions that typically underlie a measurement process;
namely, that the data behave like:
● random drawings

● from a fixed distribution

● with a fixed location

● with a fixed standard deviation

Case studies 9 and 10 show the use of EDA techniques in

distributional modeling and the analysis of a designed experiment,
respectively.

Yi = C + Ei
If the above assumptions are satisfied, the process is said to be
statistically "in control" with the core characteristic of having
"predictability". That is, probability statements can be made about the
process, not only in the past, but also in the future.
An appropriate model for an "in control" process is
Yi = C + Ei
where C is a constant (the "deterministic" or "structural" component),
and where Ei is the error term (or "random" component).

The constant C is the average value of the process--it is the primary

summary number which shows up on any report. Although C is
(assumed) fixed, it is unknown, and so a primary analysis objective of
the engineer is to arrive at an estimate of C.
This goal partitions into 4 sub-goals:
1. Is the most common estimator of C, , the best estimator for
C? What does "best" mean?
2. If is best, what is the uncertainty for . In particular, is

http://www.itl.nist.gov/div898/handbook/eda/section4/eda41.htm (1 of 4) [5/7/2002 3:15:50 PM]

1.4.1. Case Studies Introduction

the usual formula for the uncertainty of :

valid? Here, s is the standard deviation of the data and N is the

sample size.
3. If is not the best estimator for C, what is a better estimator
for C (for example, median, midrange, midmean)?
4. If there is a better estimator, , what is its uncertainty? That is,
what is ?
EDA and the routine checking of underlying assumptions provides
insight into all of the above.
1. Location and variation checks provide information as to
whether C is really constant.
2. Distributional checks indicate whether is the best estimator.
Techniques for distributional checking include histograms,
normal probability plots, and probability plot correlation
coefficient plots.
3. Randomness checks ascertain whether the usual

is valid.
4. Distributional tests assist in determining a better estimator, if
needed.
5. Simulator tools (namely bootstrapping) provide values for the
uncertainty of alternative estimators.

Assumptions If one or more of the above assumptions is not satisfied, then we use
not satisfied EDA techniques, or some mix of EDA and classical techniques, to
find a more appropriate model for the data. That is,
Yi = D + Ei
where D is the deterministic part and E is an error component.
If the data are not random, then we may investigate fitting some
simple time series models to the data. If the constant location and
scale assumptions are violated, we may need to investigate the
measurement process to see if there is an explanation.
The assumptions on the error term are still quite relevant in the sense
that for an appropriate model the error component should follow the
assumptions. The criterion for validating the model, or comparing
competing models, is framed in terms of these assumptions.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda41.htm (2 of 4) [5/7/2002 3:15:50 PM]

1.4.1. Case Studies Introduction

Multivariable Although the case studies in this chapter utilize univariate data, the
data assumptions above are relevant for multivariable data as well.
If the data are not univariate, then we are trying to find a model
Yi = F(X1, ..., Xk) + Ei
where F is some function based on one or more variables. The error
component, which is a univariate data set, of a good model should
satisfy the assumptions given above. The criterion for validating and
comparing models is based on how well the error component follows
these assumptions.
The load cell calibration case study in the process modeling chapter
shows an example of this in the regression context.

First three The first three case studies utilize data that are randomly generated
case studies from the following distributions:
utilize data ● normal distribution with mean 0 and standard deviation 1
with known
● uniform distribution with mean 0 and standard deviation
characteristics
(uniform over the interval (0,1))

random walk
●

The other univariate case studies utilize data from scientific processes.
The goal is to determine if
Yi = C + Ei
is a reasonable model. This is done by testing the underlying
assumptions. If the assumptions are satisfied, then an estimate of C
and an estimate of the uncertainty of C are computed. If the
assumptions are not satisfied, we attempt to find a model where the
error component does satisfy the underlying assumptions.

Graphical To test the underlying assumptions, each data set is analyzed using
methods that four graphical methods that are particularly suited for this purpose:
are applied to 1. run sequence plot which is useful for detecting shifts of location
the data or scale
2. lag plot which is useful for detecting non-randomness in the
data
3. histogram which is useful for trying to determine the underlying
distribution
4. normal probability plot for deciding whether the data follow the
normal distribution
There are a number of other techniques for addressing the underlying

http://www.itl.nist.gov/div898/handbook/eda/section4/eda41.htm (3 of 4) [5/7/2002 3:15:50 PM]

1.4.1. Case Studies Introduction

assumptions. However, the four plots listed above provide an

excellent opportunity for addressing all of the assumptions on a single
page of graphics.
Additional graphical techniques are used in certain case studies to
develop models that do have error components that satisfy the
underlying assumptions.

Quantitative The normal and uniform random number data sets are also analyzed
methods that with the following quantitative techniques, which are explained in
are applied to more detail in an earlier section:
the data 1. Summary statistics which include:
❍ mean

❍ standard deviation
❍ autocorrelation coefficient to test for randomness
❍ normal and uniform probability plot correlation
coefficients (ppcc) to test for a normal or uniform
distribution, respectively
❍ Wilk-Shapiro test for a normal distribution

2. Linear fit of the data as a function of time to assess drift (test

for fixed location)
3. Bartlett test for fixed variance
4. Autocorrelation plot and coefficient to test for randomness
5. Runs test to test for lack of randomness
6. Anderson-Darling test for a normal distribution
7. Grubbs test for outliers
8. Summary report
Although the graphical methods applied to the normal and uniform
random numbers are sufficient to assess the validity of the underlying
assumptions, the quantitative techniques are used to show the different
flavor of the graphical and quantitative approaches.
The remaining case studies intermix one or more of these quantitative
techniques into the analysis where appropriate.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda41.htm (4 of 4) [5/7/2002 3:15:50 PM]

1.4.2. Case Studies

1. Exploratory Data Analysis

1.4. EDA Case Studies

1.4.2. Case Studies

Univariate
Yi = C + Ei

Normal Random Uniform Random Random Walk

Numbers Numbers

Josephson Junction Beam Deflections Filter Transmittance

Cryothermometry

Standard Resistor Heat Flow Meter 1

Reliability

Airplane Glass
Failure Time

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42.htm (1 of 2) [5/7/2002 3:15:50 PM]

1.4.2. Case Studies

Multi-Factor

Ceramic Strength

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42.htm (2 of 2) [5/7/2002 3:15:50 PM]

1.4.2.1. Normal Random Numbers

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies

1.4.2.1. Normal Random Numbers

Normal This example illustrates the univariate analysis of a set of normal
Random random numbers.
Numbers

1. Background and Data

2. Graphical Output and Interpretation
3. Quantitative Output and Interpretation
4. Work This Example Yourself

http://www.itl.nist.gov/div898/handbook/eda/section4/eda421.htm [5/7/2002 3:15:51 PM]

1.4.2.1.1. Background and Data

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.1. Normal Random Numbers

1.4.2.1.1. Background and Data

Generation The normal random numbers used in this case study are from a Rand
Corporation publication.

The motivation for studying a set of normal random numbers is to

illustrate the ideal case where all four underlying assumptions hold.

Software Most general purpose statistical software programs, including Dataplot,

can generate normal random numbers.

Resulting The following is the set of normal random numbers used for this case
Data study.

-1.2760 -1.2180 -0.4530 -0.3500 0.7230

0.6760 -1.0990 -0.3140 -0.3940 -0.6330
-0.3180 -0.7990 -1.6640 1.3910 0.3820
0.7330 0.6530 0.2190 -0.6810 1.1290
-1.3770 -1.2570 0.4950 -0.1390 -0.8540
0.4280 -1.3220 -0.3150 -0.7320 -1.3480
2.3340 -0.3370 -1.9550 -0.6360 -1.3180
-0.4330 0.5450 0.4280 -0.2970 0.2760
-1.1360 0.6420 3.4360 -1.6670 0.8470
-1.1730 -0.3550 0.0350 0.3590 0.9300
0.4140 -0.0110 0.6660 -1.1320 -0.4100
-1.0770 0.7340 1.4840 -0.3400 0.7890
-0.4940 0.3640 -1.2370 -0.0440 -0.1110
-0.2100 0.9310 0.6160 -0.3770 -0.4330
1.0480 0.0370 0.7590 0.6090 -2.0430
-0.2900 0.4040 -0.5430 0.4860 0.8690
0.3470 2.8160 -0.4640 -0.6320 -1.6140
0.3720 -0.0740 -0.9160 1.3140 -0.0380
0.6370 0.5630 -0.1070 0.1310 -1.8080
-1.1260 0.3790 0.6100 -0.3640 -2.6260

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4211.htm (1 of 3) [5/7/2002 3:15:51 PM]

1.4.2.1.1. Background and Data

2.1760 0.3930 -0.9240 1.9110 -1.0400

-1.1680 0.4850 0.0760 -0.7690 1.6070
-1.1850 -0.9440 -1.6040 0.1850 -0.2580
-0.3000 -0.5910 -0.5450 0.0180 -0.4850
0.9720 1.7100 2.6820 2.8130 -1.5310
-0.4900 2.0710 1.4440 -1.0920 0.4780
1.2100 0.2940 -0.2480 0.7190 1.1030
1.0900 0.2120 -1.1850 -0.3380 -1.1340
2.6470 0.7770 0.4500 2.2470 1.1510
-1.6760 0.3840 1.1330 1.3930 0.8140
0.3980 0.3180 -0.9280 2.4160 -0.9360
1.0360 0.0240 -0.5600 0.2030 -0.8710
0.8460 -0.6990 -0.3680 0.3440 -0.9260
-0.7970 -1.4040 -1.4720 -0.1180 1.4560
0.6540 -0.9550 2.9070 1.6880 0.7520
-0.4340 0.7460 0.1490 -0.1700 -0.4790
0.5220 0.2310 -0.6190 -0.2650 0.4190
0.5580 -0.5490 0.1920 -0.3340 1.3730
-1.2880 -0.5390 -0.8240 0.2440 -1.0700
0.0100 0.4820 -0.4690 -0.0900 1.1710
1.3720 1.7690 -1.0570 1.6460 0.4810
-0.6000 -0.5920 0.6100 -0.0960 -1.3750
0.8540 -0.5350 1.6070 0.4280 -0.6150
0.3310 -0.3360 -1.1520 0.5330 -0.8330
-0.1480 -1.1440 0.9130 0.6840 1.0430
0.5540 -0.0510 -0.9440 -0.4400 -0.2120
-1.1480 -1.0560 0.6350 -0.3280 -1.2210
0.1180 -2.0450 -1.9770 -1.1330 0.3380
0.3480 0.9700 -0.0170 1.2170 -0.9740
-1.2910 -0.3990 -1.2090 -0.2480 0.4800
0.2840 0.4580 1.3070 -1.6250 -0.6290
-0.5040 -0.0560 -0.1310 0.0480 1.8790
-1.0160 0.3600 -0.1190 2.3310 1.6720
-1.0530 0.8400 -0.2460 0.2370 -1.3120
1.6030 -0.9520 -0.5660 1.6000 0.4650
1.9510 0.1100 0.2510 0.1160 -0.9570
-0.1900 1.4790 -0.9860 1.2490 1.9340
0.0700 -1.3580 -1.2460 -0.9590 -1.2970
-0.7220 0.9250 0.7830 -0.4020 0.6190
1.8260 1.2720 -0.9450 0.4940 0.0500
-1.6960 1.8790 0.0630 0.1320 0.6820
0.5440 -0.4170 -0.6660 -0.1040 -0.2530
-2.5430 -1.3330 1.9870 0.6680 0.3600
1.9270 1.1830 1.2110 1.7650 0.3500
-0.3590 0.1930 -1.0230 -0.2220 -0.6160
-0.0600 -1.3190 0.7850 -0.4300 -0.2980

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4211.htm (2 of 3) [5/7/2002 3:15:51 PM]

1.4.2.1.1. Background and Data

0.2480 -0.0880 -1.3790 0.2950 -0.1150

-0.6210 -0.6180 0.2090 0.9790 0.9060
-0.0990 -1.3760 1.0470 -0.8720 -2.2000
-1.3840 1.4250 -0.8120 0.7480 -1.0930
-0.4630 -1.2810 -2.5140 0.6750 1.1450
1.0830 -0.6670 -0.2230 -1.5920 -1.2780
0.5030 1.4340 0.2900 0.3970 -0.8370
-0.9730 -0.1200 -1.5940 -0.9960 -1.2440
-0.8570 -0.3710 -0.2160 0.1480 -2.1060
-1.4530 0.6860 -0.0750 -0.2430 -0.1700
-0.1220 1.1070 -1.0390 -0.6360 -0.8600
-0.8950 -1.4580 -0.5390 -0.1590 -0.4200
1.6320 0.5860 -0.4680 -0.3860 -0.3540
0.2030 -1.2340 2.3810 -0.3880 -0.0630
2.0720 -1.4450 -0.6800 0.2240 -0.1200
1.7530 -0.5710 1.2230 -0.1260 0.0340
-0.4350 -0.3750 -0.9850 -0.5850 -0.2030
-0.5560 0.0240 0.1260 1.2500 -0.6150
0.8760 -1.2270 -2.6470 -0.7450 1.7970
-1.2310 0.5470 -0.6340 -0.8360 -0.7190
0.8330 1.2890 -0.0220 -0.4310 0.5820
0.7660 -0.5740 -1.1530 0.5200 -1.0180
-0.8910 0.3320 -0.4530 -1.1270 2.0850
-0.7220 -1.5080 0.4890 -0.4960 -0.0250
0.6440 -0.2330 -0.1530 1.0980 0.7570
-0.0390 -0.4600 0.3930 2.0120 1.3560
0.1050 -0.1710 -0.1100 -1.1450 0.8780
-0.9090 -0.3280 1.0210 -1.6130 1.5600
-1.1920 1.7700 -0.0030 0.3690 0.0520
0.6470 1.0290 1.5260 0.2370 -1.3280
-0.0420 0.5530 0.7700 0.3240 -0.4890
-0.3670 0.3780 0.6010 -1.9960 -0.7380
0.4980 1.0720 1.5670 0.3020 1.1570
-0.7200 1.4030 0.6980 -0.3700 -0.5510

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4211.htm (3 of 3) [5/7/2002 3:15:51 PM]

1.4.2.1.2. Graphical Output and Interpretation

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.1. Normal Random Numbers

1.4.2.1.2. Graphical Output and

Interpretation
Goal The goal of this analysis is threefold:
1. Determine if the univariate model:

is appropriate and valid.

2. Determine if the typical underlying assumptions for an "in
control" measurement process are valid. These assumptions are:
1. random drawings;
2. from a fixed distribution;
3. with the distribution having a fixed location; and
4. the distribution having a fixed scale.
3. Determine if the confidence interval

is appropriate and valid where s is the standard deviation of the

original data.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4212.htm (1 of 4) [5/7/2002 3:15:53 PM]

1.4.2.1.2. Graphical Output and Interpretation

4-Plot of
Data

Interpretation The assumptions are addressed by the graphics shown above:

1. The run sequence plot (upper left) indicates that the data do not
have any significant shifts in location or scale over time. The run
sequence plot does not show any obvious outliers.
2. The lag plot (upper right) does not indicate any non-random
pattern in the data.
3. The histogram (lower left) shows that the data are reasonably
symmetric, there do not appear to be significant outliers in the
tails, and that it is reasonable to assume that the data are from
approximately a normal distribution.
4. The normal probability plot (lower right) verifies that an
assumption of normality is in fact reasonable.
From the above plots, we conclude that the underlying assumptions are
valid and the data follow approximately a normal distribution.
Therefore, the confidence interval form given previously is appropriate
for quantifying the uncertainty of the population mean. The numerical
values for this model are given in the Quantitative Output and
Interpretation section.

Individual Although it is usually not necessary, the plots can be generated

Plots individually to give more detail.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4212.htm (2 of 4) [5/7/2002 3:15:53 PM]

1.4.2.1.2. Graphical Output and Interpretation

Run
Sequence
Plot

Lag Plot

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4212.htm (3 of 4) [5/7/2002 3:15:53 PM]

1.4.2.1.2. Graphical Output and Interpretation

Histogram
(with
overlaid
Normal PDF)

Normal
Probability
Plot

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4212.htm (4 of 4) [5/7/2002 3:15:53 PM]

1.4.2.1.3. Quantitative Output and Interpretation

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.1. Normal Random Numbers

1.4.2.1.3. Quantitative Output and Interpretation

Summary As a first step in the analysis, a table of summary statistics is computed from the data. The
Statistics following table, generated by Dataplot, shows a typical set of statistics.

SUMMARY

NUMBER OF OBSERVATIONS = 500

***********************************************************************
* LOCATION MEASURES * DISPERSION MEASURES *
***********************************************************************
* MIDRANGE = 0.3945000E+00 * RANGE = 0.6083000E+01 *
* MEAN = -0.2935997E-02 * STAND. DEV. = 0.1021041E+01 *
* MIDMEAN = 0.1623600E-01 * AV. AB. DEV. = 0.8174360E+00 *
* MEDIAN = -0.9300000E-01 * MINIMUM = -0.2647000E+01 *
* = * LOWER QUART. = -0.7204999E+00 *
* = * LOWER HINGE = -0.7210000E+00 *
* = * UPPER HINGE = 0.6455001E+00 *
* = * UPPER QUART. = 0.6447501E+00 *
* = * MAXIMUM = 0.3436000E+01 *
***********************************************************************
* RANDOMNESS MEASURES * DISTRIBUTIONAL MEASURES *
***********************************************************************
* AUTOCO COEF = 0.4505888E-01 * ST. 3RD MOM. = 0.3072273E+00 *
* = 0.0000000E+00 * ST. 4TH MOM. = 0.2990314E+01 *
* = 0.0000000E+00 * ST. WILK-SHA = 0.7515639E+01 *
* = * UNIFORM PPCC = 0.9756625E+00 *
* = * NORMAL PPCC = 0.9961721E+00 *
* = * TUK -.5 PPCC = 0.8366451E+00 *
* = * CAUCHY PPCC = 0.4922674E+00 *
***********************************************************************

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4213.htm (1 of 7) [5/7/2002 3:15:54 PM]

1.4.2.1.3. Quantitative Output and Interpretation

Location One way to quantify a change in location over time is to fit a straight line to the data set,
using the index variable X = 1, 2, ..., N, with N denoting the number of observations. If
there is no significant drift in the location, the slope parameter should be zero. For this data
set, Dataplot generated the following output:

LEAST SQUARES MULTILINEAR FIT

SAMPLE SIZE N = 500
NUMBER OF VARIABLES = 1
NO REPLICATION CASE

PARAMETER ESTIMATES (APPROX. ST. DEV.) T VALUE

1 A0 0.699127E-02 (0.9155E-01) 0.7636E-01
2 A1 X -0.396298E-04 (0.3167E-03) -0.1251

RESIDUAL STANDARD DEVIATION = 1.02205

RESIDUAL DEGREES OF FREEDOM = 498

The slope parameter, A1, has a t value of -0.13 which is statistically not significant. This
indicates that the slope can in fact be considered zero.

Variation One simple way to detect a change in variation is with a Bartlett test, after dividing the data
set into several equal-sized intervals. The choice of the number of intervals is somewhat
arbitrary, although values of 4 or 8 are reasonable. Dataplot generated the following output
for the Bartlett test.

BARTLETT TEST
(STANDARD DEFINITION)
NULL HYPOTHESIS UNDER TEST--ALL SIGMA(I) ARE EQUAL

TEST:
DEGREES OF FREEDOM = 3.000000

TEST STATISTIC VALUE = 2.373660

CUTOFF: 95% PERCENT POINT = 7.814727
CUTOFF: 99% PERCENT POINT = 11.34487

CHI-SQUARE CDF VALUE = 0.501443

NULL NULL HYPOTHESIS NULL HYPOTHESIS

HYPOTHESIS ACCEPTANCE INTERVAL CONCLUSION
ALL SIGMA EQUAL (0.000,0.950) ACCEPT

In this case, the Bartlett test indicates that the standard deviations are not significantly
different in the 4 intervals.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4213.htm (2 of 7) [5/7/2002 3:15:54 PM]

1.4.2.1.3. Quantitative Output and Interpretation

Randomness
There are many ways in which data can be non-random. However, most common forms of
non-randomness can be detected with a few simple tests. The lag plot in the 4-plot above is
a simple graphical technique.
Another check is an autocorrelation plot that shows the autocorrelations for various lags.
Confidence bands can be plotted at the 95% and 99% confidence levels. Points outside this
band indicate statistically significant values (lag 0 is always 1). Dataplot generated the
following autocorrelation plot.

The lag 1 autocorrelation, which is generally the one of most interest, is 0.045. The critical
values at the 5% significance level are -0.087 and 0.087. Thus, since 0.045 is in the interval,
the lag 1 autocorrelation is not statistically significant, so there is no evidence of
non-randomness.
A common test for randomness is the runs test.

RUNS UP
STATISTIC = NUMBER OF RUNS UP
OF LENGTH EXACTLY I
I STAT EXP(STAT) SD(STAT) Z

1 98.0 104.2083 10.2792 -0.60

2 43.0 45.7167 5.2996 -0.51
3 13.0 13.1292 3.2297 -0.04
4 6.0 2.8563 1.6351 1.92
5 1.0 0.5037 0.7045 0.70
6 0.0 0.0749 0.2733 -0.27
7 0.0 0.0097 0.0982 -0.10
8 0.0 0.0011 0.0331 -0.03
9 0.0 0.0001 0.0106 -0.01
10 0.0 0.0000 0.0032 0.00
STATISTIC = NUMBER OF RUNS UP
OF LENGTH I OR MORE
I STAT EXP(STAT) SD(STAT) Z

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4213.htm (3 of 7) [5/7/2002 3:15:54 PM]

1.4.2.1.3. Quantitative Output and Interpretation
1 161.0 166.5000 6.6546 -0.83
2 63.0 62.2917 4.4454 0.16
3 20.0 16.5750 3.4338 1.00
4 7.0 3.4458 1.7786 2.00
5 1.0 0.5895 0.7609 0.54
6 0.0 0.0858 0.2924 -0.29
7 0.0 0.0109 0.1042 -0.10
8 0.0 0.0012 0.0349 -0.03
9 0.0 0.0001 0.0111 -0.01
10 0.0 0.0000 0.0034 0.00
RUNS DOWN
STATISTIC = NUMBER OF RUNS DOWN
OF LENGTH EXACTLY I
I STAT EXP(STAT) SD(STAT) Z

1 91.0 104.2083 10.2792 -1.28

2 55.0 45.7167 5.2996 1.75
3 14.0 13.1292 3.2297 0.27
4 1.0 2.8563 1.6351 -1.14
5 0.0 0.5037 0.7045 -0.71
6 0.0 0.0749 0.2733 -0.27
7 0.0 0.0097 0.0982 -0.10
8 0.0 0.0011 0.0331 -0.03
9 0.0 0.0001 0.0106 -0.01
10 0.0 0.0000 0.0032 0.00
STATISTIC = NUMBER OF RUNS DOWN
OF LENGTH I OR MORE
I STAT EXP(STAT) SD(STAT) Z

1 161.0 166.5000 6.6546 -0.83

2 70.0 62.2917 4.4454 1.73
3 15.0 16.5750 3.4338 -0.46
4 1.0 3.4458 1.7786 -1.38
5 0.0 0.5895 0.7609 -0.77
6 0.0 0.0858 0.2924 -0.29
7 0.0 0.0109 0.1042 -0.10
8 0.0 0.0012 0.0349 -0.03
9 0.0 0.0001 0.0111 -0.01
10 0.0 0.0000 0.0034 0.00
RUNS TOTAL = RUNS UP + RUNS DOWN
STATISTIC = NUMBER OF RUNS TOTAL
OF LENGTH EXACTLY I
I STAT EXP(STAT) SD(STAT) Z

1 189.0 208.4167 14.5370 -1.34

2 98.0 91.4333 7.4947 0.88
3 27.0 26.2583 4.5674 0.16
4 7.0 5.7127 2.3123 0.56
5 1.0 1.0074 0.9963 -0.01
6 0.0 0.1498 0.3866 -0.39
7 0.0 0.0193 0.1389 -0.14
8 0.0 0.0022 0.0468 -0.05
9 0.0 0.0002 0.0150 -0.01
10 0.0 0.0000 0.0045 0.00
STATISTIC = NUMBER OF RUNS TOTAL
OF LENGTH I OR MORE
I STAT EXP(STAT) SD(STAT) Z

1 322.0 333.0000 9.4110 -1.17

2 133.0 124.5833 6.2868 1.34
3 35.0 33.1500 4.8561 0.38
4 8.0 6.8917 2.5154 0.44

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4213.htm (4 of 7) [5/7/2002 3:15:54 PM]

1.4.2.1.3. Quantitative Output and Interpretation
5 1.0 1.1790 1.0761 -0.17
6 0.0 0.1716 0.4136 -0.41
7 0.0 0.0217 0.1474 -0.15
8 0.0 0.0024 0.0494 -0.05
9 0.0 0.0002 0.0157 -0.02
10 0.0 0.0000 0.0047 0.00
LENGTH OF THE LONGEST RUN UP = 5
LENGTH OF THE LONGEST RUN DOWN = 4
LENGTH OF THE LONGEST RUN UP OR DOWN = 5

NUMBER OF POSITIVE DIFFERENCES = 252

NUMBER OF NEGATIVE DIFFERENCES = 247
NUMBER OF ZERO DIFFERENCES = 0

Values in the column labeled "Z" greater than 1.96 or less than -1.96 are statistically
significant at the 5% level. The runs test does not indicate any significant non-randomness.

Distributional Probability plots are a graphical test for assessing if a particular distribution provides an
Analysis adequate fit to a data set.
A quantitative enhancement to the probability plot is the correlation coefficient of the points
on the probability plot. For this data set the correlation coefficient is 0.996. Since this is
greater than the critical value of 0.987 (this is a tabulated value), the normality assumption
is not rejected.
Chi-square and Kolmogorov-Smirnov goodness-of-fit tests are alternative methods for
assessing distributional adequacy. The Wilk-Shapiro and Anderson-Darling tests can be
used to test for normality. Dataplot generates the following output for the Anderson-Darling
normality test.

ANDERSON-DARLING 1-SAMPLE TEST

THAT THE DATA CAME FROM A NORMAL DISTRIBUTION

1. STATISTICS:
NUMBER OF OBSERVATIONS = 500
MEAN = -0.2935997E-02
STANDARD DEVIATION = 1.021041

ANDERSON-DARLING TEST STATISTIC VALUE = 1.061249

ADJUSTED TEST STATISTIC VALUE = 1.069633

2. CRITICAL VALUES:
90 % POINT = 0.6560000
95 % POINT = 0.7870000
97.5 % POINT = 0.9180000
99 % POINT = 1.092000

3. CONCLUSION (AT THE 5% LEVEL):

THE DATA DO NOT COME FROM A NORMAL DISTRIBUTION.
The Anderson-Darling test rejects the normality assumption at the 5% level but accepts it at
the 1% level.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4213.htm (5 of 7) [5/7/2002 3:15:54 PM]

1.4.2.1.3. Quantitative Output and Interpretation

Outlier A test for outliers is the Grubbs test. Dataplot generated the following output for Grubbs'
Analysis test.

GRUBBS TEST FOR OUTLIERS

(ASSUMPTION: NORMALITY)

1. STATISTICS:
NUMBER OF OBSERVATIONS = 500
MINIMUM = -2.647000
MEAN = -0.2935997E-02
MAXIMUM = 3.436000
STANDARD DEVIATION = 1.021041

GRUBBS TEST STATISTIC = 3.368068

2. PERCENT POINTS OF THE REFERENCE DISTRIBUTION

FOR GRUBBS TEST STATISTIC
0 % POINT = 0.0000000E+00
50 % POINT = 3.274338
75 % POINT = 3.461431
90 % POINT = 3.695134
95 % POINT = 3.863087
99 % POINT = 4.228033

3. CONCLUSION (AT THE 5% LEVEL):

THERE ARE NO OUTLIERS.
For this data set, Grubbs' test does not detect any outliers at the 25%, 10%, 5%, and 1%
significance levels.

Model Since the underlying assumptions were validated both graphically and analytically, we
conclude that a reasonable model for the data is:
Yi = -0.00294 + Ei
We can express the uncertainty for C as the 95% confidence interval (-0.09266,0.086779).

Univariate It is sometimes useful and convenient to summarize the above results in a report. The report
Report for the 500 normal random numbers follows.

Analysis for 500 normal random numbers

1: Sample Size = 500

2: Location
Mean = -0.00294
Standard Deviation of Mean = 0.045663
95% Confidence Interval for Mean = (-0.09266,0.086779)
Drift with respect to location? = NO

3: Variation
Standard Deviation = 1.021042
95% Confidence Interval for SD = (0.961437,1.088585)
Drift with respect to variation?
(based on Bartletts test on quarters
of the data) = NO

4: Distribution
Normal PPCC = 0.996173

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4213.htm (6 of 7) [5/7/2002 3:15:54 PM]

1.4.2.1.3. Quantitative Output and Interpretation

Data are Normal?

(as measured by Normal PPCC) = YES

5: Randomness
Autocorrelation = 0.045059
Data are Random?
(as measured by autocorrelation) = YES

6: Statistical Control
(i.e., no drift in location or scale,
data are random, distribution is
fixed, here we are testing only for
fixed normal)
Data Set is in Statistical Control? = YES

7: Outliers?
(as determined by Grubbs' test) = NO

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4213.htm (7 of 7) [5/7/2002 3:15:54 PM]

1.4.2.1.4. Work This Example Yourself

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.1. Normal Random Numbers

1.4.2.1.4. Work This Example Yourself

View This page allows you to repeat the analysis outlined in the case study
Dataplot description on the previous page using Dataplot . It is required that you
Macro for have already downloaded and installed Dataplot and configured your
this Case browser. to run Dataplot. Output from each analysis step below will be
Study displayed in one or more of the Dataplot windows. The four main
windows are the Output window, the Graphics window, the Command
History window, and the data sheet window. Across the top of the main
windows there are menus for executing Dataplot commands. Across the
bottom is a command entry window where commands can be typed in.

Data Analysis Steps Results and Conclusions

Click on the links below to start Dataplot and run this case study
The links in this column will connect you with more detailed
yourself. Each step may use results from previous steps, so please be
information about each analysis step from the case study
patient. Wait until the software verifies that the current step is
description.
complete before clicking on the next step.

1. Invoke Dataplot and read data.

1. Read in the data. 1. You have read 1 column of numbers

into Dataplot, variable Y.

2. 4-plot of the data.

1. 4-plot of Y. 1. Based on the 4-plot, there are no shifts

in location or scale, and the data seem to
follow a normal distribution.

3. Generate the individual plots.

1. Generate a run sequence plot. 1. The run sequence plot indicates that
there are no shifts of location or
scale.

2. Generate a lag plot. 2. The lag plot does not indicate any
significant patterns (which would
show the data were not random).

3. Generate a histogram with an

3. The histogram indicates that a
overlaid normal pdf.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4214.htm (1 of 2) [5/7/2002 3:15:54 PM]

1.4.2.1.4. Work This Example Yourself
normal distribution is a good
distribution for these data.
4. Generate a normal probability
plot. 4. The normal probability plot verifies
that the normal distribution is a
reasonable distribution for these data.

4. Generate summary statistics, quantitative

analysis, and print a univariate report.

1. Generate a table of summary 1. The summary statistics table displays

statistics. 25+ statistics.

2. Generate the mean, a confidence 2. The mean is -0.00294 and a 95%

interval for the mean, and compute confidence interval is (-0.093,0.087).
a linear fit to detect drift in The linear fit indicates no drift in
location. location since the slope parameter is
statistically not significant.

3. Generate the standard deviation, a 3. The standard deviation is 1.02 with

confidence interval for the standard a 95% confidence interval of (0.96,1.09).
deviation, and detect drift in variation Bartlett's test indicates no significant
by dividing the data into quarters and change in variation.
computing Barltett's test for equal
standard deviations.

4. Check for randomness by generating an 4. The lag 1 autocorrelation is 0.04.

autocorrelation plot and a runs test. From the autocorrelation plot, this is
within the 95% confidence interval
bands.

5. Check for normality by computing the 5. The normal probability plot correlation
normal probability plot correlation coefficient is 0.996. At the 5% level,
coefficient. we cannot reject the normality assumption.

6. Check for outliers using Grubbs' test. 6. Grubbs' test detects no outliers at the
5% level.

7. Print a univariate report (this assumes 7. The results are summarized in a

steps 2 thru 6 have already been run). convenient report.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4214.htm (2 of 2) [5/7/2002 3:15:54 PM]

1.4.2.2. Uniform Random Numbers

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies

1.4.2.2. Uniform Random Numbers

Uniform This example illustrates the univariate analysis of a set of uniform
Random random numbers.
Numbers

1. Background and Data

2. Graphical Output and Interpretation
3. Quantitative Output and Interpretation
4. Work This Example Yourself

http://www.itl.nist.gov/div898/handbook/eda/section4/eda422.htm [5/7/2002 3:15:55 PM]

1.4.2.2.1. Background and Data

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.2. Uniform Random Numbers

1.4.2.2.1. Background and Data

Generation The uniform random numbers used in this case study are from a Rand
Corporation publication.

The motivation for studying a set of uniform random numbers is to

illustrate the effects of a known underlying non-normal distribution.

Software Most general purpose statistical software programs, including Dataplot,

can generate uniform random numbers.

Resulting The following is the set of uniform random numbers used for this case
Data study.

.100973 .253376 .520135 .863467 .354876

.809590 .911739 .292749 .375420 .480564
.894742 .962480 .524037 .206361 .040200
.822916 .084226 .895319 .645093 .032320
.902560 .159533 .476435 .080336 .990190
.252909 .376707 .153831 .131165 .886767
.439704 .436276 .128079 .997080 .157361
.476403 .236653 .989511 .687712 .171768
.660657 .471734 .072768 .503669 .736170
.658133 .988511 .199291 .310601 .080545
.571824 .063530 .342614 .867990 .743923
.403097 .852697 .760202 .051656 .926866
.574818 .730538 .524718 .623885 .635733
.213505 .325470 .489055 .357548 .284682
.870983 .491256 .737964 .575303 .529647
.783580 .834282 .609352 .034435 .273884
.985201 .776714 .905686 .072210 .940558
.609709 .343350 .500739 .118050 .543139
.808277 .325072 .568248 .294052 .420152
.775678 .834529 .963406 .288980 .831374

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4221.htm (1 of 3) [5/7/2002 3:15:55 PM]

1.4.2.2.1. Background and Data

.670078 .184754 .061068 .711778 .886854

.020086 .507584 .013676 .667951 .903647
.649329 .609110 .995946 .734887 .517649
.699182 .608928 .937856 .136823 .478341
.654811 .767417 .468509 .505804 .776974
.730395 .718640 .218165 .801243 .563517
.727080 .154531 .822374 .211157 .825314
.385537 .743509 .981777 .402772 .144323
.600210 .455216 .423796 .286026 .699162
.680366 .252291 .483693 .687203 .766211
.399094 .400564 .098932 .050514 .225685
.144642 .756788 .962977 .882254 .382145
.914991 .452368 .479276 .864616 .283554
.947508 .992337 .089200 .803369 .459826
.940368 .587029 .734135 .531403 .334042
.050823 .441048 .194985 .157479 .543297
.926575 .576004 .088122 .222064 .125507
.374211 .100020 .401286 .074697 .966448
.943928 .707258 .636064 .932916 .505344
.844021 .952563 .436517 .708207 .207317
.611969 .044626 .457477 .745192 .433729
.653945 .959342 .582605 .154744 .526695
.270799 .535936 .783848 .823961 .011833
.211594 .945572 .857367 .897543 .875462
.244431 .911904 .259292 .927459 .424811
.621397 .344087 .211686 .848767 .030711
.205925 .701466 .235237 .831773 .208898
.376893 .591416 .262522 .966305 .522825
.044935 .249475 .246338 .244586 .251025
.619627 .933565 .337124 .005499 .765464
.051881 .599611 .963896 .546928 .239123
.287295 .359631 .530726 .898093 .543335
.135462 .779745 .002490 .103393 .598080
.839145 .427268 .428360 .949700 .130212
.489278 .565201 .460588 .523601 .390922
.867728 .144077 .939108 .364770 .617429
.321790 .059787 .379252 .410556 .707007
.867431 .715785 .394118 .692346 .140620
.117452 .041595 .660000 .187439 .242397
.118963 .195654 .143001 .758753 .794041
.921585 .666743 .680684 .962852 .451551
.493819 .476072 .464366 .794543 .590479
.003320 .826695 .948643 .199436 .168108
.513488 .881553 .015403 .545605 .014511
.980862 .482645 .240284 .044499 .908896
.390947 .340735 .441318 .331851 .623241

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4221.htm (2 of 3) [5/7/2002 3:15:55 PM]

1.4.2.2.1. Background and Data

.941509 .498943 .548581 .886954 .199437

.548730 .809510 .040696 .382707 .742015
.123387 .250162 .529894 .624611 .797524
.914071 .961282 .966986 .102591 .748522
.053900 .387595 .186333 .253798 .145065
.713101 .024674 .054556 .142777 .938919
.740294 .390277 .557322 .709779 .017119
.525275 .802180 .814517 .541784 .561180
.993371 .430533 .512969 .561271 .925536
.040903 .116644 .988352 .079848 .275938
.171539 .099733 .344088 .461233 .483247
.792831 .249647 .100229 .536870 .323075
.754615 .020099 .690749 .413887 .637919
.763558 .404401 .105182 .161501 .848769
.091882 .009732 .825395 .270422 .086304
.833898 .737464 .278580 .900458 .549751
.981506 .549493 .881997 .918707 .615068
.476646 .731895 .020747 .677262 .696229
.064464 .271246 .701841 .361827 .757687
.649020 .971877 .499042 .912272 .953750
.587193 .823431 .540164 .405666 .281310
.030068 .227398 .207145 .329507 .706178
.083586 .991078 .542427 .851366 .158873
.046189 .755331 .223084 .283060 .326481
.333105 .914051 .007893 .326046 .047594
.119018 .538408 .623381 .594136 .285121
.590290 .284666 .879577 .762207 .917575
.374161 .613622 .695026 .390212 .557817
.651483 .483470 .894159 .269400 .397583
.911260 .717646 .489497 .230694 .541374
.775130 .382086 .864299 .016841 .482774
.519081 .398072 .893555 .195023 .717469
.979202 .885521 .029773 .742877 .525165
.344674 .218185 .931393 .278817 .570568

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4221.htm (3 of 3) [5/7/2002 3:15:55 PM]

1.4.2.2.2. Graphical Output and Interpretation

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.2. Uniform Random Numbers

1.4.2.2.2. Graphical Output and

Interpretation
Goal The goal of this analysis is threefold:
1. Determine if the univariate model:

is appropriate and valid.

2. Determine if the typical underlying assumptions for an "in
control" measurement process are valid. These assumptions are:
1. random drawings;
2. from a fixed distribution;
3. with the distribution having a fixed location; and
4. the distribution having a fixed scale.
3. Determine if the confidence interval

is appropriate and valid where s is the standard deviation of the

original data.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4222.htm (1 of 7) [5/7/2002 3:15:56 PM]

1.4.2.2.2. Graphical Output and Interpretation

4-Plot of
Data

Interpretation The assumptions are addressed by the graphics shown above:

1. The run sequence plot (upper left) indicates that the data do not
have any significant shifts in location or scale over time.
2. The lag plot (upper right) does not indicate any non-random
pattern in the data.
3. The histogram shows that the frequencies are relatively flat
across the range of the data. This suggests that the uniform
distribution might provide a better distributional fit than the
normal distribution.
4. The normal probability plot verifies that an assumption of
normality is not reasonable. In this case, the 4-plot should be
followed up by a uniform probability plot to determine if it
provides a better fit to the data. This is shown below.
From the above plots, we conclude that the underlying assumptions are
valid. Therefore, the model Yi = C + Ei is valid. However, since the
data are not normally distributed, using the mean as an estimate of C
and the confidence interval cited above for quantifying its uncertainty
are not valid or appropriate.

Individual Although it is usually not necessary, the plots can be generated

Plots individually to give more detail.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4222.htm (2 of 7) [5/7/2002 3:15:56 PM]

1.4.2.2.2. Graphical Output and Interpretation

Run
Sequence
Plot

Lag Plot

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4222.htm (3 of 7) [5/7/2002 3:15:56 PM]

1.4.2.2.2. Graphical Output and Interpretation

Histogram
(with
overlaid
Normal PDF)

This plot shows that a normal distribution is a poor fit. The flatness of
the histogram suggests that a uniform distribution might be a better fit.

Histogram
(with
overlaid
Uniform
PDF)

Since the histogram from the 4-plot suggested that the uniform
distribution might be a good fit, we overlay a uniform distribution on
top of the histogram. This indicates a much better fit than a normal
distribution.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4222.htm (4 of 7) [5/7/2002 3:15:56 PM]

1.4.2.2.2. Graphical Output and Interpretation

Normal
Probability
Plot

As with the histogram, the normal probability plot shows that the
normal distribution does not fit these data well.

Uniform
Probability
Plot

Since the above plots suggested that a uniform distribution might be

appropriate, we generate a uniform probability plot. This plot shows
that the uniform distribution provides an excellent fit to the data.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4222.htm (5 of 7) [5/7/2002 3:15:56 PM]

1.4.2.2.2. Graphical Output and Interpretation

Better Model Since the data follow the underlying assumptions, but with a uniform
distribution rather than a normal distribution, we would still like to
characterize C by a typical value plus or minus a confidence interval.
In this case, we would like to find a location estimator with the
smallest variability.
The bootstrap plot is an ideal tool for this purpose. The following plots
show the bootstrap plot, with the corresponding histogram, for the
mean, median, mid-range, and median absolute deviation.

Bootstrap
Plots

Mid-Range is From the above histograms, it is obvious that for these data, the
Best mid-range is far superior to the mean or median as an estimate for
location.
Using the mean, the location estimate is 0.507 and a 95% confidence
interval for the mean is (0.482,0.534). Using the mid-range, the
location estimate is 0.499 and the 95% confidence interval for the
mid-range is (0.497,0.503).
Although the values for the location are similar, the difference in the
uncertainty intervals is quite large.
Note that in the case of a uniform distribution it is known theoretically
that the mid-range is the best linear unbiased estimator for location.
However, in many applications, the most appropriate estimator will not
be known or it will be mathematically intractable to determine a valid
condfidence interval. The bootstrap provides a method for determining

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4222.htm (6 of 7) [5/7/2002 3:15:56 PM]

1.4.2.2.2. Graphical Output and Interpretation

(and comparing) confidence intervals in these cases.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4222.htm (7 of 7) [5/7/2002 3:15:56 PM]

1.4.2.2.3. Quantitative Output and Interpretation

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.2. Uniform Random Numbers

1.4.2.2.3. Quantitative Output and Interpretation

Summary As a first step in the analysis, a table of summary statistics is computed from the data. The
Statistics following table, generated by Dataplot, shows a typical set of statistics.

SUMMARY

NUMBER OF OBSERVATIONS = 500

***********************************************************************
* LOCATION MEASURES * DISPERSION MEASURES *
***********************************************************************
* MIDRANGE = 0.4997850E+00 * RANGE = 0.9945900E+00 *
* MEAN = 0.5078304E+00 * STAND. DEV. = 0.2943252E+00 *
* MIDMEAN = 0.5045621E+00 * AV. AB. DEV. = 0.2526468E+00 *
* MEDIAN = 0.5183650E+00 * MINIMUM = 0.2490000E-02 *
* = * LOWER QUART. = 0.2508093E+00 *
* = * LOWER HINGE = 0.2505935E+00 *
* = * UPPER HINGE = 0.7594775E+00 *
* = * UPPER QUART. = 0.7591152E+00 *
* = * MAXIMUM = 0.9970800E+00 *
***********************************************************************
* RANDOMNESS MEASURES * DISTRIBUTIONAL MEASURES *
***********************************************************************
* AUTOCO COEF = -0.3098569E-01 * ST. 3RD MOM. = -0.3443941E-01 *
* = 0.0000000E+00 * ST. 4TH MOM. = 0.1796969E+01 *
* = 0.0000000E+00 * ST. WILK-SHA = -0.2004886E+02 *
* = * UNIFORM PPCC = 0.9995682E+00 *
* = * NORMAL PPCC = 0.9771602E+00 *
* = * TUK -.5 PPCC = 0.7229201E+00 *
* = * CAUCHY PPCC = 0.3591767E+00 *
***********************************************************************

Note that under the distributional measures the uniform probability plot correlation
coefficient (PPCC) value is significantly larger than the normal PPCC value. This is
evidence that the uniform distribution fits these data better than does a normal distribution.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4223.htm (1 of 6) [5/7/2002 3:15:56 PM]

1.4.2.2.3. Quantitative Output and Interpretation

Location One way to quantify a change in location over time is to fit a straight line to the data set
using the index variable X = 1, 2, ..., N, with N denoting the number of observations. If
there is no significant drift in the location, the slope parameter should be zero. For this data
set, Dataplot generated the following output:

LEAST SQUARES MULTILINEAR FIT

SAMPLE SIZE N = 500
NUMBER OF VARIABLES = 1
NO REPLICATION CASE

PARAMETER ESTIMATES (APPROX. ST. DEV.) T VALUE

1 A0 0.522923 (0.2638E-01) 19.82
2 A1 X -0.602478E-04 (0.9125E-04) -0.6603

RESIDUAL STANDARD DEVIATION = 0.2944917

RESIDUAL DEGREES OF FREEDOM = 498

The slope parameter, A1, has a t value of -0.66 which is statistically not significant. This
indicates that the slope can in fact be considered zero.

Variation One simple way to detect a change in variation is with a Bartlett test after dividing the data
set into several equal-sized intervals. However, the Bartlett test is not robust for
non-normality. Since we know this data set is not approximated well by the normal
distribution, we use the alternative Levene test. In partiuclar, we use the Levene test based
on the median rather the mean. The choice of the number of intervals is somewhat arbitrary,
although values of 4 or 8 are reasonable. Dataplot generated the following output for the
Levene test.

LEVENE F-TEST FOR SHIFT IN VARIATION

(ASSUMPTION: NORMALITY)

1. STATISTICS
NUMBER OF OBSERVATIONS = 500
NUMBER OF GROUPS = 4
LEVENE F TEST STATISTIC = 0.7983007E-01

FOR LEVENE TEST STATISTIC

0 % POINT = 0.0000000E+00
50 % POINT = 0.7897459
75 % POINT = 1.373753
90 % POINT = 2.094885
95 % POINT = 2.622929
99 % POINT = 3.821479
99.9 % POINT = 5.506884

2.905608 % Point: 0.7983007E-01

3. CONCLUSION (AT THE 5% LEVEL):

THERE IS NO SHIFT IN VARIATION.
THUS: HOMOGENEOUS WITH RESPECT TO VARIATION.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4223.htm (2 of 6) [5/7/2002 3:15:56 PM]

1.4.2.2.3. Quantitative Output and Interpretation

In this case, the Levene test indicates that the standard deviations are not significantly
different in the 4 intervals.

Randomness
There are many ways in which data can be non-random. However, most common forms of
non-randomness can be detected with a few simple tests. The lag plot in the 4-plot in the
previous section is a simple graphical technique.
Another check is an autocorrelation plot that shows the autocorrelations for various lags.
Confidence bands can be plotted using 95% and 99% confidence levels. Points outside this
band indicate statistically significant values (lag 0 is always 1). Dataplot generated the
following autocorrelation plot.

The lag 1 autocorrelation, which is generally the one of most interest, is 0.03. The critical
values at the 5% significance level are -0.087 and 0.087. This indicates that the lag 1
autocorrelation is not statistically significant, so there is no evidence of non-randomness.
A common test for randomness is the runs test.

RUNS UP
STATISTIC = NUMBER OF RUNS UP
OF LENGTH EXACTLY I
I STAT EXP(STAT) SD(STAT) Z

1 103.0 104.2083 10.2792 -0.12

2 48.0 45.7167 5.2996 0.43
3 11.0 13.1292 3.2297 -0.66
4 6.0 2.8563 1.6351 1.92
5 0.0 0.5037 0.7045 -0.71
6 0.0 0.0749 0.2733 -0.27
7 1.0 0.0097 0.0982 10.08
8 0.0 0.0011 0.0331 -0.03
9 0.0 0.0001 0.0106 -0.01
10 0.0 0.0000 0.0032 0.00
STATISTIC = NUMBER OF RUNS UP
OF LENGTH I OR MORE
I STAT EXP(STAT) SD(STAT) Z

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4223.htm (3 of 6) [5/7/2002 3:15:56 PM]

1.4.2.2.3. Quantitative Output and Interpretation

1 169.0 166.5000 6.6546 0.38

2 66.0 62.2917 4.4454 0.83
3 18.0 16.5750 3.4338 0.41
4 7.0 3.4458 1.7786 2.00
5 1.0 0.5895 0.7609 0.54
6 1.0 0.0858 0.2924 3.13
7 1.0 0.0109 0.1042 9.49
8 0.0 0.0012 0.0349 -0.03
9 0.0 0.0001 0.0111 -0.01
10 0.0 0.0000 0.0034 0.00
RUNS DOWN
STATISTIC = NUMBER OF RUNS DOWN
OF LENGTH EXACTLY I
I STAT EXP(STAT) SD(STAT) Z

1 113.0 104.2083 10.2792 0.86

2 43.0 45.7167 5.2996 -0.51
3 11.0 13.1292 3.2297 -0.66
4 1.0 2.8563 1.6351 -1.14
5 0.0 0.5037 0.7045 -0.71
6 0.0 0.0749 0.2733 -0.27
7 0.0 0.0097 0.0982 -0.10
8 0.0 0.0011 0.0331 -0.03
9 0.0 0.0001 0.0106 -0.01
10 0.0 0.0000 0.0032 0.00
STATISTIC = NUMBER OF RUNS DOWN
OF LENGTH I OR MORE
I STAT EXP(STAT) SD(STAT) Z

1 168.0 166.5000 6.6546 0.23

2 55.0 62.2917 4.4454 -1.64
3 12.0 16.5750 3.4338 -1.33
4 1.0 3.4458 1.7786 -1.38
5 0.0 0.5895 0.7609 -0.77
6 0.0 0.0858 0.2924 -0.29
7 0.0 0.0109 0.1042 -0.10
8 0.0 0.0012 0.0349 -0.03
9 0.0 0.0001 0.0111 -0.01
10 0.0 0.0000 0.0034 0.00
RUNS TOTAL = RUNS UP + RUNS DOWN
STATISTIC = NUMBER OF RUNS TOTAL
OF LENGTH EXACTLY I
I STAT EXP(STAT) SD(STAT) Z

1 216.0 208.4167 14.5370 0.52

2 91.0 91.4333 7.4947 -0.06
3 22.0 26.2583 4.5674 -0.93
4 7.0 5.7127 2.3123 0.56
5 0.0 1.0074 0.9963 -1.01
6 0.0 0.1498 0.3866 -0.39
7 1.0 0.0193 0.1389 7.06
8 0.0 0.0022 0.0468 -0.05
9 0.0 0.0002 0.0150 -0.01
10 0.0 0.0000 0.0045 0.00
STATISTIC = NUMBER OF RUNS TOTAL
OF LENGTH I OR MORE
I STAT EXP(STAT) SD(STAT) Z

1 337.0 333.0000 9.4110 0.43

2 121.0 124.5833 6.2868 -0.57
3 30.0 33.1500 4.8561 -0.65

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4223.htm (4 of 6) [5/7/2002 3:15:56 PM]

1.4.2.2.3. Quantitative Output and Interpretation
4 8.0 6.8917 2.5154 0.44
5 1.0 1.1790 1.0761 -0.17
6 1.0 0.1716 0.4136 2.00
7 1.0 0.0217 0.1474 6.64
8 0.0 0.0024 0.0494 -0.05
9 0.0 0.0002 0.0157 -0.02
10 0.0 0.0000 0.0047 0.00
LENGTH OF THE LONGEST RUN UP = 7
LENGTH OF THE LONGEST RUN DOWN = 4
LENGTH OF THE LONGEST RUN UP OR DOWN = 7

NUMBER OF POSITIVE DIFFERENCES = 263

NUMBER OF NEGATIVE DIFFERENCES = 236
NUMBER OF ZERO DIFFERENCES = 0

Values in the column labeled "Z" greater than 1.96 or less than -1.96 are statistically
significant at the 5% level. This runs test does not indicate any significant non-randomness.
There is a statistically significant value for runs of length 7. However, further examination
of the table shows that there is in fact a single run of length 7 when near 0 are expected.
This is not sufficient evidence to conclude that the data are non-random.

Distributional Probability plots are a graphical test of assessing whether a particular distribution provides
Analysis an adequate fit to a data set.
A quantitative enhancement to the probability plot is the correlation coefficient of the points
on the probability plot. For this data set the correlation coefficient, from the summary table
above, is 0.977. Since this is less than the critical value of 0.987 (this is a tabulated value),
the normality assumption is rejected.
Chi-square and Kolmogorov-Smirnov goodness-of-fit tests are alternative methods for
assessing distributional adequacy. The Wilk-Shapiro and Anderson-Darling tests can be
used to test for normality. Dataplot generates the following output for the Anderson-Darling
normality test.

ANDERSON-DARLING 1-SAMPLE TEST

THAT THE DATA CAME FROM A NORMAL DISTRIBUTION

1. STATISTICS:
NUMBER OF OBSERVATIONS = 500
MEAN = 0.5078304
STANDARD DEVIATION = 0.2943252

ANDERSON-DARLING TEST STATISTIC VALUE = 5.719849

ADJUSTED TEST STATISTIC VALUE = 5.765036

2. CRITICAL VALUES:
90 % POINT = 0.6560000
95 % POINT = 0.7870000
97.5 % POINT = 0.9180000
99 % POINT = 1.092000

3. CONCLUSION (AT THE 5% LEVEL):

THE DATA DO NOT COME FROM A NORMAL DISTRIBUTION.
The Anderson-Darling test rejects the normality assumption because the value of the test
statistic, 5.72, is larger than the critical value of 1.092 at the 1% significance level.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4223.htm (5 of 6) [5/7/2002 3:15:56 PM]

1.4.2.2.3. Quantitative Output and Interpretation

Model Based on the graphical and quantitative analysis, we use the model
Yi = C + Ei
where C is estimated by the mid-range and the uncertainty interval for C is based on a
bootstrap analysis. Specifically,
C = 0.499
95% confidence limit for C = (0.497,0.503)

Univariate It is sometimes useful and convenient to summarize the above results in a report. The report
Report for the 500 uniform random numbers follows.

Analysis for 500 uniform random numbers

1: Sample Size = 500

2: Location
Mean = 0.50783
Standard Deviation of Mean = 0.013163
95% Confidence Interval for Mean = (0.48197,0.533692)
Drift with respect to location? = NO

3: Variation
Standard Deviation = 0.294326
95% Confidence Interval for SD = (0.277144,0.313796)
Drift with respect to variation?
(based on Levene's test on quarters
of the data) = NO

4: Distribution
Normal PPCC = 0.999569
Data are Normal?
(as measured by Normal PPCC) = NO

Uniform PPCC = 0.9995

Data are Uniform?
(as measured by Uniform PPCC) = YES

5: Randomness
Autocorrelation = -0.03099
Data are Random?
(as measured by autocorrelation) = YES

6: Statistical Control
(i.e., no drift in location or scale,
data is random, distribution is
fixed, here we are testing only for
fixed uniform)
Data Set is in Statistical Control? = YES

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4223.htm (6 of 6) [5/7/2002 3:15:56 PM]

1.4.2.2.4. Work This Example Yourself

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.2. Uniform Random Numbers

1.4.2.2.4. Work This Example Yourself

View This page allows you to repeat the analysis outlined in the case study
Dataplot description on the previous page using Dataplot . It is required that you
Macro for have already downloaded and installed Dataplot and configured your
this Case browser. to run Dataplot. Output from each analysis step below will be
Study displayed in one or more of the Dataplot windows. The four main
windows are the Output window, the Graphics window, the Command
History window, and the data sheet window. Across the top of the main
windows there are menus for executing Dataplot commands. Across the
bottom is a command entry window where commands can be typed in.

Data Analysis Steps Results and Conclusions

Click on the links below to start Dataplot and run this case study
yourself. Each step may use results from previous steps, so please be The links in this column will connect you with more detailed
patient. Wait until the software verifies that the current step is information about each analysis step from the case study description.
complete before clicking on the next step.

1. Invoke Dataplot and read data.

1. Read in the data. 1. You have read 1 column of numbers

into Dataplot, variable Y.

2. 4-plot of the data.

1. 4-plot of Y. 1. Based on the 4-plot, there are no shifts

in location or scale, and the data do not
seem to follow a normal distribution.

3. Generate the individual plots.

1. Generate a run sequence plot. 1. The run sequence plot indicates that
there are no shifts of location or
scale.

2. Generate a lag plot. 2. The lag plot does not indicate any
significant patterns (which would
show the data were not random).

3. Generate a histogram with an

3. The histogram indicates that a
overlaid normal pdf.
normal distribution is not a good

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4224.htm (1 of 3) [5/7/2002 3:15:57 PM]

1.4.2.2.4. Work This Example Yourself
distribution for these data.

4. Generate a histogram with an 4. The histogram indicates that a

overlaid uniform pdf. uniform distribution is a good
distribution for these data.

5. Generate a normal probability 5. The normal probability plot verifies

plot. that the normal distribution is not a
reasonable distribution for these data.

6. Generate a uniform probability 6. The uniform probability plot verifies

plot. that the uniform distribution is a
reasonable distribution for these data.

4. Generate the bootstrap plot.

1. Generate a bootstrap plot. 1. The bootstrap plot clearly shows

the superiority of the mid-range
over the mean and median as the
location estimator of choice for
this problem.

5. Generate summary statistics, quantitative

analysis, and print a univariate report.

1. Generate a table of summary 1. The summary statistics table displays

statistics. 25+ statistics.

2. Generate the mean, a confidence 2. The mean is 0.5078 and a 95%

interval for the mean, and compute confidence interval is (0.482,0.534).
a linear fit to detect drift in The linear fit indicates no drift in
location. location since the slope parameter is
statistically not significant.

3. Generate the standard deviation, a 3. The standard deviation is 0.29 with

confidence interval for the standard a 95% confidence interval of (0.277,0.314).
deviation, and detect drift in variation Levene's test indicates no significant
by dividing the data into quarters and drift in variation.
computing Barltetts test for equal
standard deviations.

4. Check for randomness by generating an 4. The lag 1 autocorrelation is -0.03.

autocorrelation plot and a runs test. From the autocorrelation plot, this is
within the 95% confidence interval
bands.

5. Check for normality by computing the 5. The uniform probability plot correlation
normal probability plot correlation coefficient is 0.9995. This indicates that
coefficient. the uniform distribution is a good fit.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4224.htm (2 of 3) [5/7/2002 3:15:57 PM]

1.4.2.2.4. Work This Example Yourself

6. Print a univariate report (this assumes 6. The results are summarized in a

steps 2 thru 6 have already been run). convenient report.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4224.htm (3 of 3) [5/7/2002 3:15:57 PM]

1.4.2.3. Random Walk

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies

1.4.2.3. Random Walk

Random This example illustrates the univariate analysis of a set of numbers
Walk derived from a random walk.

1. Background and Data

2. Test Underlying Assumptions
3. Develop Better Model
4. Validate New Model
5. Work This Example Yourself

http://www.itl.nist.gov/div898/handbook/eda/section4/eda423.htm [5/7/2002 3:15:57 PM]

1.4.2.3.1. Background and Data

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.3. Random Walk

1.4.2.3.1. Background and Data

Generation A random walk can be generated from a set of uniform random numbers
by the formula:

where U is a set of uniform random numbers.

The motivation for studying a set of random walk data is to illustrate the
effects of a known underlying autocorrelation structure (i.e.,
non-randomness) in the data.

Software Most general purpose statistical software programs, including Dataplot,

can generate data for a random walk.

Resulting The following is the set of random walk numbers used for this case
Data study.

-0.399027
-0.645651
-0.625516
-0.262049
-0.407173
-0.097583
0.314156
0.106905
-0.017675
-0.037111
0.357631
0.820111
0.844148
0.550509
0.090709

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4231.htm (1 of 12) [5/7/2002 3:15:57 PM]

1.4.2.3.1. Background and Data

0.413625
-0.002149
0.393170
0.538263
0.070583
0.473143
0.132676
0.109111
-0.310553
0.179637
-0.067454
-0.190747
-0.536916
-0.905751
-0.518984
-0.579280
-0.643004
-1.014925
-0.517845
-0.860484
-0.884081
-1.147428
-0.657917
-0.470205
-0.798437
-0.637780
-0.666046
-1.093278
-1.089609
-0.853439
-0.695306
-0.206795
-0.507504
-0.696903
-1.116358
-1.044534
-1.481004
-1.638390
-1.270400
-1.026477
-1.123380
-0.770683
-0.510481
-0.958825
-0.531959
-0.457141

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4231.htm (2 of 12) [5/7/2002 3:15:57 PM]

1.4.2.3.1. Background and Data

-0.226603
-0.201885
-0.078000
0.057733
-0.228762
-0.403292
-0.414237
-0.556689
-0.772007
-0.401024
-0.409768
-0.171804
-0.096501
-0.066854
0.216726
0.551008
0.660360
0.194795
-0.031321
0.453880
0.730594
1.136280
0.708490
1.149048
1.258757
1.102107
1.102846
0.720896
0.764035
1.072312
0.897384
0.965632
0.759684
0.679836
0.955514
1.290043
1.753449
1.542429
1.873803
2.043881
1.728635
1.289703
1.501481
1.888335
1.408421
1.416005

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4231.htm (3 of 12) [5/7/2002 3:15:57 PM]

1.4.2.3.1. Background and Data

0.929681
1.097632
1.501279
1.650608
1.759718
2.255664
2.490551
2.508200
2.707382
2.816310
3.254166
2.890989
2.869330
3.024141
3.291558
3.260067
3.265871
3.542845
3.773240
3.991880
3.710045
4.011288
4.074805
4.301885
3.956416
4.278790
3.989947
4.315261
4.200798
4.444307
4.926084
4.828856
4.473179
4.573389
4.528605
4.452401
4.238427
4.437589
4.617955
4.370246
4.353939
4.541142
4.807353
4.706447
4.607011
4.205943

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4231.htm (4 of 12) [5/7/2002 3:15:58 PM]

1.4.2.3.1. Background and Data

3.756457
3.482142
3.126784
3.383572
3.846550
4.228803
4.110948
4.525939
4.478307
4.457582
4.822199
4.605752
5.053262
5.545598
5.134798
5.438168
5.397993
5.838361
5.925389
6.159525
6.190928
6.024970
5.575793
5.516840
5.211826
4.869306
4.912601
5.339177
5.415182
5.003303
4.725367
4.350873
4.225085
3.825104
3.726391
3.301088
3.767535
4.211463
4.418722
4.554786
4.987701
4.993045
5.337067
5.789629
5.726147
5.934353

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4231.htm (5 of 12) [5/7/2002 3:15:58 PM]

1.4.2.3.1. Background and Data

5.641670
5.753639
5.298265
5.255743
5.500935
5.434664
5.588610
6.047952
6.130557
5.785299
5.811995
5.582793
5.618730
5.902576
6.226537
5.738371
5.449965
5.895537
6.252904
6.650447
7.025909
6.770340
7.182244
6.941536
7.368996
7.293807
7.415205
7.259291
6.970976
7.319743
6.850454
6.556378
6.757845
6.493083
6.824855
6.533753
6.410646
6.502063
6.264585
6.730889
6.753715
6.298649
6.048126
5.794463
5.539049
5.290072

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4231.htm (6 of 12) [5/7/2002 3:15:58 PM]

1.4.2.3.1. Background and Data

5.409699
5.843266
5.680389
5.185889
5.451353
5.003233
5.102844
5.566741
5.613668
5.352791
5.140087
4.999718
5.030444
5.428537
5.471872
5.107334
5.387078
4.889569
4.492962
4.591042
4.930187
4.857455
4.785815
5.235515
4.865727
4.855005
4.920206
4.880794
4.904395
4.795317
5.163044
4.807122
5.246230
5.111000
5.228429
5.050220
4.610006
4.489258
4.399814
4.606821
4.974252
5.190037
5.084155
5.276501
4.917121
4.534573

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4231.htm (7 of 12) [5/7/2002 3:15:58 PM]

1.4.2.3.1. Background and Data

4.076168
4.236168
3.923607
3.666004
3.284967
2.980621
2.623622
2.882375
3.176416
3.598001
3.764744
3.945428
4.408280
4.359831
4.353650
4.329722
4.294088
4.588631
4.679111
4.182430
4.509125
4.957768
4.657204
4.325313
4.338800
4.720353
4.235756
4.281361
3.795872
4.276734
4.259379
3.999663
3.544163
3.953058
3.844006
3.684740
3.626058
3.457909
3.581150
4.022659
4.021602
4.070183
4.457137
4.156574
4.205304
4.514814

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4231.htm (8 of 12) [5/7/2002 3:15:58 PM]

1.4.2.3.1. Background and Data

4.055510
3.938217
4.180232
3.803619
3.553781
3.583675
3.708286
4.005810
4.419880
4.881163
5.348149
4.950740
5.199262
4.753162
4.640757
4.327090
4.080888
3.725953
3.939054
3.463728
3.018284
2.661061
3.099980
3.340274
3.230551
3.287873
3.497652
3.014771
3.040046
3.342226
3.656743
3.698527
3.759707
4.253078
4.183611
4.196580
4.257851
4.683387
4.224290
3.840934
4.329286
3.909134
3.685072
3.356611
2.956344
2.800432

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4231.htm (9 of 12) [5/7/2002 3:15:58 PM]

1.4.2.3.1. Background and Data

2.761665
2.744913
3.037743
2.787390
2.387619
2.424489
2.247564
2.502179
2.022278
2.213027
2.126914
2.264833
2.528391
2.432792
2.037974
1.699475
2.048244
1.640126
1.149858
1.475253
1.245675
0.831979
1.165877
1.403341
1.181921
1.582379
1.632130
2.113636
2.163129
2.545126
2.963833
3.078901
3.055547
3.287442
2.808189
2.985451
3.181679
2.746144
2.517390
2.719231
2.581058
2.838745
2.987765
3.459642
3.458684
3.870956

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4231.htm (10 of 12) [5/7/2002 3:15:58 PM]

1.4.2.3.1. Background and Data

4.324706
4.411899
4.735330
4.775494
4.681160
4.462470
3.992538
3.719936
3.427081
3.256588
3.462766
3.046353
3.537430
3.579857
3.931223
3.590096
3.136285
3.391616
3.114700
2.897760
2.724241
2.557346
2.971397
2.479290
2.305336
1.852930
1.471948
1.510356
1.633737
1.727873
1.512994
1.603284
1.387950
1.767527
2.029734
2.447309
2.321470
2.435092
2.630118
2.520330
2.578147
2.729630
2.713100
3.107260
2.876659
2.774242

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4231.htm (11 of 12) [5/7/2002 3:15:58 PM]

1.4.2.3.1. Background and Data

3.185503
3.403148
3.392646
3.123339
3.164713
3.439843
3.321929
3.686229
3.203069
3.185843
3.204924
3.102996
3.496552
3.191575
3.409044
3.888246
4.273767
3.803540
4.046417
4.071581
3.916256
3.634441
4.065834
3.844651
3.915219

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4231.htm (12 of 12) [5/7/2002 3:15:58 PM]

1.4.2.3.2. Test Underlying Assumptions

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.3. Random Walk

1.4.2.3.2. Test Underlying Assumptions

Goal The goal of this analysis is threefold:
1. Determine if the univariate model:

is appropriate and valid.

2. Determine if the typical underlying assumptions for an "in control" measurement
process are valid. These assumptions are:
1. random drawings;
2. from a fixed distribution;
3. with the distribution having a fixed location; and
4. the distribution having a fixed scale.
3. Determine if the confidence interval

is appropriate and valid, with s denoting the standard deviation of the original data.

4-Plot of Data

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4232.htm (1 of 7) [5/7/2002 3:15:58 PM]

1.4.2.3.2. Test Underlying Assumptions

Interpretation The assumptions are addressed by the graphics shown above:

1. The run sequence plot (upper left) indicates significant shifts in location over time.
2. The lag plot (upper right) indicates significant non-randomness in the data.
3. When the assumptions of randomness and constant location and scale are not satisfied,
the distributional assumptions are not meaningful. Therefore we do not attempt to make
any interpretation of the histogram (lower left) or the normal probability plot (lower
right).
From the above plots, we conclude that the underlying assumptions are seriously violated.
Therefore the Yi = C + Ei model is not valid.

When the randomness assumption is seriously violated, a time series model may be
appropriate. The lag plot often suggests a reasonable model. For example, in this case the
strongly linear appearance of the lag plot suggests a model fitting Yi versus Yi-1 might be
appropriate. When the data are non-random, it is helpful to supplement the lag plot with an
autocorrelation plot and a spectral plot. Although in this case the lag plot is enough to suggest
an appropriate model, we provide the autocorrelation and spectral plots for comparison.

Autocorrelation When the lag plot indicates significant non-randomness, it can be helpful to follow up with a
Plot an autocorrelation plot.

This autocorrelation plot shows significant autocorrelation at lags 1 through 100 in a linearly
decreasing fashion.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4232.htm (2 of 7) [5/7/2002 3:15:58 PM]

1.4.2.3.2. Test Underlying Assumptions

Spectral Plot Another useful plot for non-random data is the spectral plot.

This spectral plot shows a single dominant low frequency peak.

Quantitative Although the 4-plot above clearly shows the violation of the assumptions, we supplement the
Output graphical output with some quantitative measures.

Summary As a first step in the analysis, a table of summary statistics is computed from the data. The
Statistics following table, generated by Dataplot, shows a typical set of statistics.

SUMMARY

NUMBER OF OBSERVATIONS = 500

***********************************************************************
* LOCATION MEASURES * DISPERSION MEASURES *
***********************************************************************
* MIDRANGE = 0.2888407E+01 * RANGE = 0.9053595E+01 *
* MEAN = 0.3216681E+01 * STAND. DEV. = 0.2078675E+01 *
* MIDMEAN = 0.4791331E+01 * AV. AB. DEV. = 0.1660585E+01 *
* MEDIAN = 0.3612030E+01 * MINIMUM = -0.1638390E+01 *
* = * LOWER QUART. = 0.1747245E+01 *
* = * LOWER HINGE = 0.1741042E+01 *
* = * UPPER HINGE = 0.4682273E+01 *
* = * UPPER QUART. = 0.4681717E+01 *
* = * MAXIMUM = 0.7415205E+01 *
***********************************************************************
* RANDOMNESS MEASURES * DISTRIBUTIONAL MEASURES *
***********************************************************************
* AUTOCO COEF = 0.9868608E+00 * ST. 3RD MOM. = -0.4448926E+00 *
* = 0.0000000E+00 * ST. 4TH MOM. = 0.2397789E+01 *
* = 0.0000000E+00 * ST. WILK-SHA = -0.1279870E+02 *
* = * UNIFORM PPCC = 0.9765666E+00 *
* = * NORMAL PPCC = 0.9811183E+00 *
* = * TUK -.5 PPCC = 0.7754489E+00 *
* = * CAUCHY PPCC = 0.4165502E+00 *
***********************************************************************

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4232.htm (3 of 7) [5/7/2002 3:15:58 PM]

1.4.2.3.2. Test Underlying Assumptions

The value of the autocorrelation statistic, 0.987, is evidence of a very strong autocorrelation.

Location One way to quantify a change in location over time is to fit a straight line to the data set using
the index variable X = 1, 2, ..., N, with N denoting the number of observations. If there is no
significant drift in the location, the slope parameter should be zero. For this data set, Dataplot
generates the following output:

LEAST SQUARES MULTILINEAR FIT

SAMPLE SIZE N = 500
NUMBER OF VARIABLES = 1
NO REPLICATION CASE

PARAMETER ESTIMATES (APPROX. ST. DEV.) T VALUE

1 A0 1.83351 (0.1721 ) 10.65
2 A1 X 0.552164E-02 (0.5953E-03) 9.275

RESIDUAL STANDARD DEVIATION = 1.921416

RESIDUAL DEGREES OF FREEDOM = 498

COEF AND SD(COEF) WRITTEN OUT TO FILE DPST1F.DAT

SD(PRED),95LOWER,95UPPER,99LOWER,99UPPER
WRITTEN OUT TO FILE DPST2F.DAT
REGRESSION DIAGNOSTICS WRITTEN OUT TO FILE DPST3F.DAT
PARAMETER VARIANCE-COVARIANCE MATRIX AND
INVERSE OF X-TRANSPOSE X MATRIX
WRITTEN OUT TO FILE DPST4F.DAT
The slope parameter, A1, has a t value of 9.3 which is statistically significant. This indicates
that the slope cannot in fact be considered zero and so the conclusion is that we do not have
constant location.

Variation One simple way to detect a change in variation is with a Bartlett test after dividing the data set
into several equal-sized intervals. However, the Bartlett test is not robust for non-normality.
Since we know this data set is not approximated well by the normal distribution, we use the
alternative Levene test. In partiuclar, we use the Levene test based on the median rather the
mean. The choice of the number of intervals is somewhat arbitrary, although values of 4 or 8
are reasonable. Dataplot generated the following output for the Levene test.

LEVENE F-TEST FOR SHIFT IN VARIATION

(ASSUMPTION: NORMALITY)

1. STATISTICS
NUMBER OF OBSERVATIONS = 500
NUMBER OF GROUPS = 4
LEVENE F TEST STATISTIC = 10.45940

FOR LEVENE TEST STATISTIC

0 % POINT = 0.0000000E+00
50 % POINT = 0.7897459
75 % POINT = 1.373753
90 % POINT = 2.094885
95 % POINT = 2.622929
99 % POINT = 3.821479
99.9 % POINT = 5.506884

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4232.htm (4 of 7) [5/7/2002 3:15:58 PM]

1.4.2.3.2. Test Underlying Assumptions
99.99989 % Point: 10.45940

3. CONCLUSION (AT THE 5% LEVEL):

THERE IS A SHIFT IN VARIATION.
THUS: NOT HOMOGENEOUS WITH RESPECT TO VARIATION.

In this case, the Levene test indicates that the standard deviations are significantly different in
the 4 intervals since the test statistic of 10.46 is greater than the 95% critical value of 2.62.
Therefore we conclude that the scale is not constant.

Randomness Although the lag 1 autocorrelation coefficient above clearly shows the non-randomness, we
show the output from a runs test as well.

RUNS UP

STATISTIC = NUMBER OF RUNS UP

OF LENGTH EXACTLY I

I STAT EXP(STAT) SD(STAT) Z

1 63.0 104.2083 10.2792 -4.01

2 34.0 45.7167 5.2996 -2.21
3 17.0 13.1292 3.2297 1.20
4 4.0 2.8563 1.6351 0.70
5 1.0 0.5037 0.7045 0.70
6 5.0 0.0749 0.2733 18.02
7 1.0 0.0097 0.0982 10.08
8 1.0 0.0011 0.0331 30.15
9 0.0 0.0001 0.0106 -0.01
10 1.0 0.0000 0.0032 311.40

STATISTIC = NUMBER OF RUNS UP

OF LENGTH I OR MORE

I STAT EXP(STAT) SD(STAT) Z

1 127.0 166.5000 6.6546 -5.94

2 64.0 62.2917 4.4454 0.38
3 30.0 16.5750 3.4338 3.91
4 13.0 3.4458 1.7786 5.37
5 9.0 0.5895 0.7609 11.05
6 8.0 0.0858 0.2924 27.06
7 3.0 0.0109 0.1042 28.67
8 2.0 0.0012 0.0349 57.21
9 1.0 0.0001 0.0111 90.14
10 1.0 0.0000 0.0034 298.08

RUNS DOWN

STATISTIC = NUMBER OF RUNS DOWN

OF LENGTH EXACTLY I

I STAT EXP(STAT) SD(STAT) Z

1 69.0 104.2083 10.2792 -3.43

2 32.0 45.7167 5.2996 -2.59
3 11.0 13.1292 3.2297 -0.66
4 6.0 2.8563 1.6351 1.92
5 5.0 0.5037 0.7045 6.38
6 2.0 0.0749 0.2733 7.04

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4232.htm (5 of 7) [5/7/2002 3:15:58 PM]

1.4.2.3.2. Test Underlying Assumptions
7 2.0 0.0097 0.0982 20.26
8 0.0 0.0011 0.0331 -0.03
9 0.0 0.0001 0.0106 -0.01
10 0.0 0.0000 0.0032 0.00

STATISTIC = NUMBER OF RUNS DOWN

OF LENGTH I OR MORE

I STAT EXP(STAT) SD(STAT) Z

1 127.0 166.5000 6.6546 -5.94

2 58.0 62.2917 4.4454 -0.97
3 26.0 16.5750 3.4338 2.74
4 15.0 3.4458 1.7786 6.50
5 9.0 0.5895 0.7609 11.05
6 4.0 0.0858 0.2924 13.38
7 2.0 0.0109 0.1042 19.08
8 0.0 0.0012 0.0349 -0.03
9 0.0 0.0001 0.0111 -0.01
10 0.0 0.0000 0.0034 0.00

RUNS TOTAL = RUNS UP + RUNS DOWN

STATISTIC = NUMBER OF RUNS TOTAL

OF LENGTH EXACTLY I

I STAT EXP(STAT) SD(STAT) Z

1 132.0 208.4167 14.5370 -5.26

2 66.0 91.4333 7.4947 -3.39
3 28.0 26.2583 4.5674 0.38
4 10.0 5.7127 2.3123 1.85
5 6.0 1.0074 0.9963 5.01
6 7.0 0.1498 0.3866 17.72
7 3.0 0.0193 0.1389 21.46
8 1.0 0.0022 0.0468 21.30
9 0.0 0.0002 0.0150 -0.01
10 1.0 0.0000 0.0045 220.19

STATISTIC = NUMBER OF RUNS TOTAL

OF LENGTH I OR MORE

I STAT EXP(STAT) SD(STAT) Z

1 254.0 333.0000 9.4110 -8.39

2 122.0 124.5833 6.2868 -0.41
3 56.0 33.1500 4.8561 4.71
4 28.0 6.8917 2.5154 8.39
5 18.0 1.1790 1.0761 15.63
6 12.0 0.1716 0.4136 28.60
7 5.0 0.0217 0.1474 33.77
8 2.0 0.0024 0.0494 40.43
9 1.0 0.0002 0.0157 63.73
10 1.0 0.0000 0.0047 210.77

LENGTH OF THE LONGEST RUN UP = 10

LENGTH OF THE LONGEST RUN DOWN = 7
LENGTH OF THE LONGEST RUN UP OR DOWN = 10

NUMBER OF POSITIVE DIFFERENCES = 258

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4232.htm (6 of 7) [5/7/2002 3:15:58 PM]

1.4.2.3.2. Test Underlying Assumptions
NUMBER OF NEGATIVE DIFFERENCES = 241
NUMBER OF ZERO DIFFERENCES = 0

Values in the column labeled "Z" greater than 1.96 or less than -1.96 are statistically
significant at the 5% level. Numerous values in this column are much larger than +/-1.96, so
we conclude that the data are not random.

Distributional Since the quantitative tests show that the assumptions of randomness and constant location and
Assumptions scale are not met, the distributional measures will not be meaningful. Therefore these
quantitative tests are omitted.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4232.htm (7 of 7) [5/7/2002 3:15:58 PM]

1.4.2.3.3. Develop A Better Model

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.3. Random Walk

1.4.2.3.3. Develop A Better Model

Lag Plot Since the underlying assumptions did not hold, we need to develop a better model.
Suggests
Better The lag plot showed a distinct linear pattern. Given the definition of the lag plot, Yi versus
Model Yi-1, a good candidate model is a model of the form

Fit A linear fit of this model generated the following output.

Output
LEAST SQUARES MULTILINEAR FIT
SAMPLE SIZE N = 499
NUMBER OF VARIABLES = 1
NO REPLICATION CASE

PARAMETER ESTIMATES (APPROX. ST. DEV.) T VALUE

1 A0 0.501650E-01 (0.2417E-01) 2.075
2 A1 YIM1 0.987087 (0.6313E-02) 156.4

RESIDUAL STANDARD DEVIATION = 0.2931194

RESIDUAL DEGREES OF FREEDOM = 497

The slope parameter, A1, has a t value of 156.4 which is statistically significant. Also, the
residual standard deviation is 0.29. This can be compared to the standard deviation shown in
the summary table, which is 2.08. That is, the fit to the autoregressive model has reduced the
variability by a factor of 7.

Time This model is an example of a time series model. More extensive discussion of time series is
Series given in the Process Monitoring chapter.
Model

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4233.htm [5/7/2002 3:15:58 PM]

1.4.2.3.4. Validate New Model

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.3. Random Walk

1.4.2.3.4. Validate New Model

Plot The first step in verifying the model is to plot the predicted values from
Predicted the fit with the original data.
with Original
Data

This plot indicates a reasonably good fit.

Test In addition to the plot of the predicted values, the residual standard
Underlying deviation from the fit also indicates a significant improvement for the
Assumptions model. The next step is to validate the underlying assumptions for the
on the error component, or residuals, from this model.
Residuals

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4234.htm (1 of 4) [5/7/2002 3:15:59 PM]

1.4.2.3.4. Validate New Model

4-Plot of
Residuals

Interpretation The assumptions are addressed by the graphics shown above:

1. The run sequence plot (upper left) indicates no significant shifts
in location or scale over time.
2. The lag plot (upper right) exhibits a random appearance.
3. The histogram shows a relatively flat appearance. This indicates
that a uniform probability distribution may be an appropriate
model for the error component (or residuals).
4. The normal probability plot clearly shows that the normal
distribution is not an appropriate model for the error component.
A uniform probability plot can be used to further test the suggestion
that a uniform distribution might be a good model for the error
component.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4234.htm (2 of 4) [5/7/2002 3:15:59 PM]

1.4.2.3.4. Validate New Model

Uniform
Probability
Plot of
Residuals

Since the uniform probability plot is nearly linear, this verifies that a
uniform distribution is a good model for the error component.

Conclusions Since the residuals from our model satisfy the underlying assumptions,
we conlude that

where the Ei follow a uniform distribution is a good model for this data
set. We could simplify this model to

This has the advantage of simplicity (the current point is simply the
previous point plus a uniformly distributed error term).

Using In this case, the above model makes sense based on our definition of
Scientific and the random walk. That is, a random walk is the cumulative sum of
Engineering uniformly distributed data points. It makes sense that modeling the
Knowledge current point as the previous point plus a uniformly distributed error
term is about as good as we can do. Although this case is a bit artificial
in that we knew how the data were constructed, it is common and
desirable to use scientific and engineering knowledge of the process
that generated the data in formulating and testing models for the data.
Quite often, several competing models will produce nearly equivalent
mathematical results. In this case, selecting the model that best
approximates the scientific understanding of the process is a reasonable
choice.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4234.htm (3 of 4) [5/7/2002 3:15:59 PM]

1.4.2.3.4. Validate New Model

Time Series This model is an example of a time series model. More extensive
Model discussion of time series is given in the Process Monitoring chapter.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4234.htm (4 of 4) [5/7/2002 3:15:59 PM]

1.4.2.3.5. Work This Example Yourself

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.3. Random Walk

1.4.2.3.5. Work This Example Yourself

View This page allows you to repeat the analysis outlined in the case study
Dataplot description on the previous page using Dataplot . It is required that you
Macro for have already downloaded and installed Dataplot and configured your
this Case browser. to run Dataplot. Output from each analysis step below will be
Study displayed in one or more of the Dataplot windows. The four main
windows are the Output window, the Graphics window, the Command
History window, and the data sheet window. Across the top of the main
windows there are menus for executing Dataplot commands. Across the
bottom is a command entry window where commands can be typed in.

Data Analysis Steps Results and Conclusions

Click on the links below to start Dataplot and run this case
The links in this column will connect you with more detailed
study yourself. Each step may use results from previous steps,
information about each analysis step from the case study
so please be patient. Wait until the software verifies that the
description.
current step is complete before clicking on the next step.

1. Invoke Dataplot and read data.

1. Read in the data. 1. You have read 1 column of numbers

into Dataplot, variable Y.

2. Validate assumptions.

1. 4-plot of Y. 1. Based on the 4-plot, there are shifts

in location and scale and the data are not
random.

2. Generate a table of summary 2. The summary statistics table displays

statistics. 25+ statistics.

3. Generate a linear fit to detect 3. The linear fit indicates drift in

drift in location. location since the slope parameter
is statistically significant.

4. Detect drift in variation by 4. Levene's test indicates significant

dividing the data into quarters and drift in variation.
computing Levene's test for equal

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4235.htm (1 of 2) [5/7/2002 3:15:59 PM]

1.4.2.3.5. Work This Example Yourself
standard deviations.

5. The runs test indicates significant

5. Check for randomness by generating non-randomness.
a runs test.

3. Generate the randomness plots.

1. Generate an autocorrelation plot. 1. The autocorrelation plot shows

significant autocorrelation at lag 1.

2. Generate a spectral plot. 2. The spectral plot shows a single dominant

low frequency peak.

4. Fit Yi = A0 + A1*Yi-1 + Ei
and validate.

1. Generate the fit. 1. The residual standard deviation from the

fit is 0.29 (compared to the standard
deviation of 2.08 from the original
data).

2. Plot fitted line with original data. 2. The plot of the predicted values with
the original data indicates a good fit.

3. Generate a 4-plot of the residuals 3. The 4-plot indicates that the assumptions
from the fit. of constant location and scale are valid.
The lag plot indicates that the data are
random. However, the histogram and normal
probability plot indicate that the uniform
disribution might be a better model for
the residuals than the normal
distribution.

4. Generate a uniform probability plot

of the residuals. 4. The uniform probability plot verifies
that the residuals can be fit by a
uniform distribution.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4235.htm (2 of 2) [5/7/2002 3:15:59 PM]

1.4.2.4. Josephson Junction Cryothermometry

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies

1.4.2.4. Josephson Junction

Cryothermometry
Josephson Junction This example illustrates the univariate analysis of Josephson
Cryothermometry junction cyrothermometry.

1. Background and Data

2. Graphical Output and Interpretation
3. Quantitative Output and Interpretation
4. Work This Example Yourself

http://www.itl.nist.gov/div898/handbook/eda/section4/eda424.htm [5/7/2002 3:15:59 PM]

1.4.2.4.1. Background and Data

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.4. Josephson Junction Cryothermometry

1.4.2.4.1. Background and Data

Generation This data set was collected by Bob Soulen of NIST in October, 1971 as
a sequence of observations collected equi-spaced in time from a volt
meter to ascertain the process temperature in a Josephson junction
cryothermometry (low temperature) experiment. The response variable
is voltage counts.

Motivation The motivation for studying this data set is to illustrate the case where
there is discreteness in the measurements, but the underlying
assumptions hold. In this case, the discreteness is due to the data being
integers.

This file can be read by Dataplot with the following commands:

SKIP 25
SET READ FORMAT 5F5.0
READ SOULEN.DAT Y
SET READ FORMAT

Resulting The following are the data used for this case study.
Data
2899 2898 2898 2900 2898
2901 2899 2901 2900 2898
2898 2898 2898 2900 2898
2897 2899 2897 2899 2899
2900 2897 2900 2900 2899
2898 2898 2899 2899 2899
2899 2899 2898 2899 2899
2899 2902 2899 2900 2898
2899 2899 2899 2899 2899
2899 2900 2899 2900 2898
2901 2900 2899 2899 2899
2899 2899 2900 2899 2898
2898 2898 2900 2896 2897

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4241.htm (1 of 4) [5/7/2002 3:16:07 PM]

1.4.2.4.1. Background and Data

2899 2899 2900 2898 2900

2901 2898 2899 2901 2900
2898 2900 2899 2899 2897
2899 2898 2899 2899 2898
2899 2897 2899 2899 2897
2899 2897 2899 2897 2897
2899 2897 2898 2898 2899
2897 2898 2897 2899 2899
2898 2898 2897 2898 2895
2897 2898 2898 2896 2898
2898 2897 2896 2898 2898
2897 2897 2898 2898 2896
2898 2898 2896 2899 2898
2898 2898 2899 2899 2898
2898 2899 2899 2899 2900
2900 2901 2899 2898 2898
2900 2899 2898 2901 2897
2898 2898 2900 2899 2899
2898 2898 2899 2898 2901
2900 2897 2897 2898 2898
2900 2898 2899 2898 2898
2898 2896 2895 2898 2898
2898 2898 2897 2897 2895
2897 2897 2900 2898 2896
2897 2898 2898 2899 2898
2897 2898 2898 2896 2900
2899 2898 2896 2898 2896
2896 2896 2897 2897 2896
2897 2897 2896 2898 2896
2898 2896 2897 2896 2897
2897 2898 2897 2896 2895
2898 2896 2896 2898 2896
2898 2898 2897 2897 2898
2897 2899 2896 2897 2899
2900 2898 2898 2897 2898
2899 2899 2900 2900 2900
2900 2899 2899 2899 2898
2900 2901 2899 2898 2900
2901 2901 2900 2899 2898
2901 2899 2901 2900 2901
2898 2900 2900 2898 2900
2900 2898 2899 2901 2900
2899 2899 2900 2900 2899
2900 2901 2899 2898 2898
2899 2896 2898 2897 2898
2898 2897 2897 2897 2898

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4241.htm (2 of 4) [5/7/2002 3:16:07 PM]

1.4.2.4.1. Background and Data

2897 2899 2900 2899 2897

2898 2900 2900 2898 2898
2899 2900 2898 2900 2900
2898 2900 2898 2898 2898
2898 2898 2899 2898 2900
2897 2899 2898 2899 2898
2897 2900 2901 2899 2898
2898 2901 2898 2899 2897
2899 2897 2896 2898 2898
2899 2900 2896 2897 2897
2898 2899 2899 2898 2898
2897 2897 2898 2897 2897
2898 2898 2898 2896 2895
2898 2898 2898 2896 2898
2898 2898 2897 2897 2899
2896 2900 2897 2897 2898
2896 2897 2898 2898 2898
2897 2897 2898 2899 2897
2898 2899 2897 2900 2896
2899 2897 2898 2897 2900
2899 2900 2897 2897 2898
2897 2899 2899 2898 2897
2901 2900 2898 2901 2899
2900 2899 2898 2900 2900
2899 2898 2897 2900 2898
2898 2897 2899 2898 2900
2899 2898 2899 2897 2900
2898 2902 2897 2898 2899
2899 2899 2898 2897 2898
2897 2898 2899 2900 2900
2899 2898 2899 2900 2899
2900 2899 2899 2899 2899
2899 2898 2899 2899 2900
2902 2899 2900 2900 2901
2899 2901 2899 2899 2902
2898 2898 2898 2898 2899
2899 2900 2900 2900 2898
2899 2899 2900 2899 2900
2899 2900 2898 2898 2898
2900 2898 2899 2900 2899
2899 2900 2898 2898 2899
2899 2899 2899 2898 2898
2897 2898 2899 2897 2897
2901 2898 2897 2898 2899
2898 2897 2899 2898 2897
2898 2898 2897 2898 2899

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4241.htm (3 of 4) [5/7/2002 3:16:07 PM]

1.4.2.4.1. Background and Data

2899 2899 2899 2900 2899

2899 2897 2898 2899 2900
2898 2897 2901 2899 2901
2898 2899 2901 2900 2900
2899 2900 2900 2900 2900
2901 2900 2901 2899 2897
2900 2900 2901 2899 2898
2900 2899 2899 2900 2899
2900 2899 2900 2899 2901
2900 2900 2899 2899 2898
2899 2900 2898 2899 2899
2901 2898 2898 2900 2899
2899 2898 2897 2898 2897
2899 2899 2899 2898 2898
2897 2898 2899 2897 2897
2899 2898 2898 2899 2899
2901 2899 2899 2899 2897
2900 2896 2898 2898 2900
2897 2899 2897 2896 2898
2897 2898 2899 2896 2899
2901 2898 2898 2896 2897
2899 2897 2898 2899 2898
2898 2898 2898 2898 2898
2899 2900 2899 2901 2898
2899 2899 2898 2900 2898
2899 2899 2901 2900 2901
2899 2901 2899 2901 2899
2900 2902 2899 2898 2899
2900 2899 2900 2900 2901
2900 2899 2901 2901 2899
2898 2901 2897 2898 2901
2900 2902 2899 2900 2898
2900 2899 2900 2899 2899
2899 2898 2900 2898 2899
2899 2899 2899 2898 2900

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4241.htm (4 of 4) [5/7/2002 3:16:07 PM]

1.4.2.4.2. Graphical Output and Interpretation

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.4. Josephson Junction Cryothermometry

1.4.2.4.2. Graphical Output and

Interpretation
Goal The goal of this analysis is threefold:
1. Determine if the univariate model:

is appropriate and valid.

2. Determine if the typical underlying assumptions for an "in
control" measurement process are valid. These assumptions are:
1. random drawings;
2. from a fixed distribution;
3. with the distribution having a fixed location; and
4. the distribution having a fixed scale.
3. Determine if the confidence interval

is appropriate and valid where s is the standard deviation of the

original data.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4242.htm (1 of 4) [5/7/2002 3:16:08 PM]

1.4.2.4.2. Graphical Output and Interpretation

4-Plot of
Data

Interpretation The assumptions are addressed by the graphics shown above:

1. The run sequence plot (upper left) indicates that the data do not
have any significant shifts in location or scale over time.
2. The lag plot (upper right) does not indicate any non-random
pattern in the data.
3. The histogram (lower left) shows that the data are reasonably
symmetric, there does not appear to be significant outliers in the
tails, and that it is reasonable to assume that the data can be fit
with a normal distribution.
4. The normal probability plot (lower right) is difficult to interpret
due to the fact that there are only a few distinct values with
many repeats.
The integer data with only a few distinct values and many repeats
accounts for the discrete appearance of several of the plots (e.g., the lag
plot and the normal probability plot). In this case, the nature of the data
makes the normal probability plot difficult to interpret, especially since
each number is repeated many times. However, the histogram indicates
that a normal distribution should provide an adequate model for the
data.
From the above plots, we conclude that the underlying assumptions are
valid and the data can be reasonably approximated with a normal
distribution. Therefore, the commonly used uncertainty standard is
valid and appropriate. The numerical values for this model are given in

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4242.htm (2 of 4) [5/7/2002 3:16:08 PM]

1.4.2.4.2. Graphical Output and Interpretation

the Quantitative Output and Interpretation section.

Individual Although it is normally not necessary, the plots can be generated

Plots individually to give more detail.

Run
Sequence
Plot

Lag Plot

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4242.htm (3 of 4) [5/7/2002 3:16:08 PM]

1.4.2.4.2. Graphical Output and Interpretation

Histogram
(with
overlaid
Normal PDF)

Normal
Probability
Plot

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4242.htm (4 of 4) [5/7/2002 3:16:08 PM]

1.4.2.4.3. Quantitative Output and Interpretation

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.4. Josephson Junction Cryothermometry

1.4.2.4.3. Quantitative Output and Interpretation

Summary As a first step in the analysis, a table of summary statistics is computed from the data. The
Statistics following table, generated by Dataplot, shows a typical set of statistics.

SUMMARY

NUMBER OF OBSERVATIONS = 140

***********************************************************************
* LOCATION MEASURES * DISPERSION MEASURES *
***********************************************************************
* MIDRANGE = 0.2899000E+04 * RANGE = 0.6000000E+01 *
* MEAN = 0.2898721E+04 * STAND. DEV. = 0.1235377E+01 *
* MIDMEAN = 0.2898457E+04 * AV. AB. DEV. = 0.9642857E+00 *
* MEDIAN = 0.2899000E+04 * MINIMUM = 0.2896000E+04 *
* = * LOWER QUART. = 0.2898000E+04 *
* = * LOWER HINGE = 0.2898000E+04 *
* = * UPPER HINGE = 0.2899500E+04 *
* = * UPPER QUART. = 0.2899250E+04 *
* = * MAXIMUM = 0.2902000E+04 *
***********************************************************************
* RANDOMNESS MEASURES * DISTRIBUTIONAL MEASURES *
***********************************************************************
* AUTOCO COEF = 0.2925397E+00 * ST. 3RD MOM. = 0.1271097E+00 *
* = 0.0000000E+00 * ST. 4TH MOM. = 0.2571418E+01 *
* = 0.0000000E+00 * ST. WILK-SHA = -0.3911592E+01 *
* = * UNIFORM PPCC = 0.9580541E+00 *
* = * NORMAL PPCC = 0.9701443E+00 *
* = * TUK -.5 PPCC = 0.8550686E+00 *
* = * CAUCHY PPCC = 0.6239791E+00 *
***********************************************************************

Location One way to quantify a change in location over time is to fit a straight line to the data set using
the index variable X = 1, 2, ..., N, with N denoting the number of observations. If there is no
significant drift in the location, the slope parameter should be zero. For this data set, Dataplot
generates the following output:

LEAST SQUARES MULTILINEAR FIT

SAMPLE SIZE N = 140
NUMBER OF VARIABLES = 1
NO REPLICATION CASE

PARAMETER ESTIMATES (APPROX. ST. DEV.) T VALUE

1 A0 2898.34 (0.2074 ) 0.1398E+05
2 A1 X 0.539896E-02 (0.2552E-02) 2.116

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4243.htm (1 of 7) [5/7/2002 3:16:08 PM]

1.4.2.4.3. Quantitative Output and Interpretation

RESIDUAL STANDARD DEVIATION = 1.220212

RESIDUAL DEGREES OF FREEDOM = 138

COEF AND SD(COEF) WRITTEN OUT TO FILE DPST1F.DAT

SD(PRED),95LOWER,95UPPER,99LOWER,99UPPER
WRITTEN OUT TO FILE DPST2F.DAT
REGRESSION DIAGNOSTICS WRITTEN OUT TO FILE DPST3F.DAT
PARAMETER VARIANCE-COVARIANCE MATRIX AND
INVERSE OF X-TRANSPOSE X MATRIX
WRITTEN OUT TO FILE DPST4F.DAT
The slope parameter, A1, has a t value of 2.1 which is statistically significant (the critical value
is 1.98). However, the value of the slope is 0.0054. Given that the slope is nearly zero, the
assumption of constant location is not seriously violated even though it is (just barely)
statistically significant.

Variation One simple way to detect a change in variation is with a Bartlett test after dividing the data set
into several equal-sized intervals. However, the Bartlett test is not robust for non-normality.
Since the nature of the data (a few distinct points repeated many times) makes the normality
assumption questionable, we use the alternative Levene test. In partiuclar, we use the Levene
test based on the median rather the mean. The choice of the number of intervals is somewhat
arbitrary, although values of 4 or 8 are reasonable. Dataplot generated the following output for
the Levene test.

LEVENE F-TEST FOR SHIFT IN VARIATION

(ASSUMPTION: NORMALITY)

1. STATISTICS
NUMBER OF OBSERVATIONS = 140
NUMBER OF GROUPS = 4
LEVENE F TEST STATISTIC = 0.4128718

FOR LEVENE TEST STATISTIC

0 % POINT = 0.0000000E+00
50 % POINT = 0.7926317
75 % POINT = 1.385201
90 % POINT = 2.124494
95 % POINT = 2.671178
99 % POINT = 3.928924
99.9 % POINT = 5.737571

25.59809 % Point: 0.4128718

3. CONCLUSION (AT THE 5% LEVEL):

THERE IS NO SHIFT IN VARIATION.
THUS: HOMOGENEOUS WITH RESPECT TO VARIATION.
Since the Levene test statistic value of 0.41 is less than the 95% critical value of 2.67, we
conclude that the standard deviations are not significantly different in the 4 intervals.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4243.htm (2 of 7) [5/7/2002 3:16:08 PM]

1.4.2.4.3. Quantitative Output and Interpretation

Randomness There are many ways in which data can be non-random. However, most common forms of
non-randomness can be detected with a few simple tests. The lag plot in the previous section is
a simple graphical technique.
Another check is an autocorrelation plot that shows the autocorrelations for various lags.
Confidence bands can be plotted at the 95% and 99% confidence levels. Points outside this
band indicate statistically significant values (lag 0 is always 1). Dataplot generated the
following autocorrelation plot.

The lag 1 autocorrelation, which is generally the one of most interest, is 0.29. The critical
values at the 5% level of significance are -0.087 and 0.087. This indicates that the lag 1
autocorrelation is statistically significant, so there is some evidence for non-randomness.
A common test for randomness is the runs test.

RUNS UP

STATISTIC = NUMBER OF RUNS UP

OF LENGTH EXACTLY I

I STAT EXP(STAT) SD(STAT) Z

1 15.0 29.2083 5.4233 -2.62

2 10.0 12.7167 2.7938 -0.97
3 2.0 3.6292 1.6987 -0.96
4 4.0 0.7849 0.8573 3.75
5 2.0 0.1376 0.3683 5.06
6 0.0 0.0204 0.1425 -0.14
7 1.0 0.0026 0.0511 19.54
8 0.0 0.0003 0.0172 -0.02
9 0.0 0.0000 0.0055 -0.01
10 0.0 0.0000 0.0017 0.00

STATISTIC = NUMBER OF RUNS UP

OF LENGTH I OR MORE

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4243.htm (3 of 7) [5/7/2002 3:16:08 PM]

1.4.2.4.3. Quantitative Output and Interpretation
I STAT EXP(STAT) SD(STAT) Z

1 34.0 46.5000 3.5048 -3.57

2 19.0 17.2917 2.3477 0.73
3 9.0 4.5750 1.8058 2.45
4 7.0 0.9458 0.9321 6.49
5 3.0 0.1609 0.3976 7.14
6 1.0 0.0233 0.1524 6.41
7 1.0 0.0029 0.0542 18.41
8 0.0 0.0003 0.0181 -0.02
9 0.0 0.0000 0.0057 -0.01
10 0.0 0.0000 0.0017 0.00

RUNS DOWN

STATISTIC = NUMBER OF RUNS DOWN

OF LENGTH EXACTLY I

I STAT EXP(STAT) SD(STAT) Z

1 16.0 29.2083 5.4233 -2.44

2 10.0 12.7167 2.7938 -0.97
3 5.0 3.6292 1.6987 0.81
4 1.0 0.7849 0.8573 0.25
5 2.0 0.1376 0.3683 5.06
6 0.0 0.0204 0.1425 -0.14
7 0.0 0.0026 0.0511 -0.05
8 0.0 0.0003 0.0172 -0.02
9 0.0 0.0000 0.0055 -0.01
10 0.0 0.0000 0.0017 0.00

STATISTIC = NUMBER OF RUNS DOWN

OF LENGTH I OR MORE

I STAT EXP(STAT) SD(STAT) Z

1 34.0 46.5000 3.5048 -3.57

2 18.0 17.2917 2.3477 0.30
3 8.0 4.5750 1.8058 1.90
4 3.0 0.9458 0.9321 2.20
5 2.0 0.1609 0.3976 4.63
6 0.0 0.0233 0.1524 -0.15
7 0.0 0.0029 0.0542 -0.05
8 0.0 0.0003 0.0181 -0.02
9 0.0 0.0000 0.0057 -0.01
10 0.0 0.0000 0.0017 0.00

RUNS TOTAL = RUNS UP + RUNS DOWN

STATISTIC = NUMBER OF RUNS TOTAL

OF LENGTH EXACTLY I

I STAT EXP(STAT) SD(STAT) Z

1 31.0 58.4167 7.6697 -3.57

2 20.0 25.4333 3.9510 -1.38
3 7.0 7.2583 2.4024 -0.11
4 5.0 1.5698 1.2124 2.83
5 4.0 0.2752 0.5208 7.15
6 0.0 0.0407 0.2015 -0.20

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4243.htm (4 of 7) [5/7/2002 3:16:08 PM]

1.4.2.4.3. Quantitative Output and Interpretation
7 1.0 0.0052 0.0722 13.78
8 0.0 0.0006 0.0243 -0.02
9 0.0 0.0001 0.0077 -0.01
10 0.0 0.0000 0.0023 0.00

STATISTIC = NUMBER OF RUNS TOTAL

OF LENGTH I OR MORE

I STAT EXP(STAT) SD(STAT) Z

1 68.0 93.0000 4.9565 -5.04

2 37.0 34.5833 3.3202 0.73
3 17.0 9.1500 2.5537 3.07
4 10.0 1.8917 1.3182 6.15
5 5.0 0.3218 0.5623 8.32
6 1.0 0.0466 0.2155 4.42
7 1.0 0.0059 0.0766 12.98
8 0.0 0.0007 0.0256 -0.03
9 0.0 0.0001 0.0081 -0.01
10 0.0 0.0000 0.0024 0.00

LENGTH OF THE LONGEST RUN UP = 7

LENGTH OF THE LONGEST RUN DOWN = 5
LENGTH OF THE LONGEST RUN UP OR DOWN = 7

NUMBER OF POSITIVE DIFFERENCES = 48

NUMBER OF NEGATIVE DIFFERENCES = 49
NUMBER OF ZERO DIFFERENCES = 42
Values in the column labeled "Z" greater than 1.96 or less than -1.96 are statistically significant
at the 5% level. The runs test indicates some mild non-randomness.
Although the runs test and lag 1 autocorrelation indicate some mild non-randomness, it is not
sufficient to reject the Yi = C + Ei model. At least part of the non-randomness can be explained
by the discrete nature of the data.

Distributional Probability plots are a graphical test for assessing if a particular distribution provides an
Analysis adequate fit to a data set.
A quantitative enhancement to the probability plot is the correlation coefficient of the points on
the probability plot. For this data set the correlation coefficient is 0.970. Since this is less than
the critical value of 0.987 (this is a tabulated value), the normality assumption is rejected.
Chi-square and Kolmogorov-Smirnov goodness-of-fit tests are alternative methods for
assessing distributional adequacy. The Wilk-Shapiro and Anderson-Darling tests can be used to
test for normality. Dataplot generates the following output for the Anderson-Darling normality
test.

ANDERSON-DARLING 1-SAMPLE TEST

THAT THE DATA CAME FROM A NORMAL DISTRIBUTION

1. STATISTICS:
NUMBER OF OBSERVATIONS = 140
MEAN = 2898.721
STANDARD DEVIATION = 1.235377

ANDERSON-DARLING TEST STATISTIC VALUE = 3.839233

ADJUSTED TEST STATISTIC VALUE = 3.944029

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4243.htm (5 of 7) [5/7/2002 3:16:08 PM]

1.4.2.4.3. Quantitative Output and Interpretation
2. CRITICAL VALUES:
90 % POINT = 0.6560000
95 % POINT = 0.7870000
97.5 % POINT = 0.9180000
99 % POINT = 1.092000

3. CONCLUSION (AT THE 5% LEVEL):

THE DATA DO NOT COME FROM A NORMAL DISTRIBUTION.
The Anderson-Darling test rejects the normality assumption because the test statistic, 3.84, is
greater than the 99% critical value 1.092.
Although the data are not strictly normal, the violation of the normality assumption is not severe
enough to conclude that the Yi = C + Ei model is unreasonable. At least part of the
non-normality can be explained by the discrete nature of the data.

Outlier A test for outliers is the Grubbs test. Dataplot generated the following output for Grubbs' test.
Analysis
GRUBBS TEST FOR OUTLIERS
(ASSUMPTION: NORMALITY)

1. STATISTICS:
NUMBER OF OBSERVATIONS = 140
MINIMUM = 2896.000
MEAN = 2898.721
MAXIMUM = 2902.000
STANDARD DEVIATION = 1.235377

GRUBBS TEST STATISTIC = 2.653898

2. PERCENT POINTS OF THE REFERENCE DISTRIBUTION

FOR GRUBBS TEST STATISTIC
0 % POINT = 0.0000000E+00
50 % POINT = 2.874578
75 % POINT = 3.074733
90 % POINT = 3.320834
95 % POINT = 3.495103
99 % POINT = 3.867364

3. CONCLUSION (AT THE 5% LEVEL):

THERE ARE NO OUTLIERS.
For this data set, Grubbs' test does not detect any outliers at the 10%, 5%, and 1% significance
levels.

Model Although the randomness and normality assumptions were mildly violated, we conclude that a
reasonable model for the data is:

In addition, a 95% confidence interval for the mean value is (2898.515,2898.928).

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4243.htm (6 of 7) [5/7/2002 3:16:08 PM]

1.4.2.4.3. Quantitative Output and Interpretation

Univariate It is sometimes useful and convenient to summarize the above results in a report.
Report
Analysis for Josephson Junction Cryothermometry Data

1: Sample Size = 140

2: Location
Mean = 2898.722
Standard Deviation of Mean = 0.104409
95% Confidence Interval for Mean = (2898.515,2898.928)
Drift with respect to location? = YES
(Further analysis indicates that
the drift, while statistically
significant, is not practically
significant)

3: Variation
Standard Deviation = 1.235377
95% Confidence Interval for SD = (1.105655,1.399859)
Drift with respect to variation?
(based on Levene's test on quarters
of the data) = NO

4: Distribution
Normal PPCC = 0.970145
Data are Normal?
(as measured by Normal PPCC) = NO

5: Randomness
Autocorrelation = 0.29254
Data are Random?
(as measured by autocorrelation) = NO

6: Statistical Control
(i.e., no drift in location or scale,
data are random, distribution is
fixed, here we are testing only for
fixed normal)
Data Set is in Statistical Control? = NO

Note: Although we have violations of

the assumptions, they are mild enough,
and at least partially explained by the
discrete nature of the data, so we may model
the data as if it were in statistical
control

7: Outliers?
(as determined by Grubbs test) = NO

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4243.htm (7 of 7) [5/7/2002 3:16:08 PM]

1.4.2.4.4. Work This Example Yourself

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.4. Josephson Junction Cryothermometry

1.4.2.4.4. Work This Example Yourself

View This page allows you to repeat the analysis outlined in the case study
Dataplot description on the previous page using Dataplot . It is required that you
Macro for have already downloaded and installed Dataplot and configured your
this Case browser. to run Dataplot. Output from each analysis step below will be
Study displayed in one or more of the Dataplot windows. The four main
windows are the Output window, the Graphics window, the Command
History window, and the data sheet window. Across the top of the main
windows there are menus for executing Dataplot commands. Across the
bottom is a command entry window where commands can be typed in.

Data Analysis Steps Results and Conclusions

Click on the links below to start Dataplot and run this case study
yourself. Each step may use results from previous steps, so please be The links in this column will connect you with more detailed
patient. Wait until the software verifies that the current step is information about each analysis step from the case study description.
complete before clicking on the next step.

1. Invoke Dataplot and read data.

1. Read in the data. 1. You have read 1 column of numbers

into Dataplot, variable Y.

2. 4-plot of the data.

1. 4-plot of Y. 1. Based on the 4-plot, there are no shifts

in location or scale. Due to the nature
of the data (a few distinct points with
many repeats), the normality assumption is
questionable.

3. Generate the individual plots.

1. Generate a run sequence plot. 1. The run sequence plot indicates that
there are no shifts of location or
scale.

2. Generate a lag plot. 2. The lag plot does not indicate any
significant patterns (which would
show the data were not random).

3. Generate a histogram with an

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4244.htm (1 of 2) [5/7/2002 3:16:09 PM]

1.4.2.4.4. Work This Example Yourself
overlaid normal pdf. 3. The histogram indicates that a
normal distribution is a good
distribution for these data.
4. Generate a normal probability
plot. 4. The discrete nature of the data masks
the normality or non-normality of the
data somewhat. The plot indicates that
a normal distribution provides a rough
approximation for the data.

4. Generate summary statistics, quantitative

analysis, and print a univariate report.

1. Generate a table of summary 1. The summary statistics table displays

statistics. 25+ statistics.

2. Generate the mean, a confidence 2. The mean is 2898.27 and a 95%

interval for the mean, and compute confidence interval is (2898.52,2898.93).
a linear fit to detect drift in The linear fit indicates no meaningful drift
location. in location since the value of the slope
parameter is near zero.

3. Generate the standard deviation, a 3. The standard devaition is 1.24 with

confidence interval for the standard a 95% confidence interval of (1.11,1.40).
deviation, and detect drift in variation Levene's test indicates no significant
by dividing the data into quarters and drift in variation.
computing Levene's test for equal
standard deviations.

4. Check for randomness by generating an 4. The lag 1 autocorrelation is 0.29.

autocorrelation plot and a runs test. This indicates some mild non-randomness.

5. Check for normality by computing the 5. The normal probability plot correlation
normal probability plot correlation coefficient is 0.970. At the 5% level,
coefficient. we reject the normality assumption.

6. Check for outliers using Grubbs' test. 6. Grubbs' test detects no outliers at the
5% level.

7. Print a univariate report (this assumes 7. The results are summarized in a

steps 2 thru 6 have already been run). convenient report.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4244.htm (2 of 2) [5/7/2002 3:16:09 PM]

1.4.2.5. Beam Deflections

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies

1.4.2.5. Beam Deflections

Beam
This example illustrates the univariate analysis of beam deflection data.
Deflection

1. Background and Data

2. Test Underlying Assumptions
3. Develop a Better Model
4. Validate New Model
5. Work This Example Yourself

http://www.itl.nist.gov/div898/handbook/eda/section4/eda425.htm [5/7/2002 3:16:09 PM]

1.4.2.5.1. Background and Data

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.5. Beam Deflections

1.4.2.5.1. Background and Data

Generation This data set was collected by H. S. Lew of NIST in 1969 to measure
steel-concrete beam deflections. The response variable is the deflection
of a beam from the center point.
The motivation for studying this data set is to show how the underlying
assumptions are affected by periodic data.
This file can be read by Dataplot with the following commands:
SKIP 25
READ LEW.DAT Y

Resulting The following are the data used for this case study.
Data
-213
-564
-35
-15
141
115
-420
-360
203
-338
-431
194
-220
-513
154
-125
-559
92
-21
-579

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4251.htm (1 of 6) [5/7/2002 3:16:09 PM]

1.4.2.5.1. Background and Data

-52
99
-543
-175
162
-457
-346
204
-300
-474
164
-107
-572
-8
83
-541
-224
180
-420
-374
201
-236
-531
83
27
-564
-112
131
-507
-254
199
-311
-495
143
-46
-579
-90
136
-472
-338
202
-287
-477
169
-124
-568

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4251.htm (2 of 6) [5/7/2002 3:16:09 PM]

1.4.2.5.1. Background and Data

17
48
-568
-135
162
-430
-422
172
-74
-577
-13
92
-534
-243
194
-355
-465
156
-81
-578
-64
139
-449
-384
193
-198
-538
110
-44
-577
-6
66
-552
-164
161
-460
-344
205
-281
-504
134
-28
-576
-118
156
-437

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4251.htm (3 of 6) [5/7/2002 3:16:09 PM]

1.4.2.5.1. Background and Data

-381
200
-220
-540
83
11
-568
-160
172
-414
-408
188
-125
-572
-32
139
-492
-321
205
-262
-504
142
-83
-574
0
48
-571
-106
137
-501
-266
190
-391
-406
194
-186
-553
83
-13
-577
-49
103
-515
-280
201
300

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4251.htm (4 of 6) [5/7/2002 3:16:09 PM]

1.4.2.5.1. Background and Data

-506
131
-45
-578
-80
138
-462
-361
201
-211
-554
32
74
-533
-235
187
-372
-442
182
-147
-566
25
68
-535
-244
194
-351
-463
174
-125
-570
15
72
-550
-190
172
-424
-385
198
-218
-536
96

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4251.htm (5 of 6) [5/7/2002 3:16:09 PM]

1.4.2.5.1. Background and Data

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4251.htm (6 of 6) [5/7/2002 3:16:09 PM]

1.4.2.5.2. Test Underlying Assumptions

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.5. Beam Deflections

1.4.2.5.2. Test Underlying Assumptions

Goal The goal of this analysis is threefold:
1. Determine if the univariate model:

is appropriate and valid.

2. Determine if the typical underlying assumptions for an "in control" measurement process
are valid. These assumptions are:
1. random drawings;
2. from a fixed distribution;
3. with the distribution having a fixed location; and
4. the distribution having a fixed scale.
3. Determine if the confidence interval

is appropriate and valid where s is the standard deviation of the original data.

4-Plot of Data

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4252.htm (1 of 8) [5/7/2002 3:16:10 PM]

1.4.2.5.2. Test Underlying Assumptions

Interpretation The assumptions are addressed by the graphics shown above:

1. The run sequence plot (upper left) indicates that the data do not have any significant
shifts in location or scale over time.
2. The lag plot (upper right) shows that the data are not random. The lag plot further
indicates the presence of a few outliers.
3. When the randomness assumption is thus seriously violated, the histogram (lower left)
and normal probability plot (lower right) are ignored since determining the distribution of
the data is only meaningful when the data are random.
From the above plots we conclude that the underlying randomness assumption is not valid.
Therefore, the model

is not appropriate.
We need to develop a better model. Non-random data can frequently be modeled using time
series mehtodology. Specifically, the circular pattern in the lag plot indicates that a sinusoidal
model might be appropriate. The sinusoidal model will be developed in the next section.

Individual The plots can be generated individually for more detail. In this case, only the run sequence plot
Plots and the lag plot are drawn since the distributional plots are not meaningful.

Run Sequence
Plot

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4252.htm (2 of 8) [5/7/2002 3:16:10 PM]

1.4.2.5.2. Test Underlying Assumptions

Lag Plot

We have drawn some lines and boxes on the plot to better isolate the outliers. The following
output helps identify the points that are generating the outliers on the lag plot.

****************************************************
** print y index xplot yplot subset yplot > 250 **
****************************************************

VARIABLES--Y INDEX XPLOT YPLOT

300.00 158.00 -506.00 300.00

****************************************************
** print y index xplot yplot subset xplot > 250 **
****************************************************

VARIABLES--Y INDEX XPLOT YPLOT

201.00 157.00 300.00 201.00

********************************************************
** print y index xplot yplot subset yplot -100 to 0
subset xplot -100 to 0 **
********************************************************

VARIABLES--Y INDEX XPLOT YPLOT

-35.00 3.00 -15.00 -35.00

*********************************************************
** print y index xplot yplot subset yplot 100 to 200
subset xplot 100 to 200 **
*********************************************************

VARIABLES--Y INDEX XPLOT YPLOT

141.00 5.00 115.00 141.00

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4252.htm (3 of 8) [5/7/2002 3:16:10 PM]

1.4.2.5.2. Test Underlying Assumptions

That is, the third, fifth, and 158th points appear to be outliers.

Autocorrelation When the lag plot indicates significant non-randomness, it can be helpful to follow up with a an
Plot autocorrelation plot.

This autocorrelation plot shows a distinct cyclic pattern. As with the lag plot, this suggests a
sinusoidal model.

Spectral Plot Another useful plot for non-random data is the spectral plot.

This spectral plot shows a single dominant peak at a frequency of 0.3. This frequency of 0.3
will be used in fitting the sinusoidal model in the next section.

Quantitative Although the lag plot, autocorrelation plot, and spectral plot clearly show the violation of the
Output randomness assumption, we supplement the graphical output with some quantitative measures.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4252.htm (4 of 8) [5/7/2002 3:16:10 PM]

1.4.2.5.2. Test Underlying Assumptions

Summary As a first step in the analysis, a table of summary statistics is computed from the data. The
Statistics following table, generated by Dataplot, shows a typical set of statistics.

SUMMARY

NUMBER OF OBSERVATIONS = 200

***********************************************************************
* LOCATION MEASURES * DISPERSION MEASURES *
***********************************************************************
* MIDRANGE = -0.1395000E+03 * RANGE = 0.8790000E+03 *
* MEAN = -0.1774350E+03 * STAND. DEV. = 0.2773322E+03 *
* MIDMEAN = -0.1797600E+03 * AV. AB. DEV. = 0.2492250E+03 *
* MEDIAN = -0.1620000E+03 * MINIMUM = -0.5790000E+03 *
* = * LOWER QUART. = -0.4510000E+03 *
* = * LOWER HINGE = -0.4530000E+03 *
* = * UPPER HINGE = 0.9400000E+02 *
* = * UPPER QUART. = 0.9300000E+02 *
* = * MAXIMUM = 0.3000000E+03 *
***********************************************************************
* RANDOMNESS MEASURES * DISTRIBUTIONAL MEASURES *
***********************************************************************
* AUTOCO COEF = -0.3073048E+00 * ST. 3RD MOM. = -0.5010057E-01 *
* = 0.0000000E+00 * ST. 4TH MOM. = 0.1503684E+01 *
* = 0.0000000E+00 * ST. WILK-SHA = -0.1883372E+02 *
* = * UNIFORM PPCC = 0.9925535E+00 *
* = * NORMAL PPCC = 0.9540811E+00 *
* = * TUK -.5 PPCC = 0.7313794E+00 *
* = * CAUCHY PPCC = 0.4408355E+00 *
***********************************************************************

Location One way to quantify a change in location over time is to fit a straight line to the data set using
the index variable X = 1, 2, ..., N, with N denoting the number of observations. If there is no
significant drift in the location, the slope parameter should be zero. For this data set, Dataplot
generates the following output:

LEAST SQUARES MULTILINEAR FIT

SAMPLE SIZE N = 200
NUMBER OF VARIABLES = 1
NO REPLICATION CASE

PARAMETER ESTIMATES (APPROX. ST. DEV.) T VALUE

1 A0 -178.175 ( 39.47 ) -4.514
2 A1 X 0.736593E-02 (0.3405 ) 0.2163E-01

RESIDUAL STANDARD DEVIATION = 278.0313

RESIDUAL DEGREES OF FREEDOM = 198
The slope parameter, A1, has a t value of 0.022 which is statistically not significant. This
indicates that the slope can in fact be considered zero.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4252.htm (5 of 8) [5/7/2002 3:16:10 PM]

1.4.2.5.2. Test Underlying Assumptions

Variation One simple way to detect a change in variation is with a Bartlett test after dividing the data set
into several equal-sized intervals. However, the Bartlett the non-randomness of this data does
not allows us to assume normality, we use the alternative Levene test. In partiuclar, we use the
Levene test based on the median rather the mean. The choice of the number of intervals is
somewhat arbitrary, although values of 4 or 8 are reasonable. Dataplot generated the following
output for the Levene test.

LEVENE F-TEST FOR SHIFT IN VARIATION

(ASSUMPTION: NORMALITY)

1. STATISTICS
NUMBER OF OBSERVATIONS = 200
NUMBER OF GROUPS = 4
LEVENE F TEST STATISTIC = 0.9378599E-01

FOR LEVENE TEST STATISTIC

0 % POINT = 0.0000000E+00
50 % POINT = 0.7914120
75 % POINT = 1.380357
90 % POINT = 2.111936
95 % POINT = 2.650676
99 % POINT = 3.883083
99.9 % POINT = 5.638597

3.659895 % Point: 0.9378599E-01

3. CONCLUSION (AT THE 5% LEVEL):

THERE IS NO SHIFT IN VARIATION.
THUS: HOMOGENEOUS WITH RESPECT TO VARIATION.
In this case, the Levene test indicates that the standard deviations are significantly different in
the 4 intervals since the test statistic of 13.2 is greater than the 95% critical value of 2.6.
Therefore we conclude that the scale is not constant.

Randomness A runs test is used to check for randomness

RUNS UP

STATISTIC = NUMBER OF RUNS UP

OF LENGTH EXACTLY I

I STAT EXP(STAT) SD(STAT) Z

1 63.0 104.2083 10.2792 -4.01

2 34.0 45.7167 5.2996 -2.21
3 17.0 13.1292 3.2297 1.20
4 4.0 2.8563 1.6351 0.70
5 1.0 0.5037 0.7045 0.70
6 5.0 0.0749 0.2733 18.02
7 1.0 0.0097 0.0982 10.08
8 1.0 0.0011 0.0331 30.15
9 0.0 0.0001 0.0106 -0.01
10 1.0 0.0000 0.0032 311.40

STATISTIC = NUMBER OF RUNS UP

OF LENGTH I OR MORE

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4252.htm (6 of 8) [5/7/2002 3:16:10 PM]

1.4.2.5.2. Test Underlying Assumptions
I STAT EXP(STAT) SD(STAT) Z

1 127.0 166.5000 6.6546 -5.94

2 64.0 62.2917 4.4454 0.38
3 30.0 16.5750 3.4338 3.91
4 13.0 3.4458 1.7786 5.37
5 9.0 0.5895 0.7609 11.05
6 8.0 0.0858 0.2924 27.06
7 3.0 0.0109 0.1042 28.67
8 2.0 0.0012 0.0349 57.21
9 1.0 0.0001 0.0111 90.14
10 1.0 0.0000 0.0034 298.08

RUNS DOWN

STATISTIC = NUMBER OF RUNS DOWN

OF LENGTH EXACTLY I

I STAT EXP(STAT) SD(STAT) Z

1 69.0 104.2083 10.2792 -3.43

2 32.0 45.7167 5.2996 -2.59
3 11.0 13.1292 3.2297 -0.66
4 6.0 2.8563 1.6351 1.92
5 5.0 0.5037 0.7045 6.38
6 2.0 0.0749 0.2733 7.04
7 2.0 0.0097 0.0982 20.26
8 0.0 0.0011 0.0331 -0.03
9 0.0 0.0001 0.0106 -0.01
10 0.0 0.0000 0.0032 0.00

STATISTIC = NUMBER OF RUNS DOWN

OF LENGTH I OR MORE

I STAT EXP(STAT) SD(STAT) Z

1 127.0 166.5000 6.6546 -5.94

2 58.0 62.2917 4.4454 -0.97
3 26.0 16.5750 3.4338 2.74
4 15.0 3.4458 1.7786 6.50
5 9.0 0.5895 0.7609 11.05
6 4.0 0.0858 0.2924 13.38
7 2.0 0.0109 0.1042 19.08
8 0.0 0.0012 0.0349 -0.03
9 0.0 0.0001 0.0111 -0.01
10 0.0 0.0000 0.0034 0.00

RUNS TOTAL = RUNS UP + RUNS DOWN

STATISTIC = NUMBER OF RUNS TOTAL

OF LENGTH EXACTLY I

I STAT EXP(STAT) SD(STAT) Z

1 132.0 208.4167 14.5370 -5.26

2 66.0 91.4333 7.4947 -3.39
3 28.0 26.2583 4.5674 0.38
4 10.0 5.7127 2.3123 1.85
5 6.0 1.0074 0.9963 5.01
6 7.0 0.1498 0.3866 17.72
7 3.0 0.0193 0.1389 21.46

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4252.htm (7 of 8) [5/7/2002 3:16:10 PM]

1.4.2.5.2. Test Underlying Assumptions
8 1.0 0.0022 0.0468 21.30
9 0.0 0.0002 0.0150 -0.01
10 1.0 0.0000 0.0045 220.19

STATISTIC = NUMBER OF RUNS TOTAL

OF LENGTH I OR MORE

I STAT EXP(STAT) SD(STAT) Z

1 254.0 333.0000 9.4110 -8.39

2 122.0 124.5833 6.2868 -0.41
3 56.0 33.1500 4.8561 4.71
4 28.0 6.8917 2.5154 8.39
5 18.0 1.1790 1.0761 15.63
6 12.0 0.1716 0.4136 28.60
7 5.0 0.0217 0.1474 33.77
8 2.0 0.0024 0.0494 40.43
9 1.0 0.0002 0.0157 63.73
10 1.0 0.0000 0.0047 210.77

LENGTH OF THE LONGEST RUN UP = 10

LENGTH OF THE LONGEST RUN DOWN = 7
LENGTH OF THE LONGEST RUN UP OR DOWN = 10

NUMBER OF POSITIVE DIFFERENCES = 258

NUMBER OF NEGATIVE DIFFERENCES = 241
NUMBER OF ZERO DIFFERENCES = 0

Values in the column labeled "Z" greater than 1.96 or less than -1.96 are statistically significant
at the 5% level. Numerous values in this column are much larger than +/-1.96, so we conclude
that the data are not random.

Distributional Since the quantitative tests show that the assumptions of constant scale and non-randomness are
Assumptions not met, the distributional measures will not be meaningful. Therefore these quantitative tests
are omitted.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4252.htm (8 of 8) [5/7/2002 3:16:10 PM]

1.4.2.5.3. Develop a Better Model

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.5. Beam Deflections

1.4.2.5.3. Develop a Better Model

Sinusoidal The lag plot and autocorrelation plot in the previous section strongly suggested a
Model sinusoidal model might be appropriate. The basic sinusoidal model is:

where C is constant defining a mean level, is an amplitude for the sine

function, is the frequency, Ti is a time variable, and is the phase. This
sinusoidal model can be fit using non-linear least squares.

To obtain a good fit, sinusoidal models require good starting values for C, the
amplitude, and the frequency.

Good Starting A good starting value for C can be obtained by calculating the mean of the data.
Value for C If the data show a trend, i.e., the assumption of constant location is violated, we
can replace C with a linear or quadratic least squares fit. That is, the model
becomes

or

Since our data did not have any meaningful change of location, we can fit the
simpler model with C equal to the mean. From the summary output in the
previous page, the mean is -177.44.

Good Starting The starting value for the frequency can be obtained from the spectral plot,
Value for which shows the dominant frequency is about 0.3.
Frequency

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4253.htm (1 of 4) [5/7/2002 3:16:10 PM]

1.4.2.5.3. Develop a Better Model

Complex The complex demodulation phase plot can be used to refine this initial estimate
Demodulation for the frequency.
Phase Plot
For the complex demodulation plot, if the lines slope from left to right, the
frequency should be increased. If the lines slope from right to left, it should be
decreased. A relatively flat (i.e., horizontal) slope indicates a good frequency.
We could generate the demodulation phase plot for 0.3 and then use trial and
error to obtain a better estimate for the frequency. To simplify this, we generate
16 of these plots on a single page starting with a frequency of 0.28, increasing in
increments of 0.0025, and stopping at 0.3175.

Interpretation The plots start with lines sloping from left to right but gradually change to a right
to left slope. The relatively flat slope occurs for frequency 0.3025 (third row,
second column). The complex demodulation phase plot restricts the range from
to . This is why the plot appears to show some breaks.

Good Starting The complex demodulation amplitude plot is used to find a good starting value
Values for for the amplitude. In addition, this plot indicates whether or not the amplitude is
Amplitude constant over the entire range of the data or if it varies. If the plot is essentially
flat, i.e., zero slope, then it is reasonable to assume a constant amplitude in the
non-linear model. However, if the slope varies over the range of the plot, we
may need to adjust the model to be:

That is, we replace with a function of time. A linear fit is specified in the
model above, but this can be replaced with a more elaborate function if needed.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4253.htm (2 of 4) [5/7/2002 3:16:10 PM]

1.4.2.5.3. Develop a Better Model

Complex
Demodulation
Amplitude
Plot

The complex demodulation amplitude plot for this data shows that:
1. The amplitude is fixed at approximately 390.
2. There is a short start-up effect.
3. There is a change in amplitude at around x=160 that should be
investigated for an outlier.
In terms of a non-linear model, the plot indicates that fitting a single constant for
should be adequate for this data set.

Fit Output Using starting estimates of 0.3025 for the frequency, 390 for the amplitude, and
-177.44 for C, Dataplot generated the following output for the fit.

LEAST SQUARES NON-LINEAR FIT

SAMPLE SIZE N = 200
MODEL--Y =C + AMP*SIN(2*3.14159*FREQ*T + PHASE)
NO REPLICATION CASE

ITERATION CONVERGENCE
RESIDUAL * PARAMETER
NUMBER MEASURE
STANDARD * ESTIMATES
DEVIATION *
----------------------------------*-----------
1-- 0.10000E-01 0.52903E+03 *-0.17743E+03 0.39000E+03 0.30250E+00 0.10000E+01
2-- 0.50000E-02 0.22218E+03 *-0.17876E+03-0.33137E+03 0.30238E+00 0.71471E+00
3-- 0.25000E-02 0.15634E+03 *-0.17886E+03-0.24523E+03 0.30233E+00 0.14022E+01
4-- 0.96108E-01 0.15585E+03 *-0.17879E+03-0.36177E+03 0.30260E+00 0.14654E+01

FINAL PARAMETER ESTIMATES (APPROX. ST. DEV.) T VALUE

1 C -178.786 ( 11.02 ) -16.22
2 AMP -361.766 ( 26.19 ) -13.81

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4253.htm (3 of 4) [5/7/2002 3:16:10 PM]

1.4.2.5.3. Develop a Better Model

3 FREQ 0.302596 (0.1510E-03) 2005.

4 PHASE 1.46536 (0.4909E-01) 29.85

RESIDUAL STANDARD DEVIATION = 155.8484

RESIDUAL DEGREES OF FREEDOM = 196

Model From the fit output, our proposed model is:

We will evaluate the adequacy of this model in the next section.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4253.htm (4 of 4) [5/7/2002 3:16:10 PM]

1.4.2.5.4. Validate New Model

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.5. Beam Deflections

1.4.2.5.4. Validate New Model

4-Plot of The first step in evaluating the fit is to generate a 4-plot of the
Residuals residuals.

Interpretation The assumptions are addressed by the graphics shown above:

1. The run sequence plot (upper left) indicates that the data do not
have any significant shifts in location. There does seem to be
some shifts in scale. A start-up effect was detected previously by
the complex demodulation amplitude plot. There does appear to
be a few outliers.
2. The lag plot (upper right) shows that the data are random. The
outliers also appear in the lag plot.
3. The histogram (lower left) and the normal probability plot
(lower right) do not show any serious non-normality in the
residuals. However, the bend in the left portion of the normal
probability plot shows some cause for concern.
The 4-plot indicates that this fit is reasonably good. However, we will
attempt to improve the fit by removing the outliers.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4254.htm (1 of 3) [5/7/2002 3:16:11 PM]

1.4.2.5.4. Validate New Model

Fit Output Dataplot generated the following fit output after removing 3 outliers.
with Outliers
Removed

LEAST SQUARES NON-LINEAR FIT

SAMPLE SIZE N = 197
MODEL--Y =C + AMP*SIN(2*3.14159*FREQ*T + PHASE)
NO REPLICATION CASE

ITERATION CONVERGENCE
RESIDUAL * PARAMETER
NUMBER MEASURE
STANDARD * ESTIMATES
DEVIATION *
----------------------------------*-----------
1-- 0.10000E-01 0.14834E+03 *-0.17879E+03-0.36177E+03 0.30260E+00 0.14654E+01

2-- 0.37409E+02 0.14834E+03 *-0.17879E+03-0.36176E+03 0.30260E+00 0.14653E+01

FINAL PARAMETER ESTIMATES (APPROX. ST. DEV.) T VALUE

1 C -178.788 ( 10.57 ) -16.91
2 AMP -361.759 ( 25.45 ) -14.22
3 FREQ 0.302597 (0.1457E-03) 2077.
4 PHASE 1.46533 (0.4715E-01) 31.08

RESIDUAL STANDARD DEVIATION = 148.3398

RESIDUAL DEGREES OF FREEDOM = 193

New The original fit, with a residual standard deviation of 155.84, was:
Fit to
Edited
Data The new fit, with a residual standard deviation of 148.34, is:

There is minimal change in the parameter estimates and about a 5% reduction in

the residual standard deviation. In this case, removing the residuals has a modest
benefit in terms of reducing the variability of the model.

4-Plot
for
New
Fit

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4254.htm (2 of 3) [5/7/2002 3:16:11 PM]

1.4.2.5.4. Validate New Model

This plot shows that the underlying assumptions are satisfied and therefore the
new fit is a good descriptor of the data.
In this case, it is a judgment call whether to use the fit with or without the
outliers removed.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4254.htm (3 of 3) [5/7/2002 3:16:11 PM]

1.4.2.5.5. Work This Example Yourself

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.5. Beam Deflections

1.4.2.5.5. Work This Example Yourself

View This page allows you to repeat the analysis outlined in the case study
Dataplot description on the previous page using Dataplot . It is required that you
Macro for have already downloaded and installed Dataplot and configured your
this Case browser. to run Dataplot. Output from each analysis step below will be
Study displayed in one or more of the Dataplot windows. The four main
windows are the Output window, the Graphics window, the Command
History window, and the data sheet window. Across the top of the main
windows there are menus for executing Dataplot commands. Across the
bottom is a command entry window where commands can be typed in.

Data Analysis Steps Results and Conclusions

Click on the links below to start Dataplot and run this case
study yourself. Each step may use results from previous steps, The links in this column will connect you with more detailed
so please be patient. Wait until the software verifies that the information about each analysis step from the case study description.
current step is complete before clicking on the next step.

1. Invoke Dataplot and read data.

1. Read in the data. 1. You have read 1 column of numbers

into Dataplot, variable Y.

2. Validate assumptions.

1. 4-plot of Y. 1. Based on the 4-plot, there are no

obvious shifts in location and scale,
but the data are not random.

2. Based on the run sequence plot, there

2. Generate a run sequence plot. are no obvious shifts in location and
scale.

3. Based on the lag plot, the data

are not random.
3. Generate a lag plot.
4. The autocorrelation plot shows
significant autocorrelation at lag 1.
4. Generate an autocorrelation plot.
5. The spectral plot shows a single dominant

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4255.htm (1 of 3) [5/7/2002 3:16:11 PM]

1.4.2.5.5. Work This Example Yourself
low frequency peak.

5. Generate a spectral plot. 6. The summary statistics table displays

25+ statistics.

7. The linear fit indicates no drift in

6. Generate a table of summary location since the slope parameter
statistics. is not statistically significant.

8. Levene's test indicates no

7. Generate a linear fit to detect significant drift in variation.
drift in location.

8. Detect drift in variation by 9. The runs test indicates significant

dividing the data into quarters and non-randomness.
computing Levene's test statistic for
equal standard deviations.

9. Check for randomness by generating

a runs test.

3. Fit
Yi = C + A*SIN(2*PI*omega*ti+phi).
1. Complex demodulation phase plot
1. Generate a complex demodulation indicates a starting frequency
phase plot. of 0.3025.

2. Generate a complex demodulation 2. Complex demodulation amplitude

amplitude plot. plot indicates an amplitude of
390 (but there is a short start-up
effect).

3. Fit the non-linear model. 3. Non-linear fit generates final

parameter estimates. The
residual standard deviation from
the fit is 155.85 (compared to the
standard deviation of 277.73 from
the original data).

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4255.htm (2 of 3) [5/7/2002 3:16:11 PM]

1.4.2.5.5. Work This Example Yourself

4. Validate fit.

1. Generate a 4-plot of the residuals 1. The 4-plot indicates that the assumptions
from the fit. of constant location and scale are valid.
The lag plot indicates that the data are
random. The histogram and normal
probability plot indicate that the residuals
that the normality assumption for the
residuals are not seriously violated,
although there is a bend on the probablity
plot that warrants attention.

2. Generate a nonlinear fit with

outliers removed. 2. The fit after removing 3 outliers shows
some marginal improvement in the model
(a 5% reduction in the residual standard
deviation).

3. Generate a 4-plot of the residuals

from the fit with the outliers 3. The 4-plot of the model fit after
removed. 3 outliers removed shows marginal
improvement in satisfying model
assumptions.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4255.htm (3 of 3) [5/7/2002 3:16:11 PM]

1.4.2.6. Filter Transmittance

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies

1.4.2.6. Filter Transmittance

Filter This example illustrates the univariate analysis of filter transmittance
Transmittance data.

1. Background and Data

2. Graphical Output and Interpretation
3. Quantitative Output and Interpretation
4. Work This Example Yourself

http://www.itl.nist.gov/div898/handbook/eda/section4/eda426.htm [5/7/2002 3:16:11 PM]

1.4.2.6.1. Background and Data

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.6. Filter Transmittance

1.4.2.6.1. Background and Data

Generation This data set was collected by NIST chemist Radu Mavrodineaunu in
the 1970's from an automatic data acquisition system for a filter
transmittance experiment. The response variable is transmittance.
The motivation for studying this data set is to show how the underlying
autocorrelation structure in a relatively small data set helped the
scientist detect problems with his automatic data acquisition system.
This file can be read by Dataplot with the following commands:
SKIP 25
READ MAVRO.DAT Y

Resulting The following are the data used for this case study.
Data
2.00180
2.00170
2.00180
2.00190
2.00180
2.00170
2.00150
2.00140
2.00150
2.00150
2.00170
2.00180
2.00180
2.00190
2.00190
2.00210
2.00200
2.00160
2.00140

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4261.htm (1 of 2) [5/7/2002 3:16:12 PM]

1.4.2.6.1. Background and Data

2.00130
2.00130
2.00150
2.00150
2.00160
2.00150
2.00140
2.00130
2.00140
2.00150
2.00140
2.00150
2.00160
2.00150
2.00160
2.00190
2.00200
2.00200
2.00210
2.00220
2.00230
2.00240
2.00250
2.00270
2.00260
2.00260
2.00260
2.00270
2.00260
2.00250
2.00240

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4261.htm (2 of 2) [5/7/2002 3:16:12 PM]

1.4.2.6.2. Graphical Output and Interpretation

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.6. Filter Transmittance

1.4.2.6.2. Graphical Output and

Interpretation
Goal The goal of this analysis is threefold:
1. Determine if the univariate model:

is appropriate and valid.

2. Determine if the typical underlying assumptions for an "in
control" measurement process are valid. These assumptions are:
1. random drawings;
2. from a fixed distribution;
3. with the distribution having a fixed location; and
4. the distribution having a fixed scale.
3. Determine if the confidence interval

is appropriate and valid where s is the standard deviation of the

original data.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4262.htm (1 of 4) [5/7/2002 3:16:12 PM]

1.4.2.6.2. Graphical Output and Interpretation

4-Plot of
Data

Interpretation The assumptions are addressed by the graphics shown above:

1. The run sequence plot (upper left) indicates a significant shift in
location around x=35.
2. The linear appearance in the lag plot (upper right) indicates a
non-random pattern in the data.
3. Since the lag plot indicates significant non-randomness, we do
not make any interpretation of either the histogram (lower left)
or the normal probability plot (lower right).
The serious violation of the non-randomness assumption means that
the univariate model

is not valid. Given the linear appearance of the lag plot, the first step
might be to consider a model of the type

However, in this case discussions with the scientist revealed that

non-randomness was entirely unexpected. An examination of the
experimental process revealed that the sampling rate for the automatic
data acquisition system was too fast. That is, the equipment did not
have sufficient time to reset before the next sample started, resulting in
the current measurement being contaminated by the previous
measurement. The solution was to rerun the experiment allowing more
time between samples.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4262.htm (2 of 4) [5/7/2002 3:16:12 PM]

1.4.2.6.2. Graphical Output and Interpretation

Simple graphical techniques can be quite effective in revealing

unexpected results in the data. When this occurs, it is important to
investigate whether the unexpected result is due to problems in the
experiment and data collection or is indicative of unexpected
underlying structure in the data. This determination cannot be made on
the basis of statistics alone. The role of the graphical and statistical
analysis is to detect problems or unexpected results in the data.
Resolving the issues requires the knowledge of the scientist or
engineer.

Individual Although it is generally unnecessary, the plots can be generated

Plots individually to give more detail. Since the lag plot indicates significant
non-randomness, we omit the distributional plots.

Run
Sequence
Plot

Lag Plot

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4262.htm (3 of 4) [5/7/2002 3:16:12 PM]

1.4.2.6.2. Graphical Output and Interpretation

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4262.htm (4 of 4) [5/7/2002 3:16:12 PM]

1.4.2.6.3. Quantitative Output and Interpretation

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.6. Filter Transmittance

1.4.2.6.3. Quantitative Output and Interpretation

Summary As a first step in the analysis, a table of summary statistics is computed from the data. The
Statistics following table, generated by Dataplot, shows a typical set of statistics.

SUMMARY

NUMBER OF OBSERVATIONS = 50

***********************************************************************
* LOCATION MEASURES * DISPERSION MEASURES *
***********************************************************************
* MIDRANGE = 0.2002000E+01 * RANGE = 0.1399994E-02 *
* MEAN = 0.2001856E+01 * STAND. DEV. = 0.4291329E-03 *
* MIDMEAN = 0.2001638E+01 * AV. AB. DEV. = 0.3480196E-03 *
* MEDIAN = 0.2001800E+01 * MINIMUM = 0.2001300E+01 *
* = * LOWER QUART. = 0.2001500E+01 *
* = * LOWER HINGE = 0.2001500E+01 *
* = * UPPER HINGE = 0.2002100E+01 *
* = * UPPER QUART. = 0.2002175E+01 *
* = * MAXIMUM = 0.2002700E+01 *
***********************************************************************
* RANDOMNESS MEASURES * DISTRIBUTIONAL MEASURES *
***********************************************************************
* AUTOCO COEF = 0.9379919E+00 * ST. 3RD MOM. = 0.6191616E+00 *
* = 0.0000000E+00 * ST. 4TH MOM. = 0.2098746E+01 *
* = 0.0000000E+00 * ST. WILK-SHA = -0.4995516E+01 *
* = * UNIFORM PPCC = 0.9666610E+00 *
* = * NORMAL PPCC = 0.9558001E+00 *
* = * TUK -.5 PPCC = 0.8462552E+00 *
* = * CAUCHY PPCC = 0.6822084E+00 *
***********************************************************************

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4263.htm (1 of 6) [5/7/2002 3:16:12 PM]

1.4.2.6.3. Quantitative Output and Interpretation

Location One way to quantify a change in location over time is to fit a straight line to the data set using
the index variable X = 1, 2, ..., N, with N denoting the number of observations. If there is no
significant drift in the location, the slope parameter should be zero. For this data set, Dataplot
generates the following output:

LEAST SQUARES MULTILINEAR FIT

SAMPLE SIZE N = 50
NUMBER OF VARIABLES = 1
NO REPLICATION CASE

PARAMETER ESTIMATES (APPROX. ST. DEV.) T VALUE

1 A0 2.00138 (0.9695E-04) 0.2064E+05
2 A1 X 0.184685E-04 (0.3309E-05) 5.582

RESIDUAL STANDARD DEVIATION = 0.3376404E-03

RESIDUAL DEGREES OF FREEDOM = 48
The slope parameter, A1, has a t value of 5.6, which is statistically significant. The value of the
slope parameter is 0.0000185. Although this number is nearly zero, we need to take into
account that the original scale of the data is from about 2.0012 to 2.0028. In this case, we
conclude that there is a drift in location, although by a relatively minor amount.

Variation One simple way to detect a change in variation is with a Bartlett test after dividing the data set
into several equal sized intervals. However, the Bartlett test is not robust for non-normality.
Since the normality assumption is questionable for these data, we use the alternative Levene
test. In partiuclar, we use the Levene test based on the median rather the mean. The choice of
the number of intervals is somewhat arbitrary, although values of 4 or 8 are reasonable.
Dataplot generated the following output for the Levene test.

LEVENE F-TEST FOR SHIFT IN VARIATION

(ASSUMPTION: NORMALITY)

1. STATISTICS
NUMBER OF OBSERVATIONS = 50
NUMBER OF GROUPS = 4
LEVENE F TEST STATISTIC = 0.9714893

FOR LEVENE TEST STATISTIC

0 % POINT = 0.0000000E+00
50 % POINT = 0.8004835
75 % POINT = 1.416631
90 % POINT = 2.206890
95 % POINT = 2.806845
99 % POINT = 4.238307
99.9 % POINT = 6.424733

58.56597 % Point: 0.9714893

3. CONCLUSION (AT THE 5% LEVEL):

THERE IS NO SHIFT IN VARIATION.
THUS: HOMOGENEOUS WITH RESPECT TO VARIATION.
In this case, since the Levene test statistic value of 0.971 is less than the critical value of 2.806
at the 5% level, we conclude that there is no evidence of a change in variation.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4263.htm (2 of 6) [5/7/2002 3:16:12 PM]

1.4.2.6.3. Quantitative Output and Interpretation

Randomness There are many ways in which data can be non-random. However, most common forms of
non-randomness can be detected with a few simple tests. The lag plot in the 4-plot in the
previous seciton is a simple graphical technique.
One check is an autocorrelation plot that shows the autocorrelations for various lags.
Confidence bands can be plotted at the 95% and 99% confidence levels. Points outside this
band indicate statistically significant values (lag 0 is always 1). Dataplot generated the
following autocorrelation plot.

The lag 1 autocorrelation, which is generally the one of most interest, is 0.93. The critical
values at the 5% level are -0.277 and 0.277. This indicates that the lag 1 autocorrelation is
statistically significant, so there is strong evidence of non-randomness.
A common test for randomness is the runs test.

RUNS UP

STATISTIC = NUMBER OF RUNS UP

OF LENGTH EXACTLY I

I STAT EXP(STAT) SD(STAT) Z

1 1.0 10.4583 3.2170 -2.94

2 3.0 4.4667 1.6539 -0.89
3 1.0 1.2542 0.9997 -0.25
4 0.0 0.2671 0.5003 -0.53
5 0.0 0.0461 0.2132 -0.22
6 0.0 0.0067 0.0818 -0.08
7 0.0 0.0008 0.0291 -0.03
8 1.0 0.0001 0.0097 103.06
9 0.0 0.0000 0.0031 0.00
10 1.0 0.0000 0.0009 1087.63

STATISTIC = NUMBER OF RUNS UP

OF LENGTH I OR MORE

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4263.htm (3 of 6) [5/7/2002 3:16:12 PM]

1.4.2.6.3. Quantitative Output and Interpretation

I STAT EXP(STAT) SD(STAT) Z

1 7.0 16.5000 2.0696 -4.59

2 6.0 6.0417 1.3962 -0.03
3 3.0 1.5750 1.0622 1.34
4 2.0 0.3208 0.5433 3.09
5 2.0 0.0538 0.2299 8.47
6 2.0 0.0077 0.0874 22.79
7 2.0 0.0010 0.0308 64.85
8 2.0 0.0001 0.0102 195.70
9 1.0 0.0000 0.0032 311.64
10 1.0 0.0000 0.0010 1042.19

RUNS DOWN

STATISTIC = NUMBER OF RUNS DOWN

OF LENGTH EXACTLY I

I STAT EXP(STAT) SD(STAT) Z

1 3.0 10.4583 3.2170 -2.32

2 0.0 4.4667 1.6539 -2.70
3 3.0 1.2542 0.9997 1.75
4 1.0 0.2671 0.5003 1.46
5 1.0 0.0461 0.2132 4.47
6 0.0 0.0067 0.0818 -0.08
7 0.0 0.0008 0.0291 -0.03
8 0.0 0.0001 0.0097 -0.01
9 0.0 0.0000 0.0031 0.00
10 0.0 0.0000 0.0009 0.00

STATISTIC = NUMBER OF RUNS DOWN

OF LENGTH I OR MORE

I STAT EXP(STAT) SD(STAT) Z

1 8.0 16.5000 2.0696 -4.11

2 5.0 6.0417 1.3962 -0.75
3 5.0 1.5750 1.0622 3.22
4 2.0 0.3208 0.5433 3.09
5 1.0 0.0538 0.2299 4.12
6 0.0 0.0077 0.0874 -0.09
7 0.0 0.0010 0.0308 -0.03
8 0.0 0.0001 0.0102 -0.01
9 0.0 0.0000 0.0032 0.00
10 0.0 0.0000 0.0010 0.00

RUNS TOTAL = RUNS UP + RUNS DOWN

STATISTIC = NUMBER OF RUNS TOTAL

OF LENGTH EXACTLY I

I STAT EXP(STAT) SD(STAT) Z

1 4.0 20.9167 4.5496 -3.72

2 3.0 8.9333 2.3389 -2.54
3 4.0 2.5083 1.4138 1.06
4 1.0 0.5341 0.7076 0.66
5 1.0 0.0922 0.3015 3.01

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4263.htm (4 of 6) [5/7/2002 3:16:12 PM]

1.4.2.6.3. Quantitative Output and Interpretation
6 0.0 0.0134 0.1157 -0.12
7 0.0 0.0017 0.0411 -0.04
8 1.0 0.0002 0.0137 72.86
9 0.0 0.0000 0.0043 0.00
10 1.0 0.0000 0.0013 769.07

STATISTIC = NUMBER OF RUNS TOTAL

OF LENGTH I OR MORE

I STAT EXP(STAT) SD(STAT) Z

1 15.0 33.0000 2.9269 -6.15

2 11.0 12.0833 1.9745 -0.55
3 8.0 3.1500 1.5022 3.23
4 4.0 0.6417 0.7684 4.37
5 3.0 0.1075 0.3251 8.90
6 2.0 0.0153 0.1236 16.05
7 2.0 0.0019 0.0436 45.83
8 2.0 0.0002 0.0145 138.37
9 1.0 0.0000 0.0045 220.36
10 1.0 0.0000 0.0014 736.94

LENGTH OF THE LONGEST RUN UP = 10

LENGTH OF THE LONGEST RUN DOWN = 5
LENGTH OF THE LONGEST RUN UP OR DOWN = 10

NUMBER OF POSITIVE DIFFERENCES = 23

NUMBER OF NEGATIVE DIFFERENCES = 18
NUMBER OF ZERO DIFFERENCES = 8

Values in the column labeled "Z" greater than 1.96 or less than -1.96 are statistically significant
at the 5% level. Due to the number of values that are much larger than the 1.96 cut-off, we
conclude that the data are not random.

Distributional Since we rejected the randomness assumption, the distributional tests are not meaningful.
Analysis Therefore, these quantitative tests are omitted. We also omit Grubbs' outlier test since it also
assumes the data are approximately normally distributed.

Univariate It is sometimes useful and convenient to summarize the above results in a report.
Report

Analysis for filter transmittance data

1: Sample Size = 50

2: Location
Mean = 2.001857
Standard Deviation of Mean = 0.00006
95% Confidence Interval for Mean = (2.001735,2.001979)
Drift with respect to location? = NO

3: Variation
Standard Deviation = 0.00043
95% Confidence Interval for SD = (0.000359,0.000535)
Change in variation?
(based on Levene's test on quarters
of the data) = NO

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4263.htm (5 of 6) [5/7/2002 3:16:12 PM]

1.4.2.6.3. Quantitative Output and Interpretation
4: Distribution
Distributional tests omitted due to
non-randomness of the data

5: Randomness
Lag One Autocorrelation = 0.937998
Data are Random?
(as measured by autocorrelation) = NO

6: Statistical Control
(i.e., no drift in location or scale,
data are random, distribution is
fixed, here we are testing only for
normal)
Data Set is in Statistical Control? = NO

7: Outliers?
(Grubbs' test omitted) = NO

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4263.htm (6 of 6) [5/7/2002 3:16:12 PM]

1.4.2.6.4. Work This Example Yourself

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.6. Filter Transmittance

1.4.2.6.4. Work This Example Yourself

View This page allows you to repeat the analysis outlined in the case study
Dataplot description on the previous page using Dataplot . It is required that you
Macro for have already downloaded and installed Dataplot and configured your
this Case browser. to run Dataplot. Output from each analysis step below will be
Study displayed in one or more of the Dataplot windows. The four main
windows are the Output window, the Graphics window, the Command
History window, and the data sheet window. Across the top of the main
windows there are menus for executing Dataplot commands. Across the
bottom is a command entry window where commands can be typed in.

Data Analysis Steps Results and Conclusions

Click on the links below to start Dataplot and run this case study
The links in this column will connect you with more detailed
yourself. Each step may use results from previous steps, so please
information about each analysis step from the case study
be patient. Wait until the software verifies that the current step is
description.
complete before clicking on the next step.

1. Invoke Dataplot and read data.

1. Read in the data. 1. You have read 1 column of numbers

into Dataplot, variable Y.

2. 4-plot of the data.

1. 4-plot of Y. 1. Based on the 4-plot, there is a shift

in location and the data are not random.

3. Generate the individual plots.

1. Generate a run sequence plot. 1. The run sequence plot indicates that
there is a shift in location.

2. Generate a lag plot. 2. The strong linear pattern of the lag

plot indicates significant
non-randomness.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4264.htm (1 of 2) [5/7/2002 3:16:13 PM]

1.4.2.6.4. Work This Example Yourself

4. Generate summary statistics, quantitative

analysis, and print a univariate report.

1. Generate a table of summary 1. The summary statistics table displays

statistics. 25+ statistics.

2. Compute a linear fit based on 2. The linear fit indicates a slight drift in
quarters of the data to detect location since the slope parameter is
drift in location. statistically significant, but small.

3. Compute Levene's test based on 3. Levene's test indicates no significant

quarters of the data to detect drift in variation.
changes in variation.

4. Check for randomness by generating an 4. The lag 1 autocorrelation is 0.94.

autocorrelation plot and a runs test. This is outside the 95% confidence
interval bands which indicates significant
non-randomness.

5. Print a univariate report (this assumes 5. The results are summarized in a

steps 2 thru 4 have already been run). convenient report.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4264.htm (2 of 2) [5/7/2002 3:16:13 PM]

1.4.2.7. Standard Resistor

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies

1.4.2.7. Standard Resistor

Standard This example illustrates the univariate analysis of standard resistor data.
Resistor

1. Background and Data

2. Graphical Output and Interpretation
3. Quantitative Output and Interpretation
4. Work This Example Yourself

http://www.itl.nist.gov/div898/handbook/eda/section4/eda427.htm [5/7/2002 3:16:13 PM]

1.4.2.7.1. Background and Data

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.7. Standard Resistor

1.4.2.7.1. Background and Data

Generation This data set was collected by Ron Dziuba of NIST over a 5-year period
from 1980 to 1985. The response variable is resistor values.
The motivation for studying this data set is to illustrate data that violate
the assumptions of constant location and scale.
This file can be read by Dataplot with the following commands:
SKIP 25
COLUMN LIMITS 10 80
READ DZIUBA1.DAT Y
COLUMN LIMITS

Resulting The following are the data used for this case study.
Data
27.8680
27.8929
27.8773
27.8530
27.8876
27.8725
27.8743
27.8879
27.8728
27.8746
27.8863
27.8716
27.8818
27.8872
27.8885
27.8945
27.8797
27.8627
27.8870

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (1 of 23) [5/7/2002 3:16:13 PM]

1.4.2.7.1. Background and Data

27.8895
27.9138
27.8931
27.8852
27.8788
27.8827
27.8939
27.8558
27.8814
27.8479
27.8479
27.8848
27.8809
27.8479
27.8611
27.8630
27.8679
27.8637
27.8985
27.8900
27.8577
27.8848
27.8869
27.8976
27.8610
27.8567
27.8417
27.8280
27.8555
27.8639
27.8702
27.8582
27.8605
27.8900
27.8758
27.8774
27.9008
27.8988
27.8897
27.8990
27.8958
27.8830
27.8967
27.9105
27.9028
27.8977

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (2 of 23) [5/7/2002 3:16:13 PM]

1.4.2.7.1. Background and Data

27.8953
27.8970
27.9190
27.9180
27.8997
27.9204
27.9234
27.9072
27.9152
27.9091
27.8882
27.9035
27.9267
27.9138
27.8955
27.9203
27.9239
27.9199
27.9646
27.9411
27.9345
27.8712
27.9145
27.9259
27.9317
27.9239
27.9247
27.9150
27.9444
27.9457
27.9166
27.9066
27.9088
27.9255
27.9312
27.9439
27.9210
27.9102
27.9083
27.9121
27.9113
27.9091
27.9235
27.9291
27.9253
27.9092

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (3 of 23) [5/7/2002 3:16:13 PM]

1.4.2.7.1. Background and Data

27.9117
27.9194
27.9039
27.9515
27.9143
27.9124
27.9128
27.9260
27.9339
27.9500
27.9530
27.9430
27.9400
27.8850
27.9350
27.9120
27.9260
27.9660
27.9280
27.9450
27.9390
27.9429
27.9207
27.9205
27.9204
27.9198
27.9246
27.9366
27.9234
27.9125
27.9032
27.9285
27.9561
27.9616
27.9530
27.9280
27.9060
27.9380
27.9310
27.9347
27.9339
27.9410
27.9397
27.9472
27.9235
27.9315

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (4 of 23) [5/7/2002 3:16:13 PM]

1.4.2.7.1. Background and Data

27.9368
27.9403
27.9529
27.9263
27.9347
27.9371
27.9129
27.9549
27.9422
27.9423
27.9750
27.9339
27.9629
27.9587
27.9503
27.9573
27.9518
27.9527
27.9589
27.9300
27.9629
27.9630
27.9660
27.9730
27.9660
27.9630
27.9570
27.9650
27.9520
27.9820
27.9560
27.9670
27.9520
27.9470
27.9720
27.9610
27.9437
27.9660
27.9580
27.9660
27.9700
27.9600
27.9660
27.9770
27.9110
27.9690

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (5 of 23) [5/7/2002 3:16:13 PM]

1.4.2.7.1. Background and Data

27.9698
27.9616
27.9371
27.9700
27.9265
27.9964
27.9842
27.9667
27.9610
27.9943
27.9616
27.9397
27.9799
28.0086
27.9709
27.9741
27.9675
27.9826
27.9676
27.9703
27.9789
27.9786
27.9722
27.9831
28.0043
27.9548
27.9875
27.9495
27.9549
27.9469
27.9744
27.9744
27.9449
27.9837
27.9585
28.0096
27.9762
27.9641
27.9854
27.9877
27.9839
27.9817
27.9845
27.9877
27.9880
27.9822

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (6 of 23) [5/7/2002 3:16:13 PM]

1.4.2.7.1. Background and Data

27.9836
28.0030
27.9678
28.0146
27.9945
27.9805
27.9785
27.9791
27.9817
27.9805
27.9782
27.9753
27.9792
27.9704
27.9794
27.9814
27.9794
27.9795
27.9881
27.9772
27.9796
27.9736
27.9772
27.9960
27.9795
27.9779
27.9829
27.9829
27.9815
27.9811
27.9773
27.9778
27.9724
27.9756
27.9699
27.9724
27.9666
27.9666
27.9739
27.9684
27.9861
27.9901
27.9879
27.9865
27.9876
27.9814

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (7 of 23) [5/7/2002 3:16:13 PM]

1.4.2.7.1. Background and Data

27.9842
27.9868
27.9834
27.9892
27.9864
27.9843
27.9838
27.9847
27.9860
27.9872
27.9869
27.9602
27.9852
27.9860
27.9836
27.9813
27.9623
27.9843
27.9802
27.9863
27.9813
27.9881
27.9850
27.9850
27.9830
27.9866
27.9888
27.9841
27.9863
27.9903
27.9961
27.9905
27.9945
27.9878
27.9929
27.9914
27.9914
27.9997
28.0006
27.9999
28.0004
28.0020
28.0029
28.0008
28.0040
28.0078

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (8 of 23) [5/7/2002 3:16:13 PM]

1.4.2.7.1. Background and Data

28.0065
27.9959
28.0073
28.0017
28.0042
28.0036
28.0055
28.0007
28.0066
28.0011
27.9960
28.0083
27.9978
28.0108
28.0088
28.0088
28.0139
28.0092
28.0092
28.0049
28.0111
28.0120
28.0093
28.0116
28.0102
28.0139
28.0113
28.0158
28.0156
28.0137
28.0236
28.0171
28.0224
28.0184
28.0199
28.0190
28.0204
28.0170
28.0183
28.0201
28.0182
28.0183
28.0175
28.0127
28.0211
28.0057

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (9 of 23) [5/7/2002 3:16:13 PM]

1.4.2.7.1. Background and Data

28.0180
28.0183
28.0149
28.0185
28.0182
28.0192
28.0213
28.0216
28.0169
28.0162
28.0167
28.0167
28.0169
28.0169
28.0161
28.0152
28.0179
28.0215
28.0194
28.0115
28.0174
28.0178
28.0202
28.0240
28.0198
28.0194
28.0171
28.0134
28.0121
28.0121
28.0141
28.0101
28.0114
28.0122
28.0124
28.0171
28.0165
28.0166
28.0159
28.0181
28.0200
28.0116
28.0144
28.0141
28.0116
28.0107

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (10 of 23) [5/7/2002 3:16:13 PM]

1.4.2.7.1. Background and Data

28.0169
28.0105
28.0136
28.0138
28.0114
28.0122
28.0122
28.0116
28.0025
28.0097
28.0066
28.0072
28.0066
28.0068
28.0067
28.0130
28.0091
28.0088
28.0091
28.0091
28.0115
28.0087
28.0128
28.0139
28.0095
28.0115
28.0101
28.0121
28.0114
28.0121
28.0122
28.0121
28.0168
28.0212
28.0219
28.0221
28.0204
28.0169
28.0141
28.0142
28.0147
28.0159
28.0165
28.0144
28.0182
28.0155

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (11 of 23) [5/7/2002 3:16:13 PM]

1.4.2.7.1. Background and Data

28.0155
28.0192
28.0204
28.0185
28.0248
28.0185
28.0226
28.0271
28.0290
28.0240
28.0302
28.0243
28.0288
28.0287
28.0301
28.0273
28.0313
28.0293
28.0300
28.0344
28.0308
28.0291
28.0287
28.0358
28.0309
28.0286
28.0308
28.0291
28.0380
28.0411
28.0420
28.0359
28.0368
28.0327
28.0361
28.0334
28.0300
28.0347
28.0359
28.0344
28.0370
28.0355
28.0371
28.0318
28.0390
28.0390

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (12 of 23) [5/7/2002 3:16:13 PM]

1.4.2.7.1. Background and Data

28.0390
28.0376
28.0376
28.0377
28.0345
28.0333
28.0429
28.0379
28.0401
28.0401
28.0423
28.0393
28.0382
28.0424
28.0386
28.0386
28.0373
28.0397
28.0412
28.0565
28.0419
28.0456
28.0426
28.0423
28.0391
28.0403
28.0388
28.0408
28.0457
28.0455
28.0460
28.0456
28.0464
28.0442
28.0416
28.0451
28.0432
28.0434
28.0448
28.0448
28.0373
28.0429
28.0392
28.0469
28.0443
28.0356

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (13 of 23) [5/7/2002 3:16:13 PM]

1.4.2.7.1. Background and Data

28.0474
28.0446
28.0348
28.0368
28.0418
28.0445
28.0533
28.0439
28.0474
28.0435
28.0419
28.0538
28.0538
28.0463
28.0491
28.0441
28.0411
28.0507
28.0459
28.0519
28.0554
28.0512
28.0507
28.0582
28.0471
28.0539
28.0530
28.0502
28.0422
28.0431
28.0395
28.0177
28.0425
28.0484
28.0693
28.0490
28.0453
28.0494
28.0522
28.0393
28.0443
28.0465
28.0450
28.0539
28.0566
28.0585

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (14 of 23) [5/7/2002 3:16:13 PM]

1.4.2.7.1. Background and Data

28.0486
28.0427
28.0548
28.0616
28.0298
28.0726
28.0695
28.0629
28.0503
28.0493
28.0537
28.0613
28.0643
28.0678
28.0564
28.0703
28.0647
28.0579
28.0630
28.0716
28.0586
28.0607
28.0601
28.0611
28.0606
28.0611
28.0066
28.0412
28.0558
28.0590
28.0750
28.0483
28.0599
28.0490
28.0499
28.0565
28.0612
28.0634
28.0627
28.0519
28.0551
28.0696
28.0581
28.0568
28.0572
28.0529

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (15 of 23) [5/7/2002 3:16:13 PM]

1.4.2.7.1. Background and Data

28.0421
28.0432
28.0211
28.0363
28.0436
28.0619
28.0573
28.0499
28.0340
28.0474
28.0534
28.0589
28.0466
28.0448
28.0576
28.0558
28.0522
28.0480
28.0444
28.0429
28.0624
28.0610
28.0461
28.0564
28.0734
28.0565
28.0503
28.0581
28.0519
28.0625
28.0583
28.0645
28.0642
28.0535
28.0510
28.0542
28.0677
28.0416
28.0676
28.0596
28.0635
28.0558
28.0623
28.0718
28.0585
28.0552

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (16 of 23) [5/7/2002 3:16:13 PM]

1.4.2.7.1. Background and Data

28.0684
28.0646
28.0590
28.0465
28.0594
28.0303
28.0533
28.0561
28.0585
28.0497
28.0582
28.0507
28.0562
28.0715
28.0468
28.0411
28.0587
28.0456
28.0705
28.0534
28.0558
28.0536
28.0552
28.0461
28.0598
28.0598
28.0650
28.0423
28.0442
28.0449
28.0660
28.0506
28.0655
28.0512
28.0407
28.0475
28.0411
28.0512
28.1036
28.0641
28.0572
28.0700
28.0577
28.0637
28.0534
28.0461

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (17 of 23) [5/7/2002 3:16:13 PM]

1.4.2.7.1. Background and Data

28.0701
28.0631
28.0575
28.0444
28.0592
28.0684
28.0593
28.0677
28.0512
28.0644
28.0660
28.0542
28.0768
28.0515
28.0579
28.0538
28.0526
28.0833
28.0637
28.0529
28.0535
28.0561
28.0736
28.0635
28.0600
28.0520
28.0695
28.0608
28.0608
28.0590
28.0290
28.0939
28.0618
28.0551
28.0757
28.0698
28.0717
28.0529
28.0644
28.0613
28.0759
28.0745
28.0736
28.0611
28.0732
28.0782

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (18 of 23) [5/7/2002 3:16:14 PM]

1.4.2.7.1. Background and Data

28.0682
28.0756
28.0857
28.0739
28.0840
28.0862
28.0724
28.0727
28.0752
28.0732
28.0703
28.0849
28.0795
28.0902
28.0874
28.0971
28.0638
28.0877
28.0751
28.0904
28.0971
28.0661
28.0711
28.0754
28.0516
28.0961
28.0689
28.1110
28.1062
28.0726
28.1141
28.0913
28.0982
28.0703
28.0654
28.0760
28.0727
28.0850
28.0877
28.0967
28.1185
28.0945
28.0834
28.0764
28.1129
28.0797

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (19 of 23) [5/7/2002 3:16:14 PM]

1.4.2.7.1. Background and Data

28.0707
28.1008
28.0971
28.0826
28.0857
28.0984
28.0869
28.0795
28.0875
28.1184
28.0746
28.0816
28.0879
28.0888
28.0924
28.0979
28.0702
28.0847
28.0917
28.0834
28.0823
28.0917
28.0779
28.0852
28.0863
28.0942
28.0801
28.0817
28.0922
28.0914
28.0868
28.0832
28.0881
28.0910
28.0886
28.0961
28.0857
28.0859
28.1086
28.0838
28.0921
28.0945
28.0839
28.0877
28.0803
28.0928

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (20 of 23) [5/7/2002 3:16:14 PM]

1.4.2.7.1. Background and Data

28.0885
28.0940
28.0856
28.0849
28.0955
28.0955
28.0846
28.0871
28.0872
28.0917
28.0931
28.0865
28.0900
28.0915
28.0963
28.0917
28.0950
28.0898
28.0902
28.0867
28.0843
28.0939
28.0902
28.0911
28.0909
28.0949
28.0867
28.0932
28.0891
28.0932
28.0887
28.0925
28.0928
28.0883
28.0946
28.0977
28.0914
28.0959
28.0926
28.0923
28.0950
28.1006
28.0924
28.0963
28.0893
28.0956

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (21 of 23) [5/7/2002 3:16:14 PM]

1.4.2.7.1. Background and Data

28.0980
28.0928
28.0951
28.0958
28.0912
28.0990
28.0915
28.0957
28.0976
28.0888
28.0928
28.0910
28.0902
28.0950
28.0995
28.0965
28.0972
28.0963
28.0946
28.0942
28.0998
28.0911
28.1043
28.1002
28.0991
28.0959
28.0996
28.0926
28.1002
28.0961
28.0983
28.0997
28.0959
28.0988
28.1029
28.0989
28.1000
28.0944
28.0979
28.1005
28.1012
28.1013
28.0999
28.0991
28.1059
28.0961

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (22 of 23) [5/7/2002 3:16:14 PM]

1.4.2.7.1. Background and Data

28.0981
28.1045
28.1047
28.1042
28.1146
28.1113
28.1051
28.1065
28.1065
28.0985
28.1000
28.1066
28.1041
28.0954
28.1090

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4271.htm (23 of 23) [5/7/2002 3:16:14 PM]

1.4.2.7.2. Graphical Output and Interpretation

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.7. Standard Resistor

1.4.2.7.2. Graphical Output and

Interpretation
Goal The goal of this analysis is threefold:
1. Determine if the univariate model:

is appropriate and valid.

2. Determine if the typical underlying assumptions for an "in
control" measurement process are valid. These assumptions are:
1. random drawings;
2. from a fixed distribution;
3. with the distribution having a fixed location; and
4. the distribution having a fixed scale.
3. Determine if the confidence interval

is appropriate and valid where s is the standard deviation of the

original data.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4272.htm (1 of 4) [5/7/2002 3:16:14 PM]

1.4.2.7.2. Graphical Output and Interpretation

4-Plot of
Data

Interpretation The assumptions are addressed by the graphics shown above:

1. The run sequence plot (upper left) indicates significant shifts in
both location and variation. Specifically, the location is
increasing with time. The variability seems greater in the first
and last third of the data than it does in the middle third.
2. The lag plot (upper right) shows a significant non-random
pattern in the data. Specifically, the strong linear appearance of
this plot is indicative of a model that relates Yt to Yt-1.
3. The distributional plots, the histogram (lower left) and the
normal probability plot (lower right), are not interpreted since
the randomness assumption is so clearly violated.
The serious violation of the non-randomness assumption means that
the univariate model

is not valid. Given the linear appearance of the lag plot, the first step
might be to consider a model of the type

However, discussions with the scientist revealed the following:

1. the drift with respect to location was expected.
2. the non-constant variability was not expected.
The scientist examined the data collection device and determined that
the non-constant variation was a seasonal effect. The high variability

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4272.htm (2 of 4) [5/7/2002 3:16:14 PM]

1.4.2.7.2. Graphical Output and Interpretation

data in the first and last thirds was collected in winter while the more
stable middle third was collected in the summer. The seasonal effect
was determined to be caused by the amount of humidity affecting the
measurement equipment. In this case, the solution was to modify the
test equipment to be less sensitive to enviromental factors.
Simple graphical techniques can be quite effective in revealing
unexpected results in the data. When this occurs, it is important to
investigate whether the unexpected result is due to problems in the
experiment and data collection, or is it in fact indicative of an
unexpected underlying structure in the data. This determination cannot
be made on the basis of statistics alone. The role of the graphical and
statistical analysis is to detect problems or unexpected results in the
data. Resolving the issues requires the knowledge of the scientist or
engineer.

Individual Although it is generally unnecessary, the plots can be generated

Plots individually to give more detail. Since the lag plot indicates significant
non-randomness, we omit the distributional plots.

Run
Sequence
Plot

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4272.htm (3 of 4) [5/7/2002 3:16:14 PM]

1.4.2.7.2. Graphical Output and Interpretation

Lag Plot

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4272.htm (4 of 4) [5/7/2002 3:16:14 PM]

1.4.2.7.3. Quantitative Output and Interpretation

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.7. Standard Resistor

1.4.2.7.3. Quantitative Output and Interpretation

Summary As a first step in the analysis, a table of summary statistics is computed from the data. The
Statistics following table, generated by Dataplot, shows a typical set of statistics.

SUMMARY

NUMBER OF OBSERVATIONS = 1000

***********************************************************************
* LOCATION MEASURES * DISPERSION MEASURES *
***********************************************************************
* MIDRANGE = 0.2797325E+02 * RANGE = 0.2905006E+00 *
* MEAN = 0.2801634E+02 * STAND. DEV. = 0.6349404E-01 *
* MIDMEAN = 0.2802659E+02 * AV. AB. DEV. = 0.5101655E-01 *
* MEDIAN = 0.2802910E+02 * MINIMUM = 0.2782800E+02 *
* = * LOWER QUART. = 0.2797905E+02 *
* = * LOWER HINGE = 0.2797900E+02 *
* = * UPPER HINGE = 0.2806295E+02 *
* = * UPPER QUART. = 0.2806293E+02 *
* = * MAXIMUM = 0.2811850E+02 *
***********************************************************************
* RANDOMNESS MEASURES * DISTRIBUTIONAL MEASURES *
***********************************************************************
* AUTOCO COEF = 0.9721591E+00 * ST. 3RD MOM. = -0.6936395E+00 *
* = 0.0000000E+00 * ST. 4TH MOM. = 0.2689681E+01 *
* = 0.0000000E+00 * ST. WILK-SHA = -0.4216419E+02 *
* = * UNIFORM PPCC = 0.9689648E+00 *
* = * NORMAL PPCC = 0.9718416E+00 *
* = * TUK -.5 PPCC = 0.7334843E+00 *
* = * CAUCHY PPCC = 0.3347875E+00 *
***********************************************************************

The autocorrelation coefficient of 0.972 is evidence of significant non-randomness.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4273.htm (1 of 6) [5/7/2002 3:16:14 PM]

1.4.2.7.3. Quantitative Output and Interpretation

Location One way to quantify a change in location over time is to fit a straight line to the data set using
the index variable X = 1, 2, ..., N, with N denoting the number of observations. If there is no
significant drift in the location, the slope parameter estimate should be zero. For this data set,
Dataplot generates the following output:

LEAST SQUARES MULTILINEAR FIT

SAMPLE SIZE N = 1000
NUMBER OF VARIABLES = 1
NO REPLICATION CASE

PARAMETER ESTIMATES (APPROX. ST. DEV.) T VALUE

1 A0 27.9114 (0.1209E-02) 0.2309E+05
2 A1 X 0.209670E-03 (0.2092E-05) 100.2

RESIDUAL STANDARD DEVIATION = 0.1909796E-01

RESIDUAL DEGREES OF FREEDOM = 998

COEF AND SD(COEF) WRITTEN OUT TO FILE DPST1F.DAT

SD(PRED),95LOWER,95UPPER,99LOWER,99UPPER
WRITTEN OUT TO FILE DPST2F.DAT
REGRESSION DIAGNOSTICS WRITTEN OUT TO FILE DPST3F.DAT
PARAMETER VARIANCE-COVARIANCE MATRIX AND
INVERSE OF X-TRANSPOSE X MATRIX
WRITTEN OUT TO FILE DPST4F.DAT
The slope parameter, A1, has a t value of 100 which is statistically significant. The value of the
slope parameter estimate is 0.00021. Although this number is nearly zero, we need to take into
account that the original scale of the data is from about 27.8 to 28.2. In this case, we conclude
that there is a drift in location.

Variation One simple way to detect a change in variation is with a Bartlett test after dividing the data set
into several equal-sized intervals. However, the Bartlett test is not robust for non-normality.
Since the normality assumption is questionable for these data, we use the alternative Levene
test. In partiuclar, we use the Levene test based on the median rather the mean. The choice of
the number of intervals is somewhat arbitrary, although values of 4 or 8 are reasonable.
Dataplot generated the following output for the Levene test.

LEVENE F-TEST FOR SHIFT IN VARIATION

(ASSUMPTION: NORMALITY)

1. STATISTICS
NUMBER OF OBSERVATIONS = 1000
NUMBER OF GROUPS = 4
LEVENE F TEST STATISTIC = 140.8509

FOR LEVENE TEST STATISTIC

0 % POINT = 0.0000000E+00
50 % POINT = 0.7891988
75 % POINT = 1.371589
90 % POINT = 2.089303
95 % POINT = 2.613852
99 % POINT = 3.801369
99.9 % POINT = 5.463994

100.0000 % Point: 140.8509

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4273.htm (2 of 6) [5/7/2002 3:16:14 PM]

1.4.2.7.3. Quantitative Output and Interpretation

3. CONCLUSION (AT THE 5% LEVEL):

THERE IS A SHIFT IN VARIATION.
THUS: NOT HOMOGENEOUS WITH RESPECT TO VARIATION.
In this case, since the Levene test statistic value of 140.9 is greater than the 5% significance
level critical value of 2.6, we conclude that there is significant evidence of nonconstant
variation.

Randomness
There are many ways in which data can be non-random. However, most common forms of
non-randomness can be detected with a few simple tests. The lag plot in the 4-plot in the
previous section is a simple graphical technique.
One check is an autocorrelation plot that shows the autocorrelations for various lags.
Confidence bands can be plotted at the 95% and 99% confidence levels. Points outside this
band indicate statistically significant values (lag 0 is always 1). Dataplot generated the
following autocorrelation plot.

The lag 1 autocorrelation, which is generally the one of greatest interest, is 0.97. The critical
values at the 5% significance level are -0.062 and 0.062. This indicates that the lag 1
autocorrelation is statistically significant, so there is strong evidence of non-randomness.
A common test for randomness is the runs test.

RUNS UP

STATISTIC = NUMBER OF RUNS UP

OF LENGTH EXACTLY I

I STAT EXP(STAT) SD(STAT) Z

1 178.0 208.3750 14.5453 -2.09

2 90.0 91.5500 7.5002 -0.21
3 29.0 26.3236 4.5727 0.59
4 16.0 5.7333 2.3164 4.43
5 2.0 1.0121 0.9987 0.99
6 0.0 0.1507 0.3877 -0.39

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4273.htm (3 of 6) [5/7/2002 3:16:14 PM]

1.4.2.7.3. Quantitative Output and Interpretation
7 0.0 0.0194 0.1394 -0.14
8 0.0 0.0022 0.0470 -0.05
9 0.0 0.0002 0.0150 -0.02
10 0.0 0.0000 0.0046 0.00

STATISTIC = NUMBER OF RUNS UP

OF LENGTH I OR MORE

I STAT EXP(STAT) SD(STAT) Z

1 315.0 333.1667 9.4195 -1.93

2 137.0 124.7917 6.2892 1.94
3 47.0 33.2417 4.8619 2.83
4 18.0 6.9181 2.5200 4.40
5 2.0 1.1847 1.0787 0.76
6 0.0 0.1726 0.4148 -0.42
7 0.0 0.0219 0.1479 -0.15
8 0.0 0.0025 0.0496 -0.05
9 0.0 0.0002 0.0158 -0.02
10 0.0 0.0000 0.0048 0.00

RUNS DOWN

STATISTIC = NUMBER OF RUNS DOWN

OF LENGTH EXACTLY I

I STAT EXP(STAT) SD(STAT) Z

1 195.0 208.3750 14.5453 -0.92

2 81.0 91.5500 7.5002 -1.41
3 32.0 26.3236 4.5727 1.24
4 4.0 5.7333 2.3164 -0.75
5 1.0 1.0121 0.9987 -0.01
6 1.0 0.1507 0.3877 2.19
7 0.0 0.0194 0.1394 -0.14
8 0.0 0.0022 0.0470 -0.05
9 0.0 0.0002 0.0150 -0.02
10 0.0 0.0000 0.0046 0.00

STATISTIC = NUMBER OF RUNS DOWN

OF LENGTH I OR MORE

I STAT EXP(STAT) SD(STAT) Z

1 314.0 333.1667 9.4195 -2.03

2 119.0 124.7917 6.2892 -0.92
3 38.0 33.2417 4.8619 0.98
4 6.0 6.9181 2.5200 -0.36
5 2.0 1.1847 1.0787 0.76
6 1.0 0.1726 0.4148 1.99
7 0.0 0.0219 0.1479 -0.15
8 0.0 0.0025 0.0496 -0.05
9 0.0 0.0002 0.0158 -0.02
10 0.0 0.0000 0.0048 0.00

RUNS TOTAL = RUNS UP + RUNS DOWN

STATISTIC = NUMBER OF RUNS TOTAL

OF LENGTH EXACTLY I

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4273.htm (4 of 6) [5/7/2002 3:16:14 PM]

1.4.2.7.3. Quantitative Output and Interpretation

I STAT EXP(STAT) SD(STAT) Z

1 373.0 416.7500 20.5701 -2.13

2 171.0 183.1000 10.6068 -1.14
3 61.0 52.6472 6.4668 1.29
4 20.0 11.4667 3.2759 2.60
5 3.0 2.0243 1.4123 0.69
6 1.0 0.3014 0.5483 1.27
7 0.0 0.0389 0.1971 -0.20
8 0.0 0.0044 0.0665 -0.07
9 0.0 0.0005 0.0212 -0.02
10 0.0 0.0000 0.0065 -0.01

STATISTIC = NUMBER OF RUNS TOTAL

OF LENGTH I OR MORE

I STAT EXP(STAT) SD(STAT) Z

1 629.0 666.3333 13.3212 -2.80

2 256.0 249.5833 8.8942 0.72
3 85.0 66.4833 6.8758 2.69
4 24.0 13.8361 3.5639 2.85
5 4.0 2.3694 1.5256 1.07
6 1.0 0.3452 0.5866 1.12
7 0.0 0.0438 0.2092 -0.21
8 0.0 0.0049 0.0701 -0.07
9 0.0 0.0005 0.0223 -0.02
10 0.0 0.0000 0.0067 -0.01

LENGTH OF THE LONGEST RUN UP = 5

LENGTH OF THE LONGEST RUN DOWN = 6
LENGTH OF THE LONGEST RUN UP OR DOWN = 6

NUMBER OF POSITIVE DIFFERENCES = 505

NUMBER OF NEGATIVE DIFFERENCES = 469
NUMBER OF ZERO DIFFERENCES = 25

Values in the column labeled "Z" greater than 1.96 or less than -1.96 are statistically significant
at the 5% level. Due to the number of values that are larger than the 1.96 cut-off, we conclude
that the data are not random. However, in this case the evidence from the runs test is not nearly
as strong as it is from the autocorrelation plot.

Distributional Since we rejected the randomness assumption, the distributional tests are not meaningful.
Analysis Therefore, these quantitative tests are omitted. Since the Grubbs' test for outliers also assumes
the approximate normality of the data, we omit Grubbs' test as well.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4273.htm (5 of 6) [5/7/2002 3:16:14 PM]

1.4.2.7.3. Quantitative Output and Interpretation

Univariate It is sometimes useful and convenient to summarize the above results in a report.
Report

Analysis for resistor case study

1: Sample Size = 1000

2: Location
Mean = 28.01635
Standard Deviation of Mean = 0.002008
95% Confidence Interval for Mean = (28.0124,28.02029)
Drift with respect to location? = NO

3: Variation
Standard Deviation = 0.063495
95% Confidence Interval for SD = (0.060829,0.066407)
Change in variation?
(based on Levene's test on quarters
of the data) = YES

4: Randomness
Autocorrelation = 0.972158
Data Are Random?
(as measured by autocorrelation) = NO

5: Distribution
Distributional test omitted due to
non-randomness of the data

6: Statistical Control
(i.e., no drift in location or scale,
data are random, distribution is
fixed)
Data Set is in Statistical Control? = NO

7: Outliers?
(Grubbs' test omitted due to
non-randomness of the data

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4273.htm (6 of 6) [5/7/2002 3:16:14 PM]

1.4.2.7.4. Work This Example Yourself

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.7. Standard Resistor

1.4.2.7.4. Work This Example Yourself

View This page allows you to repeat the analysis outlined in the case study
Dataplot description on the previous page using Dataplot . It is required that you
Macro for have already downloaded and installed Dataplot and configured your
this Case browser. to run Dataplot. Output from each analysis step below will be
Study displayed in one or more of the Dataplot windows. The four main
windows are the Output window, the Graphics window, the Command
History window, and the data sheet window. Across the top of the main
windows there are menus for executing Dataplot commands. Across the
bottom is a command entry window where commands can be typed in.

Data Analysis Steps Results and Conclusions

Click on the links below to start Dataplot and run this case study
yourself. Each step may use results from previous steps, so please be
patient. Wait until the software verifies that the current step is
complete before clicking on the next step.
The links in this column will connect you with more detailed information about
NOTE: This case study has 1,000 points. For better performance, it each analysis step from the case study description.
is highly recommended that you check the "No Update" box on the
Spreadsheet window for this case study. This will suppress
subsequent updating of the Spreadsheet window as the data are
created or modified.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4274.htm (1 of 3) [5/7/2002 3:16:15 PM]

1.4.2.7.4. Work This Example Yourself

1. Invoke Dataplot and read data.

1. Read in the data. 1. You have read 1 column of numbers

into Dataplot, variable Y.

2. 4-plot of the data.

1. 4-plot of Y. 1. Based on the 4-plot, there are shifts

in location and variation and the data
are not random.

3. Generate the individual plots.

1. Generate a run sequence plot. 1. The run sequence plot indicates that
there are shifts of location and
variation.

2. Generate a lag plot. 2. The lag plot shows a strong linear

pattern, which indicates significant
non-randomness.

4. Generate summary statistics, quantitative

analysis, and print a univariate report.

1. Generate a table of summary 1. The summary statistics table displays

statistics. 25+ statistics.

2. Generate the sample mean, a confidence 2. The mean is 28.0163 and a 95%
interval for the population mean, and confidence interval is (28.0124,28.02029).
compute a linear fit to detect drift in The linear fit indicates drift in

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4274.htm (2 of 3) [5/7/2002 3:16:15 PM]

1.4.2.7.4. Work This Example Yourself
location. location since the slope parameter
estimate is statistically significant.

3. Generate the sample standard deviation, 3. The standard deviation is 0.0635 with
a confidence interval for the population a 95% confidence interval of (0.060829,0.066407).
standard deviation, and detect drift in Levene's test indicates significant
variation by dividing the data into change in variation.
quarters and computing Levene's test for
equal standard deviations.

4. Check for randomness by generating an 4. The lag 1 autocorrelation is 0.97.

autocorrelation plot and a runs test. From the autocorrelation plot, this is
outside the 95% confidence interval
bands, indicating significant non-randomness.

5. Print a univariate report (this assumes 5. The results are summarized in a

steps 2 thru 5 have already been run). convenient report.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4274.htm (3 of 3) [5/7/2002 3:16:15 PM]

1.4.2.8. Heat Flow Meter 1

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies

1.4.2.8. Heat Flow Meter 1

Heat Flow This example illustrates the univariate analysis of standard resistor data.
Meter
Calibration
and Stability

1. Background and Data

2. Graphical Output and Interpretation
3. Quantitative Output and Interpretation
4. Work This Example Yourself

http://www.itl.nist.gov/div898/handbook/eda/section4/eda428.htm [5/7/2002 3:16:15 PM]

1.4.2.8.1. Background and Data

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.8. Heat Flow Meter 1

1.4.2.8.1. Background and Data

Generation This data set was collected by Bob Zarr of NIST in January, 1990 from
a heat flow meter calibration and stability analysis. The response
variable is a calibration factor.
The motivation for studying this data set is to illustrate a well-behaved
process where the underlying assumptions hold and the process is in
statistical control.
This file can be read by Dataplot with the following commands:
SKIP 25
READ ZARR13.DAT Y

Resulting The following are the data used for this case study.
Data
9.206343
9.299992
9.277895
9.305795
9.275351
9.288729
9.287239
9.260973
9.303111
9.275674
9.272561
9.288454
9.255672
9.252141
9.297670
9.266534
9.256689
9.277542
9.248205

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4281.htm (1 of 5) [5/7/2002 3:16:16 PM]

1.4.2.8.1. Background and Data

9.252107
9.276345
9.278694
9.267144
9.246132
9.238479
9.269058
9.248239
9.257439
9.268481
9.288454
9.258452
9.286130
9.251479
9.257405
9.268343
9.291302
9.219460
9.270386
9.218808
9.241185
9.269989
9.226585
9.258556
9.286184
9.320067
9.327973
9.262963
9.248181
9.238644
9.225073
9.220878
9.271318
9.252072
9.281186
9.270624
9.294771
9.301821
9.278849
9.236680
9.233988
9.244687
9.221601
9.207325
9.258776
9.275708

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4281.htm (2 of 5) [5/7/2002 3:16:16 PM]

1.4.2.8.1. Background and Data

9.268955
9.257269
9.264979
9.295500
9.292883
9.264188
9.280731
9.267336
9.300566
9.253089
9.261376
9.238409
9.225073
9.235526
9.239510
9.264487
9.244242
9.277542
9.310506
9.261594
9.259791
9.253089
9.245735
9.284058
9.251122
9.275385
9.254619
9.279526
9.275065
9.261952
9.275351
9.252433
9.230263
9.255150
9.268780
9.290389
9.274161
9.255707
9.261663
9.250455
9.261952
9.264041
9.264509
9.242114
9.239674
9.221553

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4281.htm (3 of 5) [5/7/2002 3:16:16 PM]

1.4.2.8.1. Background and Data

9.241935
9.215265
9.285930
9.271559
9.266046
9.285299
9.268989
9.267987
9.246166
9.231304
9.240768
9.260506
9.274355
9.292376
9.271170
9.267018
9.308838
9.264153
9.278822
9.255244
9.229221
9.253158
9.256292
9.262602
9.219793
9.258452
9.267987
9.267987
9.248903
9.235153
9.242933
9.253453
9.262671
9.242536
9.260803
9.259825
9.253123
9.240803
9.238712
9.263676
9.243002
9.246826
9.252107
9.261663
9.247311
9.306055

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4281.htm (4 of 5) [5/7/2002 3:16:16 PM]

1.4.2.8.1. Background and Data

9.237646
9.248937
9.256689
9.265777
9.299047
9.244814
9.287205
9.300566
9.256621
9.271318
9.275154
9.281834
9.253158
9.269024
9.282077
9.277507
9.284910
9.239840
9.268344
9.247778
9.225039
9.230750
9.270024
9.265095
9.284308
9.280697
9.263032
9.291851
9.252072
9.244031
9.283269
9.196848
9.231372
9.232963
9.234956
9.216746
9.274107
9.273776

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4281.htm (5 of 5) [5/7/2002 3:16:16 PM]

1.4.2.8.2. Graphical Output and Interpretation

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.8. Heat Flow Meter 1

1.4.2.8.2. Graphical Output and

Interpretation
Goal The goal of this analysis is threefold:
1. Determine if the univariate model:

is appropriate and valid.

2. Determine if the typical underlying assumptions for an "in
control" measurement process are valid. These assumptions are:
1. random drawings;
2. from a fixed distribution;
3. with the distribution having a fixed location; and
4. the distribution having a fixed scale.
3. Determine if the confidence interval

is appropriate and valid where s is the standard deviation of the

original data.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4282.htm (1 of 4) [5/7/2002 3:16:16 PM]

1.4.2.8.2. Graphical Output and Interpretation

4-Plot of
Data

Interpretation The assumptions are addressed by the graphics shown above:

1. The run sequence plot (upper left) indicates that the data do not
have any significant shifts in location or scale over time.
2. The lag plot (upper right) does not indicate any non-random
pattern in the data.
3. The histogram (lower left) shows that the data are reasonably
symmetric, there does not appear to be significant outliers in the
tails, and it seems reasonable to assume that the data are from
approximately a normal distribution.
4. The normal probability plot (lower right) verifies that an
assumption of normality is in fact reasonable.

Individual Although it is generally unnecessary, the plots can be generated

Plots individually to give more detail.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4282.htm (2 of 4) [5/7/2002 3:16:16 PM]

1.4.2.8.2. Graphical Output and Interpretation

Run
Sequence
Plot

Lag Plot

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4282.htm (3 of 4) [5/7/2002 3:16:16 PM]

1.4.2.8.2. Graphical Output and Interpretation

Histogram
(with
overlaid
Normal PDF)

Normal
Probability
Plot

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4282.htm (4 of 4) [5/7/2002 3:16:16 PM]

1.4.2.8.3. Quantitative Output and Interpretation

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.8. Heat Flow Meter 1

1.4.2.8.3. Quantitative Output and Interpretation

Summary As a first step in the analysis, a table of summary statistics is computed from the data. The
Statistics following table, generated by Dataplot, shows a typical set of statistics.

SUMMARY

NUMBER OF OBSERVATIONS = 195

***********************************************************************
* LOCATION MEASURES * DISPERSION MEASURES *
***********************************************************************
* MIDRANGE = 0.9262411E+01 * RANGE = 0.1311255E+00 *
* MEAN = 0.9261460E+01 * STAND. DEV. = 0.2278881E-01 *
* MIDMEAN = 0.9259412E+01 * AV. AB. DEV. = 0.1788945E-01 *
* MEDIAN = 0.9261952E+01 * MINIMUM = 0.9196848E+01 *
* = * LOWER QUART. = 0.9246826E+01 *
* = * LOWER HINGE = 0.9246496E+01 *
* = * UPPER HINGE = 0.9275530E+01 *
* = * UPPER QUART. = 0.9275708E+01 *
* = * MAXIMUM = 0.9327973E+01 *
***********************************************************************
* RANDOMNESS MEASURES * DISTRIBUTIONAL MEASURES *
***********************************************************************
* AUTOCO COEF = 0.2805789E+00 * ST. 3RD MOM. = -0.8537455E-02 *
* = 0.0000000E+00 * ST. 4TH MOM. = 0.3049067E+01 *
* = 0.0000000E+00 * ST. WILK-SHA = 0.9458605E+01 *
* = * UNIFORM PPCC = 0.9735289E+00 *
* = * NORMAL PPCC = 0.9989640E+00 *
* = * TUK -.5 PPCC = 0.8927904E+00 *
* = * CAUCHY PPCC = 0.6360204E+00 *
***********************************************************************

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4283.htm (1 of 7) [5/7/2002 3:16:17 PM]

1.4.2.8.3. Quantitative Output and Interpretation

Location One way to quantify a change in location over time is to fit a straight line to the data set using
the index variable X = 1, 2, ..., N, with N denoting the number of observations. If there is no
significant drift in the location, the slope parameter should be zero. For this data set, Dataplot
generates the following output:

LEAST SQUARES MULTILINEAR FIT

SAMPLE SIZE N = 195
NUMBER OF VARIABLES = 1
NO REPLICATION CASE

PARAMETER ESTIMATES (APPROX. ST. DEV.) T VALUE

1 A0 9.26699 (0.3253E-02) 2849.
2 A1 X -0.564115E-04 (0.2878E-04) -1.960

RESIDUAL STANDARD DEVIATION = 0.2262372E-01

RESIDUAL DEGREES OF FREEDOM = 193
The slope parameter, A1, has a t value of -1.96 which is (barely) statistically significant since
it is essentially equal to the 95% level cutoff of -1.96. However, notice that the value of the
slope parameter estimate is -0.00056. This slope, even though statistically significant, can
essentially be considered zero.

Variation One simple way to detect a change in variation is with a Bartlett test after dividing the data set
into several equal-sized intervals. The choice of the number of intervals is somewhat arbitrary,
although values of 4 or 8 are reasonable. Dataplot generated the following output for the
Bartlett test.

BARTLETT TEST
(STANDARD DEFINITION)
NULL HYPOTHESIS UNDER TEST--ALL SIGMA(I) ARE EQUAL

TEST:
DEGREES OF FREEDOM = 3.000000

TEST STATISTIC VALUE = 3.147338

CUTOFF: 95% PERCENT POINT = 7.814727
CUTOFF: 99% PERCENT POINT = 11.34487

CHI-SQUARE CDF VALUE = 0.630538

NULL NULL HYPOTHESIS NULL HYPOTHESIS

HYPOTHESIS ACCEPTANCE INTERVAL CONCLUSION
ALL SIGMA EQUAL (0.000,0.950) ACCEPT

In this case, since the Bartlett test statistic of 3.14 is less than the critical value at the 5%
significance level of 7.81, we conclude that the standard deviations are not significantly
different in the 4 intervals. That is, the assumption of constant scale is valid.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4283.htm (2 of 7) [5/7/2002 3:16:17 PM]

1.4.2.8.3. Quantitative Output and Interpretation

Randomness There are many ways in which data can be non-random. However, most common forms of
non-randomness can be detected with a few simple tests. The lag plot in the previous section is
a simple graphical technique.
Another check is an autocorrelation plot that shows the autocorrelations for various lags.
Confidence bands can be plotted at the 95% and 99% confidence levels. Points outside this
band indicate statistically significant values (lag 0 is always 1). Dataplot generated the
following autocorrelation plot.

The lag 1 autocorrelation, which is generally the one of greatest interest, is 0.281. The critical
values at the 5% significance level are -0.087 and 0.087. This indicates that the lag 1
autocorrelation is statistically significant, so there is evidence of non-randomness.
A common test for randomness is the runs test.

RUNS UP

STATISTIC = NUMBER OF RUNS UP

OF LENGTH EXACTLY I

I STAT EXP(STAT) SD(STAT) Z

1 35.0 40.6667 6.4079 -0.88

2 8.0 17.7583 3.3021 -2.96
3 12.0 5.0806 2.0096 3.44
4 3.0 1.1014 1.0154 1.87
5 0.0 0.1936 0.4367 -0.44
6 0.0 0.0287 0.1692 -0.17
7 0.0 0.0037 0.0607 -0.06
8 0.0 0.0004 0.0204 -0.02
9 0.0 0.0000 0.0065 -0.01
10 0.0 0.0000 0.0020 0.00

STATISTIC = NUMBER OF RUNS UP

OF LENGTH I OR MORE

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4283.htm (3 of 7) [5/7/2002 3:16:17 PM]

1.4.2.8.3. Quantitative Output and Interpretation

I STAT EXP(STAT) SD(STAT) Z

1 58.0 64.8333 4.1439 -1.65

2 23.0 24.1667 2.7729 -0.42
3 15.0 6.4083 2.1363 4.02
4 3.0 1.3278 1.1043 1.51
5 0.0 0.2264 0.4716 -0.48
6 0.0 0.0328 0.1809 -0.18
7 0.0 0.0041 0.0644 -0.06
8 0.0 0.0005 0.0215 -0.02
9 0.0 0.0000 0.0068 -0.01
10 0.0 0.0000 0.0021 0.00

RUNS DOWN

STATISTIC = NUMBER OF RUNS DOWN

OF LENGTH EXACTLY I

I STAT EXP(STAT) SD(STAT) Z

1 33.0 40.6667 6.4079 -1.20

2 18.0 17.7583 3.3021 0.07
3 3.0 5.0806 2.0096 -1.04
4 3.0 1.1014 1.0154 1.87
5 1.0 0.1936 0.4367 1.85
6 0.0 0.0287 0.1692 -0.17
7 0.0 0.0037 0.0607 -0.06
8 0.0 0.0004 0.0204 -0.02
9 0.0 0.0000 0.0065 -0.01
10 0.0 0.0000 0.0020 0.00

STATISTIC = NUMBER OF RUNS DOWN

OF LENGTH I OR MORE

I STAT EXP(STAT) SD(STAT) Z

1 58.0 64.8333 4.1439 -1.65

2 25.0 24.1667 2.7729 0.30
3 7.0 6.4083 2.1363 0.28
4 4.0 1.3278 1.1043 2.42
5 1.0 0.2264 0.4716 1.64
6 0.0 0.0328 0.1809 -0.18
7 0.0 0.0041 0.0644 -0.06
8 0.0 0.0005 0.0215 -0.02
9 0.0 0.0000 0.0068 -0.01
10 0.0 0.0000 0.0021 0.00

RUNS TOTAL = RUNS UP + RUNS DOWN

STATISTIC = NUMBER OF RUNS TOTAL

OF LENGTH EXACTLY I

I STAT EXP(STAT) SD(STAT) Z

1 68.0 81.3333 9.0621 -1.47

2 26.0 35.5167 4.6698 -2.04
3 15.0 10.1611 2.8420 1.70
4 6.0 2.2028 1.4360 2.64
5 1.0 0.3871 0.6176 0.99

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4283.htm (4 of 7) [5/7/2002 3:16:17 PM]

1.4.2.8.3. Quantitative Output and Interpretation
6 0.0 0.0574 0.2392 -0.24
7 0.0 0.0074 0.0858 -0.09
8 0.0 0.0008 0.0289 -0.03
9 0.0 0.0001 0.0092 -0.01
10 0.0 0.0000 0.0028 0.00

STATISTIC = NUMBER OF RUNS TOTAL

OF LENGTH I OR MORE

I STAT EXP(STAT) SD(STAT) Z

1 116.0 129.6667 5.8604 -2.33

2 48.0 48.3333 3.9215 -0.09
3 22.0 12.8167 3.0213 3.04
4 7.0 2.6556 1.5617 2.78
5 1.0 0.4528 0.6669 0.82
6 0.0 0.0657 0.2559 -0.26
7 0.0 0.0083 0.0911 -0.09
8 0.0 0.0009 0.0305 -0.03
9 0.0 0.0001 0.0097 -0.01
10 0.0 0.0000 0.0029 0.00

LENGTH OF THE LONGEST RUN UP = 4

LENGTH OF THE LONGEST RUN DOWN = 5
LENGTH OF THE LONGEST RUN UP OR DOWN = 5

NUMBER OF POSITIVE DIFFERENCES = 98

NUMBER OF NEGATIVE DIFFERENCES = 95
NUMBER OF ZERO DIFFERENCES = 1

Values in the column labeled "Z" greater than 1.96 or less than -1.96 are statistically
significant at the 5% level. The runs test does indicate some non-randomness.
Although the autocorrelation plot and the runs test indicate some mild non-randomness, the
violation of the randomness assumption is not serious enough to warrant developing a more
sophisticated model. It is common in practice that some of the assumptions are mildly violated
and it is a judgement call as to whether or not the violations are serious enough to warrant
developing a more sophisticated model for the data.

Distributional Probability plots are a graphical test for assessing if a particular distribution provides an
Analysis adequate fit to a data set.
A quantitative enhancement to the probability plot is the correlation coefficient of the points
on the probability plot. For this data set the correlation coefficient is 0.996. Since this is
greater than the critical value of 0.987 (this is a tabulated value), the normality assumption is
not rejected.
Chi-square and Kolmogorov-Smirnov goodness-of-fit tests are alternative methods for
assessing distributional adequacy. The Wilk-Shapiro and Anderson-Darling tests can be used
to test for normality. Dataplot generates the following output for the Anderson-Darling
normality test.

ANDERSON-DARLING 1-SAMPLE TEST

THAT THE DATA CAME FROM A NORMAL DISTRIBUTION

1. STATISTICS:

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4283.htm (5 of 7) [5/7/2002 3:16:17 PM]

1.4.2.8.3. Quantitative Output and Interpretation
NUMBER OF OBSERVATIONS = 195
MEAN = 9.261460
STANDARD DEVIATION = 0.2278881E-01

ANDERSON-DARLING TEST STATISTIC VALUE = 0.1264954

ADJUSTED TEST STATISTIC VALUE = 0.1290070

2. CRITICAL VALUES:
90 % POINT = 0.6560000
95 % POINT = 0.7870000
97.5 % POINT = 0.9180000
99 % POINT = 1.092000

3. CONCLUSION (AT THE 5% LEVEL):

THE DATA DO COME FROM A NORMAL DISTRIBUTION.

The Anderson-Darling test also does not reject the normality assumption because the test
statistic, 0.129, is less than the critical value at the 5% significance level of 0.918.

Outlier A test for outliers is the Grubbs' test. Dataplot generated the following output for Grubbs' test.
Analysis

GRUBBS TEST FOR OUTLIERS

(ASSUMPTION: NORMALITY)

1. STATISTICS:
NUMBER OF OBSERVATIONS = 195
MINIMUM = 9.196848
MEAN = 9.261460
MAXIMUM = 9.327973
STANDARD DEVIATION = 0.2278881E-01

GRUBBS TEST STATISTIC = 2.918673

2. PERCENT POINTS OF THE REFERENCE DISTRIBUTION

FOR GRUBBS TEST STATISTIC
0 % POINT = 0.0000000E+00
50 % POINT = 2.984294
75 % POINT = 3.181226
90 % POINT = 3.424672
95 % POINT = 3.597898
99 % POINT = 3.970215

3. CONCLUSION (AT THE 5% LEVEL):

THERE ARE NO OUTLIERS.

For this data set, Grubbs' test does not detect any outliers at the 25%, 10%, 5%, and 1%
significance levels.

Model Since the underlying assumptions were validated both graphically and analytically, with a mild
violation of the randomness assumption, we conclude that a reasonable model for the data is:

We can express the uncertainty for C, here estimated by 9.26146, as the 95% confidence
interval (9.258242,9.26479).

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4283.htm (6 of 7) [5/7/2002 3:16:17 PM]

1.4.2.8.3. Quantitative Output and Interpretation

Univariate It is sometimes useful and convenient to summarize the above results in a report. The report
Report for the heat flow meter data follows.

Analysis for heat flow meter data

1: Sample Size = 195

2: Location
Mean = 9.26146
Standard Deviation of Mean = 0.001632
95% Confidence Interval for Mean = (9.258242,9.264679)
Drift with respect to location? = NO

3: Variation
Standard Deviation = 0.022789
95% Confidence Interval for SD = (0.02073,0.025307)
Drift with respect to variation?
(based on Bartlett's test on quarters
of the data) = NO

4: Randomness
Autocorrelation = 0.280579
Data are Random?
(as measured by autocorrelation) = NO

5: Distribution
Normal PPCC = 0.998965
Data are Normal?
(as measured by Normal PPCC) = YES

6: Statistical Control
(i.e., no drift in location or scale,
data are random, distribution is
fixed, here we are testing only for
fixed normal)
Data Set is in Statistical Control? = YES

7: Outliers?
(as determined by Grubbs' test) = NO

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4283.htm (7 of 7) [5/7/2002 3:16:17 PM]

1.4.2.8.4. Work This Example Yourself

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.8. Heat Flow Meter 1

1.4.2.8.4. Work This Example Yourself

View This page allows you to repeat the analysis outlined in the case study
Dataplot description on the previous page using Dataplot . It is required that you
Macro for have already downloaded and installed Dataplot and configured your
this Case browser. to run Dataplot. Output from each analysis step below will be
Study displayed in one or more of the Dataplot windows. The four main
windows are the Output window, the Graphics window, the Command
History window, and the data sheet window. Across the top of the main
windows there are menus for executing Dataplot commands. Across the
bottom is a command entry window where commands can be typed in.

Data Analysis Steps Results and Conclusions

Click on the links below to start Dataplot and run this case study
yourself. Each step may use results from previous steps, so please be The links in this column will connect you with more detailed information
patient. Wait until the software verifies that the current step is about each analysis step from the case study description.
complete before clicking on the next step.

1. Invoke Dataplot and read data.

1. Read in the data. 1. You have read 1 column of numbers

into Dataplot, variable Y.

2. 4-plot of the data.

1. 4-plot of Y. 1. Based on the 4-plot, there are no shifts

in location or scale, and the data seem to
follow a normal distribution.

3. Generate the individual plots.

1. Generate a run sequence plot. 1. The run sequence plot indicates that
there are no shifts of location or
scale.

2. Generate a lag plot. 2. The lag plot does not indicate any
significant patterns (which would
show the data were not random).

3. Generate a histogram with an

3. The histogram indicates that a
overlaid normal pdf.
normal distribution is a good
distribution for these data.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4284.htm (1 of 2) [5/7/2002 3:16:17 PM]

1.4.2.8.4. Work This Example Yourself

4. Generate a normal probability 4. The normal probability plot verifies

plot. that the normal distribution is a
reasonable distribution for these data.

4. Generate summary statistics, quantitative

analysis, and print a univariate report.

1. Generate a table of summary 1. The summary statistics table displays

statistics. 25+ statistics.

2. Generate the mean, a confidence 2. The mean is 9.261 and a 95%

interval for the mean, and compute confidence interval is (9.258,9.265).
a linear fit to detect drift in The linear fit indicates no drift in
location. location since the slope parameter
estimate is essentially zero.

3. Generate the standard deviation, a 3. The standard deviation is 0.023 with

confidence interval for the standard a 95% confidence interval of (0.0207,0.0253).
deviation, and detect drift in variation Bartlett's test indicates no significant
by dividing the data into quarters and change in variation.
computing Bartlett's test for equal
standard deviations.

4. Check for randomness by generating an 4. The lag 1 autocorrelation is 0.28.

autocorrelation plot and a runs test. From the autocorrelation plot, this is
statistically significant at the 95%
level.

5. Check for normality by computing the 5. The normal probability plot correlation
normal probability plot correlation coefficient is 0.999. At the 5% level,
coefficient. we cannot reject the normality assumption.

6. Check for outliers using Grubbs' test. 6. Grubbs' test detects no outliers at the
5% level.

7. Print a univariate report (this assumes 7. The results are summarized in a

steps 2 thru 6 have already been run). convenient report.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4284.htm (2 of 2) [5/7/2002 3:16:17 PM]

1.4.2.9. Airplane Glass Failure Time

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies

1.4.2.9. Airplane Glass Failure Time

Airplane This example illustrates the univariate analysis of airplane glass failure
Glass time data.
Failure
Time

1. Background and Data

2. Graphical Output and Interpretation
3. Weibull Analysis
4. Lognormal Analysis
5. Gamma Analysis
6. Power Normal Analysis
7. Power Lognormal Analysis
8. Work This Example Yourself

http://www.itl.nist.gov/div898/handbook/eda/section4/eda429.htm [5/7/2002 3:16:17 PM]

1.4.2.9.1. Background and Data

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.9. Airplane Glass Failure Time

1.4.2.9.1. Background and Data

Generation This data set was collected by Ed Fuller of NIST in December, 1993.
The response variable is time to failure for airplane glass under test.

Purpose of The goal of this case study is to find a good distributional model for the
Analysis data. Once a good distributional model has been determined, various
percent points for glass failure will be computed.
Since the data are failure times, this case study is a form of reliability
analysis. The assessing product reliability chapter contains a more
complete discussion of reliabilty methods. This case study is meant to
complement that chapter by showing the use of graphical techniques in
one aspect of reliability modeling.
Failure times are basically extreme values that do not follow a normal
distribution; non-parametric methods (techniques that do not rely on a
specific distribution) are frequently recommended for developing
confidence intervals for failure data. One problem with this approach is
that sample sizes are often small due to the expense involved in
collecting the data, and non-parametric methods do not work well for
small sample sizes. For this reason, a parametric method based on a
specific distributional model of the data is preferred if the data can be
shown to follow a specific distribution. Parametric models typically
have greater efficiency at the cost of more specific assumptions about
the data, but, it is important to verify that the distributional assumption
is indeed valid. If the distributional assumption is not justified, then the
conclusions drawn from the model may not be valid.
This file can be read by Dataplot with the following commands:
SKIP 25
READ FULLER2.DAT Y

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4291.htm (1 of 2) [5/7/2002 3:16:17 PM]

1.4.2.9.1. Background and Data

Resulting The following are the data used for this case study.
Data
18.830
20.800
21.657
23.030
23.230
24.050
24.321
25.500
25.520
25.800
26.690
26.770
26.780
27.050
27.670
29.900
31.110
33.200
33.730
33.760
33.890
34.760
35.750
35.910
36.980
37.080
37.090
39.580
44.045
45.290
45.381

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4291.htm (2 of 2) [5/7/2002 3:16:17 PM]

1.4.2.9.2. Graphical Output and Interpretation

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.9. Airplane Glass Failure Time

1.4.2.9.2. Graphical Output and Interpretation

Goal The goal of this analysis is to determine a good distributional model for these
failure time data. A secondary goal is to provide estimates for various percent
points of the data. Percent points provide an answer to questions of the type "At
what time do we expect 5% of the airplane glass to have failed?".

Initial Plots of the

Data The first step is to generate a histogram to get an overall feel for the data.

The histogram shows the following:

● The failure times range between slightly greater than 15 to slightly less
than 50.
● There are modes at approximately 28 and 38 with a gap in-between.

● The data are somewhat symmetric, but with a gap in the middle.

We next generate a normal probability plot.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4292.htm (1 of 7) [5/7/2002 3:16:18 PM]

1.4.2.9.2. Graphical Output and Interpretation

The normal probability plot has a correlation coefficient of 0.980. We can use
this number as a reference baseline when comparing the performance of other
distributional fits.

Other Potential There is a large number of distributions that would be distributional model
Distributions candidates for the data. However, we will restrict ourselves to consideration of
the following distributional models because these have proven to be useful in
reliability studies.
1. Normal distribution
2. Exponential distribution
3. Weibull distribution
4. Lognormal distribution
5. Gamma distribution
6. Power normal distribution
7. Power lognormal distribution

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4292.htm (2 of 7) [5/7/2002 3:16:18 PM]

1.4.2.9.2. Graphical Output and Interpretation

Approach There are two basic questions that need to be addressed.

1. Does a given distributional model provide an adequate fit to the data?
2. Of the candidate distributional models, is there one distribution that fits
the data better than the other candidate distributional models?
The use of probability plots and probability plot correlation coefficient (PPCC)
plots provide answers to both of these questions.

If the distribution does not have a shape parameter, we simply generate a

probability plot.
1. If we fit a straight line to the points on the probability plot, the intercept
and slope of that line provide estimates of the location and scale
parameters, respectively.
2. Our critierion for the "best fit" distribution is the one with the most linear
probability plot. The correlation coefficient of the fitted line of the points
on the probability plot, referred to as the PPCC value, provides a measure
of the linearity of the probability plot, and thus a measure of how well the
distribution fits the data. The PPCC values for multiple distributions can
be compared to address the second question above.
If the distribution does have a shape parameter, then we are actually addressing
a family of distributions rather than a single distribution. We first need to find
the optimal value of the shape parameter. The PPCC plot can be used to
determine the optimal parameter. We will use the PPCC plots in two stages. The
first stage will be over a broad range of parameter values while the second stage
will be in the neighborhood of the largest values. Although we could go further
than two stages, for practical purposes two stages is sufficient. After
determining an optimal value for the shape parameter, we use the probability
plot as above to obtain estimates of the location and scale parameters and to
determine the PPCC value. This PPCC value can be compared to the PPCC
values obtained from other distributional models.

Analyses for We analyzed the data using the approach described above for the following
Specific distributional models:
Distributions 1. Normal distribution - from the 4-plot above, the PPCC value was 0.980.
2. Exponential distribution - the exponential distribution is a special case of
the Weibull with shape parameter equal to 1. If the Weibull analysis
yields a shape parameter close to 1, then we would consider using the
simpler exponential model.
3. Weibull distribution
4. Lognormal distribution
5. Gamma distribution
6. Power normal distribution
7. Power lognormal distribution

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4292.htm (3 of 7) [5/7/2002 3:16:18 PM]

1.4.2.9.2. Graphical Output and Interpretation

Summary of The results are summarized below.

Results Normal Distribution
Max PPCC = 0.980
Estimate of location = 30.81
Estimate of scale = 7.38
Weibull Distribution
Max PPCC = 0.988
Estimate of shape = 2.13
Estimate of location = 15.9
Estimate of scale = 16.92
Lognormal Distribution
Max PPCC = 0.986
Estimate of shape = 0.18
Estimate of location = -9.96
Estimate of scale = 40.17
Gamma Distribution
Max PPCC = 0.987
Estimate of shape = 11.8
Estimate of location = 5.19
Estimate of scale = 2.17
Power Normal Distribution
Max PPCC = 0.987
Estimate of shape = 0.11
Estimate of location = 20.9
Estimate of scale = 3.3
Power Lognormal Distribution
Max PPCC = 0.988
Estimate of shape = 50
Estimate of location = 13.5
Estimate of scale = 150.8
These results indicate that several of these distributions provide an adequate
distributional model for the data. We choose the 3-parameter Weibull
distribution as the most appropriate model because it provides the best balance
between simplicity and best fit.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4292.htm (4 of 7) [5/7/2002 3:16:18 PM]

1.4.2.9.2. Graphical Output and Interpretation

Percent Point The final step in this analysis is to compute percent point estimates for the 1%,
Estimates 2.5%, 5%, 95%, 97.5%, and 99% percent points. A percent point estimate is an
estimate of the time by which a given percentage of the units will have failed.
For example, the 5% point is the time at which we estimate 5% of the units will
have failed.
To calculate these values, we use the Weibull percent point function with the
appropriate estimates of the shape, location, and scale parameters. The Weibull
percent point function can be computed in many general purpose statistical
software programs, including Dataplot.

Dataplot generated the following estimates for the percent points:

Estimated percent points using Weibull Distribution

PERCENT POINT FAILURE TIME

0.01 17.86
0.02 18.92
0.05 20.10
0.95 44.21
0.97 47.11
0.99 50.53

Quantitative Although it is generally unnecessary, we can include quantitative measures of

Measures of distributional goodness-of-fit. Three of the commonly used measures are:
Goodness of Fit 1. Chi-square goodness-of-fit.
2. Kolmogorov-Smirnov goodness-of-fit.
3. Anderson-Darling goodness-of-fit.
In this case, the sample size of 31 precludes the use of the chi-square test since
the chi-square approximation is not valid for small sample sizes. Specifically,
the smallest expected frequency should be at least 5. Although we could
combine classes, we will instead use one of the other tests. The
Kolmogorov-Smirnov test requires a fully specified distribution. Since we need
to use the data to estimate the shape, location, and scale parameters, we do not
use this test here. The Anderson-Darling test is a refinement of the
Kolmogorov-Smirnov test. We run this test for the normal, lognormal, and
Weibull distributions.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4292.htm (5 of 7) [5/7/2002 3:16:18 PM]

1.4.2.9.2. Graphical Output and Interpretation

Normal
Anderson-Darling **************************************
** Anderson-Darling normal test y **
Output **************************************

ANDERSON-DARLING 1-SAMPLE TEST

THAT THE DATA CAME FROM A NORMAL DISTRIBUTION

1. STATISTICS:
NUMBER OF OBSERVATIONS = 31
MEAN = 30.81142
STANDARD DEVIATION = 7.253381

ANDERSON-DARLING TEST STATISTIC VALUE = 0.5321903

ADJUSTED TEST STATISTIC VALUE = 0.5870153

2. CRITICAL VALUES:
90 % POINT = 0.6160000
95 % POINT = 0.7350000
97.5 % POINT = 0.8610000
99 % POINT = 1.021000

3. CONCLUSION (AT THE 5% LEVEL):

THE DATA DO COME FROM A NORMAL DISTRIBUTION.

Lognormal
Anderson-Darling *****************************************
** Anderson-Darling lognormal test y **
Output *****************************************

ANDERSON-DARLING 1-SAMPLE TEST

THAT THE DATA CAME FROM A LOGNORMAL DISTRIBUTION

1. STATISTICS:
NUMBER OF OBSERVATIONS = 31
MEAN = 3.401242
STANDARD DEVIATION = 0.2349026

ANDERSON-DARLING TEST STATISTIC VALUE = 0.3888340

ADJUSTED TEST STATISTIC VALUE = 0.4288908

2. CRITICAL VALUES:
90 % POINT = 0.6160000
95 % POINT = 0.7350000
97.5 % POINT = 0.8610000
99 % POINT = 1.021000

3. CONCLUSION (AT THE 5% LEVEL):

THE DATA DO COME FROM A LOGNORMAL DISTRIBUTION.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4292.htm (6 of 7) [5/7/2002 3:16:18 PM]

1.4.2.9.2. Graphical Output and Interpretation

Weibull
Anderson-Darling ***************************************
** Anderson-Darling Weibull test y **
Output ***************************************

ANDERSON-DARLING 1-SAMPLE TEST

THAT THE DATA CAME FROM A WEIBULL DISTRIBUTION

1. STATISTICS:
NUMBER OF OBSERVATIONS = 31
MEAN = 30.81142
STANDARD DEVIATION = 7.253381
SHAPE PARAMETER = 4.635379
SCALE PARAMETER = 33.67423

ANDERSON-DARLING TEST STATISTIC VALUE = 0.5973396

ADJUSTED TEST STATISTIC VALUE = 0.6187967

2. CRITICAL VALUES:
90 % POINT = 0.6370000
95 % POINT = 0.7570000
97.5 % POINT = 0.8770000
99 % POINT = 1.038000

3. CONCLUSION (AT THE 5% LEVEL):

THE DATA DO COME FROM A WEIBULL DISTRIBUTION.
Note that for the Weibull distribution, the Anderson-Darling test is actually
testing the 2-parameter Weibull distribution (based on maximum likelihood
estimates), not the 3-parameter Weibull distribution. However, passing the
2-parameter Weibull distribution does give evidence that the Weibull is an
appropriate distributional model even though we used a different parameter
estimation method.

Conclusions The Anderson-Darling test passes all three of these distributions.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4292.htm (7 of 7) [5/7/2002 3:16:18 PM]

1.4.2.9.3. Weibull Analysis

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.9. Airplane Glass Failure Time

1.4.2.9.3. Weibull Analysis

Plots for The following plots were generated for a Weibull distribution.
Weibull
Distribution

Conclusions We can make the following conclusions from these plots.

1. The optimal value, in the sense of having the most linear
probability plot, of the shape parameter gamma is 2.13.
2. At the optimal value of the shape parameter, the PPCC value is
0.988.
3. At the optimal value of the shape parameter, the estimate of the
location parameter is 15.90 and the estimate of the scale
parameter is 16.92.
4. Fine tuning the estimate of gamma (from 2 to 2.13) has minimal
impact on the PPCC value.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4293.htm (1 of 3) [5/7/2002 3:16:18 PM]

1.4.2.9.3. Weibull Analysis

Alternative The Weibull plot and the Weibull hazard plot are alternative graphical
Plots analysis procedures to the PPCC plots and probability plots.
These two procedures, especially the Weibull plot, are very commonly
employed. That not withstanding, the disadvantage of these two
procedures is that they both assume that the location parameter (i.e., the
lower bound) is zero and that we are fitting a 2-parameter Weibull
instead of a 3-parameter Weibull. The advantage is that there is an
extensive literature on these methods and they have been designed to
work with either censored or uncensored data.

Weibull Plot

This Weibull plot shows the following

1. The Weibull plot is approximately linear indicating that the
2-parameter Weibull provides an adequate fit to the data.
2. The estimate of the shape parameter is 5.28 and the estimate of
the scale parameter is 33.32.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4293.htm (2 of 3) [5/7/2002 3:16:18 PM]

1.4.2.9.3. Weibull Analysis

Weibull
Hazard Plot

The construction and interpretation of the Weibull hazard plot is

discussed in the Assessing Product Reliability chapter.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4293.htm (3 of 3) [5/7/2002 3:16:18 PM]

1.4.2.9.4. Lognormal Analysis

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.9. Airplane Glass Failure Time

1.4.2.9.4. Lognormal Analysis

Plots for The following plots were generated for a lognormal distribution.
Lognormal
Distribution

Conclusions We can make the following conclusions from these plots.

1. The optimal value, in the sense of having the most linear
probability plot, of the shape parameter is 0.18.
2. At the optimal value of the shape parameter, the PPCC value is
0.986.
3. At the optimal value of the shape parameter, the estimate of the
location parameter is -9.96 and the estimate of the scale parameter
is 40.17.
4. Fine tuning the estimate of the shape parameter (from 0.2 to 0.18)
has minimal impact on the PPCC value.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4294.htm (1 of 2) [5/7/2002 3:16:18 PM]

1.4.2.9.4. Lognormal Analysis

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4294.htm (2 of 2) [5/7/2002 3:16:18 PM]

1.4.2.9.5. Gamma Analysis

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.9. Airplane Glass Failure Time

1.4.2.9.5. Gamma Analysis

Plots for The following plots were generated for a gamma distribution.
Gamma
Distribution

Conclusions We can make the following conclusions from these plots.

1. The optimal value, in the sense of having the most linear
probability plot, of the shape parameter is 11.8.
2. At the optimal value of the shape parameter, the PPCC value is
0.987.
3. At the optimal value of the shape parameter, the estimate of the
location parameter is 5.19 and the estimate of the scale parameter
is 2.17.
4. Fine tuning the estimate of (from 12 to 11.8) has some impact
on the PPCC value (from 0.978 to 0.987).

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4295.htm (1 of 2) [5/7/2002 3:16:19 PM]

1.4.2.9.5. Gamma Analysis

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4295.htm (2 of 2) [5/7/2002 3:16:19 PM]

1.4.2.9.6. Power Normal Analysis

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.9. Airplane Glass Failure Time

1.4.2.9.6. Power Normal Analysis

Plots for The following plots were generated for a power normal distribution.
Power
Normal
Distribution

Conclusions We can make the following conclusions from these plots.

1. A reasonable value, in the sense of having the most linear
probability plot, of the shape parameter p is 0.11.
2. At the this value of the shape parameter, the PPCC value is 0.987.
3. At the optimal value of the shape parameter, the estimate of the
location parameter is 20.9 and the estimate of the scale parameter
is 3.3.
4. Fine tuning the estimate of p (from 1 to 0.11) results in a slight
improvement of the the computed PPCC value (from 0.980 to
0.987).

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4296.htm (1 of 2) [5/7/2002 3:16:19 PM]

1.4.2.9.6. Power Normal Analysis

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4296.htm (2 of 2) [5/7/2002 3:16:19 PM]

1.4.2.9.7. Power Lognormal Analysis

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.9. Airplane Glass Failure Time

1.4.2.9.7. Power Lognormal Analysis

Plots for The following plots were generated for a power lognormal distribution.
Power
Lognormal
Distribution

Conclusions We can make the following conclusions from these plots.

1. A reasonable value, in the sense of having the most linear
probability plot, of the shape parameter p is 100 (i.e., p is
asymptotically increasing).
2. At this value of the shape parameter, the PPCC value is 0.987.
3. At this value of the shape parameter, the estimate of the location
parameter is 12.01 and the estimate of the scale parameter is
212.92.
4. Fine tuning the estimate of p (from 50 to 100) has minimal impact
on the PPCC value.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4297.htm (1 of 2) [5/7/2002 3:16:19 PM]

1.4.2.9.7. Power Lognormal Analysis

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4297.htm (2 of 2) [5/7/2002 3:16:19 PM]

1.4.2.9.8. Work This Example Yourself

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.9. Airplane Glass Failure Time

1.4.2.9.8. Work This Example Yourself

View This page allows you to repeat the analysis outlined in the case study
Dataplot description on the previous page using Dataplot . It is required that you
Macro for have already downloaded and installed Dataplot and configured your
this Case browser. to run Dataplot. Output from each analysis step below will be
Study displayed in one or more of the Dataplot windows. The four main
windows are the Output window, the Graphics window, the Command
History window, and the data sheet window. Across the top of the main
windows there are menus for executing Dataplot commands. Across the
bottom is a command entry window where commands can be typed in.

Data Analysis Steps Results and Conclusions

Click on the links below to start Dataplot and run this case study
yourself. Each step may use results from previous steps, so please be The links in this column will connect you with more detailed information
patient. Wait until the software verifies that the current step is about each analysis step from the case study description.
complete before clicking on the next step.

1. Invoke Dataplot and read data.

1. Read in the data. 1. You have read 1 column of numbers

into Dataplot, variable Y.

2. 4-plot of the data.

1. 4-plot of Y. 1. The failure times are in the range 15 to

50. The histogram and normal probability
plot indicate a normal distribution fits
the data reasonably well, but we can probably
do better.

3. Generate the Weibull analysis.

1. Generate 2 iterations of the 1. The Weibull analysis results in a

Weibull PPCC plot, a Weibull maximum PPCC value of 0.988.
probability plot, and estimate
some percent points.

2. Generate a Weibull plot. 2. The Weibull plot permits the

estimation of a 2-parameter Weibull
model.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4298.htm (1 of 2) [5/7/2002 3:16:25 PM]

1.4.2.9.8. Work This Example Yourself
3. Generate a Weibull hazard plot. 3. The Weibull hazard plot is
approximately linear, indicating
that the Weibull provides a good
distributional model for these data.

4. Generate the lognormal analysis.

1. Generate 2 iterations of the 1. The lognormal analysis results in

lognormal PPCC plot and a a maximum PPCC value of 0.986.
lognormal probability plot.

5. Generate the gamma analysis.

1. Generate 2 iterations of the 1. The gamma analysis results in

gamma PPCC plot and a a maximum PPCC value of 0.987.
gamma probability plot.

6. Generate the power normal analysis.

1. Generate 2 iterations of the 1. The power normal analysis results

power normal PPCC plot and a in a maximum PPCC value of 0.988.
power normal probability plot.

7. Generate the power lognormal analysis.

1. Generate 2 iterations of the 1. The power lognormal analysis

power lognormal PPCC plot and results in a maximum PPCC value
a power lognormal probability of 0.987.
plot.

8. Generate quantitative goodness of fit tests

1. Generate Anderson-Darling test 1. The Anderson-Darling normality

for normality. test indicates the normal
distribution provides an adequate
fit to the data.

2. Generate Anderson-Darling test 2. The Anderson-Darling lognormal

for lognormal distribution. test indicates the lognormal
distribution provides an adequate
fit to the data.

3. Generate Anderson-Darling test

3. The Anderson-Darling Weibull
for Weibull distribution.
test indicates the lognormal
distribution provides an adequate
fit to the data.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda4298.htm (2 of 2) [5/7/2002 3:16:25 PM]

1.4.2.10. Ceramic Strength

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies

1.4.2.10. Ceramic Strength

Ceramic This case study analyzes the effect of machining factors on the strength
Strength of ceramics.

1. Background and Data

2. Analysis of the Response Variable
3. Analysis of Batch Effect
4. Analysis of Lab Effect
5. Analysis of Primary Factors
6. Work This Example Yourself

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a.htm [5/7/2002 3:16:26 PM]

1.4.2.10.1. Background and Data

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.10. Ceramic Strength

1.4.2.10.1. Background and Data

Generation The data for this case study were collected by Said Jahanmir of the NIST
Ceramics Division in 1996 in connection with a NIST/industry ceramics
consortium for strength optimization of ceramic strength
The motivation for studying this data set is to illustrate the analysis of multiple
factors from a designed experiment
This case study will utilize only a subset of a full study that was conducted by
Lisa Gill and James Filliben of the NIST Statistical Engineering Division
The response variable is a measure of the strength of the ceramic material
(bonded Si nitrate). The complete data set contains the following variables:
1. Factor 1 = Observation ID, i.e., run number (1 to 960)
2. Factor 2 = Lab (1 to 8)
3. Factor 3 = Bar ID within lab (1 to 30)
4. Factor 4 = Test number (1 to 4)
5. Response Variable = Strength of Ceramic
6. Factor 5 = Table speed (2 levels: 0.025 and 0.125)
7. Factor 6 = Down feed rate (2 levels: 0.050 and 0.125)
8. Factor 7 = Wheel grit size (2 levels: 150 and 80)
9. Factor 8 = Direction (2 levels: longitudinal and transverse)
10. Factor 9 = Treatment (1 to 16)
11. Factor 10 = Set of 15 within lab (2 levels: 1 and 2)
12. Factor 11 = Replication (2 levels: 1 and 2)
13. Factor 12 = Bar Batch (1 and 2)
The four primary factors of interest are:
1. Table speed (X1)
2. Down feed rate (X2)
3. Wheel grit size (X3)

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a1.htm (1 of 13) [5/7/2002 3:16:26 PM]

1.4.2.10.1. Background and Data

4. Direction (X4)
For this case study, we are using only half the data. Specifically, we are using
the data with the direction longitudinal. Therefore, we have only three primary
factors
In addtion, we are interested in the nuisance factors
1. Lab
2. Batch
The complete file can be read into Dataplot with the following commands:

DIMENSION 20 VARIABLES
SKIP 50
READ JAHANMI2.DAT RUN RUN LAB BAR SET Y X1 TO X8 BATCH

Purpose of The goals of this case study are:

Analysis 1. Determine which of the four primary factors has the strongest effect on
the strength of the ceramic material
2. Estimate the magnitude of the effects
3. Determine the optimal settings for the primary factors
4. Determine if the nuisance factors (lab and batch) have an effect on the
ceramic strength
This case study is an example of a designed experiment. The Process
Improvement chapter contains a detailed discussion of the construction and
analysis of designed experiments. This case study is meant to complement the
material in that chapter by showing how an EDA approach (emphasizing the use
of graphical techniques) can be used in the analysis of designed experiments

Resulting The following are the data used for this case study
Data
Run Lab Batch Y X1 X2 X3
1 1 1 608.781 -1 -1 -1
2 1 2 569.670 -1 -1 -1
3 1 1 689.556 -1 -1 -1
4 1 2 747.541 -1 -1 -1
5 1 1 618.134 -1 -1 -1
6 1 2 612.182 -1 -1 -1
7 1 1 680.203 -1 -1 -1
8 1 2 607.766 -1 -1 -1
9 1 1 726.232 -1 -1 -1
10 1 2 605.380 -1 -1 -1
11 1 1 518.655 -1 -1 -1
12 1 2 589.226 -1 -1 -1

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a1.htm (2 of 13) [5/7/2002 3:16:26 PM]

1.4.2.10.1. Background and Data

13 1 1 740.447 -1 -1 -1
14 1 2 588.375 -1 -1 -1
15 1 1 666.830 -1 -1 -1
16 1 2 531.384 -1 -1 -1
17 1 1 710.272 -1 -1 -1
18 1 2 633.417 -1 -1 -1
19 1 1 751.669 -1 -1 -1
20 1 2 619.060 -1 -1 -1
21 1 1 697.979 -1 -1 -1
22 1 2 632.447 -1 -1 -1
23 1 1 708.583 -1 -1 -1
24 1 2 624.256 -1 -1 -1
25 1 1 624.972 -1 -1 -1
26 1 2 575.143 -1 -1 -1
27 1 1 695.070 -1 -1 -1
28 1 2 549.278 -1 -1 -1
29 1 1 769.391 -1 -1 -1
30 1 2 624.972 -1 -1 -1
61 1 1 720.186 -1 1 1
62 1 2 587.695 -1 1 1
63 1 1 723.657 -1 1 1
64 1 2 569.207 -1 1 1
65 1 1 703.700 -1 1 1
66 1 2 613.257 -1 1 1
67 1 1 697.626 -1 1 1
68 1 2 565.737 -1 1 1
69 1 1 714.980 -1 1 1
70 1 2 662.131 -1 1 1
71 1 1 657.712 -1 1 1
72 1 2 543.177 -1 1 1
73 1 1 609.989 -1 1 1
74 1 2 512.394 -1 1 1
75 1 1 650.771 -1 1 1
76 1 2 611.190 -1 1 1
77 1 1 707.977 -1 1 1
78 1 2 659.982 -1 1 1
79 1 1 712.199 -1 1 1
80 1 2 569.245 -1 1 1
81 1 1 709.631 -1 1 1
82 1 2 725.792 -1 1 1
83 1 1 703.160 -1 1 1
84 1 2 608.960 -1 1 1
85 1 1 744.822 -1 1 1
86 1 2 586.060 -1 1 1
87 1 1 719.217 -1 1 1
88 1 2 617.441 -1 1 1

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a1.htm (3 of 13) [5/7/2002 3:16:26 PM]

1.4.2.10.1. Background and Data

89 1 1 619.137 -1 1 1
90 1 2 592.845 -1 1 1
151 2 1 753.333 1 1 1
152 2 2 631.754 1 1 1
153 2 1 677.933 1 1 1
154 2 2 588.113 1 1 1
155 2 1 735.919 1 1 1
156 2 2 555.724 1 1 1
157 2 1 695.274 1 1 1
158 2 2 702.411 1 1 1
159 2 1 504.167 1 1 1
160 2 2 631.754 1 1 1
161 2 1 693.333 1 1 1
162 2 2 698.254 1 1 1
163 2 1 625.000 1 1 1
164 2 2 616.791 1 1 1
165 2 1 596.667 1 1 1
166 2 2 551.953 1 1 1
167 2 1 640.898 1 1 1
168 2 2 636.738 1 1 1
169 2 1 720.506 1 1 1
170 2 2 571.551 1 1 1
171 2 1 700.748 1 1 1
172 2 2 521.667 1 1 1
173 2 1 691.604 1 1 1
174 2 2 587.451 1 1 1
175 2 1 636.738 1 1 1
176 2 2 700.422 1 1 1
177 2 1 731.667 1 1 1
178 2 2 595.819 1 1 1
179 2 1 635.079 1 1 1
180 2 2 534.236 1 1 1
181 2 1 716.926 1 -1 -1
182 2 2 606.188 1 -1 -1
183 2 1 759.581 1 -1 -1
184 2 2 575.303 1 -1 -1
185 2 1 673.903 1 -1 -1
186 2 2 590.628 1 -1 -1
187 2 1 736.648 1 -1 -1
188 2 2 729.314 1 -1 -1
189 2 1 675.957 1 -1 -1
190 2 2 619.313 1 -1 -1
191 2 1 729.230 1 -1 -1
192 2 2 624.234 1 -1 -1
193 2 1 697.239 1 -1 -1
194 2 2 651.304 1 -1 -1

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a1.htm (4 of 13) [5/7/2002 3:16:26 PM]

1.4.2.10.1. Background and Data

195 2 1 728.499 1 -1 -1
196 2 2 724.175 1 -1 -1
197 2 1 797.662 1 -1 -1
198 2 2 583.034 1 -1 -1
199 2 1 668.530 1 -1 -1
200 2 2 620.227 1 -1 -1
201 2 1 815.754 1 -1 -1
202 2 2 584.861 1 -1 -1
203 2 1 777.392 1 -1 -1
204 2 2 565.391 1 -1 -1
205 2 1 712.140 1 -1 -1
206 2 2 622.506 1 -1 -1
207 2 1 663.622 1 -1 -1
208 2 2 628.336 1 -1 -1
209 2 1 684.181 1 -1 -1
210 2 2 587.145 1 -1 -1
271 3 1 629.012 1 -1 1
272 3 2 584.319 1 -1 1
273 3 1 640.193 1 -1 1
274 3 2 538.239 1 -1 1
275 3 1 644.156 1 -1 1
276 3 2 538.097 1 -1 1
277 3 1 642.469 1 -1 1
278 3 2 595.686 1 -1 1
279 3 1 639.090 1 -1 1
280 3 2 648.935 1 -1 1
281 3 1 439.418 1 -1 1
282 3 2 583.827 1 -1 1
283 3 1 614.664 1 -1 1
284 3 2 534.905 1 -1 1
285 3 1 537.161 1 -1 1
286 3 2 569.858 1 -1 1
287 3 1 656.773 1 -1 1
288 3 2 617.246 1 -1 1
289 3 1 659.534 1 -1 1
290 3 2 610.337 1 -1 1
291 3 1 695.278 1 -1 1
292 3 2 584.192 1 -1 1
293 3 1 734.040 1 -1 1
294 3 2 598.853 1 -1 1
295 3 1 687.665 1 -1 1
296 3 2 554.774 1 -1 1
297 3 1 710.858 1 -1 1
298 3 2 605.694 1 -1 1
299 3 1 701.716 1 -1 1
300 3 2 627.516 1 -1 1

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a1.htm (5 of 13) [5/7/2002 3:16:26 PM]

1.4.2.10.1. Background and Data

301 3 1 382.133 1 1 -1
302 3 2 574.522 1 1 -1
303 3 1 719.744 1 1 -1
304 3 2 582.682 1 1 -1
305 3 1 756.820 1 1 -1
306 3 2 563.872 1 1 -1
307 3 1 690.978 1 1 -1
308 3 2 715.962 1 1 -1
309 3 1 670.864 1 1 -1
310 3 2 616.430 1 1 -1
311 3 1 670.308 1 1 -1
312 3 2 778.011 1 1 -1
313 3 1 660.062 1 1 -1
314 3 2 604.255 1 1 -1
315 3 1 790.382 1 1 -1
316 3 2 571.906 1 1 -1
317 3 1 714.750 1 1 -1
318 3 2 625.925 1 1 -1
319 3 1 716.959 1 1 -1
320 3 2 682.426 1 1 -1
321 3 1 603.363 1 1 -1
322 3 2 707.604 1 1 -1
323 3 1 713.796 1 1 -1
324 3 2 617.400 1 1 -1
325 3 1 444.963 1 1 -1
326 3 2 689.576 1 1 -1
327 3 1 723.276 1 1 -1
328 3 2 676.678 1 1 -1
329 3 1 745.527 1 1 -1
330 3 2 563.290 1 1 -1
361 4 1 778.333 -1 -1 1
362 4 2 581.879 -1 -1 1
363 4 1 723.349 -1 -1 1
364 4 2 447.701 -1 -1 1
365 4 1 708.229 -1 -1 1
366 4 2 557.772 -1 -1 1
367 4 1 681.667 -1 -1 1
368 4 2 593.537 -1 -1 1
369 4 1 566.085 -1 -1 1
370 4 2 632.585 -1 -1 1
371 4 1 687.448 -1 -1 1
372 4 2 671.350 -1 -1 1
373 4 1 597.500 -1 -1 1
374 4 2 569.530 -1 -1 1
375 4 1 637.410 -1 -1 1
376 4 2 581.667 -1 -1 1

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a1.htm (6 of 13) [5/7/2002 3:16:26 PM]

1.4.2.10.1. Background and Data

377 4 1 755.864 -1 -1 1
378 4 2 643.449 -1 -1 1
379 4 1 692.945 -1 -1 1
380 4 2 581.593 -1 -1 1
381 4 1 766.532 -1 -1 1
382 4 2 494.122 -1 -1 1
383 4 1 725.663 -1 -1 1
384 4 2 620.948 -1 -1 1
385 4 1 698.818 -1 -1 1
386 4 2 615.903 -1 -1 1
387 4 1 760.000 -1 -1 1
388 4 2 606.667 -1 -1 1
389 4 1 775.272 -1 -1 1
390 4 2 579.167 -1 -1 1
421 4 1 708.885 -1 1 -1
422 4 2 662.510 -1 1 -1
423 4 1 727.201 -1 1 -1
424 4 2 436.237 -1 1 -1
425 4 1 642.560 -1 1 -1
426 4 2 644.223 -1 1 -1
427 4 1 690.773 -1 1 -1
428 4 2 586.035 -1 1 -1
429 4 1 688.333 -1 1 -1
430 4 2 620.833 -1 1 -1
431 4 1 743.973 -1 1 -1
432 4 2 652.535 -1 1 -1
433 4 1 682.461 -1 1 -1
434 4 2 593.516 -1 1 -1
435 4 1 761.430 -1 1 -1
436 4 2 587.451 -1 1 -1
437 4 1 691.542 -1 1 -1
438 4 2 570.964 -1 1 -1
439 4 1 643.392 -1 1 -1
440 4 2 645.192 -1 1 -1
441 4 1 697.075 -1 1 -1
442 4 2 540.079 -1 1 -1
443 4 1 708.229 -1 1 -1
444 4 2 707.117 -1 1 -1
445 4 1 746.467 -1 1 -1
446 4 2 621.779 -1 1 -1
447 4 1 744.819 -1 1 -1
448 4 2 585.777 -1 1 -1
449 4 1 655.029 -1 1 -1
450 4 2 703.980 -1 1 -1
541 5 1 715.224 -1 -1 -1
542 5 2 698.237 -1 -1 -1

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a1.htm (7 of 13) [5/7/2002 3:16:26 PM]

1.4.2.10.1. Background and Data

543 5 1 614.417 -1 -1 -1
544 5 2 757.120 -1 -1 -1
545 5 1 761.363 -1 -1 -1
546 5 2 621.751 -1 -1 -1
547 5 1 716.106 -1 -1 -1
548 5 2 472.125 -1 -1 -1
549 5 1 659.502 -1 -1 -1
550 5 2 612.700 -1 -1 -1
551 5 1 730.781 -1 -1 -1
552 5 2 583.170 -1 -1 -1
553 5 1 546.928 -1 -1 -1
554 5 2 599.771 -1 -1 -1
555 5 1 734.203 -1 -1 -1
556 5 2 549.227 -1 -1 -1
557 5 1 682.051 -1 -1 -1
558 5 2 605.453 -1 -1 -1
559 5 1 701.341 -1 -1 -1
560 5 2 569.599 -1 -1 -1
561 5 1 759.729 -1 -1 -1
562 5 2 637.233 -1 -1 -1
563 5 1 689.942 -1 -1 -1
564 5 2 621.774 -1 -1 -1
565 5 1 769.424 -1 -1 -1
566 5 2 558.041 -1 -1 -1
567 5 1 715.286 -1 -1 -1
568 5 2 583.170 -1 -1 -1
569 5 1 776.197 -1 -1 -1
570 5 2 345.294 -1 -1 -1
571 5 1 547.099 1 -1 1
572 5 2 570.999 1 -1 1
573 5 1 619.942 1 -1 1
574 5 2 603.232 1 -1 1
575 5 1 696.046 1 -1 1
576 5 2 595.335 1 -1 1
577 5 1 573.109 1 -1 1
578 5 2 581.047 1 -1 1
579 5 1 638.794 1 -1 1
580 5 2 455.878 1 -1 1
581 5 1 708.193 1 -1 1
582 5 2 627.880 1 -1 1
583 5 1 502.825 1 -1 1
584 5 2 464.085 1 -1 1
585 5 1 632.633 1 -1 1
586 5 2 596.129 1 -1 1
587 5 1 683.382 1 -1 1
588 5 2 640.371 1 -1 1

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a1.htm (8 of 13) [5/7/2002 3:16:26 PM]

1.4.2.10.1. Background and Data

589 5 1 684.812 1 -1 1
590 5 2 621.471 1 -1 1
591 5 1 738.161 1 -1 1
592 5 2 612.727 1 -1 1
593 5 1 671.492 1 -1 1
594 5 2 606.460 1 -1 1
595 5 1 709.771 1 -1 1
596 5 2 571.760 1 -1 1
597 5 1 685.199 1 -1 1
598 5 2 599.304 1 -1 1
599 5 1 624.973 1 -1 1
600 5 2 579.459 1 -1 1
601 6 1 757.363 1 1 1
602 6 2 761.511 1 1 1
603 6 1 633.417 1 1 1
604 6 2 566.969 1 1 1
605 6 1 658.754 1 1 1
606 6 2 654.397 1 1 1
607 6 1 664.666 1 1 1
608 6 2 611.719 1 1 1
609 6 1 663.009 1 1 1
610 6 2 577.409 1 1 1
611 6 1 773.226 1 1 1
612 6 2 576.731 1 1 1
613 6 1 708.261 1 1 1
614 6 2 617.441 1 1 1
615 6 1 739.086 1 1 1
616 6 2 577.409 1 1 1
617 6 1 667.786 1 1 1
618 6 2 548.957 1 1 1
619 6 1 674.481 1 1 1
620 6 2 623.315 1 1 1
621 6 1 695.688 1 1 1
622 6 2 621.761 1 1 1
623 6 1 588.288 1 1 1
624 6 2 553.978 1 1 1
625 6 1 545.610 1 1 1
626 6 2 657.157 1 1 1
627 6 1 752.305 1 1 1
628 6 2 610.882 1 1 1
629 6 1 684.523 1 1 1
630 6 2 552.304 1 1 1
631 6 1 717.159 -1 1 -1
632 6 2 545.303 -1 1 -1
633 6 1 721.343 -1 1 -1
634 6 2 651.934 -1 1 -1

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a1.htm (9 of 13) [5/7/2002 3:16:26 PM]

1.4.2.10.1. Background and Data

635 6 1 750.623 -1 1 -1
636 6 2 635.240 -1 1 -1
637 6 1 776.488 -1 1 -1
638 6 2 641.083 -1 1 -1
639 6 1 750.623 -1 1 -1
640 6 2 645.321 -1 1 -1
641 6 1 600.840 -1 1 -1
642 6 2 566.127 -1 1 -1
643 6 1 686.196 -1 1 -1
644 6 2 647.844 -1 1 -1
645 6 1 687.870 -1 1 -1
646 6 2 554.815 -1 1 -1
647 6 1 725.527 -1 1 -1
648 6 2 620.087 -1 1 -1
649 6 1 658.796 -1 1 -1
650 6 2 711.301 -1 1 -1
651 6 1 690.380 -1 1 -1
652 6 2 644.355 -1 1 -1
653 6 1 737.144 -1 1 -1
654 6 2 713.812 -1 1 -1
655 6 1 663.851 -1 1 -1
656 6 2 696.707 -1 1 -1
657 6 1 766.630 -1 1 -1
658 6 2 589.453 -1 1 -1
659 6 1 625.922 -1 1 -1
660 6 2 634.468 -1 1 -1
721 7 1 694.430 1 1 -1
722 7 2 599.751 1 1 -1
723 7 1 730.217 1 1 -1
724 7 2 624.542 1 1 -1
725 7 1 700.770 1 1 -1
726 7 2 723.505 1 1 -1
727 7 1 722.242 1 1 -1
728 7 2 674.717 1 1 -1
729 7 1 763.828 1 1 -1
730 7 2 608.539 1 1 -1
731 7 1 695.668 1 1 -1
732 7 2 612.135 1 1 -1
733 7 1 688.887 1 1 -1
734 7 2 591.935 1 1 -1
735 7 1 531.021 1 1 -1
736 7 2 676.656 1 1 -1
737 7 1 698.915 1 1 -1
738 7 2 647.323 1 1 -1
739 7 1 735.905 1 1 -1
740 7 2 811.970 1 1 -1

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a1.htm (10 of 13) [5/7/2002 3:16:26 PM]

1.4.2.10.1. Background and Data

741 7 1 732.039 1 1 -1
742 7 2 603.883 1 1 -1
743 7 1 751.832 1 1 -1
744 7 2 608.643 1 1 -1
745 7 1 618.663 1 1 -1
746 7 2 630.778 1 1 -1
747 7 1 744.845 1 1 -1
748 7 2 623.063 1 1 -1
749 7 1 690.826 1 1 -1
750 7 2 472.463 1 1 -1
811 7 1 666.893 -1 1 1
812 7 2 645.932 -1 1 1
813 7 1 759.860 -1 1 1
814 7 2 577.176 -1 1 1
815 7 1 683.752 -1 1 1
816 7 2 567.530 -1 1 1
817 7 1 729.591 -1 1 1
818 7 2 821.654 -1 1 1
819 7 1 730.706 -1 1 1
820 7 2 684.490 -1 1 1
821 7 1 763.124 -1 1 1
822 7 2 600.427 -1 1 1
823 7 1 724.193 -1 1 1
824 7 2 686.023 -1 1 1
825 7 1 630.352 -1 1 1
826 7 2 628.109 -1 1 1
827 7 1 750.338 -1 1 1
828 7 2 605.214 -1 1 1
829 7 1 752.417 -1 1 1
830 7 2 640.260 -1 1 1
831 7 1 707.899 -1 1 1
832 7 2 700.767 -1 1 1
833 7 1 715.582 -1 1 1
834 7 2 665.924 -1 1 1
835 7 1 728.746 -1 1 1
836 7 2 555.926 -1 1 1
837 7 1 591.193 -1 1 1
838 7 2 543.299 -1 1 1
839 7 1 592.252 -1 1 1
840 7 2 511.030 -1 1 1
901 8 1 740.833 -1 -1 1
902 8 2 583.994 -1 -1 1
903 8 1 786.367 -1 -1 1
904 8 2 611.048 -1 -1 1
905 8 1 712.386 -1 -1 1
906 8 2 623.338 -1 -1 1

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a1.htm (11 of 13) [5/7/2002 3:16:26 PM]

1.4.2.10.1. Background and Data

907 8 1 738.333 -1 -1 1
908 8 2 679.585 -1 -1 1
909 8 1 741.480 -1 -1 1
910 8 2 665.004 -1 -1 1
911 8 1 729.167 -1 -1 1
912 8 2 655.860 -1 -1 1
913 8 1 795.833 -1 -1 1
914 8 2 715.711 -1 -1 1
915 8 1 723.502 -1 -1 1
916 8 2 611.999 -1 -1 1
917 8 1 718.333 -1 -1 1
918 8 2 577.722 -1 -1 1
919 8 1 768.080 -1 -1 1
920 8 2 615.129 -1 -1 1
921 8 1 747.500 -1 -1 1
922 8 2 540.316 -1 -1 1
923 8 1 775.000 -1 -1 1
924 8 2 711.667 -1 -1 1
925 8 1 760.599 -1 -1 1
926 8 2 639.167 -1 -1 1
927 8 1 758.333 -1 -1 1
928 8 2 549.491 -1 -1 1
929 8 1 682.500 -1 -1 1
930 8 2 684.167 -1 -1 1
931 8 1 658.116 1 -1 -1
932 8 2 672.153 1 -1 -1
933 8 1 738.213 1 -1 -1
934 8 2 594.534 1 -1 -1
935 8 1 681.236 1 -1 -1
936 8 2 627.650 1 -1 -1
937 8 1 704.904 1 -1 -1
938 8 2 551.870 1 -1 -1
939 8 1 693.623 1 -1 -1
940 8 2 594.534 1 -1 -1
941 8 1 624.993 1 -1 -1
942 8 2 602.660 1 -1 -1
943 8 1 700.228 1 -1 -1
944 8 2 585.450 1 -1 -1
945 8 1 611.874 1 -1 -1
946 8 2 555.724 1 -1 -1
947 8 1 579.167 1 -1 -1
948 8 2 574.934 1 -1 -1
949 8 1 720.872 1 -1 -1
950 8 2 584.625 1 -1 -1
951 8 1 690.320 1 -1 -1
952 8 2 555.724 1 -1 -1

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a1.htm (12 of 13) [5/7/2002 3:16:26 PM]

1.4.2.10.1. Background and Data

953 8 1 677.933 1 -1 -1
954 8 2 611.874 1 -1 -1
955 8 1 674.600 1 -1 -1
956 8 2 698.254 1 -1 -1
957 8 1 611.999 1 -1 -1
958 8 2 748.130 1 -1 -1
959 8 1 530.680 1 -1 -1
960 8 2 689.942 1 -1 -1

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a1.htm (13 of 13) [5/7/2002 3:16:26 PM]

1.4.2.10.2. Analysis of the Response Variable

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.10. Ceramic Strength

1.4.2.10.2. Analysis of the Response Variable

Numerical As a first step in the analysis, a table of summary statistics is computed for the response
Summary variable. The following table, generated by Dataplot, shows a typical set of statistics.

SUMMARY

NUMBER OF OBSERVATIONS = 480

***********************************************************************
* LOCATION MEASURES * DISPERSION MEASURES *
***********************************************************************
* MIDRANGE = 0.5834740E+03 * RANGE = 0.4763600E+03 *
* MEAN = 0.6500773E+03 * STAND. DEV. = 0.7463826E+02 *
* MIDMEAN = 0.6426155E+03 * AV. AB. DEV. = 0.6184948E+02 *
* MEDIAN = 0.6466275E+03 * MINIMUM = 0.3452940E+03 *
* = * LOWER QUART. = 0.5960515E+03 *
* = * LOWER HINGE = 0.5959740E+03 *
* = * UPPER HINGE = 0.7084220E+03 *
* = * UPPER QUART. = 0.7083415E+03 *
* = * MAXIMUM = 0.8216540E+03 *
***********************************************************************
* RANDOMNESS MEASURES * DISTRIBUTIONAL MEASURES *
***********************************************************************
* AUTOCO COEF = -0.2290508E+00 * ST. 3RD MOM. = -0.3682922E+00 *
* = 0.0000000E+00 * ST. 4TH MOM. = 0.3220554E+01 *
* = 0.0000000E+00 * ST. WILK-SHA = 0.3877698E+01 *
* = * UNIFORM PPCC = 0.9756916E+00 *
* = * NORMAL PPCC = 0.9906310E+00 *
* = * TUK -.5 PPCC = 0.8357126E+00 *
* = * CAUCHY PPCC = 0.5063868E+00 *
***********************************************************************

From the above output, the mean strength is 650.08 and the standard deviation of the
strength is 74.64.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a2.htm (1 of 2) [5/7/2002 3:16:27 PM]

1.4.2.10.2. Analysis of the Response Variable

4-Plot The next step is generate a 4-plot of the response variable.

This 4-plot shows:

1. The run sequence plot (upper left corner) shows that the location and scale are
relatively constant. It also shows a few outliers on the low side. Most of the points
are in the range 500 to 750. However, there are about half a dozen points in the 300
to 450 range that may require special attention.
A run sequence plot is useful for designed experiments in that it can reveal time
effects. Time is normally a nuisance factor. That is, the time order on which runs are
made should not have a significant effect on the response. If a time effect does
appear to exist, this means that there is a potential bias in the experiment that needs
to be investigated and resolved.
2. The lag plot (the upper right corner) does not show any significant structure. This is
another tool for detecting any potential time effect.
3. The histogram (the lower left corner) shows the response appears to be reasonably
symmetric, but with a bimodal distribution.
4. The normal probability plot (the lower right corner) shows some curvature
indicating that distributions other than the normal may provide a better fit.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a2.htm (2 of 2) [5/7/2002 3:16:27 PM]

1.4.2.10.3. Analysis of the Batch Effect

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.10. Ceramic Strength

1.4.2.10.3. Analysis of the Batch Effect

Batch is a The two nuisance factors in this experiment are the batch number and the lab. There are
Nuisance Factor 2 batches and 8 labs. Ideally, these factors will have minimal effect on the response
variable.
We will investigate the batch factor first.

Bihistogram

This bihistogram shows the following.

1. There does appear to be a batch effect.
2. The batch 1 responses are centered at 700 while the batch 2 responses are
centered at 625. That is, the batch effect is approximately 75 units.
3. The variability is comparable for the 2 batches.
4. Batch 1 has some skewness in the lower tail. Batch 2 has some skewness in the
center of the distribution, but not as much in the tails compared to batch 1.
5. Both batches have a few low-lying points.
Although we could stop with the bihistogram, we will show a few other commonly used
two-sample graphical techniques for comparison.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a3.htm (1 of 5) [5/7/2002 3:16:27 PM]

1.4.2.10.3. Analysis of the Batch Effect

Quantile-Quantile
Plot

This q-q plot shows the following.

1. Except for a few points in the right tail, the batch 1 values have higher quantiles
than the batch 2 values. This implies that batch 1 has a greater location value than
batch 2.
2. The q-q plot is not linear. This implies that the difference between the batches is
not explained simply by a shift in location. That is, the variation and/or skewness
varies as well. From the bihistogram, it appears that the skewness in batch 2 is the
most likely explanation for the non-linearity in the q-q plot.

Box Plot

This box plot shows the following.

1. The median for batch 1 is approximately 700 while the median for batch 2 is
approximately 600.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a3.htm (2 of 5) [5/7/2002 3:16:27 PM]

1.4.2.10.3. Analysis of the Batch Effect

2. The spread is reasonably similar for both batches, maybe slightly larger for batch
1.
3. Both batches have a number of outliers on the low side. Batch 2 also has a few
outliers on the high side. Box plots are a particularly effective method for
identifying the presence of outliers.

Block Plots A block plot is generated for each of the eight labs, with "1" and "2" denoting the batch
numbers. In the first plot, we do not include any of the primary factors. The next 3
block plots include one of the primary factors. Note that each of the 3 primary factors
(table speed = X1, down feed rate = X2, wheel grit size = X3) has 2 levels. With 8 labs
and 2 levels for the primary factor, we would expect 16 separate blocks on these plots.
The fact that some of these blocks are missing indicates that some of the combinations
of lab and primary factor are empty.

These block plots show the following.

1. The mean for batch 1 is greater than the mean for batch 2 in all of the cases
above. This is strong evidence that the batch effect is real and consistent across
labs and primary factors.

Quantitative We can confirm some of the conclusions drawn from the above graphics by using
Techniques quantitative techniques. The two sample t-test can be used to test whether or not the
means from the two batches are equal and the F-test can be used to test whether or not
the standard deviations from the two batches are equal.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a3.htm (3 of 5) [5/7/2002 3:16:27 PM]

1.4.2.10.3. Analysis of the Batch Effect

Two Sample The following is the Dataplot output from the two sample t-test.
T-Test
T-TEST
(2-SAMPLE)
NULL HYPOTHESIS UNDER TEST--POPULATION MEANS MU1 = MU2

SAMPLE 1:
NUMBER OF OBSERVATIONS = 240
MEAN = 688.9987
STANDARD DEVIATION = 65.54909
STANDARD DEVIATION OF MEAN = 4.231175

SAMPLE 2:
NUMBER OF OBSERVATIONS = 240
MEAN = 611.1559
STANDARD DEVIATION = 61.85425
STANDARD DEVIATION OF MEAN = 3.992675

IF ASSUME SIGMA1 = SIGMA2:

POOLED STANDARD DEVIATION = 63.72845
DIFFERENCE (DELTA) IN MEANS = 77.84271
STANDARD DEVIATION OF DELTA = 5.817585
T-TEST STATISTIC VALUE = 13.38059
DEGREES OF FREEDOM = 478.0000
T-TEST STATISTIC CDF VALUE = 1.000000

IF NOT ASSUME SIGMA1 = SIGMA2:

STANDARD DEVIATION SAMPLE 1 = 65.54909
STANDARD DEVIATION SAMPLE 2 = 61.85425
BARTLETT CDF VALUE = 0.629618
DIFFERENCE (DELTA) IN MEANS = 77.84271
STANDARD DEVIATION OF DELTA = 5.817585
T-TEST STATISTIC VALUE = 13.38059
EQUIVALENT DEG. OF FREEDOM = 476.3999
T-TEST STATISTIC CDF VALUE = 1.000000

ALTERNATIVE- ALTERNATIVE-
ALTERNATIVE- HYPOTHESIS HYPOTHESIS
HYPOTHESIS ACCEPTANCE INTERVAL CONCLUSION
MU1 <> MU2 (0,0.025) (0.975,1) ACCEPT
MU1 < MU2 (0,0.05) REJECT
MU1 > MU2 (0.95,1) ACCEPT
The t-test indicates that the mean for batch 1 is larger than the mean for batch 2 (at the
5% confidence level).

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a3.htm (4 of 5) [5/7/2002 3:16:27 PM]

1.4.2.10.3. Analysis of the Batch Effect

F-Test The following is the Dataplot output from the F-test.

F-TEST
NULL HYPOTHESIS UNDER TEST--SIGMA1 = SIGMA2
ALTERNATIVE HYPOTHESIS UNDER TEST--SIGMA1 NOT EQUAL SIGMA2

SAMPLE 1:
NUMBER OF OBSERVATIONS = 240
MEAN = 688.9987
STANDARD DEVIATION = 65.54909

SAMPLE 2:
NUMBER OF OBSERVATIONS = 240
MEAN = 611.1559
STANDARD DEVIATION = 61.85425

TEST:
STANDARD DEV. (NUMERATOR) = 65.54909
STANDARD DEV. (DENOMINATOR) = 61.85425
F-TEST STATISTIC VALUE = 1.123037
DEG. OF FREEDOM (NUMER.) = 239.0000
DEG. OF FREEDOM (DENOM.) = 239.0000
F-TEST STATISTIC CDF VALUE = 0.814808

NULL NULL HYPOTHESIS NULL HYPOTHESIS

HYPOTHESIS ACCEPTANCE INTERVAL CONCLUSION
SIGMA1 = SIGMA2 (0.000,0.950) ACCEPT
The F-test indicates that the standard deviations for the two batches are not significantly
different at the 5% confidence level.

Conclusions We can draw the following conclusions from the above analysis.
1. There is in fact a significant batch effect. This batch effect is consistent across
labs and primary factors.
2. The magnitude of the difference is on the order of 75 to 100 (with batch 2 being
smaller than batch 1). The standard deviations do not appear to be significantly
different.
3. There is some skewness in the batches.
This batch effect was completely unexpected by the scientific investigators in this
study.
Note that although the quantitative techniques support the conclusions of unequal
means and equal standard deviations, they do not show the more subtle features of the
data such as the presence of outliers and the skewness of the batch 2 data.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a3.htm (5 of 5) [5/7/2002 3:16:27 PM]

1.4.2.10.4. Analysis of the Lab Effect

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.10. Ceramic Strength

1.4.2.10.4. Analysis of the Lab Effect

Box Plot The next matter is to determine if there is a lab effect. The first step is to
generate a box plot for the ceramic strength based on the lab.

This box plot shows the following.

1. There is minor variation in the medians for the 8 labs.
2. The scales are relatively constant for the labs.
3. Two of the labs (3 and 5) have outliers on the low side.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a4.htm (1 of 3) [5/7/2002 3:16:27 PM]

1.4.2.10.4. Analysis of the Lab Effect

Box Plot for Given that the previous section showed a distinct batch effect, the next
Batch 1 step is to generate the box plots for the two batches separately.

This box plot shows the following.

1. Each of the labs has a median in the 650 to 700 range.
2. The variability is relatively constant across the labs.
3. Each of the labs has at least one outlier on the low side.

Box Plot for

Batch 2

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a4.htm (2 of 3) [5/7/2002 3:16:27 PM]

1.4.2.10.4. Analysis of the Lab Effect

This box plot shows the following.

1. The medians are in the range 550 to 600.
2. There is a bit more variability, across the labs, for batch2
compared to batch 1.
3. Six of the eight labs show outliers on the high side. Three of the
labs show outliers on the low side.

Conclusions We can draw the following conclusions about a possible lab effect from
the above box plots.
1. The batch effect (of approximately 75 to 100 units) on location
dominates any lab effects.
2. It is reasonable to treat the labs as homogeneous.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a4.htm (3 of 3) [5/7/2002 3:16:27 PM]

1.4.2.10.5. Analysis of Primary Factors

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.10. Ceramic Strength

1.4.2.10.5. Analysis of Primary Factors

Main effects The first step in analyzing the primary factors is to determine which
factors are the most significant. The dex scatter plot, dex mean plot, and
the dex standard deviation plots will be the primary tools, with "dex"
being short for "design of experiments".
Since the previous pages showed a significant batch effect but a minimal
lab effect, we will generate separate plots for batch 1 and batch 2.
However, the labs will be treated as equivalent.

Dex Scatter
Plot for
Batch 1

This dex scatter plot shows the following for batch 1.

1. Most of the points are between 500 and 800.
2. There are about a dozen or so points between 300 and 500.
3. Except for the outliers on the low side (i.e., the points between
300 and 500), the distribution of the points is comparable for the

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a5.htm (1 of 7) [5/7/2002 3:16:28 PM]

1.4.2.10.5. Analysis of Primary Factors

3 primary factors in terms of location and spread.

Dex Mean
Plot for
Batch 1

This dex mean plot shows the following for batch 1.

1. The table speed factor (X1) is the most significant factor with an
effect, the difference between the two points, of approximately 35
units.
2. The wheel grit factor (X3) is the next most significant factor with
an effect of approximately 10 units.
3. The feed rate factor (X2) has minimal effect.

Dex SD Plot
for Batch 1

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a5.htm (2 of 7) [5/7/2002 3:16:28 PM]

1.4.2.10.5. Analysis of Primary Factors

This dex standard deviation plot shows the following for batch 1.
1. The table speed factor (X1) has a significant difference in
variability between the levels of the factor. The difference is
approximately 20 units.
2. The wheel grit factor (X3) and the feed rate factor (X2) have
minimal differences in variability.

Dex Scatter
Plot for
Batch 2

This dex scatter plot shows the following for batch 2.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a5.htm (3 of 7) [5/7/2002 3:16:28 PM]

1.4.2.10.5. Analysis of Primary Factors

1. Most of the points are between 450 and 750.

2. There are a few outliers on both the low side and the high side.
3. Except for the outliers (i.e., the points less than 450 or greater
than 750), the distribution of the points is comparable for the 3
primary factors in terms of location and spread.

Dex Mean
Plot for
Batch 2

This dex mean plot shows the following for batch 2.

1. The feed rate (X2) and wheel grit (X3) factors have an
approximately equal effect of about 15 or 20 units.
2. The table speed factor (X1) has a minimal effect.

Dex SD Plot
for Batch 2

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a5.htm (4 of 7) [5/7/2002 3:16:28 PM]

1.4.2.10.5. Analysis of Primary Factors

This dex standard deviation plot shows the following for batch 2.
1. The difference in the standard deviations is roughly comparable
for the three factors (slightly less for the feed rate factor).

Interaction The above plots graphically show the main effects. An additonal
Effects concern is whether or not there any significant interaction effects.
Main effects and 2-term interaction effects are discussed in the chapter
on Process Improvement.

In the following dex interaction plots, the labels on the plot give the
variables and the estimated effect. For example, factor 1 is TABLE
SPEED and it has an estimated effect of 30.77 (it is actually -30.77 if
the direction is taken into account).

DEX
Interaction
Plot for
Batch 1

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a5.htm (5 of 7) [5/7/2002 3:16:28 PM]

1.4.2.10.5. Analysis of Primary Factors

The ranked list of factors for batch 1 is:

1. Table speed (X1) with an estimated effect of -30.77.
2. The interaction of table speed (X1) and wheel grit (X3) with an
estimated effect of -20.25.
3. The interaction of table speed (X1) and feed rate (X2) with an
estimated effect of 9.7.
4. Wheel grit (X3) with an estimated effect of -7.18.
5. Down feed (X2) and the down feed interaction with wheel grit
(X3) are essentially zero.

DEX
Interaction
Plot for
Batch 2

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a5.htm (6 of 7) [5/7/2002 3:16:28 PM]

1.4.2.10.5. Analysis of Primary Factors

The ranked list of factors for batch 2 is:

1. Down feed (X2) with an estimated effect of 18.22.
2. The interaction of table speed (X1) and wheel grit (X3) with an
estimated effect of -16.71.
3. Wheel grit (X3) with an estimated effect of -14.71
4. Remaining main effect and 2-factor interaction effects are
essentially zero.

Conclusions From the above plots, we can draw the following overall conclusions.
1. The batch effect (of approximately 75 units) is the dominant
primary factor.
2. The most important factors differ from batch to batch. See the
above text for the ranked list of factors with the estimated effects.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a5.htm (7 of 7) [5/7/2002 3:16:28 PM]

1.4.2.10.6. Work This Example Yourself

1. Exploratory Data Analysis

1.4. EDA Case Studies
1.4.2. Case Studies
1.4.2.10. Ceramic Strength

1.4.2.10.6. Work This Example Yourself

View This page allows you to use Dataplot to repeat the analysis outlined in
Dataplot the case study description on the previous page. It is required that you
Macro for have already downloaded and installed Dataplot and configured your
this Case browser. to run Dataplot. Output from each analysis step below will be
Study displayed in one or more of the Dataplot windows. The four main
windows are the Output window, the Graphics window, the Command
History window, and the data sheet window. Across the top of the main
windows there are menus for executing Dataplot commands. Across the
bottom is a command entry window where commands can be typed in.

Data Analysis Steps Results and Conclusions

Click on the links below to start Dataplot and run this case
study yourself. Each step may use results from previous The links in this column will connect you with more
steps, so please be patient. Wait until the software verifies detailed information about each analysis step from the case
that the current step is complete before clicking on the next study description.
step.

1. Invoke Dataplot and read data.

1. Read in the data. 1. You have read 1 column of numbers

into Dataplot, variable Y.

2. Plot of the response variable

1. Numerical summary of Y. 1. The summary shows the mean strength

is 650.08 and the standard deviation
of the strength is 74.64.

2. 4-plot of Y. 2. The 4-plot shows no drift in

the location and scale and a
bimodal distribution.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a6.htm (1 of 3) [5/7/2002 3:16:28 PM]

1.4.2.10.6. Work This Example Yourself

3. Determine if there is a batch effect.

1. Generate a bihistogram based on 1. The bihistogram shows a distinct

the 2 batches. batch effect of approximately
75 units.

2. Generate a q-q plot. 2. The q-q plot shows that batch 1

and batch 2 do not come from a
common distribution.

3. Generate a box plot. 3. The box plot shows that there is

a batch effect of approximately
75 to 100 units and there are
some outliers.
4. Generate block plots.
4. The block plot shows that the batch
effect is consistent across labs
and levels of the primary factor.

5. Perform a 2-sample t-test for

equal means. 5. The t-test confirms the batch
effect with respect to the means.

6. Perform an F-test for equal

standard deviations. 6. The F-test does not indicate any
significant batch effect with
respect to the standard deviations.

4. Determine if there is a lab effect.

1. Generate a box plot for the labs 1. The box plot does not show a
with the 2 batches combined. significant lab effect.

2. Generate a box plot for the labs 2. The box plot does not show a
for batch 1 only. significant lab effect for batch 1.

3. Generate a box plot for the labs 3. The box plot does not show a
for batch 2 only. significant lab effect for batch 2.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a6.htm (2 of 3) [5/7/2002 3:16:28 PM]

1.4.2.10.6. Work This Example Yourself

5. Analysis of primary factors.

1. Generate a dex scatter plot for 1. The dex scatter plot shows the
batch 1. range of the points and the
presence of outliers.

2. Generate a dex mean plot for 2. The dex mean plot shows that
batch 1. table speed is the most
significant factor for batch 1.

3. Generate a dex sd plot for 3. The dex sd plot shows that

batch 1. table speed has the most
variability for batch 1.

4. Generate a dex scatter plot for 4. The dex scatter plot shows
batch 2. the range of the points and
the presence of outliers.

5. Generate a dex mean plot for

5. The dex mean plot shows that
batch 2.
feed rate and wheel grit are
the most significant factors
for batch 2.
6. Generate a dex sd plot for
batch 2. 6. The dex sd plot shows that
the variability is comparable
for all 3 factors for batch 2.
7. Generate a dex interaction
effects matrix plot for 7. The dex interaction effects
batch 1. matrix plot provides a ranked
list of factors with the
estimated effects.
8. Generate a dex interaction
effects matrix plot for 8. The dex interaction effects
batch 2. matrix plot provides a ranked
list of factors with the
estimated effects.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda42a6.htm (3 of 3) [5/7/2002 3:16:28 PM]

1.4.3. References For Chapter 1: Exploratory Data Analysis

1. Exploratory Data Analysis

1.4. EDA Case Studies

1.4.3. References For Chapter 1:

Exploratory Data Analysis
Anscombe, Francis (1973), Graphs in Statistical Analysis, The American Statistician,
pp. 195-199.

Anscombe, Francis and Tukey, J. W. (1963), The Examination and Analysis of

Residuals, Technometrics, pp. 141-160.

Bloomfield, Peter (1976), Fourier Analysis of Time Series, John Wiley and Sons.

Box, G. E. P. and Cox, D. R. (1964), An Analysis of Transformations, Journal of the

Royal Statistical Society, 211-243, discussion 244-252.

Box, G. E. P., Hunter, W. G., and Hunter, J. S. (1978), Statistics for Experimenters: An
Introduction to Design, Data Analysis, and Model Building, John Wiley and Sons.

Box, G. E. P., and Jenkins, G. (1976), Time Series Analysis: Forecasting and Control,
Holden-Day.

Bradley, (1968). Distribution-Free Statistical Tests, Chapter 12.

Brown, M. B. and Forsythe, A. B. (1974), Journal of the American Statistical

Association, 69, 364-367.

Chakravarti, Laha, and Roy, (1967). Handbook of Methods of Applied Statistics, Volume
I, John Wiley and Sons, pp. 392-394.

Chambers, John, William Cleveland, Beat Kleiner, and Paul Tukey, (1983), Graphical
Methods for Data Analysis, Wadsworth.

Cleveland, William (1985), Elements of Graphing Data, Wadsworth.

Cleveland, William and Marylyn McGill, Editors (1988), Dynamic Graphics for
Statistics, Wadsworth.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda43.htm (1 of 4) [5/7/2002 3:16:29 PM]

1.4.3. References For Chapter 1: Exploratory Data Analysis

Cleveland, William (1993), Visualizing Data, Hobart Press.

Draper and Smith, (1981). Applied Regression Analysis, 2nd ed., John Wiley and Sons.

du Toit, Steyn, and Stumpf (1986), Graphical Exploratory Data Analysis,

Springer-Verlag.

Evans, Hastings, and Peacock (2000), Statistical Distributions, 3rd. Ed., John Wiley and
Sons.

Everitt, Brian (1978), Multivariate Techniques for Multivariate Data, North-Holland.

Efron and Gong (February 1983), A Leisurely Look at the Bootstrap, the Jackknife, and
Cross Validation, The American Statistician.

Filliben, J. J. (February 1975), The Probability Plot Correlation Coefficient Test for
Normality , Technometrics, pp. 111-117.

Gill, Lisa (April 1997), Summary Analysis: High Performance Ceramics Experiment to
Characterize the Effect of Grinding Parameters on Sintered Reaction Bonded Silicon
Nitride, Reaction Bonded Silicon Nitride, and Sintered Silicon Nitride , presented at the
NIST - Ceramic Machining Consortium, 10th Program Review Meeting, April 10, 1997.

Granger and Hatanaka (1964). Spectral Analysis of Economic Time Series, Princeton
University Press.

Grubbs, Frank (February 1969), Procedures for Detecting Outlying Observations in

Samples, Technometrics, Vol. 11, No. 1, pp. 1-21.

Harris, Robert L. (1996), Information Graphics, Management Graphics.

Jenkins and Watts, (1968), Spectral Analysis and Its Applications, Holden-Day.

Johnson, Kotz, and Balakrishnan, (1994), Continuous Univariate Distributions, Volumes

I and II, 2nd. Ed., John Wiley and Sons.

Johnson, Kotz, and Kemp, (1992), Univariate Discrete Distributions, 2nd. Ed., John
Wiley and Sons.

Kuo, Way and Pierson, Marcia Martens, Eds. (1993), Quality Through Engineering
Design", specifically, the article Filliben, Cetinkunt, Yu, and Dommenz (1993),
Exploratory Data Analysis Techniques as Applied to a High-Precision Turning Machine,
Elsevier, New York, pp. 199-223.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda43.htm (2 of 4) [5/7/2002 3:16:29 PM]

1.4.3. References For Chapter 1: Exploratory Data Analysis

Levene, H. (1960). In Contributions to Probability and Statistics: Essays in Honor of

Harold Hotelling, I. Olkin et al. eds., Stanford University Press, pp. 278-292.

McNeil, Donald (1977), Interactive Data Analysis, John Wiley and Sons.

Mosteller, Frederick and Tukey, John (1977), Data Analysis and Regression,
Addison-Wesley.

Nelson, Wayne (1982), Applied Life Data Analysis, Addison-Wesley.

Neter, Wasserman, and Kunter (1990). Applied Linear Statistical Models, 3rd ed., Irwin.

Nelson, Wayne and Doganaksoy, Necip (1992), A Computer Program POWNOR for
Fitting the Power-Normal and -Lognormal Models to Life or Strength Data from
Specimens of Various Sizes, NISTIR 4760, U.S. Department of Commerce, National
Institute of Standards and Technology.

The RAND Corporation (1955), A Million Random Digits with 100,000 Normal
Deviates, Free Press.

Ryan, Thomas (1997). Modern Regression Methods, John Wiley.

Scott, David (1992), Multivariate Density Estimation: Theory, Practice, and

Visualization , John Wiley and Sons.

Snedecor, George W. and Cochran, William G. (1989), Statistical Methods, Eighth

Edition, Iowa State University Press.

Stefansky, W. (1972), Rejecting Outliers in Factorial Designs, Technometrics, Vol. 14,

pp. 469-479.

Stephens, M. A. (1974). EDF Statistics for Goodness of Fit and Some Comparisons,
Journal of the American Statistical Association, Vol. 69, pp. 730-737.

Stephens, M. A. (1976). Asymptotic Results for Goodness-of-Fit Statistics with Unknown

Parameters, Annals of Statistics, Vol. 4, pp. 357-369.

Stephens, M. A. (1977). Goodness of Fit for the Extreme Value Distribution,

Biometrika, Vol. 64, pp. 583-588.

Stephens, M. A. (1977). Goodness of Fit with Special Reference to Tests for

Exponentiality , Technical Report No. 262, Department of Statistics, Stanford
University, Stanford, CA.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda43.htm (3 of 4) [5/7/2002 3:16:29 PM]

1.4.3. References For Chapter 1: Exploratory Data Analysis

Stephens, M. A. (1979). Tests of Fit for the Logistic Distribution Based on the Empirical
Distribution Function, Biometrika, Vol. 66, pp. 591-595.

Tukey, John (1977), Exploratory Data Analysis, Addison-Wesley.

Tufte, Edward (1983), The Visual Display of Quantitative Information, Graphics Press.

Velleman, Paul and Hoaglin, David (1981), The ABC's of EDA: Applications, Basics,
and Computing of Exploratory Data Analysis, Duxbury.

Wainer, Howard (1981), Visual Revelations, Copernicus.

http://www.itl.nist.gov/div898/handbook/eda/section4/eda43.htm (4 of 4) [5/7/2002 3:16:29 PM]