Statistical Tools For Process Validation

ISPE Process Validation Conference
12 – 14 September 2017
Bethesda, MD
PROCESS VALIDATION –
STATISTICAL TOOL
OVERVIEW
Maneesha Altekar, Katherine Giacoletti

ISPE Statistics Validation Conference
September 2017
Statistical Methods Used in Process Validation

Stages 2 & 3 (PPQ & CPV)
Stage 2 (PPQ) Stage 3 (CPV)
• Statistical Intervals • Acceptance Sampling
• Confidence Intervals • Attributes & Variable
• Tolerance Intervals • Operating Characteristics
• (Prediction Intervals – not often • Control Charts
used)
• Types of charts
• Visualization Tools
• Setting Limits
•
Scatterplots, run charts, box plots,
• Run rules & assumptions
histograms, variability charts
• Capability
• Sampling Strategies
• Cpk & Ppk
• Types of sampling
• Discrete data?
• Sample size
• Assumptions
• Capability (preliminary)
• Variance Components Analysis
Beyond the scope of
• ASTM
today’s session
Connecting Pharmaceutical Knowledge ispe.org
1
Bethesda, MD
Basic Statistical Concepts – Not Covered

This “refresher” assumes an understanding of some
basic statistical terms & concepts:
• Sample vs. Population
• Summary statistics – mean, standard deviation, range, median,
etc.
• Scatter plots, box plots, histograms
… and at least a high-level familiarity with statistical
methods such as:
• Statistical intervals
• Acceptance sampling
• Control charts
Connecting Pharmaceutical Knowledge ispe.org 3
How to get in trouble with statistics
This Photo by Unknown Author is licensed under CC BY-SA
… or better yet, let’s find out

how to avoid trouble!
2
Bethesda, MD
The critical assumptions for many

statistics that often are overlooked
• Randomness & Independence – the most critical
assumptions for many statistical tools
• Randomness means a lack of a pattern
• Technically, it means that the data arise from a (single, common)
probability distribution (e.g. Normal, binomial)
• Independence means the value of any given data point does

not depend on the value of any others
• Lack of randomness/independence makes many
statistical calculations invalid
• … meaning they can lead to misleading conclusions & the
wrong decisions

Examples of non-randomness: Location
Effect
• Location effect – consistent pattern of
lower values at end of a batch. The
repeated pattern indicates that this is likely
non-random
• Calculations of the mean and SD, control
limits, statistical intervals, etc. will all be
misleading 105.0
1 2 3 4
102.5
Individual Value
… and how to detect 100.0
97.5
Graph your data!! 95.0
1
If data are collected from multiple 1 5 9 13 17 21

1
25 29 33
Observation
batches and/or multiple locations, plot 1 2 3 4
them in time order to look for non-

random patterns
3
Bethesda, MD

Examples of non-randomness: Two Populations
• Two sides of press resulting in different tablet
weights
• Therefore different API content
• Other related examples: multiple filling lines, different filling
nozzles, etc.
• Combining them to calculate any statistics will be
invalid and misleading

statistics that often are overlooked Time Series Plot of Sorted CU (%LC)
106 Sorted
Batch
Time Series Plot of Sorted weight 1
104 2
1500
Two populations: how to detect 1400

Sorted
Batch
1021
2
3
Sorted CU (%LC)
Graph your data!! 1300
1200
100
98
• Hint for 2 populations: Run/control chart

Sorted weight
1100
96
1000
will show alternating high/low values 900

94
• To confirm: side-by-side individual 800
700
92
1 6 12 18 4
Sorted Loc
10 16 2 8 14 20
value/box plots, run/control charts by 600

1 6 12 18 4 10 16 2 8 14 20 Individual Value Plot of weight
Sorted Loc 1300
press side… even just a histogram 1200
(sometimes) – but again, if data are 1100

weight
collected with ordering or in groups, you 1000
must look for signs of non-randomness 900
before doing any statistics Histogram of weight

800
Loc 12 34 567 891011121314151617181920 12345 67891011121314151617181920 12 34567891011121314151617181920 123 45 67 891011121314151617181920 1234 56 7891011121314151617181920 123 45 67891011121314151617181920
20 Press Side L R L R L R
Histogram of CU (%LC) Batch 1 2 3
25 15
Individual Value Plot of CU (%LC)

Frequency
106
20 10
104
Frequency
15 5 102
100
CU (%LC)
10 0 98
900 975 1050 1125 1200
weight
96
5
94
92
0 Loc 12 34 5 67 8 9101112131415161718190 12 34 5 67 89101112131415161718190 12 3 45 6 78 910111213141516171819 0 12 3 45 67 8 9101112131415161718190 123 4 56 78 910111213141516171819 0 123 45 6 78 910111213141516171819 0
2 2 2 2 2 2
94 96 98 100 102 104
Press Side L R L R L R
CU (%LC)
Batch 1 2 3
4
Bethesda, MD
What about Normality?
“Normality” – that data come from a Normal

distribution – is an assumption of many statistical
tests and calculations
Distribution Plot
Normal, Mean=100, StDev=2
0.20
0.15
Density
0.10
0.05
0.00
95.0 97.5 100.0 102.5 105.0 107.5
X
• A Normal distribution is symmetrical around the

mean and has a specific steepness and thickness of
“tails” (how fast the ends taper off on either side)
• In fact, failure to meet these precise conditions has
little practical impact on most statistics
• Tests for Normality (Anderson-Darling, etc.) are not
good at detecting non-Normality for small samples
(for which it may matter more) and overly sensitive for
large samples (for which it matters even less)
5
Bethesda, MD
Why are you testing your data for normality?

• For large sample sizes the normality tests
often give a meaningful answer to a
meaningless question
• For small samples they give a meaningless
answer to a meaningful question
- Greg Snow, Intermountain Healthcare

Histogram of CU (%LC)
What is important for valid statistics 25
20
is that the data are

Frequency
15
• Unimodal
10
• Symmetrical, at least roughly 0

94 96 98
CU (%LC)
100 102 104
• Continuous (not “chunky”)
If one or more of these things is not true, ask

• Is the thing being measured expected to be non-Normal?
• If not, what is causing it? Multiple populations? Non-
randomness?
• What is the practical implication on the statistics to be done?
6
Bethesda, MD
Representativeness: If I need randomness,

isn’t a simple random sample always best?
Types of sampling:
• Simple Random Sample: every unit has an equal
chance of being selected
• Stratified Random Sample: a random sample is
selected from each pre-designated group or
“stratum”
• Systematic Sample: every nth unit is sampled
Representativeness: If I need randomness,

isn’t a simple random sample always best?
• Another critical assumption for making valid inferences
about a population from a sample, is that the sample be
representative
• The best sampling strategy is the one that guarantees a
representative sample
• Covers the full range of values in the population
• In the same proportions as the population
• Earlier in the lifecycle, stratified or systematic sampling to
learn about homogeneity may justify simple random sample
later
7
Bethesda, MD
Statistical intervals – misconceptions &

abuses
• Why use a statistical interval?
• Quantifies the uncertainty of an estimate (e.g. the mean)
• Gives a range of plausible values
• This is only true for tolerance intervals (with some caveats) and
Bayesian intervals
• What “confidence” really means
• The confidence level is the probability of the interval
including the true value.
• Think of it as the probability of being right.

abuses
Ok, so we have to do a little math… Here is the formula for a
Confidence Interval:
Estimate ± Margin of Error Standard
s
x  t  deviation -
n Point estimate
Point estimate of of variation, σ
mean, µ
Confidence factor, t , depends on sample Sample size

size and the level of risk, α (the
percentage of time that the confidence
interval will not cover the true mean, µ)
16
8
Bethesda, MD

abuses
Summary Report for Precipitation
Anderson-Darling Normality Test
A-Squared 0.35
P-Value 0.402 90% CI = [2.33, 3.67]
Mean 3.0000
StDev
Variance
1.1547
1.3333
95% CI = [2.17, 3.83]
Skewness
Kurtosis
-0.0000000
0.0803571 99% CI = [1.81, 4.19]
N 10
Minimum 1.0000
1st Quartile 2.0000
Median 3.0000
3rd Quartile 4.0000
Maximum 5.0000
95% Confidence Interval for Mean
1 2 3 4 5 2.1740 3.8260
95% Confidence Interval for Median
2.0000 4.0000
95% Confidence Interval for StDev
0.7942 2.1080
95% Confidence Intervals
Mean
Median
2.0 2.5 3.0 3.5 4.0

abuses
• Confidence, variability, sample size & uncertainty
• A wider interval does not mean that data are likely to be as
extreme as the interval boundaries
• Keep in mind “probability of being right” – that’s what
statistical confidence is, and why intervals get wider with
• Less information (smaller n)
• Higher variability
• Higher confidence
When using them to make decisions, interpretation must
keep these things in mind
9
Bethesda, MD
Putting some of this together – Statistics

for Stage 2 PPQ Planning & Analysis
• Stage 2 Planning
• Sample Size (intra-batch)
• Quantitative attribute: statistical interval within specs,
given assumptions from Stage 1
• Qualitative attribute: statistical interval for probability of
non-conformance – tie to routine AQL
• ASTM (CU), ANSI with routine AQL as RQL or
tightened/higher level than routine

• Stage 2 Planning
• Number of batches – most companies doing risk-based
• Careful with capability or other estimates of or based on between-
batch variability in Stage 2 – why?
Statistical cautions
• Estimate of between batch standard deviation very unreliable
with small n
• Translates to very unstable capability estimate
Practical cautions
• Even with large n to give statistically stable capability estimate, it
would have little practical use – haven’t seen long-term sources
of common cause variability yet (more on this later)
10
Bethesda, MD

• Stage 2 Analysis & Conclusions
• Use statistics to help understand risk
• Keep goal in mind – PPQ should not fail due to
statistics
• TIs or CIs outside of specifications – why?
• LOOK AT THE DATA
Key Messages
• The most critical assumptions for valid statistical
analyses are randomness, independence, and
representativeness.
• Normality is not as important as the data being
unimodal, symmetrical (or close to), and
continuous (not “chunky”).
• Lack of apparent Normality is often a hint that there is a lack of
randomness or independence
• Graph your data
• To check assumptions
• To see what you expect statistical analyses to show
• To look for unexpected responses, relationships, etc.
• When using statistical intervals to make decisions,
understand the impact of sample size, variability, and
confidence level on the width (i.e., uncertainty) – and
what confidence really means
11
Bethesda, MD
Acceptance Sampling
• The process of selecting a representative part of a larger
quantity for inspection or analysis to determine if the larger
quantity can be approved as acceptable.1
• Why should we care about it?

– FDA has made it clear that it expects people to not simply
execute a sampling plan, but also to understand and justify
• what it means (AQL, RQL, OC curves)
• zero defects in sample does not mean zero defects in batch
• its implications (when acceptance criteria are not met)
• its risks (how well can you confidently characterize the process, are
the risks acceptable), etc.
1 The Handbook of Applied Acceptance Sampling, K. Stephens
Acceptance Sampling – “Fit for Purpose”

Sampling used for different purposes throughout the PV lifecycle
• Stage 1: Estimate defect rate, factors that affect it, batch
uniformity
• Stage 2: Confirm desired reliability levels, tied to severity
of harm
• Stage 3: Monitor for change, establish inter-batch
consistency
12
Bethesda, MD
Acceptance Sampling
• Two types of data
• Attribute - each result is Pass/Fail, qualitative
• Variable - each result is numerical, quantitative
• Attribute: Characteristic or property of an item
– Glass vial (cracked, not cracked)
– Cap color (correct, incorrect)
– Printing on box (legible, Illegible)
• Variable: Numerical measurement
– Dimensions of a component
– Injection force for a device
Acceptance Sampling
5 things to know about a sampling plan
• Lot size
• Sample size
• AQL – acceptance quality limit
• RQL – rejectable quality limit, also LTPD or lot tolerance
percent defective
• Acceptance number
13
Bethesda, MD
Acceptance Sampling, cont.

• AQL - the quality level that is the worst tolerable, as a
process average. Often it is defined as the percent
defective with a 95% chance of acceptance.
• Associated with producer’s risk
• RQL – the worst quality level that would be acceptable in a
single lot. Often it is defined as the percent defective with a
10% chance of acceptance.
• Associated with consumer’s risk
• Acceptance number – if number of observed defects is at or
below this number, lot will be accepted

• Risk acceptance
TRUE DEFECTIVE RATE:

< AQL > RQL (LTPD)
Consumer’s
risk
DECISION:
Accept lot
  Type II Error
() – 10%
Reject lot
 Type I Error
() – 5% 
Supplier’s
risk
The probabilities for Acceptance/Rejection associated with a sampling plan

can be graphically displayed in an Operating Characteristic Curve
14
Bethesda, MD
Operating Characteristic (OC) Curve

Operating Characteristic (OC) Curves are often used to illustrate the
performance of the sampling plan. These curves provide a way to
compare the performance of different plans.
1.0
High probability
means lots will
0.8
typically be found
Steepness of the
Prob. of Acceptance (Pa)
acceptable.
curve indicates the
0.6
discrimination
sampling plan.
0.4
Low probability
means lots will
0.2
typically be found
unacceptable.
0.0
0 2 4 6 8 10
Percent Defective

Increasing the Sample Size Improves the Discriminatory
ability of the test
OC (Operating Characteristic) Curves for AQL 0f 0.1% for various batch sizes
1
0.9 Risk of rejecting good quality = 1- Pr (acc Lot)
0.8 N, acc #
P
13;0
r 0.7
o 50;1
b 0.6 80;2
125;3
o 0.5
200;5
f
315;7
0.4
A 500;10
c 0.3 800;14
c 1250;21
0.2
Risk of accepting poor quality

0.1
0
0 2 4 6 8 10 12 14 16
Defect Rate (%)
15
Bethesda, MD

• During PPQ (Stage 2), sampling plans based on RQL
• Emphasis on consumer protection (beta risk)
• Tightened or increased confidence/coverage vs routine
levels
• During CPV (Stage 3), sampling plans based on AQL/RQL
• Monitor for shifts
• Adjust confidence/coverage to allow reduced sampling
(focus on alpha risk)

• Attribute Acceptance Sampling Standards
– ANSI/ASQ Z1.4-2003 (R2013)
– ISO 2859-1:1999
– MIL-STD-105E (1989), cancelled 1995
• Variable Acceptance Sampling Standards
– ANSI/ASQ Z1.9
– ISO 3951
• These plans are AQL based and designed to control the
producer’s risk
16
Bethesda, MD
How to use the ANSI Attribute Sampling Table
• Need 3 Values
– AQL (selected from choices above)
• Percent Nonconforming
– Batch Size Being Inspected
– General Inspection Level
• Typically use Level II
• Determine sample size code letter
• Table I
• Determine Sample Size / AC
• Single sampling plan
• Normal Inspection
• Table IIA
Determining the Sampling Plan
Sampling plan: inspect 315 units, if more than 5

defectives are observed reject the lot.
17
Bethesda, MD
Statistical Conclusions
• Sample plan: sample 315, Reject if 6 or more defectives
observed
• AQL = 0.65%
• P(X≥6|N=315 and p=0.0065) = 1.8%
• The Producers risk that a batch with an acceptable quality (worst
tolerable) will be rejected is 1.8%. False Reject rate is 1.8%
• P(X<6|N=315 and p=0.0331) = 5.0%
• The Consumers risk is that “poor” quality batches will be
accepted is 5% for p =0.0331. False Accept rate is 5% for p =
0.0331
• OC curve will visually show this for ALL p’s

Can be compared to determine an appropriate sampling
scheme
18
Bethesda, MD
Key Messages
• Sampling carries risks – understand what it is and be able to
justify it in the sampling plan
• AQLs and RQLs should be chosen according to risk to
patient, GMP compliance (customer complaints)
• Sampling plans should not
– be the same throughout the lifecycle
– chosen blindly from ANSI / ISO tables
• Sampling plans should
– be linked to risk management
– evolve through the lifecycle
– evaluated and chosen based on their operating
charateristics
Questions?
19
Bethesda, MD
Control Charts
• Statistical methodology to graphically monitor results, and
ensure that process is stable and in statistical control
• Why should we care?

– FDA has made it clear that companies need to
demonstrate that their process continues to remain in a
validated state, and to do so with the use of data and
statistics
– CPV is not a checkbox, intended to motivate continuous
improvement
Control Charts
• Graphically monitor process against pre-established limits
• Ensure process is stable and in statistical control
– Only common cause variation
• Quickly detect special cause variation
20
Bethesda, MD
Control Charts
Control Chart displaying only Common
Cause Variation
Control Chart displaying Special
Avg Dissolution Cause Variation
104 UCL=104.04
Avg Dissolution
102
104 UCL=104.04
100 102
_
X=98.47 100
98
_
X=98.47
98
96
96
94
LCL=92.90 94 55
5
92 5 LCL=92.90
5
1 3 5 7 9 11 13 15 17 19 21 92 1 1
Batch 1
90
1 5 9 13 17 21 25 29 33 37 41
Batch
Control Charts
• Data can be continuous or discrete
• Continuous – assay, dissolution, tablet weight
• Discrete – number of defects, proportion
defective
• Not all control charts apply to all types of data
21
Bethesda, MD
Various Control Charts – Continuous data

• X-Bar and R charts (Normal dist)
• Multiple measurements (reported values) per
batch
• Batch means, range,
• Example, tablet weights, dissolution, CU
• X and MR chart (Normal dist)

• Single measurement (reported value) per batch
• Example, water content, pH, assay
Various Control Charts – Discrete data

• P chart
• Proportion of defective units (Binomial dist)
• NP chart
• Number of defective units (Binomial dist)
• C chart
• Number of defects (Poisson dist)
• U chart
• Number of defects per unit (Poisson dist)
22
Bethesda, MD
Control Charts – Discrete data
I Chart of Number of Defects per Month

2.5
UCL=2.270
2.0
1.5
1.0
Correct
0.5 X=0.469
0.0
-0.5
chart
C Chart of Number of Defects per Month
-1.0 2.5 UCL=2.523

LCL=-1.333
-1.5
Jan 12 Apr 12 Jul 12 Oct 12 Jan 13 Apr 13 Jul 13 Oct 13 Jan 14 Apr 14 Jul 14 2.0
Month
1.5
Common 1.0
mistake 0.5
_
C=0.469
0.0 LCL=0
Jan 12 Apr 12 Jul 12 Oct 12 Jan 13 Apr 13 Jul 13 Oct 13 Jan 14 Apr 14 Jul 14
Month
Control Charts – Discrete data
Number of Broken Tablets

Normal
25
20
15
I Chart
10 Number of Broken Tablets
30
29
5
25
0
12 15 18 21 24 27
20 20
Good Enough 15
11
10
1 5 10 15 20 25 30 35 40 45
Batch
23
Bethesda, MD
Control Charts – Calculating Limits

• Data from historical batches – 25-30 most recent
• If not enough, can calculate preliminary limits
• Check that data are from a “stable” process

• Allow for variability from raw material batches
• Remove any special cause result
Control Charts – Calculating Limits
I Chart of Assay
105
UCL=104.45
104
103
102
101
_
100 X=100.19
99
I Chart of Assay
98 103
UCL=102.7
97
96 102
LCL=95.93
1 8 15 22 29 36 43 50 57 64
Batch 101
_
X=100.3
100
99
98 LCL=97.9
1 5 9 13 17 21 25 29 33 37
Batches
24
Bethesda, MD
Control Charts – Responding to Signals

• Western Electric rules - Decision rules for detecting
an out-of-control process
I Chart of Results
115
UCL=113.64
110
105 2
2 I Chart of Result
100 _
X=98.04 115
UCL=113.64
95
110
90 Tested on 14/07/15
2
105
85
LCL=82.44
80 100 _
X=98.04
15 15 15 15 15 15 15 15 15 15
20 20 20 20 20 20 20 20 20 20
1/ 2/ 4/ 5/ 5/ 5/ 6/ 6/ 7/ 8/ 95
/0 /0 /0 /0 /0 /0 /0 /0 /0 /0
01 05 28 11 11 29 17 17 14 16
Date Tested 90
85
LCL=82.44
80
1 6 11 16 21 26 31 36 41 46
Batch
Key Messages
• Process not always truly stable
• Shift up or down with raw material batches
• Successive batch results are not always
independent
• Testing performed in groups
• Batches manufactured in campaigns
• Data not always normal
• CQA truly non-normal or is there an outlier?
• Response to signal should be risk based
• Signal should trigger response, but commensurate with
risk to patient, process, compliance
25
Bethesda, MD
Questions?
Process Capability
• Process Capability is a measure of a process’ ability to
meet expectations. This is typically expressed in the
form of an Index, usually Cpk or Ppk.
• Motivate continuous process improvement
26
Bethesda, MD
Process Capability
• Expectations are typically specifications, but may be ANY set of
limits deemed important by the business.
• Process Capability is intended to describe what the process is
expected TO DO, not what it did.
• Process capability < 1 can be expected to produce OOS
results with some probability
• A process must be Stable and Predictable (in a state of
statistical control) to use capability measures.
 USL  x x  LSL 
Cpk  min  ,   Ppk
 3s 3s 
Process Capability
• Assumes normality, stability
• at least approximately
• Impacted by number of results used in calculation
• how well are mean and SD estimated from data
• 1-sided lower confidence bound is a better indicator
of process capability
27
Bethesda, MD
Interpreting Cpk and Ppk
LSL USL LSL USL
Cpk or Ppk < 1 Cpk or Ppk > 1

Process not Process highly
capable. capable.
-3 -2 -1 ܺത 1 2 3 -3 -2 -1 ܺത 1 2 3
LSL USL LSL USL
Cpk or Ppk ~ 1 Cpk or Ppk < 1

Process on target Process off target
-3 -2 -1 ܺത 1 2 3 -3 -2 -1 ܺത 1 2 3
55
Interpreting Cpk and Ppk
Sigma Level Ppk/Cpk % OOS

2 0.67 4.55
3 1.00 0.27
4 1.33 0.01
5 1.67 0.0001
6 2.00 0.0000002
For a two-sided specification
56
28
Bethesda, MD
Cpk versus Ppk (If only this was Easy!)

Here are some direct quotes from the web showing HOW
confusing it is out there:
• Cpk is for short term, Ppk is for long term.
• Cpk tells you what the process is capable of doing in the future. Ppk
tells you how the process has performed in the past.
• Cpk includes only common cause variation, whereas Ppk includes
both common and special cause variation.
• Cpk uses as “estimated sigma” value in its formula. Ppk uses the
“actual sigma” value in its formula.
• Unlike Cpk which uses time as a factor (because of its control chart
roots), Ppk ignores time.
• Ppk is an estimate of process capability during its initial set-up,
before it has been brought into a state of statistical control.
Difference Between Cpk and Ppk

• There is none! During this conference, regardless of
whether presentations use Cpk or Ppk it means the same
thing. Same formula, same intent; maybe different names.
• Potential differences DO exist in how “s” is determined
R S MR
, , , S , S within  Sbetween
d 2 c4 d 2
• In this conference pay attention to how the standard
deviation is estimated.
29
Bethesda, MD
Process Capability Example

Particle size data from a high volume drug substance. 341 lots included in
the analysis.
PS MV
USL
200
150
100
50 LSL
Key Messages
• Not just a number!
• Motivate process improvement
• Requires process to be stable and normal
• Critical to look at the confidence interval!
30
Bethesda, MD
Questions?
Final thoughts . . .
• FDA (and other regulators) expects companies to
demonstrate that their processes remain in a
validated state throughout the lifecycle
. . . and do so with the use of statistical methods
• Everyone in the company needs to understand
their data (not just statisticians), the risks
assumed, the interpretation of results
. . . know when you can do it yourself and when you
need a professional statistician’s help
31

Statistical Tools For Process Validation

Uploaded by

Copyright:

Available Formats

Statistical Tools For Process Validation

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Statistical Tools For Process Validation

Uploaded by

Copyright:

Available Formats

ISPE Process Validation Conference

Maneesha Altekar, Katherine Giacoletti

Statistical Methods Used in Process Validation

Connecting Pharmaceutical Knowledge ispe.org

Basic Statistical Concepts – Not Covered

Connecting Pharmaceutical Knowledge ispe.org 3

How to get in trouble with statistics

This Photo by Unknown Author is licensed under CC BY-SA

… or better yet, let’s find out

The critical assumptions for many

• Independence means the value of any given data point does

Connecting Pharmaceutical Knowledge ispe.org 5

The critical assumptions for many

… and how to detect 100.0

If data are collected from multiple 1 5 9 13 17 21

batches and/or multiple locations, plot 1 2 3 4

them in time order to look for non-

Connecting Pharmaceutical Knowledge ispe.org 6

The critical assumptions for many

Connecting Pharmaceutical Knowledge ispe.org 7

The critical assumptions for many

Two populations: how to detect 1400

Graph your data!! 1300

• Hint for 2 populations: Run/control chart

will show alternating high/low values 900

• To confirm: side-by-side individual 800

value/box plots, run/control charts by 600

press side… even just a histogram 1200

(sometimes) – but again, if data are 1100

collected with ordering or in groups, you 1000

must look for signs of non-randomness 900

before doing any statistics Histogram of weight

Histogram of CU (%LC) Batch 1 2 3

Individual Value Plot of CU (%LC)

Connecting Pharmaceutical Knowledge ispe.org 8

What about Normality?

“Normality” – that data come from a Normal

Connecting Pharmaceutical Knowledge ispe.org 9

What about Normality?

• A Normal distribution is symmetrical around the

Connecting Pharmaceutical Knowledge ispe.org 10

What about Normality?

Why are you testing your data for normality?

- Greg Snow, Intermountain Healthcare

Connecting Pharmaceutical Knowledge ispe.org 11

What about Normality?

What is important for valid statistics 25

is that the data are

• Symmetrical, at least roughly 0

• Continuous (not “chunky”)

If one or more of these things is not true, ask

Connecting Pharmaceutical Knowledge ispe.org 12

Representativeness: If I need randomness,

Connecting Pharmaceutical Knowledge ispe.org 13

Representativeness: If I need randomness,

Connecting Pharmaceutical Knowledge ispe.org 14

Statistical intervals – misconceptions &

Connecting Pharmaceutical Knowledge ispe.org 15

Statistical intervals – misconceptions &