Multivariate Analysis

Many statistical techniques focus on just one or two variables Multivariate analysis (MVA) techniques allow more than two variables to be analysed at once
Multiple regression is not typically included under this heading, but can be thought of as a multivariate analysis

Outline of Lectures
We will cover
Why MVA is useful and important
Simpsons Paradox

Some commonly used techniques

Principal components Cluster analysis Correspondence analysis Others if time permits

Market segmentation methods An overview of MVA methods and their niches

Simpsons Paradox
Example: 44% of male applicants are admitted by a university, but only 33% of female applicants Does this mean there is unfair discrimination? University investigates and breaks down figures for Engineering and English programmes

Male
Accept 35 Refuse 45 entry Total 80

Female
20 40

60

Simpsons Paradox
No relationship between sex and acceptance for either programme
So no evidence of discrimination

Engineering
Accept Refuse entry Total English Accept Refuse entry Total

Male
30 30 60 Male 5 15 20

Female
10 10 20 Female 10 30 40

Why?
More females apply for the English programme, but it it hard to get into More males applied to Engineering, which has a higher acceptance rate than English

Must look deeper than single cross-tab to find this out

Another Example
A study of graduates salaries showed negative association between economists starting salary and the level of the degree
i.e. PhDs earned less than Masters degree holders, who in turn earned less than those with just a Bachelors degree Why?

The data was split into three employment sectors

Teaching, government and private industry Each sector showed a positive relationship Employer type was confounded with degree level

Simpsons Paradox
In each of these examples, the bivariate analysis (cross-tabulation or correlation) gave misleading results Introducing another variable gave a better understanding of the data
It even reversed the initial conclusions

Many Variables
Commonly have many relevant variables in market research surveys
E.g. one not atypical survey had ~2000 variables Typically researchers pore over many crosstabs However it can be difficult to make sense of these, and the crosstabs may be misleading

MVA can help summarise the data

E.g. factor analysis and segmentation based on agreement ratings on 20 attitude statements

MVA can also reduce the chance of obtaining spurious results

Multivariate Analysis Methods

Two general types of MVA technique
Analysis of dependence
Where one (or more) variables are dependent variables, to be explained or predicted by others
E.g. Multiple regression, PLS, MDA

Analysis of interdependence
No variables thought of as dependent Look at the relationships among variables, objects or cases
E.g. cluster analysis, factor analysis

Principal Components
Identify underlying dimensions or principal components of a distribution Helps understand the joint or common variation among a set of variables Probably the most commonly used method of deriving factors in factor analysis (before rotation)

Principal Components
The first principal component is identified as the vector (or equivalently the linear combination of variables) on which the most data variation can be projected The 2nd principal component is a vector perpendicular to the first, chosen so that it contains as much of the remaining variation as possible And so on for the 3rd principal component, the 4th, the 5th etc.

Principal Components - Examples

Ellipse, ellipsoid, sphere Rugby ball Pen Frying pan Banana CD Book

Multivariate Normal Distribution

Generalisation of the univariate normal Determined by the mean (vector) and covariance matrix

X ~ N ,

E.g. Standard bivariate normal

X ~ N 0,0 , I 2 , 1 p( x) e 2

x2 y2 2

Example Crime Rates by State

Crime Rates per 100,000 Population by State
Obs 1 2 3 4 5 State
Alabama Alaska Arizona Arkansas California

Murder
14.2 10.8 9.5 8.8 11.5

Rape
25.2 51.6 34.2 27.6 49.4

Robbery
96.8 96.8 138.2 83.2 287.0

Assault
278.3 284.0 312.3 203.4 358.0

Burglary
1135.5 1331.7 2346.1 972.6 2139.4

Larceny
1881.9 3369.8 4467.4 1862.1 3499.8

Auto_Theft
280.7 753.3 439.5 183.4 663.5

...

...

...

...

...

...

...

The PRINCOMP Procedure Observations Variables Simple Statistics Murder Mean StD
7.444000000 3.866768941 50 7

Rape
25.73400000 10.75962995

Robbery
124.0920000 88.3485672

Assault
211.3000000 100.2530492

Burglary
1291.904000 432.455711

Larceny
2671.288000 725.908707

Auto_Theft
377.5260000 193.3944175

Correlation Matrix Murder Murder Rape Robbery

1.0000 0.6012 0.4837

Rape
0.6012 1.0000 0.5919

Robbery
0.4837 0.5919 1.0000

Assault
0.6486 0.7403 0.5571

Burglary
0.3858 0.7121 0.6372

Larceny
0.1019 0.6140 0.4467

Auto_Theft
0.0688 0.3489 0.5907

Assault
Burglary Larceny Auto_Theft

0.6486
0.3858 0.1019 0.0688

0.7403
0.7121 0.6140 0.3489

0.5571
0.6372 0.4467 0.5907

1.0000
0.6229 0.4044 0.2758

0.6229
1.0000 0.7921 0.5580

0.4044
0.7921 1.0000 0.4442

0.2758
0.5580 0.4442 1.0000

Eigenvalues of the Correlation Matrix Eigenvalue 1 2 3 4 5 6 7

4.11495951 1.23872183 0.72581663 0.31643205 0.25797446 0.22203947 0.12405606

Difference
2.87623768 0.51290521 0.40938458 0.05845759 0.03593499 0.09798342

Proportion
0.5879 0.1770 0.1037 0.0452 0.0369 0.0317 0.0177

Cumulative
0.5879 0.7648 0.8685 0.9137 0.9506 0.9823 1.0000

Eigenvectors Prin1 Murder Rape Robbery Assault Burglary Larceny Auto_Theft

0.300279 0.431759 0.396875 0.396652 0.440157 0.357360 0.295177

Prin2
-.629174 -.169435 0.042247 -.343528 0.203341 0.402319 0.502421

Prin3
0.178245 -.244198 0.495861 -.069510 -.209895 -.539231 0.568384

Prin4
-.232114 0.062216 -.557989 0.629804 -.057555 -.234890 0.419238

Prin5
0.538123 0.188471 -.519977 -.506651 0.101033 0.030099 0.369753

Prin6
0.259117 -.773271 -.114385 0.172363 0.535987 0.039406 -.057298

Prin7
0.267593 -.296485 -.003903 0.191745 -.648117 0.601690 0.147046

2-3 components explain 76%-87% of the variance First principal component has uniform variable weights, so is a general crime level indicator Second principal component appears to contrast violent versus property crimes Third component is harder to interpret

Cluster Analysis
Techniques for identifying separate groups of similar cases
Similarity of cases is either specified directly in a distance matrix, or defined in terms of some distance function

Also used to summarise data by defining segments of similar cases in the data
This use of cluster analysis is known as dissection

Clustering Techniques
Two main types of cluster analysis methods
Hierarchical cluster analysis
Each cluster (starting with the whole dataset) is divided into two, then divided again, and so on

Iterative methods
k-means clustering (PROC FASTCLUS) Analogous non-parametric density estimation method

Also other methods

Overlapping clusters Fuzzy clusters

Applications
Market segmentation is usually conducted using some form of cluster analysis to divide people into segments
Other methods such as latent class models or archetypal analysis are sometimes used instead

It is also possible to cluster other items such as products/SKUs, image attributes, brands

Tandem Segmentation
One general method is to conduct a factor analysis, followed by a cluster analysis This approach has been criticised for losing information and not yielding as much discrimination as cluster analysis alone However it can make it easier to design the distance function, and to interpret the results

Tandem k-means Example

proc factor data=datafile n=6 rotate=varimax round reorder flag=.54 scree out=scores; var reasons1-reasons15 usage1-usage10; run;
proc fastclus data=scores maxc=4 seed=109162319 maxiter=50; var factor1-factor6; run;

Have used the default unweighted Euclidean distance function, which is not sensible in every context Also note that k-means results depend on the initial cluster centroids (determined here by the seed) Typically k-means is very prone to local maxima
Run at least 20 times to ensure reasonable maximum

Selected Outputs
19th run of 5 segments Cluster Summary Maximum Distance RMS Std from Seed Nearest Distance Between Cluster Frequency Deviation to Observation Cluster Cluster Centroids 1 433 0.9010 4.5524 4 2.0325 2 471 0.8487 4.5902 4 1.8959 3 505 0.9080 5.3159 4 2.0486 4 870 0.6982 4.2724 2 1.8959 5 433 0.9300 4.9425 4 2.0308

Selected Outputs
19th run of 5 segments FASTCLUS Procedure: Replace=RANDOM Radius=0 Maxclusters=5 Maxiter=100 Converge=0.02 Statistics for Variables Variable Total STD Within STD R-Squared RSQ/(1-RSQ) FACTOR1 1.000000 0.788183 0.379684 0.612082 FACTOR2 1.000000 0.893187 0.203395 0.255327 FACTOR3 1.000000 0.809710 0.345337 0.527503 FACTOR4 1.000000 0.733956 0.462104 0.859095 FACTOR5 1.000000 0.948424 0.101820 0.113363 FACTOR6 1.000000 0.838418 0.298092 0.424689 OVER-ALL 1.000000 0.838231 0.298405 0.425324 Pseudo F Statistic = 287.84 Approximate Expected Over-All R-Squared = 0.37027 Cubic Clustering Criterion = -26.135 WARNING: The two above values are invalid for correlated variables.

Selected Outputs
19th run of 5 segments Cluster Means Cluster FACTOR1 FACTOR2 FACTOR3 FACTOR4 FACTOR5 FACTOR6 1 -0.17151 0.86945 -0.06349 0.08168 0.14407 1.17640 2 -0.96441 -0.62497 -0.02967 0.67086 -0.44314 0.05906 3 -0.41435 0.09450 0.15077 -1.34799 -0.23659 -0.35995 4 0.39794 -0.00661 0.56672 0.37168 0.39152 -0.40369 5 0.90424 -0.28657 -1.21874 0.01393 -0.17278 -0.00972

Cluster Standard Deviations Cluster FACTOR1 FACTOR2 FACTOR3 FACTOR4 FACTOR5 FACTOR6 1 0.95604 0.79061 0.95515 0.81100 1.08437 0.76555 2 0.79216 0.97414 0.88440 0.71032 0.88449 0.82223 3 0.89084 0.98873 0.90514 0.74950 0.92269 0.97107 4 0.59849 0.74758 0.56576 0.58258 0.89372 0.74160 5 0.80602 1.03771 0.86331 0.91149 1.00476 0.93635

Cluster Analysis Options

There are several choices of how to form clusters in hierarchical cluster analysis
Single linkage Average linkage Density linkage Wards method Many others

Wards method (like k-means) tends to form equal sized, roundish clusters Average linkage generally forms roundish clusters with equal variance Density linkage can identify clusters of different shapes

FASTCLUS

Density Linkage

Cluster Analysis Issues

Distance definition
Weighted Euclidean distance often works well, if weights are chosen intelligently

Cluster shape
Shape of clusters found is determined by method, so choose method appropriately

Hierarchical methods usually take more computation time than kmeans However multiple runs are more important for k-means, since it can be badly affected by local minima Adjusting for response styles can also be worthwhile
Some people give more positive responses overall than others Clusters may simply reflect these response styles unless this is adjusted for, e.g. by standardising responses across attributes for each respondent

MVA - FASTCLUS
PROC FASTCLUS in SAS tries to minimise the root mean square difference between the data points and their corresponding cluster means
Iterates until convergence is reached on this criterion However it often reaches a local minimum Can be useful to run many times with different seeds and choose the best set of clusters based on this RMS criterion

See http://www.clustan.com/kmeans_critique.html for more k-means issues

Iteration History from FASTCLUS

Relative Change in Cluster Seeds Iteration Criterion 1 2 3 4 5 1 0.9645 1.0436 0.7366 0.6440 0.6343 0.5666 2 0.8596 0.3549 0.1727 0.1227 0.1246 0.0731 3 0.8499 0.2091 0.1047 0.1047 0.0656 0.0584 4 0.8454 0.1534 0.0701 0.0785 0.0276 0.0439 5 0.8430 0.1153 0.0640 0.0727 0.0331 0.0276 6 0.8414 0.0878 0.0613 0.0488 0.0253 0.0327 7 0.8402 0.0840 0.0547 0.0522 0.0249 0.0340 8 0.8392 0.0657 0.0396 0.0440 0.0188 0.0286 9 0.8386 0.0429 0.0267 0.0324 0.0149 0.0223 10 0.8383 0.0197 0.0139 0.0170 0.0119 0.0173 Convergence criterion is satisfied. Criterion Based on Final Seeds = 0.83824

Results from Different Initial Seeds

19th run of 5 segments

Cluster Means
Cluster FACTOR1 FACTOR2 FACTOR3 FACTOR4 FACTOR5 FACTOR6 1 -0.17151 0.86945 -0.06349 0.08168 0.14407 1.17640 2 -0.96441 -0.62497 -0.02967 0.67086 -0.44314 0.05906 3 -0.41435 0.09450 0.15077 -1.34799 -0.23659 -0.35995 4 0.39794 -0.00661 0.56672 0.37168 0.39152 -0.40369 5 0.90424 -0.28657 -1.21874 0.01393 -0.17278 -0.00972

20th run of 5 segments Cluster Means Cluster FACTOR1 FACTOR2 FACTOR3 FACTOR4 FACTOR5 FACTOR6 1 0.08281 -0.76563 0.48252 -0.51242 -0.55281 0.64635 2 0.39409 0.00337 0.54491 0.38299 0.64039 -0.26904 3 -0.12413 0.30691 -0.36373 -0.85776 -0.31476 -0.94927 4 0.63249 0.42335 -1.27301 0.18563 0.15973 0.77637 5 -1.20912 0.21018 -0.07423 0.75704 -0.26377 0.13729

Howard-Harris Approach
Provides automatic approach to choosing seeds for kmeans clustering Chooses initial seeds by fixed procedure
Takes variable with highest variance, splits the data at the mean, and calculates centroids of the resulting two groups Applies k-means with these centroids as initial seeds This yields a 2 cluster solution Choose the cluster with the higher within-cluster variance Choose the variable with the highest variance within that cluster, split the cluster as above, and repeat to give a 3 cluster solution Repeat until have reached a set number of clusters

I believe this approach is used by the ESPRI software package (after variables are standardised by their range)

Another Clustering Method

One alternative approach to identifying clusters is to fit a finite mixture model
Assume the overall distribution is a mixture of several normal distributions Typically this model is fit using some variant of the EM algorithm
E.g. weka.clusterers.EM method in WEKA data mining package See WEKA tutorial for an example using Fishers iris data

Advantages of this method include:

Probability model allows for statistical tests Handles missing data within model fitting process Can extend this approach to define clusters based on model parameters, e.g. regression coefficients

Also known as latent class modeling

=max.
Reason 1 Reason 2

Cluster Means
Cluster 1 4.55 4.32 Cluster 2 2.65 4.32 Cluster 3 4.21 4.12

=min.
Cluster 4 4.50 4.02

Reason 3
Reason 4 Reason 5 Reason 6 Reason 7 Reason 8 Reason 9 Reason 10 Reason 11

4.43
3.85 4.10 4.50 3.93 4.09 4.17 4.12 4.58

3.28
3.89 3.77 4.57 4.10 3.17 4.27 3.75 3.79

3.90
2.15 2.19 4.09 1.94 2.30 3.51 2.66 3.84

4.06
3.35 3.80 4.28 3.66 3.77 3.82 3.47 4.37

Reason 12
Reason 13 Reason 14 Reason 15

3.51
4.14 3.96 4.19

2.78
3.95 3.75 2.42

1.86
3.06 2.06 2.93

2.60
3.45 3.83 4.04

Cluster Means
=max.
Cluster 1 Usage 1 Usage 2 Usage 3 Usage 4 Usage 5 Usage 6 Usage 7 Usage 8 Usage 9 Usage 10 3.43 3.91 3.07 3.85 3.86 3.87 3.88 3.71 4.09 4.58

=min.
Cluster 2 3.66 3.94 2.95 3.02 3.55 4.25 3.29 2.88 3.38 4.26 Cluster 3 3.48 3.86 2.61 2.62 3.52 4.14 2.78 2.58 3.19 4.00 Cluster 4 4.00 4.26 3.13 2.50 3.56 4.56 2.59 2.34 2.68 3.91

Cluster Means
Cluster 1 Usage 1 Usage 2 Usage 3 Usage 4 Usage 5 Usage 6 Usage 7 Usage 8 Usage 9 Usage 10 3.43 3.91 3.07 3.85 3.86 3.87 3.88 3.71 4.09 4.58 Cluster 2 3.66 3.94 2.95 3.02 3.55 4.25 3.29 2.88 3.38 4.26 Cluster 3 3.48 3.86 2.61 2.62 3.52 4.14 2.78 2.58 3.19 4.00 Cluster 4 4.00 4.26 3.13 2.50 3.56 4.56 2.59 2.34 2.68 3.91

Correspondence Analysis
Provides a graphical summary of the interactions in a table Also known as a perceptual map
But so are many other charts

Can be very useful

E.g. to provide overview of cluster results

However the correct interpretation is less than intuitive, and this leads many researchers astray

Four Clusters (imputed, normalised)

Usage 9 Usage 7 Usage 8 Usage 4 Reason 2 Cluster 3 Usage 10 Reason 9 Reason 13 Usage 5 Reason 10 Usage 6 Usage 1 Reason 3 Reason 5 Reason 14 Cluster 4 Usage 2 Cluster 1 Usage 3 Reason 11 Reason 12 Reason 7 Reason 4 Cluster 2

Reason 6

Reason 1

Reason 8

25.3%

Reason 15

53.8% = Correlation < 0.50 2D Fit = 79.1%

Interpretation
Correspondence analysis plots should be interpreted by looking at points relative to the origin
Points that are in similar directions are positively associated Points that are on opposite sides of the origin are negatively associated Points that are far from the origin exhibit the strongest associations

Also the results reflect relative associations, not just which rows are highest or lowest overall

Software for Correspondence Analysis

Earlier chart was created using a specialised package called BRANDMAP Can also do correspondence analysis in most major statistical packages For example, using PROC CORRESP in SAS:
*---Perform Simple Correspondence AnalysisExample 1 in SAS OnlineDoc; proc corresp all data=Cars outc=Coor; tables Marital, Origin; run; *---Plot the Simple Correspondence Analysis Results---; %plotit(data=Coor, datatype=corresp)

Cars by Marital Status

Canonical Discriminant Analysis

Predicts a discrete response from continuous predictor variables Aims to determine which of g groups each respondent belongs to, based on the predictors Finds the linear combination of the predictors with the highest correlation with group membership
Called the first canonical variate

Repeat to find further canonical variates that are uncorrelated with the previous ones
Produces maximum of g-1 canonical variates

CDA Plot

Canonical Var 2

Canonical Var 1

Discriminant Analysis
Discriminant analysis also refers to a wider family of techniques
Still for discrete response, continuous predictors Produces discriminant functions that classify observations into groups
These can be linear or quadratic functions Can also be based on non-parametric techniques

Often train on one dataset, then test on another

CHAID
Chi-squared Automatic Interaction Detection For discrete response and many discrete predictors
Common situation in market research

Produces a tree structure

Nodes get purer, more different from each other

Uses a chi-squared test statistic to determine best variable to split on at each node
Also tries various ways of merging categories, making a Bonferroni adjustment for multiple tests Stops when no more statistically significant splits can be found

Example of CHAID Output

Titanic Survival Example

Adults (20%) / / Men / \ / \ / Children (45%) / All passengers \ \ 3rd class or crew (46%) \ / \ / Women \ \ 1st or 2nd class passenger (93%)

CHAID Software
Available in SAS Enterprise Miner (if you have enough money)
Was provided as a free macro until SAS decided to market it as a data mining technique TREEDISC.SAS still available on the web, although apparently not on the SAS web site

Also implemented in at least one standalone package Developed in 1970s Other tree-based techniques available
Will discuss these later

TREEDISC Macro
%treedisc(data=survey2, depvar=bs, nominal=c o p q x ae af ag ai: aj al am ao ap aw bf_1 bf_2 ck cn:, ordinal=lifestag t u v w y ab ah ak, ordfloat=ac ad an aq ar as av, options=list noformat read,maxdepth=3, trace=medium, draw=gr, leaf=50, outtree=all);

Need to specify type of each variable

Nominal, Ordinal, Ordinal with a floating value

Partial Least Squares (PLS)

Multivariate generalisation of regression
Have model of form Y=XB+E Also extract factors underlying the predictors These are chosen to explain both the response variation and the variation among predictors

Results are often more powerful than principal components regression PLS also refers to a more general technique for fitting general path models, not discussed here

Structural Equation Modeling (SEM)

General method for fitting and testing path analysis models, based on covariances Also known as LISREL Implemented in SAS in PROC CALIS Fits specified causal structures (path models) that usually involve factors or latent variables
Confirmatory analysis

SEM Example: Relationship between Academic and Job Success

SAS Code
data jobfl (type=cov); input _type_ $ _name_ $ act cgpa entry salary promo; cards; n 500 500 500 500 500 cov act 1.024 cov cgpa 0.792 1.077 cov entry 0.567 0.537 0.852 cov salary 0.445 0.424 0.518 0.670 cov promo 0.434 0.389 0.475 0.545 0.716 ; proc calis data=jobfl cov stderr; lineqs act = 1*F1 + e1, cgpa = p2f1*F1 + e2, entry = p3f1*F1 + e3, salary = 1*F2 + e4, promo = p5f1*F2 + e5; std e1 = vare1, e2 = vare2, e3 = vare3, e4 = vare4, e5 = vare5, F1 = varF1, F2 = varF2; cov f1 f2 = covf1f2; var act cgpa entry salary promo; run;

Results

All parameters are statistically significant, with a high correlation being found between the latent traits of academic and job success However the overall chi-squared value for the model is 111.3, with 4 d.f., so the model does not fit the observed covariances perfectly

Latent Variable Models

Have seen that both latent trait and latent class models can be useful
Latent traits for factor analysis and SEM Latent class for probabilistic segmentation

Mplus software can now fit combined latent trait and latent class models
Appears very powerful Subsumes a wide range of multivariate analyses

Broader MVA Issues

Preliminaries
EDA is usually very worthwhile
Univariate summaries, e.g. histograms Scatterplot matrix Multivariate profiles, spider-web plots

Missing data
Establish amount (by variable, and overall) and pattern (across individuals) Think about reasons for missing data Treat missing data appropriately e.g. impute, or build into model fitting

MVA Issues
Preliminaries (continued)
Check for outliers
Large values of Mahalonobis D2

Testing results
Some methods provide statistical tests But others do not
Cross-validation gives a useful check on the results
Leave-1-out cross-validation Split-sample training and test datasets Sometimes 3 groups needed For model building, training and testing