I L L I N O I S UNIVERSITY OF ILLINOIS AT URBANA-CHAMPAIGN

Transcription

1 Canonical Edps/Soc 584 and Psych 594 Applied Multivariate Statistics Carolyn J. Anderson Department of Educational Psychology I L L I N O I S UNIVERSITY OF ILLINOIS AT URBANA-CHAMPAIGN Canonical Slide 1 of 52

2 Outline Introduction Canonical and Tests on & Covariance Matrices (i.e., multiple linear combinations) Matrix Describing the relationship between sets (i.e., specific questions asked and answer in canonical analysis) Some ideas on dealing with More than two sets.. Reading: J&W Reference: Morrison (2005) Canonical Slide 2 of 52

3 Introduction We ended MANOVA talking about checking hypotheses and also made assumptions about equality of covariance matrices in discriminant analysis. There are other tests on covariance matrices that are interesting. For example Single sample: H o : Σ = Σ o (where Σ o is some specified matrix) versus H a : Σ Σ o. Tests for special structures, e.g., 1 ρ ρ ρ 1 ρ H o : Σ = σ ρ ρ 1 When might you want to test this one? Canonical Slide 3 of 52

4 More Tests Population Matrix 1 ρ ρ ρ 1 ρ H o : R = ρ ρ 1 H o : Σ 1 = Σ 2 = = Σ K versus not all are equal. Simultaneously test equality of µ and Σ from K samples. Testing the independence of sets of variables, which in what Canonical analysis deals with. Partition the covariance matrix into two sets ( ) Σ 11 Σ 12 Σ = Σ 21 Σ 22 and test H o : Σ 12 = 0. Canonical Slide 4 of 52

5 Sets of Variables In Canonical correlation analysis, we re concerned with whether two sets of variables are related or not. For example: Teacher ratings: X 1 relaxed, X 2 (motivated), X 3 (organized) and Student Achievement: Y 1 (reading), Y 2 (language), Y 3 (math) Psychological health and Performance or Behavioral measures Job performance and Job satisfaction WAIS sub-tests (e.g., digitspan, vocabulary) and Various measures of experience (e.g., age, education, etc) Canonical Slide 5 of 52

6 Goal of Canonical (Due to Hotelling about 1935) Suppose you have (p+q) variables in a vector and partition it into two parts ( ) X (p+q) = X 1,(p 1) X 2,(q 1) with covariance matrix Σ, which has also been partitioned ( ) Σ 11 Σ 12 } p Σ = Σ 21 Σ 22 } q }{{}}{{} p q Note that Σ 12 = Σ 21. Goal: Determine the relationship between the two sets of variables X 1 and X 2. Canonical Slide 6 of 52

7 Goal continued What linear combination of X 1, i.e., ax 1 = a 1 X 11 + a 2 X a p X 1p, is most closely related to a linear combination of X 2, b X 2 = b 1 X 21 + b 2 X b p X 2p. We want to choose a p 1 and b q 1 to maximize the correlation ρ(a X 1, b X 2 ). These linear combinations are called canonical variates. Plan: 1. Determine whether X 1 and X 2 are related. 2. If related, find linear combinations that maximize the canonical correlation. Canonical Slide 7 of 52

8 Assumptions of Wilk s Test Wilk s Test Statistic Wilk s Test continued Example: WAIS and Age Example of Wilk s Test for Relationship Method II: Canonical Covariance matrix for C X The Idea Solution that Maximizes the The Test Example: Method II Example continued Finishing the Test & Finding a 1 and b 1 Finding a 1 and b 1 and Conclusion (So Far) The More Usual Scaling of U 1 and V 1 Our Example Two methods to test whether the two sets of variables are related. We ll start with Wilk s likelihood ratio test for the independence of several sets of variables. This test pertains to K set of variables measures on n individuals. The i th set of variables consists of p i variables. There are ( 2 K) = K!/(2!(K 2)!) inter-variable covariance matrices. Σ ik which is (p i p k ) covariance matrix whose elements are equal to the covariances between variables in the i th set and the k th set. H o : Σ ik = 0 for all i k. This test is more general than what we need for canonical correlation analysis. Canonical Slide 8 of 52

9 Assumptions of Wilk s Test The requirements are Assumptions of Wilk s Test Wilk s Test Statistic Wilk s Test continued Example: WAIS and Age Example of Wilk s Test for Relationship Method II: Canonical Covariance matrix for C X The Idea Solution that Maximizes the The Test Example: Method II Example continued Finishing the Test & Finding a 1 and b 1 Finding a 1 and b 1 and Conclusion (So Far) The More Usual Scaling of U 1 and V 1 Our Example 1. The within set covariance matrices are positive definite; that is Σ 11, Σ 22,..., Σ KK are all positive definite. 2. A sample of n observations has been drawn from a (single) population and measures taken for p = K i=1 p i variables. For each set of variables, compute Sp p, so we have all the within set covariance matrices and between set matrices: S 11 S 12 S 1K S 21 S 22 S 2K Sp p = S K1 S K2 S KK Canonical Slide 9 of 52

10 Wilk s Test Statistic Assumptions of Wilk s Test Wilk s Test Statistic Wilk s Test continued Example: WAIS and Age Example of Wilk s Test for Relationship Method II: Canonical Covariance matrix for C X The Idea Solution that Maximizes the The Test Example: Method II Example continued Finishing the Test & Finding a 1 and b 1 Finding a 1 and b 1 and Conclusion (So Far) The More Usual Scaling of U 1 and V 1 Our Example Wilk s test statistic equals V = det(s) det(s 11 ) det(s 22 ) det(s KK ) = det(r) det(r 11 ) det(r 22 ) det(r KK ) where R is the correlation matrix and the R ii are the within set correlation matrices. The scale of the variables is not important, so we can use either S or R. The distribution of V is very complicates; however, Box (1949) gave a good approximation of V s sampling distribution. When H o is true and n is large, then (n 1) c 1 c = f(n 1) (2τ 3 + 3τ 2 ) f = (1/2)τ 2 ln(v ) χ 2 f τ 2 = ( K i=1 p i) 2 K i=1 (p i) 2 and τ 3 = ( K i=1 p i) 3 K i=1 (p i) 3. Canonical Slide 10 of 52

11 Wilk s Test continued Reject for large values of V. Assumptions of Wilk s Test Wilk s Test Statistic Wilk s Test continued Example: WAIS and Age Example of Wilk s Test for Relationship Method II: Canonical Covariance matrix for C X The Idea Solution that Maximizes the The Test Example: Method II Example continued Finishing the Test & Finding a 1 and b 1 Finding a 1 and b 1 and Conclusion (So Far) The More Usual Scaling of U 1 and V 1 Our Example For canonical correlation analysis where K 2, the test statistic simplifies to V = det(s) det(s 11 ) det(s 22 ) = det(s 11 S 12 S 1 22 S 21) det(s 11 ) = det(s 22 S 21 S 1 11 S 12) det(s 22 ) = det(i S 1 11 S 12S 1 22 S 21) = det(i S 1 22 S 21S 1 11 S 12) Time for an example Canonical Slide 11 of 52

12 Example: WAIS and Age Assumptions of Wilk s Test Wilk s Test Statistic Wilk s Test continued Example: WAIS and Age Example of Wilk s Test for Relationship Method II: Canonical Covariance matrix for C X The Idea Solution that Maximizes the The Test Example: Method II Example continued Finishing the Test & Finding a 1 and b 1 Finding a 1 and b 1 and Conclusion (So Far) The More Usual Scaling of U 1 and V 1 Our Example The data (from Morrison (1990), pp ) are from an investigation of the relationship between the Wechsler Adult Intelligence Scale (WAIS) and age. Participants: n = 933 white men and women aged Two sets of variables: Set 1: p = 2 X 1 = digit span sub-test of WAIS X 2 = vocabulary sub-test of WAIS Set 2: q = 2 X 3 = chronological age X 4 = years of formal education Sample correlation matrx: R = ( R 11 R 12 R 21 R 22 ) = Canonical Slide 12 of 52

13 Example of Wilk s Test for Relationship Testing H o : Σ 12 = 0: Method I Assumptions of Wilk s Test Wilk s Test Statistic Wilk s Test continued Example: WAIS and Age Example of Wilk s Test for Relationship Method II: Canonical Covariance matrix for C X The Idea Solution that Maximizes the The Test Example: Method II Example continued Finishing the Test & Finding a 1 and b 1 Finding a 1 and b 1 and Conclusion (So Far) The More Usual Scaling of U 1 and V 1 Our Example V = R R 11 R 22 =.4015 (.7975)(.9159) =.5497 When n is large and the null hypothesis is true (n 1) c lnv is approximately distributed as χ 2 f random variable where τ 2 = (p + q) 2 (p 2 + q 2 ) = (2 + 2) 2 ( ) = 8 τ 3 = (p + q) 3 (p 3 + q 3 ) = (2 + 2) 3 ( ) = 48 f = 1 2 τ 2 = (.5)(8) = 4 1/c = f(n 1) (2τ 3 + 3τ 2 ) =.9973 Since (n 1) c ln V = (.9973)(933 1) ln(.5497) = is much larger than a χ 2 4(.05), we reject the hypothesis; the data support the conclusion that the two sets of variables are related. (p value<<.00001). Canonical Slide 13 of 52

14 Method II: Canonical Another approach to testing H o : Σ 12 = 0. Assumptions of Wilk s Test Wilk s Test Statistic Wilk s Test continued Example: WAIS and Age Example of Wilk s Test for Relationship Method II: Canonical Covariance matrix for C X The Idea Solution that Maximizes the The Test Example: Method II Example continued Finishing the Test & Finding a 1 and b 1 Finding a 1 and b 1 and Conclusion (So Far) The More Usual Scaling of U 1 and V 1 Our Example Find the vectors a and b that maximize the correlation ρ(a X 1, b X 2 ) Define C which is (p + q) 2 matrix ( ) a 0 } p rows C = 0 b } q rows Consider the linear combination C X, ( ) ( ) ( C a 0 X 1 X = 0 b = X 2 The next piece that we need is cov(c X)... a X 1 b X 2 ) Canonical Slide 14 of 52

15 Covariance matrix for C X cov(c X) = C ΣC = = ( ( ) ( a 0 Σ 11 Σ 12 0 b Σ 21 Σ 22 ) a Σ 11 a a Σ 12 b b Σ 21 a b Σ 22 b ) ( a 0 0 b ) Assumptions of Wilk s Test Wilk s Test Statistic Wilk s Test continued Example: WAIS and Age Example of Wilk s Test for Relationship Method II: Canonical Covariance matrix for C X The Idea Solution that Maximizes the The Test Example: Method II Example continued Finishing the Test & Finding a 1 and b 1 Finding a 1 and b 1 and Conclusion (So Far) The More Usual Scaling of U 1 and V 1 Our Example and the correlation between a X 1 and b X 2 is ρ(a X 1, b X 2 ) = a Σ 12 b a Σ 11 a b Σ 22 b If Σ 12 = 0, then a Σ 12 b = 0 for all possible choices of a and b. The correlation is estimated by a S 12 b a S 11 a b S 22 b Canonical Slide 15 of 52

16 The Idea Assumptions of Wilk s Test Wilk s Test Statistic Wilk s Test continued Example: WAIS and Age Example of Wilk s Test for Relationship Method II: Canonical Covariance matrix for C X The Idea Solution that Maximizes the The Test Example: Method II Example continued Finishing the Test & Finding a 1 and b 1 Finding a 1 and b 1 and Conclusion (So Far) The More Usual Scaling of U 1 and V 1 Our Example The idea underlying this method for testing whether there is a relationship between two sets of variables is that if the correlation (in the population) is 0, then let s find the maximum possible value of the correlation ρ(a X 1, b X 2 ) and use it as a test statistic for the null hypothesis H o : Σ 12 = 0. To simply the problem, we ll constrain a and b such that var(a X 1 ) = a S 11 a = 1 and var(b X i ) = b S 22 b = 1 Our maximization problem is now to the find the a and b max a,b (a S 12 b) This is the canonical correlation (subject to the constraint that the variances of the linear combinations equal 1). Canonical Slide 16 of 52

17 Solution that Maximizes the Assumptions of Wilk s Test Wilk s Test Statistic Wilk s Test continued Example: WAIS and Age Example of Wilk s Test for Relationship Method II: Canonical Covariance matrix for C X The Idea Solution that Maximizes the The Test Example: Method II Example continued Finishing the Test & Finding a 1 and b 1 Finding a 1 and b 1 and Conclusion (So Far) The More Usual Scaling of U 1 and V 1 Our Example The largest sample correlation is the square root of the the largest eigenvalue ( root ) of S 1 11 S 12S 1 22 S 21 or equivalently S 1 22 S 21S 1 11 S 12 The matrix products S 1 11 S 12S 1 22 S 21 and S 1 22 S 21S 1 11 S 12 have the same characteristic roots (eigenvalues), which we ll call c 1, c 2,...,c r. Assume that we ve ordered the roots: c 1 c r. The eigenvector of S 1 11 S 12S 1 22 S 21 associated with c 1 gives use a 1. The eigenvector of S 1 22 S 21S 1 11 S 12 corresponding to c 1 corresponds to b 1. Setting a = a 1 and b = b 1 yields the maximum: c1 = max a,b (a S 12 b). The sample correlation between U 1 = a 1 X 1 and V 1 = b 1 X 2 equals ± c 1 (you have to determine the sign). Canonical Slide 17 of 52

18 The Test Assumptions of Wilk s Test Wilk s Test Statistic Wilk s Test continued Example: WAIS and Age Example of Wilk s Test for Relationship Method II: Canonical Covariance matrix for C X The Idea Solution that Maximizes the The Test Example: Method II Example continued Finishing the Test & Finding a 1 and b 1 Finding a 1 and b 1 and Conclusion (So Far) The More Usual Scaling of U 1 and V 1 Our Example Formal Test of H o : Σ 12 = 0: Consider c 1 = the largest root (eigenvalue). Reject H o : Σ 12 = 0 if c 1 > θ α;s,m.n where θ is the (1 α) 100% percentile point of the greatest root distribution with parameters s = min(p, q), m = 1 2 ( p q 1), n = 1 (n p q 2) 2 There are charts and tables of upper percentile points of the largest root distribution in various multivariate statistics texts (e.g., Morrison). For online via jstor.org: (Assuming connection to uiuc via netid) D. L. Heck (1960) Charts of Some Upper Percentage Points of the Distribution of the Largest Characteristic Root. The Annals of Mathematical Statistics, Vol. 31, No. 3, pp K. C. Sreedharan Pillai, Celia G. Bantegui (1959). On the Distribution of the Largest of Six Roots a Matrix in Multivariate On the Distribution of the Largest of Six Roots a Matrix Canonical Slide 18 of 52 in Multivariate. Biometrika, vol 46, pp )

19 Example: Method II Testing H o : Σ 12 = 0: Method II the largest root distribution Assumptions of Wilk s Test Wilk s Test Statistic Wilk s Test continued Example: WAIS and Age Example of Wilk s Test for Relationship Method II: Canonical Covariance matrix for C X The Idea Solution that Maximizes the The Test Example: Method II Example continued Finishing the Test & Finding a 1 and b 1 Finding a 1 and b 1 and Conclusion (So Far) The More Usual Scaling of U 1 and V 1 Our Example We first find the roots of R 1 11 R 12R 1 the roots of R 1 22 R 21R 1 R 1 11 = ( and multiplying the matrices gives us ( 22 R 21 (which are equal to 11 R 12). So we need ) ( ) R 1 22 = R 1 11 R 12R 1 22 R 21 = The roots of this matrix product are the solution of the equation (.0937 c) (.3730 c) = 0 (.0937 c)(.3730 c) (.2130)(.0873) = 0 ) Canonical Slide 19 of 52

20 Example continued Assumptions of Wilk s Test Wilk s Test Statistic Wilk s Test continued Example: WAIS and Age Example of Wilk s Test for Relationship Method II: Canonical Covariance matrix for C X The Idea Solution that Maximizes the The Test Example: Method II Example continued Finishing the Test & Finding a 1 and b 1 Finding a 1 and b 1 and Conclusion (So Far) The More Usual Scaling of U 1 and V 1 Our Example (.0937 c)(.3730 c) (.2130)(.0873) = 0 c c = 0 (c.4285)(c.0381) = 0 So c 1 =.4285 and c 2 = The largest correlation between a linear combination of variables in set 1 (U 1 = a X 1 ) and a linear combination of variables in set 2 (V 1 = b X 2 ) equals c1 =.4285 = a Σb =.654 To test whether H o : Σ 12 = 0, we have s = min(p, q) = min(2, 2) = 2 m = (1/2)( p q 1) =.5( 2 2 1) =.5 n = (1/2)(n p q 2) =.5( ) = where p = number of variables in set 1, and q = number of variables in set 2. Canonical Slide 20 of 52

21 Finishing the Test & Finding a 1 and b 1 Assumptions of Wilk s Test Wilk s Test Statistic Wilk s Test continued Example: WAIS and Age Example of Wilk s Test for Relationship Method II: Canonical Covariance matrix for C X The Idea Solution that Maximizes the The Test Example: Method II Example continued Finishing the Test & Finding a 1 and b 1 Finding a 1 and b 1 and Conclusion (So Far) The More Usual Scaling of U 1 and V 1 Our Example Using the chart (page 628 of Heck (1960)) of the greatest root distribution, we find θ 2,.5,463.5 (.01) =.02. Since c 1 =.4285 >.02, we reject H o ; there is a dependency between the sets of variables. The linear combination that Maximizes the correlation. To compute a 1, we use R 1 11 R 12R 1 22 R 21 ( R 1 11 R 12R R 21a 1 = c 1 a 1 ) ( ) a 11 a 21 Two equations, two unknowns: = a a 12 = a a 12 = 0 For convenience, we ll set a 12 = 1, and solve for a 11 = (.0547/.2130)a 12 =.26. ( a 11 a 21 ) Canonical Slide 21 of 52

22 Finding a 1 and b 1 Assumptions of Wilk s Test Wilk s Test Statistic Wilk s Test continued Example: WAIS and Age Example of Wilk s Test for Relationship Method II: Canonical Covariance matrix for C X The Idea Solution that Maximizes the The Test Example: Method II Example continued Finishing the Test & Finding a 1 and b 1 Finding a 1 and b 1 and Conclusion (So Far) The More Usual Scaling of U 1 and V 1 Our Example Any vector that is proportional to (i.e., is a multiple of) a = (.26, 1) is a solution and gives us the correct linear combination. To compute b 1, we use R 1 22 R 21R 1 11 R 12 ( R 1 22 R 21R R 12b 1 = c 1 b 1 ) ( ) b 11 b 12 Two Equations, Two Unknowns = b b 12 = b b 12 = 0 ( b 11 b 12 We ll set b 12 = 1 and solve for b 11 = (.0076/.0378)b 12 =.20. Any vector proportional to (a multiple of) b = (.20, 1) is a solution and gives us the correct linear combination. ) Canonical Slide 22 of 52

23 and Conclusion (So Far) Since the correlation matrix, R, was used, to solve for the vectors a 1 and b 1, we use the standardized scores (i.e., z-scores) rather than the original (raw) variables. U 1 =.26z (digit span) z (vocabulary) V 1 =.20z (age) z (years of formal education) Assumptions of Wilk s Test Wilk s Test Statistic Wilk s Test continued Example: WAIS and Age Example of Wilk s Test for Relationship Method II: Canonical Covariance matrix for C X The Idea Solution that Maximizes the The Test Example: Method II Example continued Finishing the Test & Finding a 1 and b 1 Finding a 1 and b 1 and Conclusion (So Far) The More Usual Scaling of U 1 and V 1 Our Example Interpretation/: The correlation between equals U 1 and V 1, which equals c1 =.4285 =.654 = (a 1R 12 b 1 )/( a 1 R 11a 1 b 1 R 22 b 1 ), is the largest possible one for any linear combination of the variables in sets 1 and 2. U 1 : places four times more weight on vocabulary than on digit span... long term versus short term memory. V 1 : places five times more weight on years of formal education than on chronological age. The major link between the two sets of variables is due to education and vocabulary. Canonical Slide 23 of 52

24 The More Usual Scaling of U 1 and V 1 (What I did wasn t the typical way to scale a 1 and b 1 ). Assumptions of Wilk s Test Wilk s Test Statistic Wilk s Test continued Example: WAIS and Age Example of Wilk s Test for Relationship Method II: Canonical Covariance matrix for C X The Idea Solution that Maximizes the The Test Example: Method II Example continued Finishing the Test & Finding a 1 and b 1 Finding a 1 and b 1 and Conclusion (So Far) The More Usual Scaling of U 1 and V 1 Our Example The standard or typical way that a 1 and b 1 are scaled is such that the variances of U 1 = a 1 X 1 and V 1 = b 1 X 2 equal 1. Since any multiple of a 1 and/or b 1 is a solution, we just need to multiply the vectors by an appropriate constant. For example, and now a 1 = a 1 a 1 R 11 a var(u 1 ) = var(a 1X 1 ) = a 1 R 11 a 1 ( ( ) a 1 a = )R a R 11 a a 1 R 11 a = 1 Canonical Slide 24 of 52

25 Our Example a 1R 11 a 1 = (0.26, 1.00) ( ) ( ) = So a 1 = (0.26, 1.00) = (.2279,.8765) Assumptions of Wilk s Test Wilk s Test Statistic Wilk s Test continued Example: WAIS and Age Example of Wilk s Test for Relationship Method II: Canonical Covariance matrix for C X The Idea Solution that Maximizes the The Test Example: Method II Example continued Finishing the Test & Finding a 1 and b 1 Finding a 1 and b 1 and Conclusion (So Far) The More Usual Scaling of U 1 and V 1 Our Example As a check: var(u 1 ) = (.2279,.8765) ( Doing the same thing for b: ( b R 22 b 1 = (0.20, 1.00) So b 1 = ) ( ) ( ( 1 )b 1 = (.2081, ) ) ) = 1.00 = Canonical Slide 25 of 52

26 Assumptions and Solution Back to our Example The Canonical Variates al Structure of Canonical Variates So far we ve only looked at the largest correlation possible between linear combinations of the variables from two sets; however, there are more c i s. There are more linear combinations: U 1 = a 1X 1 V 1 = b 1X 2 U 2 = a 2X 1 V 2 = b 2X 2.. U r = a rx 1 V r = b rx 2 With the property that the sample correlation between U 1 and V 1 is the largest, the sample correlation between U 2 and V 2 is the largest among all linear combinations uncorrelated with U 1 and V 1, etc. That is, for all i k, cov(u i, U k ) = a is 11 a k = 0 cov(v i, V k ) = b is 22 b k = 0 cov(u i, V k ) = a is 12 b k = 0 cov(u k, V i ) = a ks 12 b i = 0 Canonical Slide 26 of 52

27 Assumptions and Solution Assumptions and Solution Back to our Example The Canonical Variates al Structure of Canonical Variates Assume 1. The elements of Σ (p+q) (p+q) are finite. 2. Σ is full rank; that is, rank = p + q. 3. The first r min(p, q) characteristic roots of Σ 1 11 Σ 12Σ 1 22 Σ 21 are distinct. Then a i and b i are estimated from the data by solving (S 12 S 1 22 S 21 c i S 11 )a i = 0 and (S 21 S 1 11 S 12 c i S 22 )b i = 0 where c i is the i th root of the equation. det(s 12 S 1 22 S 21 c i S 11 ) = 0 or equivalently det(s 21 S 1 11 S 12 c i S 22 ) = 0 c i = squared sample correlation between U i and V i Canonical Slide 27 of 52

28 Back to our Example Assumptions and Solution Back to our Example The Canonical Variates al Structure of Canonical Variates The linear combination that gives the next largest correlation and is orthogonal to the first one. Using the second root of the matrix product, c 2 =.0381, which is the second largest squared correlation, repeat the process we did previously to get the vectors a 2 and b 2 : a 2 = ( 1.00, 0.64) and b 2 = (1.00, 0.10) or using the more typically scaling, we get a 2 = ( ,.7001) and b 2 = (1.0249,.1025) Statistical Hypothesis Test: H o : ρ(u 2, V 2 ) = 0 vs H a : ρ(u 2, V 2 ) 0: (n 1 12 ) (p + q + 1) ln(1 c 2 ) = (932.5(5)) ln(1.0381) = If the null hypothesis is true (and n large), then this statistic is approximately distributed at χ 2 pq. Since χ 2 4(.05) = 9.488, we reject the null and conclude that the second canonical correlation is not zero. Canonical Slide 28 of 52

29 The Canonical Variates To find a 2 and b 2, we use the same process that we used to find a 1 and b 1, which gives us and U 2 = 1.00Z ( digit span) Z ( vocabulary) V 2 = 1.00Z ( age) Z ( years of formal education) Assumptions and Solution Back to our Example The Canonical Variates al Structure of Canonical Variates U 2 is a weighted contrast between digit span (performance) and vocabulary (verbal) sub-tests. V 2 is nearly all age. There s a widening in the gap between performance with advancing age. As people get older, there s a larger difference between accumulated knowledge (vocabulary) and performance skills. Canonical Slide 29 of 52

30 al Structure of Canonical Variates Note: The following covariances (correlations) are all zero: Assumptions and Solution Back to our Example The Canonical Variates al Structure of Canonical Variates cov(u 1, U 2 ) = a 1 R 11 a 2 = 0 cov(v 1, V 2 ) = b 1 R 22 b 2 = 0 cov(u 1, V 2 ) = a 1 R 12 b 2 = 0 cov(v 1, U 2 ) = b 1 R 21 a 2 = 0 In other words,the sample correlation matrix for the canonical variates is U 1 U 2 V 1 V 2 U U V V which is much simpler than the sample correlation matrix back on page 12. Canonical Slide 30 of 52

31 Showing Why this Works Or the how to do this in IML, R, and/or MATLAB: We symmetrize the matrix so that our solution are the eigenvalues and vectors of 1/2 (S 11 S 12 S 1 22 S 21S 1/2 11 )e i = c i e i where S 1/2 11 is the inverse of the square root matrix. Then the combination vectors that you want equal a i = S 1/2 11 e i b i = 1 S 1 22 S 21a i ci OR You can find the eigenvalues and eigenvectors of (S 1/2 22 S 21 S 1 11 S 12S 1/2 22 )f i = c i f i Then the combination vectors that you want equal b i = S 1/2 22 f i a i = 1 S 1 11 S 21b i ci Canonical Slide 31 of 52

32 Showing Why this Works Showing Why this Works S 1/2 11 ( ) S 1/2 11 S 12 S 1 22 S 21S 1/2 11 e = c e ( ) S 1/2 11 S 12 S 1 22 S 21S 1/2 11 e = c S 1/2 11 e ( S 1 11 S 12S 1 22 S ) 1/2 21 S 11 e }{{} a = c S 1/2 11 e }{{} a ( S 1 11 S 12S 1 22 S 21) a = c a This is what we did for discriminant analysis to find eigenvalues and vectors of a non-symmetric matrix. Canonical Slide 32 of 52

33 1. To what extent can one set of two (or more) variables be predicted by or explained by another set of two or more variables? 2. What contributions does a single variable make to the explanatory power of the set of variable to which the variable belongs? of What CCA Does Example: The Simplification of R Question 1 Question 2 The Structure Coefficients Question 3 Relationship between Index and Structure 3. To what extent does a single variable contribute to predicting or explaining the composite of the variables in the variable set to which the variable does not belong? We ll talk about how to answer each of these. Canonical Slide 33 of 52

34 of What CCA Does Started with sample R (or S), which in our example is X 1 X 2 X 3 X 4 set 1 X digit span X vocabulary set 2 X age X years of education of What CCA Does Example: The Simplification of R Question 1 Question 2 The Structure Coefficients Question 3 Relationship between Index and Structure We found linear transformation, canonical variates, of the original variables within sets to maximize the between set correlations: where U i = a ix 1 and V i = b ix 2 (R 1 11 R 12R 1 22 R 21)a i = c i a i and (R 1 22 R 21R 1 11 R 12)b i = c i b i And scaled them so that the variance of U i and V i equal 1 a ir 11 a i = a ir 22 b i = 1 Canonical Slide 34 of 52

35 Example: The Simplification of R of What CCA Does Example: The Simplification of R Question 1 Question 2 The Structure Coefficients Question 3 Relationship between Index and Structure X 1 X 2 X 3 X 4 U 1 U 2 V 1 V 2 X X X X U U V V Canonical Slide 35 of 52

36 Question 1 X 1 X 2 X 3 X 4 U 1 U 2 V 1 V 2 of What CCA Does Example: The Simplification of R Question 1 Question 2 The Structure Coefficients Question 3 Relationship between Index and Structure X 1 R 11 R 12 X 2 X 3 R 21 R 22 X 4 U 1 Structure Index c U 2 coefficients coefficients c2 V 1 Index Structure c V 2 coefficients coefficients 0.00 c Question 1: To what extent can one set of two or more variables be explained by another set of two or more variables? Answer: The (first) canonical correlation c 1. Canonical Slide 36 of 52

37 Question 2 Question 2: What is the contribution between a variable and the canonical (composite) variable for it s set? of What CCA Does Example: The Simplification of R Question 1 Question 2 The Structure Coefficients Question 3 Relationship between Index and Structure Answer: The structure coefficients. Variables Within Sets X 1 = X 11 X 12. X 1p X 2 = Structure coefficients equal, X 21 X 22. X 2q Set 1: }{{} corr (X 1, U i ) = corr(x 1, a ix 1 ) vector Canonical Variates U i = a i X 1 V i = b ix 2 Set 2: }{{} corr (X 2, V i ) = corr(x 2, b ix 2 ) vector Canonical Slide 37 of 52

38 The Structure Coefficients Set 1:corr (p) (X 1, U i ) = corr(x 1, a ix 1 ) 1/ s / s 22 0 = /sqrts pp = diag(1/ s ii )S 11 a i S 11a i of What CCA Does Example: The Simplification of R Question 1 Question 2 The Structure Coefficients Question 3 Relationship between Index and Structure When we use R rather than S to find U i and V i, then set 1: corr(z 1, U i ) (p 1) = R 11 a i and set 2: corr(z 2, V i ) q 1) = R 22 b i Canonical Slide 38 of 52

39 Question 3 Question 3 To what extent does a single variable contribute to explaining the canonical (composite) variable in the set of variables to which it does not belong? Answer: The correlation between it and the canonical variate of the other set of variables: index coefficients. of What CCA Does Example: The Simplification of R Question 1 Question 2 The Structure Coefficients Question 3 Relationship between Index and Structure corr(x 1, V i ) (p 1) = corr(x 1, ī X 2 ) = diag(1/ s 11(ii) )Σ 12 b i and corr(x 2, U i ) (q 1) = corr(x 2, i X 1 ) = diag(1/ s 22(ii) )Σ 21 a i Canonical Slide 39 of 52

40 Relationship between Index and Structure a i can be written as a linear combination of b i (and visa versa). of What CCA Does Example: The Simplification of R Question 1 Question 2 The Structure Coefficients Question 3 Relationship between Index and Structure a i = 1 c i Σ 1 11 Σ 12b i and b i = 1 c i Σ a i Re-arrange terms a bit ci }{{} Σ 11a i = Σ 12 b i So the Index Coefficients corr(x 1, V i ) (p 1) = diag(1/ σ 11(ii) )Σ 12 b i = D 1/2 1 ( c i Σ 11 a i ) = ci D 1/2 1 Σ 11 a i }{{}}{{} cov(u i,v i ) corr(x 1,U i ) ci }{{} Σ 22b i = Σ 21 a i Index coefficient = cov(u i, V i ) (Structure Coefficient) }{{} canonical correlation and corr(x 2, U i ) = cov(u i, v i )corr(x 2, V i ) Canonical Slide 40 of 52

41 Edited Output Canonical s, etc Canonical Coefficients Structure Coefficients Index Coefficients If you have a correlation matrix, you can input it as as data set using: data corrmat (type=corr); input TYPE $ NAME $ x1 x2 x3 x4; list; datalines; N CORR x CORR x CORR x CORR x * If you do not input N, default is to assume that N=10,000; proc cancorr data=corrmat simple corr; var x1 x2; with x3 x4; title Canonical of WAIS ; run; Canonical Slide 41 of 52

42 Edited Output Edited Output Canonical s, etc Canonical Coefficients Structure Coefficients Index Coefficients s Among the VAR Variables x1 x2 x <- R_11 x s Among the WITH Variables x3 x4 x <- R_22 x s Between the VAR Variables and the WITH Variables x3 x4 x <- R_12 x Canonical Slide 42 of 52

43 Canonical s, etc Adjusted Approximate Squared Canonical Canonical Standard Canonical Error rho1 = rho2 = Edited Output Canonical s, etc Canonical Coefficients Structure Coefficients Index Coefficients Test of H0: The canonical correlations in the current row and all that follow are zero Likelihood Approximate Ratio F Value Num DF Den DF Pr > F < <.0001 Note: First line above tests H_o: Sigma_12 = rho1 = rho2 = 0 Second line above tests H_o: rho2 = 0 Canonical Slide 43 of 52

44 Canonical Coefficients Edited Output Canonical s, etc Canonical Coefficients Structure Coefficients Index Coefficients Raw Canonical Coefficients for the VAR Variables V1 x <- columns = a_i V2 x Raw Canonical Coefficients for the WITH Variables W1 x <- columns = b_i W2 x Notes: * Since we input a correlation matrix, "raw" are same as standardized coefficients. * "V" is same as lecture notes "U". * "W" is same as lecture notes "V". Canonical Slide 44 of 52

45 Structure Coefficients s Between the VAR Variables and Their Canonical Variables V1 V2 x x Edited Output Canonical s, etc Canonical Coefficients Structure Coefficients Index Coefficients s Between the WITH Variables and Their Canonical Variables W1 W2 x x Canonical Slide 45 of 52

46 Index Coefficients s Between the VAR Variables and the Canonical Variables of the WITH Variables W1 W2 x x Edited Output Canonical s, etc Canonical Coefficients Structure Coefficients Index Coefficients s Between the WITH Variables and the Canonical Variables of the VAR Variables V1 V2 x x Canonical Slide 46 of 52

47 For M > 2, let p= M m=1 p m, X 1 } p 1 X 2 } p 2 X = and Σ =.. } p M X m Σ 11 Σ 12 Σ 1M Σ 21 Σ 22 Σ 2M Σ M1 Σ M2 Σ MM Consider the linear combinations Z 11 = a 11X 1 Set up for Horst s Suggestions Kettenring s Suggestions GENVAR Z 12 = a 12X 2 Z 1M = a 1MX M where a 1m is the (p m 1) vector for the m th canonical variable. The (estimated) covariance matrix of (Z 1 1, Z 1 2,...,Z 1 M) is ˆ Φ(1) (M M), which equals.... Canonical Slide 47 of 52

48 Set up for ˆ Φ(1) (M M) = 1 ˆφ12 (1) ˆφ13 (1) ˆφ1M (1) ˆφ 21 (1) 1 ˆφ23 (1) ˆφ2M (1) ˆφ M1 (1) ˆφM2 (1) ˆφM3 (1) 1 Set up for Horst s Suggestions Kettenring s Suggestions GENVAR where ˆφ ii(1) = a 1i Σ iia 1i = 1 and ˆφ ik(1) = a 1i Σ ika 1k. In the two set case, we only had one off diagonal element that could maximize, i.e., ˆφ 12 (1). There are (at least) five ways to generalized canonical correlation to a multiple set problems. For M = 2 they are all equivalent to what we ve talked about. Canonical Slide 48 of 52

49 Horst s Suggestions SUMCOR: Horst (1965) Factor analysis of data matrices Horst suggested maximizing max a 1,...,a M i<k ˆφ ik (1) That is, sum of all off-diagonal elements of the correlation matrix between the M canonical variates Z 1,...,Z M ), i.e., ˆΦ(1). Set up for Horst s Suggestions Kettenring s Suggestions GENVAR MAXVAR: Horst also suggest maximizing the variance of a linear combination of Z = (Z 11, Z 12,...,Z 1M, which is the first principal component of ˆΦ(1). The variance of this maximal linear combination is the largest eigenvalue of ˆΦ(1); that is, λ 1. Canonical Slide 49 of 52

50 Kettenring s Suggestions Kettenring (1971), Biometrika SSQCOR: Find the linear combinations that maximize max a 1,...,a M ( i<k ˆφ 2 ik(1)) = M λ 2 ik M i=1 Set up for Horst s Suggestions Kettenring s Suggestions GENVAR MINVAR: Kettenring also suggested minimizing the smallest eigenvalue λ 1M ; that is, the variance of the minimal linear combination. Canonical Slide 50 of 52

51 GENVAR Steel (1951) in the Annals of Mathematical Statistics suggested minimizing the generalized sample variance det( ˆΦ(1)) = M m=1 λ ij Once the first canonical variate has been found, all five methods call all be extended to find Z 2, Z 3,...,Z M such that they are orthogonal to ones previously found. Set up for Horst s Suggestions Kettenring s Suggestions GENVAR Canonical Slide 51 of 52

52 /Discusion /Discusion Discriminant analysis is a special case of canonical correlation analysis where one set of variables is a dummy coded variable that defines populations. i.e., For individual j in group i, X 1j (1 p) = (0,..., }{{} 1,...0). i th The other set of variables are q continuous/numerical ones. Discriminant analysis and MANOVA use the same matrix W 1 B or E 1 H. (Binary or Multinomial) Logistic regression is the flip side of discriminant analysis and MANOVA (i.e., interchange response and predictor roles). Conditional Guassian distribution or the location model. See handout on similarities and differences between PCA, MANOVA, DA, and canonical correlation analysis. Canonical Slide 52 of 52