Lecture 7: Analysis of Factors and Canonical Correlations

Transcription

1 1/33 Lecture 7: Analysis of Factors and Canonical Correlations Måns Thulin Department of Mathematics, Uppsala University Multivariate Methods 6/5 2011

2 2/33 Outline Factor analysis Basic idea: factor the covariance matrix Methods for factoring Canonical correlation analysis Basic idea: correlations between sets Finding the canonical correlations

3 3/33 Factor analysis Assume that we have a data set with many variables and that it is reasonable to believe that all these, to some extent, depend on a few underlying but unobservable factors. The purpose of factor analysis is to find dependencies on such factors and to use this to reduce the dimensionality of the data set. In particular, the covariance matrix is described by the factors.

4 4/33 Factor analysis: an early example C. Spearman (1904), General Intelligence, Objectively Determined and Measured, The American Journal of Psychology. Children s performance in mathematics (X 1 ), French (X 2 ) and English (X 3 ) was measured. Correlation matrix: R = Assume the following model: X 1 = λ 1 f + ɛ 1, X 2 = λ 2 f + ɛ 2, X 3 = λ 3 f + ɛ 3 where f is an underlying common factor ( general ability ), λ 1, λ 2, λ 3 are factor loadings and ɛ 1, ɛ 2, ɛ 3 are random disturbance terms.

5 5/33 Factor analysis: an early example Model: X i = λ i f + ɛ i, i = 1, 2, 3 with the unobservable factor f = General ability The variation of ɛ i consists of two parts: a part that represents the extent to which an individual s matehematics ability, say, differs from her general ability a measurement error due to the experimental setup, since examination is only an approximate measure of her ability in the subject The relative sizes of λ i f and ɛ i tell us to which extent variation between individuals can be described by the factor.

6 6/33 Factor analysis: linear model In factor analysis, a linear model is assumed: X (p 1) L (p m) F (m 1) ɛ (p 1) X = µ + LF + ɛ = observable random vector = matrix of factor loadings = common factors; unobserved values of factors which describe major features of members of the population = errors/specific factors; measurement error and variation not accounted for by the common factors µ i = mean of variable i ɛ i = ith specific factor F j = jth common factor l ij = loading of the ith variable on the jth factor The above model differs from ordinary linear models in that the independent variables LF are unobservable and random.

7 7/33 Factor analysis: linear model We assume that the unobservable random vectors F and ɛ satisfy the following conditions: F and ɛ are independent. E(ɛ) = 0 and Cov(ɛ) = Ψ, where Ψ is a diagonal matrix. E(F) = 0 and Cov(F) = I. Thus, the factors are assumed to be uncorrelated. This is called the orthogonal factor model.

8 8/33 Factor analysis: factoring Σ Given the model X = µ + LF + ɛ, what is Σ? See blackboard! We find that Cov(X) = LL + Ψ where V(X i ) = l 2 i1 + + l 2 im + ψ i and Cov(X i, X k ) = l i1 l k1 + + l im l km. Furthermore, Cov(X, F) = L, so that Cov(X i, F j ) = l ij. If m = the number of factors is much smaller than p =the number of measured attributes, the covariance of X can be described by the pm elements of LL and the p nonzero elements of Ψ, rather than the (p 2 + p)/2 elements of Σ.

9 9/33 Factor analysis: factoring Σ The marginal variance σ ii can also be partitioned into two parts: The ith communality: The proportion of the variance at the ith measurement X i contributed by the factors F 1, F 2,..., F m. The uniqueness or specific variance: The remaining proportion of the variance of the ith measurement, associated with ɛ i. From Cov(X) = LL + Ψ we have that σ ii = l 2 i1 + l 2 i2 + + l 2 im + }{{} ψ }{{} i Communality,hi 2 Specific variance so σ ii = h 2 i + ψ i, i = 1, 2,..., p.

10 10/33 Orthogonal transformations Consider an orthogonal matrix T. See blackboard! We find that factor loadings L are determined only up to an orthogonal matrix T. With L = LT and F = T F, the pairs (L, F ) and (L, F) give equally valid decompositions. The communalities given by the diagonal elements of LL = (L )(L ) are also unaffected by the choice of T.

11 11/33 Factor analysis: estimation The factor model is determined uniquely up to an orthogonal transformation of the factors. How then can L and Ψ be estimated? Is some L = LT better than others? Idea: Find some L and Ψ and then consider various L. Methods for estimation: Principal component method Principal factor method Maximum likelihood method

12 12/33 Factor analysis: principal component method Let (λ i, e i ) be the eigenvalue-eigenvector pairs of Σ, with λ 1 λ 2... λ p > 0. From the spectral theorem (and the principal components lecture) we know that Σ = λ 1 e 1 e 1 + λ 2e 2 e λ pe p e p. Let L = ( λ 1 e 1, λ 2 e 2,..., λ p e p ). Then Σ = λ 1 e 1 e λ pe p e p = LL = LL + 0. Thus L is given by λ i times the coefficients of the principal components, and Ψ = 0.

13 Factor analysis: principal component method Now, if λ m+1, λ m+2,..., λ p are small, then the first m principal components explain most of Σ. Thus, with L m = ( λ 1 e 1,..., λ m e m ), Σ L m L m. With specific factors, this becomes where Ψ i = σ ii m i=1 l 2 ij. Σ L m L m + Ψ As estimators for the factor loadings and specific variances, we take ( L = L ) m = ˆλ 1 ê 1,..., ˆλ m ê m where (ˆλ i, ê i ) are the eigenvalue-eigenvector pairs of the sample covariance matrix S, and m 2 Ψ = s ii l ij. i=1 13/33

14 14/33 Factor analysis: principal component method In many cases, the correlation matrix R (which is also the covariance matrix of the standardized data) is used instead of S, to avoid problems related to measurements being in different scales. Example: consumer-preference data p. 491 in J&W. See R code! Is this a good example? Can factor analysis be applied to ordinal data? Example: stock price data p. 493 in J&W. See R code!

15 15/33 Factor analysis: other estimation methods Principal factor method: Modification of principal component approach. Uses initial estimates of the specific variances. Maximum likelihood method: Assuming normal data, the maximum likelihood estimators of L and Ψ are derived. In general the estimators must be calculated by numerical maximization of a function of matrices.

16 Factor analysis: factor rotation Given an orthogonal matrix T, a different but equally valid possible estimator of the factor loadings is L = LT. We would like to find a L that gives a nice and simple interpretation of the corresponding factors. Ideally: a pattern of loadings such that each variable loads highly on a single factor and has small to moderate loadings on the remaining factors. The varimax criterion: Define l ij = ˆl ij /ĥ i. This scaling gives variables with smaller communalities more influence. Select the orthogonal transformation T that makes ( V = 1 m p p ) 2 l 4 ij 1 l 2 ij p p i=1 i=1 i=1 as large as possible. 16/33

17 17/33 Factor analysis: factor rotation Example: consumer-preference data p. 508 in J&W. See R code! Example: stock price data p. 510 in J&W. See R code!

18 18/33 Factor analysis: limitations of orthogonal factor model The factor model is determined uniquely up to an orthogonal transformation of the factors. Linearity: the linear covariance approximation LL + Ψ may not be appropriate. The factor model is most useful when m is small, but in many cases mp + p parameters are not adequate and Σ is not close to LL + Ψ.

19 19/33 Factor analysis: further topics Some further topics of interest are: Tests for the number of common factors. Factor scores. Oblique factor model In which Cov(F) is not diagonal; the factors are correlated.

20 20/33 Canonical correlations Canonical correlation analysis CCA is a means of assessing the relationship between two sets of variables. The idea is to study the correlation between a linear combination of the variables in one set and a linear combination of the variables in another set.

21 21/33 Canonical correlations: the exams problem In a classical CCA problem exams in five topics are considered. Exams are closed-book (C) or open-book (O): Mechanics (C), Vectors (C), Algebra (O), Analysis (O), Statistics (O). Question: how highly correlated is a student s performance in closed-book exams with her performance in open-book exams?

22 22/33 Canonical correlations: basic idea Given a data set X, partition the collection of variables into two sets: X (1) (p 1) and X (2) (q 1). (Assume p q.) For these random vectors: E(X (1) ) = µ (1), Cov(X (1) ) = Σ 11 E(X (2) ) = µ (2), Cov(X (2) ) = Σ 22 and Cov(X (1), X (2) ) = Σ 12 = Σ 21.

23 23/33 Canonical correlations: basic idea Introduce the linear combinations U = a X (1), V = b X (2). For these, Var(U) = a Σ 11 a, Var(V ) = b Σ 22 b and Cov(U, V ) = a Σ 12 b. Goal: Seek a and b such that is as large as possible. Corr(U, V ) = a Σ 12 b a Σ 11 a b Σ 22 b

24 Canonical correlations: some definitions The first pair of canonical variables is the pair of linear combinations U 1 and V 1 having unit variances, which maximize the correlation Corr(U, V ) = a Σ 12 b a Σ 11 a b Σ 22 b The second pair of canonical variables is the pair of linear combinations U 2 and V 2 having unit variances, which maximize the correlation among all choices that are uncorrelated with the first pair of canonical variables. The kth pair of canonical variables is the pair of linear combinations U k and V k having unit variances, which maximize the correlation among all choices that are uncorrelated with the previous k 1 canonical variable pairs. The correlation between the kth pair of canonical variables is called the kth canonical correlation. 24/33

25 25/33 Canonical correlations: solution Result The kth pair of canonical variables is given by We have that and U k = e k Σ 1/2 11 X (1) and V k = f k Σ 1/2 22 X (2). Var(U k ) = Var(V k ) = 1 Cov(U k, V k ) = ρ k, where ρ 2 1 ρ ρ 2 p and (ρ 2 k, e k) and (ρ 2 k, f k) are the eigenvalue-eigenvectors pairs of Σ 1/2 11 Σ 12 Σ 1/2 22 Σ 21 Σ 1/2 11 and Σ 1/2 22 Σ 21 Σ 1/2 11 Σ 12 Σ 1/2 22, respectively.

26 26/33 Canonical correlations: solution Alternatively, the canonical variables coefficient vectors a and b and the corresponding correlations can be found by solving the eigenvalue equations Σ 1 11 Σ 12Σ 1 22 Σ 21 a = ρ 2 a Σ 1 22 Σ 21Σ 1 11 Σ 12 b = ρ 2 b This is often more practical to use when computing the coefficients and the correlations.

27 Canonical correlations: head-lengths of sons Example: Consider the two sons born first in n families. For these, consider the following measurements: X 1 : head length of first son; X 2 : head breadth of first son; X 3 : head length of second son; X 4 : head breadth of second son. Observations: x = , S = Correlation matrices: [ R 11 = R 12 = R 21 = See R code! ] [, R 22 = [ ] ] 27/33

28 28/33 Canonical correlations: head-lengths of sons The eigenvalues of R 1 11 R 12R 1 22 R 21 are and Thus the canonical correlations are: = and = (or are they? Remember to check signs!). From the eigenvectors we obtain the canonical correlation vectors : [ ] [ ] a 1 =, a = and b 1 = [ ] [, b 2 = ] Note: The canonical correlation ˆρ 1 = exceeds any of the individual correlations between a variable of the first set and a variable of the second set.

29 29/33 Canonical correlations: head-lengths of sons The first canonical correlation variables are { U = 0.727x (1) x (1) 2 V = 0.684x (2) x (2) 2 with Corr(U, V ) = Interpretation: The sum of length and breadth of head size of each brother. These variables are highly negatively correlated between brothers.

30 30/33 Canonical correlations: head-lengths of sons The second canonical correlation variables are { U = 0.704x (1) x (1) 2 V = 0.709x (2) x (2) 2 with Corr(U, V ) = Interpretation: Seem to measure the difference between length and breadth. Head shape of first and second brothers appear to have little correlation.

31 Canonical correlations: poor summary of variability Example: Consider the following covariance matrix (p = q = 2): [ ] Σ11 Σ Σ = 12 = Σ 21 Σ It can be shown that the first pair of canonical variates is with correlation U 1 = X (1) 2, V 1 = X (2) 1 ρ 1 = Corr(U 1, V 1 ) = However, U 1 = X (1) 2 provides a very poor summary of the variability in the first set. Most of the variability in the first set is in X (1) 1, which is uncorrelated with U 1. (The same situation is true for V 1 = X (2) 1 in the second set.) 31/33

32 32/33 Canonical correlations: further topics Some further topics of interest are: Tests for Σ 12 = 0. Sample descriptive measures. Connection to MANOVA.

33 33/33 Summary Factor analysis Basic idea: factor the covariance matrix Methods for factoring Principal components Maximum likelihood Canonical correlation analysis Basic idea: correlations between sets Finding the canonical correlations Eigenvalues and eigenvectors