1. Multivariate Analysis of Variance (MANOVA)

2. Outline Purpose and logic : page 3 Purpose and logic : page 3 Hypothesis testing : page 6 Hypothesis testing : page 6 Computations: page 11 Computations: page 11 F-Ratios: page 25 F-Ratios: page 25 Assumptions and noncentrality : page 35 Assumptions and noncentrality : page 35

3. MANOVA When ? When a research design contains two or more dependent variables we could perform multiple univariate tests or one multivariate test Why ? MANOVA does not have the problem of inflated overall type I error rate (  ) Univariate tests ignore the correlations among the variables Multivariate tests are more powerful than multiple univariate tests Assumptions Multivariate normality Absence of outliers Homogeneity of variance-covariance matrices Linearity Absence of multicollinearity

4. MANOVA If the independent variables are discrete and the dependant variables are continuous we will performed a MANOVA To GLMwhere, From MANOVA where,  = grand mean,  = treatment effect 1,  = treatment effect 2,  = interaction, e = error

5. MANOVA Example Drug ABC Male 5, 6 5, 4 9, 9 7, 6 7, 7 9, 12 6, 8 21, 15 14, 11 17, 12 12, 10 Female 7, 10 6, 6 9, 7 8, 10 10, 13 8, 7 7, 6 6, 9 16, 12 14, 9 14, 8 10, 5 The general idea behind MANOVA is the same as previously. We want to find a ratio between explained variability over unexplained variability (error)  = treatment effect 1 (rows; r = 2)  = treatment effect 2 (columns; c = 3) n i = 4N = r*c*n i =24 q = number of DV = 2 (WeightLoss, Time)

6. Analysis of Variance (ANOVA) Drug ABC Male 5, 6 5, 4 9, 9 7, 6 7, 7 9, 12 6, 8 21, 15 14, 11 17, 12 12, 10 Female 7, 10 6, 6 9, 7 8, 10 10, 13 8, 7 7, 6 6, 9 16, 12 14, 9 14, 8 10, 5 Hypothesis Are the drug mean vectors equal? Are the sex mean vectors equal? Do some drugs interact with sex to produce inordinately high or low weight decrements?

7. Analysis of Variance (ANOVA) Using the GLM approach through a coding matrix

8. Analysis of Variance (ANOVA) Then, for each subject we associate its corresponding group coding.

9. Analysis of Variance (ANOVA)

10. Canonical correlation matrix R is obtained by:

11. Error Matrix (E) In ANOVA, the error was defined as e = (1-R 2 )S cc This is a special case of the MANOVA error matrix E

12. Hypothesis variation matrix The total variation is the sum of the various hypothesis variation add to the error variation, i.e. T=E+H  +H  +H . Each matrix H is obtained by Where i  { , ,  } The full model is omitted when performing hypothesis testing (We start by testing the interaction, then the main effects, etc.)

13. Hypothesis variation matrix Interaction = M 

14. Hypothesis variation matrix Interaction

15. Here is the catch! In univariate, the statistics is based on the F-ratio distribution However, in MANOVA there is no unique statistic. Four statistics are commonly used: Hotelling-Lawley trace (HL), Pillai-Bartlett trace (PB), Wilk`s likelihood ratio (W) and Roy’s largest root (RLR).

16. Hotelling-Lawley trace (HL) The HL statistic is defined as where s = min(df i, q), i represents the tested effect (i  { , ,  }), df i is the degree of freedom associated with the hypothesis under investigation ( ,  or  ) and k is kth eigenvalue extracted from H i E -1.

17. Hotelling-Lawley trace (HL) Interaction Extracted eigenvalues

18. Hotelling-Lawley trace (HL) Interaction df  = (r-1)(c-1)=(2-1)(3-1) = 2 s = min(df , q) = min(2, 2) = 2 Trace

19. Pillai-Bartlett trace (PB) The PB statistic is defined as where s = min(df i, q), i represents the tested effect (i  { , ,  }), df i is the degree of freedom associated with the hypothesis under investigation ( ,  or  ) and k is kth eigenvalue extracted from H i E -1.

20. Pillai-Bartlett trace (PB) Interaction df  = (r-1)(c-1)=(2-1)(3-1) = 2 s = min(df , q) = min(2, 2) = 2

21. Wilk’s likelihood ratio (W) The W statistic is defined as where s = min(df i, q), i represents the tested effect (i  { , ,  }), df i is the degree of freedom associated with the hypothesis under investigation ( ,  or  ), k is kth eigenvalue extracted from H i E -1 and |E| (as well as |E+H i |) is the determinant.

22. Wilk’s likelihood ratio (W) Interaction df  = (r-1)(c-1)=(2-1)(3-1) = 2 s = min(df , q) = min(2, 2) = 2

23. Roy’s largest root (RLR) The RLR statistic is defined as where i represents the tested effect (i  { , ,  }) and k is kth eigenvalue extracted from H i E -1.

24. Roy’s largest root (RLR) Interaction

25. Multivariate F-ratio All the statistics are equivalent when s = 1. In general there is no exact formula for finding the associated p-value except on rare situations. Nevertheless, a convenient and sufficient approximation exists for all but RLR. Since RLR is the least robust, attention will be focused on the first three statistics: HL, PB and W. These three statistics’ distributions are approximated using an F distribution which has the advantage of being simple to understand

26. Multivariate F-ratio Where df 1 represents the numerator degree of freedom (df 1 = q*df i ) df 2 (m) the denominator degree of freedom for each statistic m (m  {HL i, PB i and W i })  2 m is the multivariate measure of association for each statistic m

27. Multivariate F-ratio (HL) The multivariate measure of association for HL is given by The numerator df The denominator df

28. Multivariate F-ratio (HL) Interaction The multivariate measure of association for HL is given by The numerator df The denominator df

29. Multivariate F-ratio (PB) The multivariate measure of association for PB is given by The numerator df The denominator df

30. Multivariate F-ratio (PB) Interaction The multivariate measure of association for PB is given by The numerator df The denominator df

31. Multivariate F-ratio (W) The multivariate measure of association for W is given by The numerator df The denominator df

32. Multivariate F-ratio (W) Interaction The multivariate measure of association for W is given by The numerator df The denominator df

33. Multivariate F-ratio HL (interaction,  ) PB (interaction,  ) W (interaction,  )

34. MANOVA Summary

35. MANOVA Unfortunately there is no single test that is the most powerful if the MANOVA assumptions are not met. If there is a violation of homogeneity of the covariance matrices or the multivariate normality, then the PB statistic is the most robust while RLR is the least robust statistic. If the noncentrality is concentrated (when the population centroids are largely confined to a single dimension), RLR provides the most power test.

36. MANOVA If on the other hand, the noncentrality is diffuse (when the population centroids differ almost equally in all dimensions) then PB, HT or W will all give good power. However, in most cases, power differences among the four statistics are quite small (<0.06), thus it does not matter which statistics is used.