In SPSS using the GLM:repeated measures procedure a multivariate analysis of variance is also presented. Consider nine subjects who perform a task under each of three conditions. The scores are given in the table along with the differences, d1 and d2, of C1-C2 and C1-C3 respectively.
||||<20% style="TEXT-ALIGN: center"> C1 || C2 || C3 || D1 || D2 ||||<20% style="VERTICAL-ALIGN: top; TEXT-ALIGN: center"> 2 || 1 || 2 || 1 || 0 ||||<20% style="VERTICAL-ALIGN: top; TEXT-ALIGN: center"> 1 || 2 || 1 || -1 || 0 ||||<20% style="VERTICAL-ALIGN: top; TEXT-ALIGN: center"> 2 || 3 || 2 || -1 || 0 ||||<20% style="VERTICAL-ALIGN: top; TEXT-ALIGN: center"> 3 || 2 || 3 || 1 || 0 ||||<20% style="VERTICAL-ALIGN: top; TEXT-ALIGN: center"> 2 || 3 || 4 || -1 || -2 ||||<20% style="VERTICAL-ALIGN: top; TEXT-ALIGN: center"> 1 || 4 || 5 || -3 || -4 ||||<20% style="VERTICAL-ALIGN: top; TEXT-ALIGN: center"> 3 || 5 || 3 || -2 || 0 ||||<20% style="VERTICAL-ALIGN: top; TEXT-ALIGN: center"> 2 || 6 || 4 || -4 || -2 ||||<20% style="VERTICAL-ALIGN: top; TEXT-ALIGN: center"> 1 || 4 || 5 || -3 || -4
Now, the multivariate anova test of the overall mean difference tests if the vector (d1 mean,d2 mean) equals (0,0) using the 2 by 2 (d1,d2) covariance matrix consisting of the d1 and d2 variances and their covariance.
The univariate anova test, on the other hand, takes the score and fits a multiple regression with a series of indicator variables. The indicators for the first four subjects are coded in the table. Two of these indicators represent condition and eight represent subjects. The SS(condition), for example, in the Univariate repeated measures anova represents the unexplained variance in scores which is not explained by subject differences ie 31.037 - 19.407 = 11.63.
||||<9% style="TEXT-ALIGN: center"> SCORE || INDC2 || INDC3 || INDS1 || INDS2 || INDS3 || INDS4 || INDS5 || INDS6 || INDS7 || INDS8 ||||<9% style="VERTICAL-ALIGN: top; TEXT-ALIGN: center"> 2 || 0 || 0 || 1 || 0 || 0 || 0 || 0 || 0 || 0 || 0 ||||<9% style="VERTICAL-ALIGN: top; TEXT-ALIGN: center"> 1 || 0 || 0 || 0 || 1 || 0 || 0 || 0 || 0 || 0 || 0 ||||<9% style="VERTICAL-ALIGN: top; TEXT-ALIGN: center"> 2 || 0 || 0 || 0 || 0 || 1 || 0 || 0 || 0 || 0 || 0 ||||<9% style="VERTICAL-ALIGN: top; TEXT-ALIGN: center"> 3 || 0 || 0 || 0 || 0 || 0 || 1 || 0 || 0 || 0 || 0
It can be seen from this that only the multivariate procedure assumes there is a correlation of the change in score between conditions 1 and 2 with the change in score between conditions 1 and 3 across the subjects. Any discrepancy in results between univariate and multivariate results will be due to the presence of such a correlation. This indicates a violation of the [:FAQ/sphericity:sphericity assumption.]
- [:FAQ/MANOVA/manrm/sphericity:More on Sphericity]