# MANOVA contrast results p-value in SPSS - Bonferroni correction required?

I am running a MANOVA with 2 IVs(2*3) and 4 DVs. I want to test the effect of the 3-level IV using a Helmert contrast: to compare if level 1 significantly differs from the mean of levels 2 and 3, as well as to compare if level 2 and level 3 differ significantly, on all 4 DVs. So there are 2 specific planned contrasts running on 4 DVs, constituting to 2*4=8 comparisons.

I would like to ask if the p-value shown in the output K-matrix is NOT Bonferroni corrected, such that I have to adjust the alpha value (.05 familywise) manually by dividing .05 by 8?

I assume you're running the GLM procedure (in the menus, Analyze>General Linear Model>Multivariate). If that's the case, then you are correct, the p values and confidence intervals for the contrast results are not corrected in any way for multiple comparisons or multiple dependent variables.

If you're willing to take a trip down legacy lane, it's possible to get joint multivariate confidence intervals for your contrasts from the old MANOVA procedure, which is available only using command syntax. It requires that your factor variables be numeric, with sequential integer values (such as 1,2 or 1,2,3, or 0,1, etc.). If your factors are A with two levels and B with three, and you have them coded as 1 and 2 for A and 1 to 3 for B, then with dependents named Y1, Y2, Y3, and Y4, the command might look like:

MANOVA Y1 Y2 Y3 Y4 by A(1,2) B(1,3) /CONTRAST(B)=HELMERT /CINTERVAL=JOINT MULTIVARIATE(PILLAI).

if you're fitting the contrasts averaging over the other factor, or:

MANOVA Y1 Y2 Y3 Y4 by A(1,2) B(1,3) /CONTRAST(B)=HELMERT /CINTERVAL=JOINT MULTIVARIATE(PILLAI) /DESIGN=A B.

if you're not including the interaction in the model.

There are actually five different options for the type of joint multivariate intervals, including BONFERRONI, HOTELLING, ROY, and WILKS in addition to PILLAI. I listed PILLAI in the example because I seem to recall it generally being the best option, though the existence of so many options implies none are optimal under all circumstances.

In the output, you should be able to match up parameter estimates and associated statistics from MANOVA to the contrast results in GLM, including the significance levels or p values. These are also unadjusted in MANOVA. What will differ will be the confidence interval bounds, which are adjusted both for multiple contrasts on the factor (joint) and multiple dependent variables (multivariate).

• Thank you very much! This clarifies a lot. Aug 28, 2020 at 14:09