Mixed ANCOVA - inconsistent results with estimated means?

[Jillian Ma]

Hi,

I am performing a 4-way 2*2*2*2 ANCOVA in SPSS. 2 of my independent variables (Interest, Difficulty) are within-subject and 2 are between subjects (Gender, Group). I have one covariate (Score).

When I run the General Linear Model, I am seeing different results for my within-subject variables under "Multivariate Tests/Within-Subject Effects" vs. under "Estimated Marginal Means". For example, for the Interest variable, the results under "Multivariate Tests" suggest there is no main effect of Interest (F = .033, p = .857). However, under "Estimated Marginal Means" , under the Interest variable, the results suggest there is a significant main effect of Interest (F = 65.895, p < .00005).

Why are these values inconsistent? I am seeing similarly very different results across the sections of the output for the Difficulty variable (the other within-subject variable), but not for Gender and Group (between-subjects). When I run the same analyses without the covariate, the F-ratios and p-values are always consistent in all places of the SPSS output. Why am I not getting consistent numbers with the covariate, and which is the accurate result?

Thank you for your help!

Jillian

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Kirill Orlov]

These two are very different testings.
You can see under the Multivariate tests table in the Estimated Marginal means section: "Each F tests the multivariate effect of.... These tests are based on the linearly independent pairwise comparisons among the estimated marginal means". That is, the omnibus test in the table is derived  after the pairwise comparisons (shown  in the just preceding table). While in the main section Multivariate tests table the omnibus test is not derived from pairwise comparisons, it is truly omnibus test preceding any pairwise testing, if any. This is first.

Second, pairwise comparisons of Estimated Marginal means (i.e. estimated population means from the complete model) isn't the same as would be pairwise comparisons of observed means (such as you could, for a between subject factor, request in the Post-hoc menu). Estimated marginal means are usually different from the observed means, and always different in the presence of a covariate (like in your case) because the estimated means are adjusted for the covariate in the sense that they are estimated at the covariate= mean level (by default).

20.07.2018 0:15, Jillian Ma пишет:

Hi,

I am performing a 4-way 2*2*2*2 ANCOVA in SPSS. 2 of my independent variables (Interest, Difficulty) are within-subject and 2 are between subjects (Gender, Group). I have one covariate (Score). 

When I run the General Linear Model, I am seeing different results for my within-subject variables under "Multivariate Tests/Within-Subject Effects" vs. under "Estimated Marginal Means". For example, for the Interest variable, the results under "Multivariate Tests" suggest there is no main effect of Interest (F = .033, p = .857). However, under "Estimated Marginal Means" , under the Interest variable, the results suggest there is a significant main effect of Interest (F = 65.895, p < .00005).

Why are these values inconsistent? I am seeing similarly very different results across the sections of the output for the Difficulty variable (the other within-subject variable), but not for Gender and Group (between-subjects). When I run the same analyses without the covariate, the F-ratios and p-values are always consistent in all places of the SPSS output. Why am I not getting consistent numbers with the covariate, and which is the accurate result?

Thank you for your help!

Jillian

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD