# SPSS: discrepancy between estimated marginal means plot and post-hoc tests?

I am using a mixed-Anova with one dependent variable measured across two time points (within-subject factor). There are two between-subject factors, between1 & between2, with items equally split across their respective levels.

Results indicate a significant within-subjects effect for the DV * between1, which is comprised of three groups: A, B, C.

Post hoc tests (Tukey's and Bonferroni) indicate significant effect for group A * group B: However, plots of the estimated marginal means show a greater difference in slope steepness between group C (green) and the others, from time point 1 to time point 2: Since I am measuring change in DV over time between these three groups, and if the plots are also representing this, then it seems there is a contradiction here.

If the slope of group C is so steep, then the DV has changed to a greater extent relative to groups A and B. If that is the case, given that the groups are all balanced, why would the post-hoc tests find a significant difference between groups A and B instead?

I guess I'm conceptually confused about something here, and it's probably a simple answer. Thanks in advance.

• I see no contradiction in your results so far. "significant within-subjects effect for the DV * between1 [interaction]" corresponds to the steeper slope of one of the lines (the green) on your plot. While the fact that the blue and the red lines are far away from each other corresponds to your post hoc table. – ttnphns Nov 6 '20 at 11:41
• No, post-hoc tests are only for levels of a between-subject factor. Not for within-subject factor, nor for interactions. To get means and tests between levels of the within subj factor go to "EM means" button and select Compare main effects for the factor(s) you want. – ttnphns Nov 6 '20 at 12:29
• @remnant what you are thinking about with the change in the slope is the interaction effect. Those tables with post-hoc tests are probably for the main effect. – Sextus Empiricus Nov 7 '20 at 21:42
• @remnant I am not so good in SPSS so I can't really answer this question fully. I guess that the interaction effect should be in the output, but maybe you did not specify the interaction effect as part of the model? I do not know exactly how this works for SPSS and ANOVA. I find it more intuitive to look at it as a (mixed effects) linear model (with the subject being the random effect). – Sextus Empiricus Nov 9 '20 at 8:13
• Btw, looking at the graph, I would not be expecting that the interaction effect is gonna be significant (because the difference that the effect makes looks smaller than the difference due to the contrasts A-C and B-C, which is already not significant). .... This doesn't mean the effect is not present, it means that the effect is smaller than the noise levels of your experiment. – Sextus Empiricus Nov 9 '20 at 8:19