I'm trying to predict that speed at which people complete a walking test, where they perform this test for multiple trials and overall increase their performance on each trial. They perform as many as they can, in a set amount of time, and come back on three sessions to perform it. There are two groups, intervention and control, and I hypothesize a difference in the intervention compared to controls. I'm aiming to predict their completion time on multiple days based on the continuous score of a test from day 1 that they do prior to the walking task, this is their proprioceptive sensitivity score.
I'm using nlme:
Completion_Time ~ Session * Group * Sensitivity + Age + Sex, random =~ 1 | Subject, data = Primary_analsys, na.action = omit)
Ultimately, I find that there are significant 3 way interactions:
Now, I used the effects package, and I identified different slopes for each group at the first time point (Control slopes negatively, higher sensitivity associated with worse performance; Intervention slopes positively, higher sensitivity associated with better performance).
But in general, I follow up with a single session model, assessing S1, S2, and S3 individually, looking for a Group*Sensitivity interaction. Here, I find nothing for any of the sessions, no main effects, interactions etc.
So, what gives? Am I capturing a change that occurs between sessions in the overall model? Is there anyway for my to stick with the 3 way interaction model and report those effects in a statistically sound approach? I started to use emmeans based off a similar post:
emt <- emtrends(Proprio_Predict, ~First_Sess_Proprio*Group, var = 'Session') pairs(emt)
And this identified a significant p-value, but I assume that its basically verifying that the two groups differ in their slopes (which the 3 way interaction model shows).
Sorry if this is rambling, I'm just trying to verify if there is another approach to get at what I'm looking for, or if my post-hoc analysis really is the best way, and there simply is not the effect present.
