My firet guess would be that in Model 3 has been constructed in the basis of more observations than the other models. It might thus be that Models 1 and 2 are underpowered.

Thank you Sven Greving; Indeed this totally makes sense! I was hesitant with particularly model 1 but for model 2 I have to include for the variables that were repeatedly measured at 3 and 6 months only.

Nivine Hanach If the true associations are strong, however, then explicitly modeling any association structure usually leads to more efficient GEE estimators compared with assuming independence. In this case the observed effects on the relative efficiency may be weak if specific versions of invariant covariate designs are considered or may be amplified if the number of observations within each cluster varies. The efficiency loss relative to ML estimators will often be small.

Good luck.

Thank you Pradeep Paraman for your answer. It is highly appreciated. Based on your answer, you mean that if the significant association in model 3 is strong enough then the GEE estimator would still be significant either in model 1 or 2? And perhaps the change in the number of observations within each model may have lead to efficiency loss?

Nivine Hanach Since the study can be said to be LS in nature and is time-varying prognostic factor, using the MRM, to test individual level effects of time-varying prognostic factors, from 3, 6 and 12 months before each visit. Here you identified that , the time-varying predictors tested did influence the DASm trajectory? It is therefore very clear that assuming that these time-varying predictors been significant, their effects would have been to shift the trajectory up or down.

So if the primary interest of the study is to understand how covariates impact disease course, based on the individual patient’s prognosis trajectory, the MRM is the best choice. The MRM is also more suitable when there is significant heterogeneity in individuals’ disease courses in the population.

You might want to argue that although the results of the MRM may appear much like the GEE in continuous outcomes models, but the interpretations are very different. The MRM models the average of individuals’ trajectories (subject specific) but the GEE models the trajectory from DASm averages at each time point (population average).

Hope this clarifies.

Good luck.