mixed model repeated measures

[Tommaso Salvitti]

Good morning to everybody!!
Is it correct according to you what I wrote?

For each continuous variable, data collected at the scheduled visits were analyzed by linear mixed models including as fixed factors the time (18m, 24m), group of diagnosis , and their interaction (time* group of diagnosis). The reference levels were “18m” for time, and “typical” for group of diagnosis. Interaction terms between groups of diagnosis and time were included to determine if groups of diagnosis led to different trajectories over time. Random intercepts for participants accounted for the dependence of repeated measures. Un unstructured covariance matrix was assumed. P-values for the overall effects of factors and their interactions were obtained. Contrasts were assessed to test differences, specifically among groups by time, and among times by group. Mixed models can accommodate unbalanced data patterns and use all available observations and children/subjects in the analysis. Inferences are valid under the assumption that missing observations are “missing at random”.

I use this syntax

xtmixed VARIABLEA Diagnosi_numbis##timepoint || ID:, covariance(unstructured) reml

Thanks to everybody
Tommaso

[Joseph Coveney]

Is it correct according to you what I wrote?

I don't know about all of the rest, but using the covariance(unstructured) option for the random effects equation with only a random intercept as you show will have given you a warning message:

note: single-variable random-effects specification in ID equation; covariance structure set to identity.

You must've seen it, right? So, at least the "Un unstructured covariance matrix was assumed." can't be correct.

As an aside, you might be better off fitting a MANOVA model in order to account for the temporal correlation and potential for heteroscedastic residual errors. With missing data you cannot do that with manova, but you can consider something like the following, which will accommodate at least missing-at-random or better.

xtmixed VARIABLEA i.Diagnosi_numbis##i.timepoint || ID: , ///
    noconstant residuals(unstructured, t(timepoint)) ///
    reml dfmethod(kroger)

contrast Diagnosi_numbis timepoint Diagnosi_numbis#timepoint, small

And if you're interested in whether "groups of diagnosis led to different trajectories over time", then consider also the following.

margins Diagnosi_numbis#timepoint, df(`e(df_avg)')
marginsplot

[Tommaso Salvitti]

thank you
i'll try the code but STATA say " dfmethod() is available only with mixed "
SO I CHANGE xtmixed with mixed?

[Joseph Coveney]

SO I CHANGE xtmixed with mixed?

Yes. Sorry, I copied your code in order to avoid having to type the cumbersome variable names and didn't notice that you were using obsolete command names.