standard variance components structure for GLMM

[Marc-Flo]

Hello everyone.

I'm currently working on a GLMM for a repeated measures design whith a binary outcome.

I use PROC GLIMMIX to do so and I've compared different covariance structures using information criteria.

The standard variance components structure (TYPE=VC) gives me the lowest BIC so far. Since I'm working on repeated measures, I'm a bit doubtful about how I should interpret this.

Does this mean that the different measures for a same individual are absolutely not correlated ? Therefore would a simple logistic regression gives me the same results ?

Thank you by advance.

Marc

[SteveDenham]

Could you share code?  Of interest to me would be the candidate covariance structures, and the number of repeated time points.  From that you should be able to see why VC minimized the information criteria.

SteveDenham

[Marc-Flo]

Hello Steve and thank you for your answer.

Here's the code I use:

PROC GLIMMIX DATA=suivi_file METHOD=laplace IC=q NOCLPRINT;
CLASS id_pat profil sexe prescr ass_so tble_psy;
EFFECT spl_tps=SPLINE(temps / DEGREE=1 KNOTMETHOD=LIST(70) DETAILS);
MODEL bilan(EVENT='1') = spl_tps|profil age sexe tble_psy ass_so prescr / DIST = BINARY LINK=LOGIT S INTERCEPT;
RANDOM INTERCEPT temps / SUB=id_pat TYPE=VC;
RUN;

I should precise that the number of repeated time points is variable between subject and varies from 1 to 18. Also, time points are not evenly spaced in time.

So far, my best two candidates as covariance structure are VC (AICC = 39401.8) and ANTE(1) (AICC=39402.5). Considering the low difference between these two results I might be tempted to choose the ante-dependence structure which makes more sense in a repeated measures situation.

Thank you for your future response.

[SteveDenham]

Have you considered spatial power, SP(POW)(index), as a possible structure. You would need to create a continuous variable equal to temps (call it t, for example), and then splitting your random statement into two statements, where you do not look at the covariance between the intercept and the slope.  That is:

PROC GLIMMIX DATA=suivi_file METHOD=laplace IC=q NOCLPRINT;
CLASS id_pat profil sexe prescr ass_so tble_psy;
EFFECT spl_tps=SPLINE(temps / DEGREE=1 KNOTMETHOD=LIST(70) DETAILS);
MODEL bilan(EVENT='1') = profil spl_tps*profil age sexe tble_psy ass_so prescr / DIST = BINARY LINK=LOGIT S INTERCEPT;
RANDOM INTERCEPT  / SUB=id_pat ;
RANDOM temps / sub=id_pat type=sp(pow)(t);
RUN;

Note that the spline term only appears as an interaction with profil, and not as a stand-alone term.  I got this idea from the example here .  The example goes on to show how to compare differences at various points along the spline.

With the split of the random statement into separate intercepts and slopes, you could also look at:

RANDOM temps / sub=id_pat type=csh; (remove the random intercept if you try this)

SteveDenham