Help with longitudinal, overdispersed, count data via PROC GLIMMIX

[PSB]

Hi all,

I posted this on a different SAS forum but was told my question was suited for this forum.

I'm relatively new to the world of SAS - I have some experience running PROC MIXED but have not needed to use SAS all too much throughout my graduate training. Here's a summary of my data that is being used for my dissertation:

- Approximately 50 youth, diagnosed with certain mental health disorders, who were a part of an intensive, 5-week treatment program

- The outcome of interest is aggression which is a continuous, non-negative, count variable and is showing a negative binomial distribution (highly positively skewed with var > mean). This outcome was measured across the 5 weeks - thus, each child has a weekly aggression score

- This outcome is showing a highly skewed distribution across the 5 weeks

- I am interested in whether certain pre-treatment affective states (e.g., irritability), measured continuously, predict changes in the intercept and slope/trajectory of aggression across the 5 treatment weeks

- Importantly, treatment did NOT differ across youth; they all received the same treatment

- Outside of a categorical covariate, the rest of my predictors are all continuous

The various resources I have read online suggest PROC GLIMMIX to be the ideal approach but the issue I am running into is that every single resource/example implies that GLIMMIX is best suited for clustered data where participants are separated into various conditions. Again, that is not the case for my data. I have scores nested within youth but that is it. 

I've consulted various resources and have piecemealed syntax that runs a converging model without errors (all predictors are continuous and have been centered hence the 'c'):

PROC GLIMMIX DATA = folder.data method=quad;
Class ID;
Model Y = cx1 cx2 cweek / s link=log dist=negbin;
Random intercept / sub = STPID;
Run;

I removed the random effect of week from the RANDOM statement as that led to the estimated G matrix being not positive definite. I'd appreciate any/all feedback. Apologies for the long post and sorry for any elementary mistakes I've made in this.

Thanks so much.

-Pev.

[StatDave · Accepted solution]

Since you want to assess whether the pre-treatment states affect the effect of time (week), then you probably want to include the interactions (cx1*cweek and cx2*cweek). If either is not significant then you can drop that interaction. Your model as it is assumes that cx1 and cx2 each affect the mean count linearly regardless of week. You could do similarly if any of the variables have higher-order effects (like quadratic, and so on).

[PSB]

Hi @StatDave ,

Thanks much for your reply, much appreciated; and thank you for the suggestions re: the interaction effects. I plan on adding those effects into later models. I have found additional resources that will let me decide which estimation procedures would be best for my model/data.

Thank you again. I feel a bit more at ease knowing that the model I presented is not totally ineffective for my data.

-Pev.