Is score data as count vs continuous

[lboyd]

I've done a bit of reading and gathered I need to run a Poisson regression. I transformed the data into long and ran this:
proc genmod data=long;
class time eid;
model score= time / dist = poisson
link = log;
repeated subject=eid / corr=exch corrw;
run;

The repeated subject line takes into account the paired data, correct? Do I include time in the class statement because I've seen it both ways? I used exchanged since I saw that in an example, but I know there is much more to it.

Thanks!

[StatDave · Accepted solution]

If you have just two repeated observations per subject, then all correlation structures other than CORR=IND are the same since there is only a single correlation parameter to estimate. You would specify TIME in the CLASS statement if you want the model to allow for any association of time with the response. If you leave it out of the CLASS statement, then the model is restricted to time having a linear effect. Yes, using the REPEATED statement uses a Generalized Estimating Equations (GEE) analysis with accounts for the correlation within subjects. 

[lboyd]

If I have three repeated observations per subject, what CORR can I use?

[StatDave]

You can use any of the available correlation structures. Choose the one that matches the pattern of correlations that you suspect exist among your repeated measures. The exchangeable (TYPE=EXCH) structure assumes that all pairs of repeated measures have the same correlation. The unstructured correlation matrix (TYPE=UN) allows all pairs of repeated measures to have different correlations. This is the most general structure but also requires the most correlations to be estimated and can therefore result in model fitting errors. Other structures (like TYPE=AR) allow for degradation of the correlation for pairs of repeated measures farther apart (such as when measures are done over time). 

For more, see this note.