Repeated Measures Weighting

[JonKetchup]

I am doing a log-binomial analysis with repeated measures per subject. When I use the "repeated" statement and create a nested structure per record, is each nest weighted equally or does it depend on the number of repeated measures within the nest? For example, if record1 has 1 observation, record2 has 5 observations, and record3 has 10 observations, is record 3 weighted 10x that of record1, or are all 3 records weighted equally in the model?

proc genmod data = DATA;
	class RECORD TREATMENT VAR1 VAR2 VAR3
		 / PARAM=REFERENCE;
	model OUTCOME(event='1')= TREATMENT VAR1 VAR2 VAR3
			/ dist=bin link=log type3 AGGREGATE ;
	REPEATED SUBJECT=RECORD  / TYPE=IND CORRW;
estimate 'Treatment' TREATMENT 1 /exp e;
run;

[sbxkoenk]

I moved your post to the 'Statistical Procedures' board.

You get a higher probability for a response over here (with your PROC GENMOD question).

Koen

[JonKetchup]

Thank you. I appreciate it.

[MichaelL_SAS · Accepted solution]

When you specify a repeated statement PROC GENMOD fits a GEE model, and it might be helpful to refer to the PROC GENMOD documentation section on Generalized Estimating Equations (GEEs). I don’t think this type of model nicely lends itself to a notion of clusters being “weighted” according to their size.

If you look at the estimating equations, you don’t multiply the score function for a cluster by any term that depends on the cluster size. That said, the score function for a cluster i can be written as a sum of the residual from each observation in a cluster times a column of the D_i’V_i^-1 matrix (using the notation from documentation). In that sense larger clusters contribute “more” in the sense that there are more terms in that summation, but again I don’t think it is appropriate to say it is weighted more heavily.

[JonKetchup]

Thank you, that is very helpful.