Some considerations on whether to use the GEE model - it requires a large number of clusters (schools, I assume in this case) of correlated observations and it allows you to make inferences on the population level rather than for making predictions at the individual level. Time and memory needs will increase with the number of observations in a cluster. The 0-3 response sounds like it is ordinal, so using the default link (CLOGIT, not GLOGIT) might be more appropriate as it will treat the response as ordinal. If you want to assume that the correlations within a cluster diminish over time in an autoregressive way, then you could specify the TYPE=AR option in the REPEATED statement. If you are going to do that, then you need to also specify the WITHIN= option with a variable that order the observations in the clusters - perhaps your YEAR variable - but if the years are different for different clusters, then you might want to make a variable that simply orders the observations in a common way with values 1, 2, 3, ... . Otherwise all unique years could result in the clusters being very large and make the model infeasible. Since the GEE method is robust to misspecification of the correlation structure, you could also consider using the simple independence (TYPE=IND, the default) or exchangeable (TYPE=EXCH) structure and then you don't even need to specify WITHIN=. I don't understand the need for aggregation and a FREQ statement - if schools are the clusters and your data contain multiple observations in each school with each observation representing a single unit (such as person), then you wouldn't need a FREQ statement. ... View more