Identifying levels (HLM) in a longitudinal dataset

[Elc Estrera]

I have several years of data for each student, and each student attended a specific school and is part of a different cohort. It looks like the data should be treated like this: students (level 1) are nested within schools (level 2), which are nested within cohorts (level 3). However, should the data instead be treated like this: students (level 1) are nested within cohorts (level 2), which are nested within schools (level 3)? Is there a reason to use one set-up over the other? I think I could obtain estimates under each way and see if they are the same. But if they are different, I wouldn't know which set-up to pick.

[Clyde Schechter]

They are very different, and either approach might be right, or both might be wrong, depending on the way your data were gathered.

If all of the students at a given school are part of the same cohort, then this is students nested within schools nested within cohorts.

If all of the students in a given cohort attended the same school, then this is students nested within cohorts within schools.

If, as would be more usual in educational research, students from multiple cohorts occur at each school, then this is not simply nested data, rather you have crossed cohort and school effects. As for how to represent this in an HLM analysis, if you have only a few cohorts (resp. only a few schools), it might make sense to treat cohort (resp. school) as a fixed effect. But if you have many cohorts and many schools (again, what is more commonly seen in educational research) probably you need to treat them as crossed random effects. There is a good explanation, with a simple example, of how to construct the syntax for this design in the [ME] model at page 393 (I'm assuming Stata version 14). If you have an earlier version of Stata, look for the example on crossed random effects in the -mixed- section of [ME].