1
$\begingroup$

I have a question about multi-level models with multi-level survey data. I am working with survey data that has a two-stage sampling design with primary sampling units defined as schools randomly selected at the first stage and students randomly selected from schools at the second stage.

My outcome variable is student GPA. I have access to student transcript records from the initial PSU schools as well as any schools that sample members transferred to after the base survey.

Can I incorporate these "transfer schools" into the MLM or should I restrict my analysis to the PSU schools? On the one hand, I would like to take advantage of the transfer school data because it provides a bit of analytic leverage. Students who transfer from the base school are also likely to be non-random. On the other hand, I am worried that it is violating a central principal of multi-level modeling. Students are still nested within schools, of course, but the whole point of using the two-stage sample is that the base schools were randomly determined.

Improve this question
$\endgroup$
4
  • $\begingroup$ If I understand correctly, I don't think this is so much a multilevel model question but a study design question. I would run the models on both datasets, the one only including the randomized sample and the one including the transferred students, and present them in parallel manner in your article/report. You can then perhaps gain some extra information from the possible differences between the results based on random vs. non-random data? $\endgroup$
    – Sointu
    Commented May 10 at 7:28
  • $\begingroup$ Thanks! I should have been a bit more clear. The transferred students are part of the randomized sample. The case in question would be a student who was randomly selected into the sample, but transferred to another school after the base year survey. The Level 1 (students) would still represent the random sample, but the Level 2 (schools) would no longer represent the random sample of schools. I like your idea of just doing it both ways and seeing how things shape up. The only issue is that I am not sure school-level weights would be applicable when including the transfer students. $\endgroup$
    – UT_Max
    Commented May 10 at 15:07
  • $\begingroup$ Ah, that's annoying! I can't say about the weights, but I would try some version of "parallel analyses" - maybe the 1) whole sample, 2) students who remained only, and 3) students who transfered only? But, are there enough students per school among the transfered students? Or were they scattered in many different schools? Also, are the remaining students still randomized because if transfer is not random, isn't non-transfer also not random? $\endgroup$
    – Sointu
    Commented May 10 at 17:43
  • $\begingroup$ Yea I think the parallel analysis is probably my best bet unless I get a more clear answer. The transfer students are scattered but definitely not random. That’s also part of the reason I don’t want to drop them. I can include a transfer flag as a control in the model? $\endgroup$
    – UT_Max
    Commented May 11 at 14:43

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.