Given the few number of sites I would used fixed effect dummies for the sites - this will ameliorate the problem of within-site dependency.

Hello Hanish Kodali. As you suggest, the GENLIN command (which can include GEE) can estimate an ordinal logit model, but not a multinomial model (i.e., model with Y = a nominal variable). I just checked the documentation for GENLINMIXED, though, and it can estimate a multinomial logit model. It does not include an option for GEE, but you could use a multilevel modeling approach to handling the correalted observations rather than GEE. (Your desire to use GEE suggests to me that you must have correlated observations--e.g., repeated measures for the same subject. If that is not so, please clarify why you want to use GEE.)

https://www.ibm.com/support/knowledgecenter/en/SSLVMB_26.0.0/statistics_reference_project_ddita/spss/advanced/syn_genlinmixed_target_options.html

Notice that in order to get a multinomial model, you must set the variable level for Y to nominal, because GENLINMIXED (as a newer command) uses the meta-data about variables when determining what kind of model to estimate.

HTH.

PS- Do you have access to any other statistical software?

Bruce Weaver Thank you for replying.

It is not repeated measures.

my dataset has unrelated randoms observations at each site we don't have any specific participants.

At site A, we saw 15 people with masks, 20 without, and 6 improper.

At site B, we saw 25 people with masks, 24 without, and 14 improper.

At site C, we saw 15 people with masks, 21 without, and 8 improper.

At site D, we saw 14 people with masks, 28 without, and 14 improper.

Yes, I do have access to R and SAS.

What do you suggest?

If each observation is independent of all others, why are you wanting to use GEE? Can you not simply do something like this? Y = the outcome variable, and N = the number of observations for each cell.

DATA LIST LIST / Site Y N (3F5.0).

BEGIN DATA

1 1 15

1 2 20

1 3 6

2 1 25

2 2 24

2 3 14

3 1 15

3 2 21

3 3 8

4 1 14

4 2 28

4 3 14

END DATA.

WEIGHT CASES BY N.

NOMREG Y BY Site {plus other options}.

Or did I misunderstand what you said?

Is the GEE being used to capture dependency at each site?

Bruce Weaver, I am trying to compensate for intra-correlation or dependency at each site as Thom Baguley said.

My sites are spread across the city. Each site population has its own characteristics and they might be intercorrelated, therefore I thought using GEE maybe a better solution.

What do you guys think now?

Given the few number of sites I would used fixed effect dummies for the sites - this will ameliorate the problem of within-site dependency.

If you need a citation supporting the dummy variable approach this might be useful:

McNeish, D., & Stapleton, L. M. (2016). Modeling clustered data with very few clusters. Multivariate Behavioral Research, 51, 495-518.

I've found in practice that GEE or multilevel models with small level 2 often produce similar results to that of the dummy variable approach for fixed effects with low level 2 n, but I think the dummy variable approach is probably more cautious.

Kelvyn Jones and Thom Baguley I appreciate your input. and it makes sense. With sample size it hard to tease out such a difference. I am going to work with this excellent paper, it has a great example. But the paper says 10 clusters of less GEE performs well.

I was wondering, what if we have 13 sites and at every site, we have different 80-90 observations. and we are trying to adjust for individual characteristics and then site-level characteristics. and our outcome is the same- multinomial 3 level outcome.

Here are some possibly useful links:

https://www.google.com/search?client=firefox-b-1-d&q=multinomial+logistic+regression+in+r

https://www.google.com/search?client=firefox-b-1-d&ei=8_WYX8TnLJOMtAa_8oiIBg&q=multinomial+logistic+regression+in+spss&oq=multinomial+logistic+regression+in+&gs_lcp=CgZwc3ktYWIQARgDMgcIABDJAxBDMgIIADICCAAyAggAMgIIADICCAAyAggAMgIIADICCAAyAggAOgQIABBHUMytCVjMrQlg7-EJaABwAngAgAFkiAFkkgEDMC4xmAEAoAEBqgEHZ3dzLXdpesgBCMABAQ&sclient=psy-ab

https://www.google.com/search?client=firefox-b-1-d&ei=lfaYX-GSD-6F9PwPoYulwA4&q=multilevel+models+in+r&oq=multilevel+modesls+&gs_lcp=CgZwc3ktYWIQARgAMgcIABDJAxANMgQIABANMgQIABANMgQIABANMgQIABANMgQIABANMgQIABANMgQILhANMgQIABANMgQIABANOgQIABBHOgcIABDJAxBDOgIIADoGCAAQFhAeOgQIABBDOgUIABCxAzoFCC4QsQM6CAgAELEDEIMBOggIABDJAxCRAjoFCAAQkQI6CAguEMcBEK8BOgQIABAKOgIILkoFCAgSATFKBQgJEgE1UO_6FVib7BdgrpYYaAVwAngAgAGRAogBvSuSAQcyOC4yNC4xmAEAoAEBqgEHZ3dzLXdpesgBCMABAQ&sclient=psy-ab

Hope these are of some assistance. Good luck, David Booth