3
$\begingroup$

I am looking at how some individual-level psychological variables relate to an organizational-level dependent variable. In this case, the dependent variable is the ranking of the organization (ordinal variable).

I understand that I need to use ordinal logistic regression. I came across R’s polr package mentioned here. But I am not sure if polr can account for the nested nature of my data: individuals nested in teams, and teams nested in organizations.

Is there a way to model the following: y_ordinal ~ ind_psych1+ ind_psych2 + (1 + team_id/org_id)?

Many thanks!

$\endgroup$
3
  • 1
    $\begingroup$ Have you checked using the lme4 package with the function glmer? Your question seems to arise from not knowing that nested data is modeled through multilevel / mixed models or am I missing something? $\endgroup$
    – Kuku
    Commented Apr 18, 2020 at 17:51
  • $\begingroup$ Is not lme4 for linear mixed effects only? My dependent variable is ordinal. $\endgroup$
    – SanMelkote
    Commented Apr 18, 2020 at 18:29
  • 1
    $\begingroup$ The glmer function is for generalized linear mixed models, so it should be able to fit your model (although I haven't used such function). $\endgroup$
    – Kuku
    Commented Apr 18, 2020 at 18:33

1 Answer 1

5
$\begingroup$

You have many options for modeling ordinal outcome data when your data structure is multilevel. Among the options are the clmm2 (cumulative link mixed models) function within the ordinal package. This package fits proportional odds cumulative logit models, which assume that the effect of x is the same for each cumulative odds ratio.

The proportional odds assumption can be relaxed with a non-proportional odds approach, which is available in the mixor package. The default in mixor is to fit proportional odds cumulative logit models, but with the KG= option, you tell mixor how many of your predictors you would like to relax this assumption for. You can then run a likelihood ratio test using the anova command on the two models (proportional and non-proportional odds), which are nested models.

If you are interested in Bayesian approaches, you can use brms, which uses Stan to fit a literal cornucopia of ordinal multilevel models. See this great introductory and how-to article from B$\ddot{u}$rkner & Vuorre.

$\endgroup$

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.