I ran a GLS random effects regression on some NBA data in Stata, and I was told that it was wrong because I didn't use mixed effects model. This may every well be the case, but I was quite confused by the explanation. Here is what I did using Stata:

  • Dependent variable is team win's and independent variables is different types of opponent shooting data.

  • Paneled data using xtset to account for the different NBA seasons

  • Ran robust xtreg for both fixed and random effects
  • Hausman test had a high chi^2, so I stayed with the GLS random effects regression.

If this is incorrect, please let me know so I can fixed this. I always thought that the mixed effects model was for logistic regressions when my dependent variable was categorical. I was told that I did not take into account this team grouping structure within your analysis, but the Group Variable was the teams.

  • $\begingroup$ Some people think that a mixed effects model is always the right approach. It might be, but I don't see why it should be in this case. However, it's best to try many different models and compare the results, rather than just one. $\endgroup$
    – Paul
    Jan 18, 2017 at 16:04
  • $\begingroup$ @Paul thank you so much for this answer. I do have a followup question if you do not mind. In my xtreg, I need to add i.year to take into account the different NBA seasons? $\endgroup$
    – user28062
    Jan 18, 2017 at 16:18
  • $\begingroup$ Is this a duplicate of datascience.stackexchange.com/questions/16387/… ? $\endgroup$
    – Spacedman
    Jan 18, 2017 at 16:26
  • $\begingroup$ @Spacedman it is because I am desperate for an answer. I will remove it now. My issue may be that I forgot to do something, but I am still unsure. $\endgroup$
    – user28062
    Jan 18, 2017 at 16:28

1 Answer 1


I always thought that the mixed effects model was for logistic regressions when my dependent variable was categorical

This is not true at all. We use mixed effects, in particular random intercepts when we have clustered data, such as repeated measures on individuals, or sampling within clusters such as schools, or regions, or countries. It is certainly not the only approach - instead we can use Generalised Esitmating Equations, or just fit fixed effects for the grouping variable, but when there are a large number of groups, and we can justifiably consider their "effect" to be normally distributed and uncorrelated with the residual error or the fixed effects, a mixed effects model is often very attractive.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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