4  Generalized Linear Mixed Models

Just as how generalized linear models are an extension of linear regression, generalized linear mixed models are an extension of linear mixed models. In general, analysis and interpretation proceeds in a logical fashion from GLM’s and mixed models.

We’ll briefly list some of the mixed models, and any quirks to be discussed about them. This section may be expanded in the future.

4.1 Logistic Mixed Model

There are actually two commands for logistic mixed models: melogit and meqrlogit. The former is faster, but the latter is more likely to converge. Both commands function generally identically. Note that meqrlogit is a somewhat outdated command, so it’s possible that newer features to melogit may no longer work with meqrlogit.

Separation remains a major concern amongst fixed effects, but of lesser concern amongst random intercepts (e.g. a household where everyone had a positive response would break if included as a fixed effect, but generally would run as a random intercept). The only concern is that separation in random effects can make convergence harder to achieve.

4.2 Poisson Mixed Model

Poisson mixed models can be run with the mepoisson command. A meqrpoisson command exists and has benefits just like meqrlogit, but again, is an outdated command. If over-dispersion is an issue, menbreg exists for negative binomial regression.

4.3 Ordinal Logistic Regression

These models can be run with meologit.

4.4 Multinomial Logistic Regression

This can only be fit via sem.