Written and maintained by CASRAI Editorial Board
Last updated
GLMM in research stands for generalized linear mixed model, a regression model for non-normal outcomes (binary, count, proportion) whose data are grouped or repeated, such as patients within doctors or repeated measures on the same subject. This page covers the statistical term only; searches for “glmm” in games, anime or social media slang are a different subject. After reading you should be able to decide whether a GLMM fits your data, interpret its coefficients correctly, and report it.
What does GLMM mean? Definition in one paragraph
A generalized linear mixed model combines two ideas. From a generalized linear model it takes a link function and a non-normal outcome distribution (for example logit with a binomial outcome, or log with a Poisson count). From a mixed-effects model it takes random effects that capture variation between groups (clusters). Given the random effects, the outcome follows an exponential-family distribution, and the random effects themselves are assumed to be normally distributed (Wikipedia, “Generalized linear mixed model”, read 2026-11-07).
Formula with every symbol explained
The standard form is g(E[y | u]) = Xβ + Zu.
- y: the outcome vector (for example remission yes/no, tumour count).
- u: the random effects (for example one random intercept per doctor), assumed normal with mean 0 and an estimated variance.
- g: the link function: logit for binary outcomes, log for counts, identity for normally distributed continuous outcomes (UCLA OARC, read 2026-11-07).
- X, β: the design matrix and coefficients of the fixed effects (the predictors whose effects you want to estimate).
- Z: the design matrix for the random effects.
Because the likelihood requires integrating over the random effects, p(y) = ∫ p(y | u) p(u) du, it has no general closed form. Software therefore approximates it: quasi-likelihood (fast but biased and not preferred for final inference), Gauss-Hermite quadrature (a single quadrature point equals the Laplace approximation; accuracy rises with more points but cost grows quickly with the number of random effects), or MCMC. No approximation performs well for every model and dataset, and ungrouped binary data are particularly difficult (Wikipedia; UCLA OARC).
Worked example with real numbers: from log-odds to subject-specific probabilities
UCLA OARC’s introduction uses a simulated dataset of 8,525 patients treated by 407 doctors and fits a mixed logistic model for cancer remission with a random intercept for doctor. The reported estimates are below; every conversion in the last two columns is computed here from those estimates.
| Term | Estimate (log-odds) | SE | Odds ratio = exp(estimate) |
|---|---|---|---|
| Intercept | 1.467 | 0.274 | exp(1.467) = 4.34 |
| Age (per year) | -0.056 | 0.005 | exp(-0.056) = 0.946 |
| Married (vs not) | 0.26 | 0.064 | exp(0.26) = 1.297 |
| IL-6 (per unit) | -0.053 | 0.011 | exp(-0.053) = 0.948 |
Step 1, odds ratios: each extra year of age multiplies the odds of remission by 0.946, about a 5.4% reduction, holding doctor and the other predictors fixed. Ten extra years multiply the odds by exp(-0.56) = 0.571.
Step 2, the random-intercept variance is 3.979, so the standard deviation of doctor effects is √3.979 = 1.995 log-odds. For a reference patient (all other predictors at the values that give the intercept), the probability of remission is p = 1 / (1 + exp(-η)):
| Doctor | η = 1.467 + u | Probability of remission |
|---|---|---|
| One SD below average (u = -1.995) | -0.528 | 0.371 |
| Average doctor (u = 0) | 1.467 | 0.813 |
| One SD above average (u = +1.995) | 3.462 | 0.970 |
The same fixed effects therefore imply remission probabilities from 0.37 to 0.97 depending on the doctor, which is the between-cluster heterogeneity a plain logistic regression would ignore. The same source’s Poisson model for tumour counts (random-intercept variance 0.169) uses a log link, so its coefficients are read as rate ratios: for example, a coefficient of 0.026 for age gives exp(0.026) = 1.026, a 2.6% higher expected count per year.
Interpretation: conditional versus marginal, and what the numbers mean
GLMM coefficients and odds ratios are conditional: they compare two subjects who share the same random effect (the same doctor, the same participant). They are not population-averaged effects. Marginal effects require averaging over the random-effects distribution, and most software does not report them by default (UCLA OARC). If you want population-averaged effects, generalized estimating equations target them directly.
| Quantity | How to read it |
|---|---|
| Fixed-effect estimate (logit link) | Change in log-odds per unit, within the same cluster; exponentiate for an odds ratio. |
| Fixed-effect estimate (log link) | Change in log expected count per unit; exponentiate for a rate ratio. |
| Random-intercept variance | Spread of cluster-specific baselines on the link scale; its square root is the SD shown in the worked example. |
| Intercept | Log-odds (or log-rate) for a reference subject in an average cluster (u = 0). |
GLMM compared with related models
| Model | Outcome | Clustered data? | Effects are |
|---|---|---|---|
| Linear mixed model | Continuous, normal errors | Yes (random effects) | Conditional |
| Generalized linear model | Binary, count, other exponential-family | No | Single-level |
| GLMM | Binary, count, other exponential-family | Yes (random effects) | Conditional |
| GEE | Binary, count, other | Yes (working correlation) | Marginal |
To compare nested GLMMs (for example with and without a random slope), see likelihood ratio tests for nested models, and for random slope structure see random intercepts versus random slopes.
Assumptions and common mistakes
- Too few random-effect levels. Guidance quoted in Ben Bolker’s GLMM FAQ (read 2026-11-07): more than 1,000 levels give accurate variance estimates, more than 100 are reasonable, 10 to 20 need careful validation, and fewer than 10 are safer treated as a fixed effect. See also Oberpriller, Leite and Pichler (2021).
- Ignoring overdispersion. For count and binomial data, compare the sum of squared Pearson residuals with the residual degrees of freedom (an approximate check). It does not apply to models that estimate a scale parameter and cannot be estimated for Bernoulli data. Remedies include an observation-level random effect, a negative binomial or beta-binomial family, or a quasi adjustment (Bolker FAQ).
- Relying on quasi-likelihood. Penalized quasi-likelihood is fast but biased, and the Bolker FAQ notes it is deemed unreliable in lme4 and no longer available there.
- Reading conditional estimates as population averages (see the interpretation section).
- Normal random effects are an assumption, not a result; inspect simulated-residual diagnostics, for example with the DHARMa R package.
How to report a GLMM (APA-style example)
Report the family and link, the random-effects structure, the estimation method and software, the number of groups, fixed effects with uncertainty, and the variance components. An illustrative template (numbers from the worked example above):
We fitted a generalized linear mixed model with a binomial family and logit link, with a random intercept for doctor (407 doctors, 8,525 patients), estimated by maximum likelihood with adaptive quadrature. Each additional year of age was associated with lower odds of remission (OR = 0.946, b = -0.056, SE = 0.005). The random-intercept variance was 3.979 (SD = 1.995). Overdispersion was checked and [result].
Checklist: family; link; random-effects terms; estimation method (Laplace, quadrature points, quasi-likelihood); package and version; number of levels per random factor; fixed effects with SE or confidence intervals; variance components; overdispersion check; statement that estimates are conditional.
How to fit a GLMM in R or Python
Statistical packages that fit GLMMs include R (lme4, glmmTMB, glmm, MASS::glmmPQL, and the Bayesian MCMCglmm and brms), SAS, SPSS, MATLAB (fitglme), Julia (MixedModels.jl) and Python statsmodels for binomial and Poisson outcomes (Wikipedia; Bolker FAQ). Microsoft Excel is not among the packages these sources list.
R, binomial outcome with a random intercept per doctor (lme4):
library(lme4)
m <- glmer(remission ~ Age + Married + IL6 + (1 | DID),
data = hdp, family = binomial)
summary(m)
exp(fixef(m)) # odds ratios (conditional)R, count outcome with a log link:
m2 <- glmer(tumorsize_count ~ Age + Married + IL6 + (1 | DID),
data = hdp, family = poisson)Python (statsmodels, variational Bayes fit, binomial):
from statsmodels.genmod.bayes_mixed_glm import BinomialBayesMixedGLM
vc = {"doctor": "0 + C(DID)"}
m = BinomialBayesMixedGLM.from_formula("remission ~ Age + Married + IL6", vc, df)
res = m.fit_vb()
print(res.summary())The variable names (remission, DID, hdp, df) are placeholders for your own data.
FAQ
What does GLMM stand for?
In statistics and research methods, GLMM stands for generalized linear mixed model. Other meanings of “glmm” in gaming or slang contexts are unrelated to this page.
What is a GLMM used for?
It models non-normal outcomes such as binary events and counts when observations are clustered or repeated, for example patients within doctors, students within schools, or repeated measurements on one participant. The random effects absorb between-cluster variation so standard errors are not too small.
What is the difference between a GLMM and a linear mixed model?
A linear mixed model assumes a continuous outcome with normal errors. A GLMM adds a link function and a non-normal family, which is why its likelihood needs approximation (quadrature, Laplace, quasi-likelihood or MCMC).
What are GLMM random effects?
Random effects are cluster-specific deviations (intercepts, optionally slopes) assumed to come from a normal distribution with mean zero; the model estimates their variance. See mixed-effects models for the fixed versus random distinction.
Are GLMM odds ratios population averages?
No. They are conditional on the random effect. Population-averaged odds ratios need marginalisation or a GEE model.
Is logistic regression a GLMM?
Ordinary logistic regression is a generalized linear model with no random effects. A mixed logistic model with a random intercept is a GLMM.
Sources
- Wikipedia, “Generalized linear mixed model”, en.wikipedia.org/wiki/Generalized_linear_mixed_model (read 2026-11-07).
- UCLA OARC Statistical Consulting, “Introduction to Generalized Linear Mixed Models”, stats.oarc.ucla.edu (read 2026-11-07).
- Bolker, B. et al., “GLMM FAQ”, bbolker.github.io/mixedmodels-misc/glmmFAQ.html (read 2026-11-07).
- Breslow, N. E., & Clayton, D. G. (1993). Approximate inference in generalized linear mixed models. Journal of the American Statistical Association, 88(421), 9-25.
- Stroup, W. W. (2012). Generalized Linear Mixed Models. CRC Press.
- Oberpriller, J., Leite, M. de, & Pichler, M. (2021), on random-effect levels, as cited in the Bolker GLMM FAQ.
