Mixed models
Jonathan Dushoff, Ben Bolker
1 March 2026
Introduction
Mixed models are models which combine random and fixed effects.
- Fixed effects are effects whose parameters we wish to estimate
- The slope of a phosphorus response
- The difference in shield thickness between males and females
- Random effects are effects that we wish to account for without
estimating individually
- Neuron identity
- Village of residence
Modern and classic approaches
The modern approach to mixed models involves estimating the variances of random effects and finding the conditional modes for each group (see below).
Compared to the classic ANOVA approach to mixed models, the modern approach is:
- conceptually clear
- flexible
- powerful
- computationally difficult
Example
Testing the effects of acid rain on spruce tree growth.
We have a complex manipulation (what kind of air the trees are exposed to). As a result:
- We have a small number of trees
- We take a lot of samples from each tree
Model
Treatment is a fixed effect
- We want to estimate the difference in growth with clean air and dirty air
Tree is a random effect
- We want to know what about the distribution of tree effects (so we can control for it); in this case we are not specifically interested in the difference in growth between (e.g.) tree 3 and tree 5
Random effects
Random effects are based on unordered factors (grouping variables)
- the levels of the factor are conceptualized as random samples from a larger population
- the effect of each level is therefore a random variable
- the essential parameters we estimate are not the effect of each level, but the variance of the distribution.
Is this a random effect?
Treating something as a random effect means treating the levels as interchangeable from the point of view of your scientific hypothesis.
There aren’t hard-and-fast rule about when you should model a predictor as a random effect. Here are some criteria …
philosophical questions
- are the levels chosen from a larger population?
- are the levels chosen randomly?
- do you want to be able to make predictions about new (unobserved)
levels?
- or inferences that include them?
- are you interested in the distribution of levels/variability among levels?
- are you uninterested in testing hypotheses about specific levels?
Types of analysis
- multilevel or hierarchical models
- each identifiable level could have its own random effect: e.g., country, village, household
- repeated measures
- individual identity can be associated with a random effect
- this can also be done by modeling correlation in the residuals
inferential questions
Choosing to use a random effect will affect your inferences
- Using a fixed effect means we are making inferences about the levels we have measured (only)
- Using a random effect means we are making inferences about a population represented by the levels we have measured
Examples
Influenza vaccination experiment
- Year as a fixed effect: did vaccination help on average over these years (with a particular set of flu strains, etc.)
- Year as random effect: will vaccination help on average over a wider set of years (assuming observed years are representative)
Spruce trees
- May not have enough data to do a mixed model
- Our statistical inference is then limited to the areas we studied
- Whether we can broaden our scientific inference is a scientific question (not statistical)
Practical questions
have you measured a sufficient number of levels to estimate a variance (>5, preferably >10)?
May need more for more complicated models
Otherwise
- Limit your inference (e.g. switch to fixed effects)
- Try an even more advanced approach (usually Bayesian)
Random slopes
- Are only intercepts affected by random factors, or also effects of
some predictors?
- e.g., spruce trees
- Random slopes are often conceptually appealing, but computationally
challenging
- There is a (questionable) culture of not worrying about them
- Conceptually we are thinking about interactions between a fixed effect (e.g., level of acidity) and a grouping variable (spruce trees)
Fitting
Modern mixed-model packages (see below) can fit a wide variety of models. You just need to specify which effects are random.
- Works with unbalanced designs
- Works with crossed random effects
- May require a lot of data/a lot of levels for a reliable fit
- Depends on model complexity
Too few levels
If you have something that should be a random-effect predictor, but you don’t have enough levels, you can’t fit a modern mixed model
It’s OK to treat your random effect as a fixed effect, as long as this is properly reflected in your scientific conclusions.
- The scope of your analysis covers only the sampled levels, not the population they were sampled from
Residual structure and group structure
Standard random effects: we attach a random effect to the group
- Sometimes called G-side modeling
- Group membership is binary (same or different)
- Hierarchy is possible
- You and I are in the same country, but not the same village
- But the model doesn’t understand if our villages are close together
R-side modeling: we impose a covariance structure on the residuals
- Arbitrary variance-covariance structure (time, space, etc.)
- I am very close to Ben, and sort of close to you
- Can also allow for heteroscedasticity
- Disadvantages
- Harder to combine effects
- Easier to mis-specify the model
- Difficult to implement (especially for generalized models)
How modern methods work
It’s complicated!
How modern methods work
Typically based on marginal likelihood: probability of observing outcomes integrated over different possible values of the random effects.
Balance (dispersion of RE around 0) with (dispersion of data conditional on RE)
Shrinkage: estimated values get “shrunk” toward the overall mean, especially in small-sample/extreme units
How do we do it?
Different for linear mixed models (LMMs: normally distributed response) and generalized linear mixed models (GLMMs: binomial, Poisson, etc.)
- LMMs: REML vs ML
- analogy: dividing by \(n-1\) when estimating variance
- analogy: paired \(t\)-test
- REML is the natural extension to >2 treatment levels per block
- GLMMs: PQL, Laplace, Gauss-Hermite
Or:
- Bayesian approach: put a combined prior on the parameters
Practical details in R
- Classical designs
aov + Error()term
- Modern methods
lme(nlmepackage): older, better documented, more stable, does R-side models, complex variance structures, gives denominator df/\(p\) values(g)lmer(lme4package): newer, faster, does crossed random effects, GLMMsglmmTMB: more families for GLMMs; zero-inflation etc etc- many others (see task view)