A regression model for continuous responses that combines fixed effects for population-level predictors with random effects to model correlated observations (for example, repeated measures or hierarchical sampling), assuming a linear relationship and typically Gaussian residuals conditional on random effects.

Principle

Modeling between-cluster variability with random intercepts and/or slopes permits correct partitioning of variance and yields valid inference about fixed effects and subject-specific trajectories when assumptions of linearity and conditional normality hold.

Demonstration

Illustrative scenario — Situation: Body weight measured monthly for pigs in several pens. Recognition: Measurements are repeated within the same pig and pigs are nested in pens. Action: Fit an LMM with fixed effects for time and diet, random intercepts and slopes for pig, and a random intercept for pen if needed. Consequence: Estimated time-by-diet interaction describes average growth while random effects capture individual deviations and provide improved predictions for each pig.

Misapplication

Using an LMM for markedly non-normal outcomes (e.g., counts, bounded proportions) without transformation or an appropriate link violates the model assumptions; additionally, including complex random-effects structures with too few clusters or observations per cluster leads to unstable variance estimates.

Consequence

Appropriate LMM use produces unbiased estimates of fixed effects, correct standard errors, and credible subject-level predictions; misuse can yield biased coefficients, incorrect uncertainty quantification, and poor prediction for individuals or clusters.

Reversal

When residuals are strongly non-normal or variance is heteroscedastic, alternatives such as transformation, generalized mixed models, or robust/resampling methods may be required; likewise, with very small numbers of higher-level units the random-effects approach may be unreliable and fixed-effects for those units may be preferable.

Boundary

Clearly within: continuous, approximately Gaussian outcomes measured repeatedly on animals with nested structure. Boundary case: continuous outcomes with heavy tails where mixed models with robust estimation are considered. Clearly outside: discrete outcomes (binary, counts) that require generalized mixed models.

Semantic Tension

Tension between modeling individual trajectories (random-effects focus) and estimating population-average effects (marginal models); the same LMM can be used for both but interpretation differs and must be stated.

Synthesis

LMMs formalize the decomposition of variability into within- and between-subject components and support subject-specific inference; ensuring linearity, conditional normality, and adequate cluster sizes is essential to make subject-level predictions and population inferences coherent.