MerQuR+Statistics Glossary

Statistics Glossary → Estimated marginal means (EMM)

Estimated marginal means (EMM)concept

Means for each level of a factor computed from the fitted model, with the other factors balanced and covariates held at a fixed value (usually their mean).

Observed (raw) means can be misleading when the groups are unbalanced on other variables. Estimated marginal means (EMMEANS in SPSS, emmeans in R, formerly called least-squares means) are produced from the fitted model's predictions: a model prediction is computed for each cell, and these cell predictions are then averaged with equal weight over the levels of the other factors. When there is a covariate, predictions are made at its overall mean; the 'adjusted means' of ANCOVA are exactly these.

In a balanced design EMMs equal the raw marginal means; in an unbalanced design raw means give more weight to large cells, while EMMs weight every cell equally. EMMs are only as meaningful as the model: when an interaction is present, the EMMs of a main effect average over the levels of the other factor and can mislead, so simple effects should be examined instead. In generalized models EMMs can be given on the link scale (e.g. log-odds) or back-transformed to the response scale (probabilities, counts).

An everyday example

A restaurant chain compares customer satisfaction at two branches. At branch A 80% of customers come at the weekend, at branch B only 20%, and weekend satisfaction is lower everywhere. On raw means A looks worse; on EMMs that balance the type of day, the two branches come out almost the same.

In MerQur

Offered as the 'Estimated marginal means' checkbox in two-way ANOVA (observed cell and marginal means are a separate option), and as 'EMMEANS factor' and 'EMMEANS p adjustment' (Bonferroni, Šidák, Tukey, No adjustment) in the linear mixed model, nested and crossed mixed models, GEE and GLMM. ANCOVA gives the adjusted (EMM) means and cell EMMs and runs its post-hoc comparisons on them; repeated measures ANOVA gives EMMs with confidence intervals in the multi-factor path, and mixed-design ANOVA has the 'EMM / simple-effect adjustment' option.

Analyses where this concept appears

Two-Way ANOVARepeated Measures ANOVAMixed-Design ANOVA, Split-PlotANCOVALinear Mixed Models (LMM)Nested Mixed Model (Nested LMM, Multi-Level Hierarchical Mixed Model)Crossed Mixed Model (Crossed LMM, Crossed Random Effects Mixed Model)Generalized Estimating Equations (GEE)Generalized Linear Mixed Models (GLMM)

Related concepts

Simple effectsPost hoc tests (pairwise comparisons)Covariate