Statistics Glossary → Expected mean squares (EMS)
Expected mean squares (EMS)concept
The expected value of each mean square in an ANOVA table, expressed in terms of variance components and fixed effects; it determines the correct F denominator and the variance component estimates.
In random- or mixed-effects designs, each mean square carries several sources of variance at once. In a two-way crossed design with A and B random and n observations per cell, for example, E(MS_A) = σ²_e + nσ²_AB + nbσ²_A. To test an effect, the denominator is the mean square that has the same expected value as the numerator under the null hypothesis; here the denominator for A is MS_AB, not MS_error.
Solving the EMS equations gives method-of-moments (ANOVA-method) estimates of the variance components; these can come out negative and are then usually set to zero. Mixed models have two traditions: the unrestricted model (the SPSS and SAS default) and the restricted (Cornfield-Tukey) model, which imposes sum-to-zero constraints on the interaction effects; they lead to different denominators for the random main effect. In unbalanced designs the EMS are only approximate and methods such as REML are preferred.
An everyday example
In a factory, 5 randomly chosen workers operate 4 randomly chosen machines, and each worker-machine pair produces 3 parts. If some workers are especially good or bad on certain machines (a worker × machine interaction), part of the differences between workers actually comes from that interaction. The EMS show that the worker effect should be tested against the interaction mean square, not the error term.
In MerQur
The 'EMS model (mixed design)' option of the Crossed Mixed Model chooses between the unrestricted (default; SPSS/SAS) and restricted (Cornfield-Tukey) models; the Nested Mixed Model builds its F tests from the EMS; and the ANOVA method of Variance Components Analysis uses them.
Analyses where this concept appears
Nested Mixed Model (Nested LMM, Multi-Level Hierarchical Mixed Model)Crossed Mixed Model (Crossed LMM, Crossed Random Effects Mixed Model)Variance Components Analysis (VARCOMP)