When you Should and Should Not use Mixed Models

Mixed-effects models are among the most important and influential statistical developments of the last decades. One key for their success are software tools that make it easy for users to apply them to their data. Because of the availability of these tools and pressure from reviewers and editors, many researchers feel the need to incorporate mixed models into their methodological arsenal. In my talk I provide a conceptual overview of mixed models and provide some recommendation of when and when not to use these tools. In short, I recommend that mixed models should only be used when they provide a clear benefit over well-established procedures such as repeated-measures ANOVA. The list of cases in which they provide a benefit are: Continuous within-subjects independent variables; Crossed random effect (i.e., simultaneously accounting for effect of participants and items); Nested random effects (e.g., participants tested in groups/classes); Mixed between-within factors (each participant sees subset of factor levels, but not all levels); Missing values in within-subjects factors; Within-subjects factors with many levels, multiple observations per participant and level, and substantial violations of sphericity.

Sponsorship of an event does not constitute institutional endorsement of external speakers or views presented.