Week 9
Kou Murayama — Professor, University of Tübingen
Mixed-effects modelling: Overrated strengths and underrecognized potential
Mixed-effects modelling has been widely used in psychology and neuroscience to analyze clustered data. It is used so frequently that it almost becomes a default option when data are clustered. However, I first argue that mixed-effects modelling can be substituted by simpler approaches in some situations (e.g., summary-statistics approach) and we need to understand when it is really needed. Understanding the equivalence with simpler approaches promotes our conceptual understanding of the model. I then argue that some real values of mixed-effects modelling are, on the other hand, often overlooked or not recognized. More specifically, when the data are crossed, mixed-effects modelling has significant advantages over other simpler approaches, but for several common types of data in psychology and neuroscience, researchers conventionally do not apply mixed-effects modelling with crossed random effects, jeopardizing the accuracy of statistical inference.
In his presidential address from 1957, Lee Cronbach distinguished between the correlational and the experimental traditions of psychological science. This distinction highlights why we as psychologists sometimes talk past each other, why there are separate research communities that target similar phenomena but never interact, and why psychology does not rely on one but two separate methodological toolboxes. Cronbach’s presidential address is also a plea for uniting the discipline of psychology. In my talk, I will discuss some of the (methodological) gaps that remain in the effort for a united discipline. I will draw on the example of cognitive tasks that are developed in the experimental tradition but are used to understand individual differences. One of the potential paths for a united methodology is hierarchical modeling. This approach can be used to disentangle sample noise from true individual variability in repeated measures designs common for cognitive tasks. Finally, I will present results that highlight both the promises and pitfalls of a united discipline.