Consequences of power transforms in linear mixed-effects models of chronometric data

Psycholinguists regularly use power transforms in linear mixed-effects models of chronometric data to avoid violating the statistical assumption of residual normality. Here, I extend previous demonstrations of the consequences of using power transforms to a wide range of sample sizes by analyzing word recognition megastudies and performing Monte Carlo simulations. Analyses of the megastudies revealed that stronger power transforms were associated with greater changes in fixed-effect test statistics and random-effect correlation patterns as they appear in the raw scale. The simulations reinforced these findings and revealed that models fit to transformed data tended to be more powerful in detecting main effects in smaller samples but were less powerful in detecting interactions as sample sizes increased. These results suggest that the decision to use a power transform should be motivated by hypotheses about how the predictors relate to chronometric data instead of the desire to meet the normality assumption.Master of Art