When Data Transformations are Appropriate or Even Necessary: A Response to Cohen-Shikora, Suh and Bugg (2019)

Abstract

Abstract In this paper, I argue that common data transformations used for statistical modelling are not inherently problematic. Depending on the research question, transformation can be appropriate or even necessary. The paper also discusses the often-overlooked impact of decision-related processes (e.g., rhythmic timing) on behaviour and how such biases can often unintentionally confound research designs. More narrowly, the current paper considers a recent debate about the list-level proportion congruent (llpc) effect, which is the finding that congruency effects (e.g., in the Stroop task) are reduced when most trials are incongruent relative to when most trials are congruent. The llpc effect is typically interpreted as evidence for conflict-driven attentional control (conflict monitoring). However, another view proposes that a rhythmic responding bias (temporal-learning) explains the effect. In a recent article, Cohen-Shikora, Suh, and Bugg (2019) challenged some of the evidence for the latter account. One key question they raise is whether it is appropriate to inverse transform (essentially: de-skew) response times when using linear mixed effect modelling. The authors argued that this transform is problematic and presented a series of analyses that they argued demonstrate both (a) that there are minimal concerns about temporal-learning confounds, and (b) that conflict monitoring clearly contributes to the llpc effect. The present article presents new analyses and demonstrates that neither of these two key conclusions of Cohen-Shikora and colleagues are justified. More global implications for linear mixed effect modelling are discussed, including an analysis of when data transformations should or should not be used.