Outliers in L2 Research in Applied Linguistics: A Synthesis and Data Re-Analysis

Data from self-paced reading (SPR) tasks are routinely checked for statistical outliers (Marsden, Thompson, & Plonsky, 2018). Such data points can be handled in a variety of ways (e.g., trimming, data transformation), each of which may influence study results in a different manner. This two-phase study sought, first, to systematically review outlier handling techniques found in studies that involve SPR and, second, to re-analyze raw data from SPR tasks to understand the impact of those techniques. Toward these ends, in Phase I, a sample of 104 studies that employed SPR tasks was collected and coded for different outlier treatments. As found in Marsden et al. (2018), wide variability was observed across the sample in terms of selection of time and standard deviation (SD)-based boundaries for determining what constitutes a legitimate reading time (RT). In Phase II, the raw data from the SPR studies in Phase I were requested from the authors. Nineteen usable datasets were obtained and re-analyzed using data transformations, SD boundaries, trimming, and winsorizing, in order to test their relative effectiveness for normalizing SPR reaction time data. The results suggested that, in the vast majority of cases, logarithmic transformation circumvented the need for SD boundaries, which blindly eliminate or alter potentially legitimate data. The results also indicated that choice of SD boundary had little influence on the data and revealed no meaningful difference between trimming and winsorizing, implying that blindly removing data from SPR analyses might be unnecessary. Suggestions are provided for future research involving SPR data and the handling of outliers in second language (L2) research more generally.