False Positives and Other Statistical Errors in Standard Analyses of Eye Movements in Reading

TL;DR

This paper demonstrates through simulations that neglecting multiple comparison corrections in eye movement reading studies leads to excessive false positives, and advocates for standard correction procedures like Bonferroni to improve statistical validity.

Contribution

It provides the first formal simulation-based analysis of false positive rates in eye movement research and advocates for standard correction methods to control errors.

Findings

False positives increase significantly without corrections.

Bonferroni correction effectively controls false positives.

Statistical power can be improved to avoid illusions.

Abstract

In research on eye movements in reading, it is common to analyze a number of canonical dependent measures to study how the effects of a manipulation unfold over time. Although this gives rise to the well-known multiple comparisons problem, i.e. an inflated probability that the null hypothesis is incorrectly rejected (Type I error), it is accepted standard practice not to apply any correction procedures. Instead, there appears to be a widespread belief that corrections are not necessary because the increase in false positives is too small to matter. To our knowledge, no formal argument has ever been presented to justify this assumption. Here, we report a computational investigation of this issue using Monte Carlo simulations. Our results show that, contrary to conventional wisdom, false positives are increased to unacceptable levels when no corrections are applied. Our simulations also show that counter-measures like the Bonferroni correction keep false positives in check while reducing statistical power only moderately. Hence, there is little reason why such corrections should not be made a standard requirement. Further, we discuss three statistical illusions that can arise when statistical power is low, and we show how power can be improved to prevent these illusions. In sum, our work renders a detailed picture of the various types of statistical errors than can occur in studies of reading behavior and we provide concrete guidance about how these errors can be avoided.