Prediction interval: What to expect when you're expecting . . . A replication

Abstract

A challenge when interpreting replications is determining whether the results of a replication "successfully" replicate the original study. Looking for consistency between two studies is challenging because individual studies are susceptible to many sources of error that can cause study results to deviate from each other and the population effect in unpredictable directions and magnitudes. In the current paper, we derive methods to compute a prediction interval, a range of results that can be expected in a replication due to chance (i.e., sampling error), for means and commonly used indexes of effect size: correlations and dvalues. The prediction interval is calculable based on objective study characteristics (i.e., effect size of the original study and sample sizes of the original study and planned replication) even when sample sizes across studies are unequal. The prediction interval provides an a priori method for assessing if the difference between an original and replication result is consistent with what can be expected due to sample error alone. We provide opensource software tools that allow researchers, reviewers, replicators, and editors to easily calculate prediction intervals.