Publication bias and the failure of replication in experimental psychology

Abstract

Replication of empirical findings plays a fundamental role in science. Among experimental psychologists, successful replication enhances belief in a finding, while a failure to replicate is often interpreted to mean that one of the experiments is flawed. This view is wrong. Because experimental psychology uses statistics, empirical findings should appear with predictable probabilities. In a misguided effort to demonstrate successful replication of empirical findings and avoid failures to replicate, experimental psychologists sometimes report too many positive results. Rather than strengthen confidence in an effect, too much successful replication actually indicates publication bias, which invalidates entire sets of experimental findings. Researchers cannot judge the validity of a set of biased experiments because the experiment set may consist entirely of type I errors. This article shows how an investigation of the effect sizes from reported experiments can test for publication bias by looking for too much successful replication. Simulated experiments demonstrate that the publication bias test is able to discriminate biased experiment sets from unbiased experiment sets, but it is conservative about reporting bias. The test is then applied to several studies of prominent phenomena that highlight how publication bias contaminates some findings in experimental psychology. Additional simulated experiments demonstrate that using Bayesian methods of data analysis can reduce (and in some cases, eliminate) the occurrence of publication bias. Such methods should be part of a systematic process to remove publication bias from experimental psychology and reinstate the important role of replication as a final arbiter of scientific findings.