What have we learned about learned categorical perception? A meta-analysis of 30 years of research

Much empirical research over the last thirty years suggests that learning to categorize a set of objects alters the way in which they are perceived or judged, a phenomenon known as learned categorical perception or LCP. LCP has provided key evidence for controversial claims of top-down effects of knowledge on perception and suggested a possible mechanism by which continuous perceptual input is transformed into discrete symbolic representations. LCP may consist of increased cross-category discrimination ("expansion"), decreased within-category discrimination ("compression"), or both, and some studies fail to show either type of boundary effect. To address these inconsistencies and determine the general robustness of LCP, we conducted separate meta-analyses of expansion and compression on a total of 70 experiments comprising 4,627 participants We found a moderate average effect size for expansion (d = .48) with strong evidence of publication bias, an average effect size of approximately 0 for compression with likely publication bias, and highly significant heterogeneity of effects across experiments for both expansion and compression. Moderator analyses revealed a significant effect of the task used to assess LCP, with same-different judgments tending to produce results consistent with a non-boundary effect of improved discrimination on the category-relevant dimension. However, the sensitivity of the moderator analyses was limited by low statistical power and most of the heterogeneity of effect sizes across studies remained unexplained. We discuss the implications of these results for our understanding of LCP, including suggestions for how future LCP research could be more informative and better support meta-analysis.