Category-Specific Sensitivity Allows Modeling Variability Effects in a Similarity Framework

Categorization is a core cognitive ability that enables humans to treat distinct entities as equivalent members of the same concept. Similarity-based models have been highly successful in explaining a wide range of categorization phenomena, but struggle to account for the category variability effect—the tendency to assign ambiguous items to more variable categories—because they do not explicitly represent distributional properties such as variability. We argue that explaining this effect does not require abandoning similarity-based models, but rather relaxing the assumption that model parameters are fixed across categories. Specifically, we consider two parametric extensions: allowing a bias parameter to favor high-variability categories, and allowing a sensitivity parameter to vary with each category's distributional precision. We directly compare both mechanisms in a model simulation across prototype and exemplar models, evaluating their categorization performance. Results consistently favor category-specific sensitivity over response bias, particularly for prototype models. We interpret this finding in terms of the broader cognitive science literature and extend the sensitivity account to prevalence-induced concept change, arguing that both variability and prevalence effects may reflect a common underlying mechanism in which sensitivity scales with the degree to which a category is specified. Together, these findings point towards a unified similarity-based account of distributional influences on categorization.