Paradoxical parsimony: How latent complexity favors theory simplicity

Investigating how people evaluate more or less complex causaltheories has been a focal point of research. However, previ-ous studies have either focused on token-level causation or re-stricted themselves to very small sets of explanatory variables.We provide a new approach for modeling theory selection thatforegrounds the balance between observed and latent structurein the mechanism being explained. We combine a Bayesianframework with program induction, allowing an unboundedand partially observable model space through sampling, andreflecting how a preference for simplicity emerges naturallyin this setting. Through simulation, we identify two rationalprinciples: (1) Simpler explanations should be favored as la-tent uncertainty (the number of hidden variables) increases; (2)latent structure is attributed a larger role when the observablepatterns become less compressible. We conducted a behavioralexperiment and found that human judgments tended to reflectthese principles, indicating that people are sensitive to latentuncertainty when selecting between explanations.