Conservatism in a simple probability inference task.

TLDR

Three experiments examined how prior probabilities, data quantity, diagnostic value, payoffs, and response modes influence posterior probability estimates. Participants displayed conservative probability updating that was insensitive to priors, stable across data amounts, and reduced with lower diagnostic value; learning and variance improved under payoff conditions, and estimates were closest to Bayesian when using linear payoffs or logarithmic odds, though log payoffs still reduced conservatism compared to quadratic payoffs.

Abstract

3 experiments investigated the effects on posterior probability estimates of: (1) prior probabilities, amount of data, and diagnostic impact of the data; (2) payoffs; and (3) response modes. Ss usually behaved conservatively, i.e., the difference between their prior and posterior probability estimates was less than that prescribed by Bayes' theorem. Conservatism was unaffected by prior probabilities, remained constant as the amount of data increased, and decreased as the diagnostic value of each datum decreased. More learning occurred under payoff than under nonpayoff conditions and between-S variance was less under payoff conditions. Estimates were most nearly Bayesian under the (formally inappropriate) linear payoff, but considerable overestimation resulted; the log payoff condition yielded less conservatism than the quadratic payoff. Estimates were most nearly Bayesian when Ss estimated odds on a logarithmic scale.