HUMAN PROCESSING OF EQUIVOCAL INFORMATION
MICHIGAN UNIV ANN ARBOR ENGINEERING PSYCHOLOGY LAB
Pagination or Media Count:
This report contains a series of studies investigating the abilities of subjects to revise probability estimates on the basis of new information. These studies show that subjects probability estimates are reliable, but deviate considerably from posterior probabilities calculated from Bayess theorem. These deviations are almost always in the conservative direction, i.e., low Bayesian probabilities are overestimated, and high ones are underestimated. Only when each datum is very ambiguous do subjects estimates become more extreme than Bayesian probabilities. Further, when subjects are asked to give 90 or 50 credible intervals of a posterior probability distribution, their estimates are wider than Bayesian credible intervals. This finding of conservatism has led to the design of a man-computer system that should minimize the effects of human shortcomings in making diagnoses.
- Administration and Management
- Statistics and Probability