Inferences based on uncertain data: Some experiments on the role of slope magnitude, instructions, and stimulus distribution shape on the learning of contingency relationships

Abstract

Subject sensitivity with regard to the uncertainty present in single-cue probability learning inference tasks was examined in a series of three related experiments. The first experiment varied validity sign, validity magnitude, and type of cue marginal (normal, uniform, bimodal). Bimodal distributions tended to be learned most quickly, particularly at low validities, followed by uniform, followed by normal. Surprisingly, however, uniform conditions tended to have higher asymptotic residual variance than normal or bimodal conditions. In the second experiment, special instructions on the nature of uncertainty relationships were given to subjects. Although performance did improve, subjects did not maximize their behavior. Experiment III consisted of two transfer paradigms, from a bimodal to normal distribution and vice versa. Results again indicated that bimodal distributions tended to result in quicker learning of the basic relation (slope), providing greater absolute transfer to test trials than normal distributions. However, relative (true) transfer affects were highest for normal distribution groups. Arguments for the dependence of learning rate on SceSx evolved from these findings.