Covariation detection biases in sufficient and necessary situations.

Abstract

In 4 experiments, university students played video games in which one action or cause covaried with an outcome. Judgments on sufficient and necessary causes were observed. On the basis of the obtained judgments, different computational models, Cheng and Novick's (1990a, 1992) probabilistic contrast ($\Delta$P rule) and the Rescorla-Wagner (1972) model were evaluated. In Experiments 1 and 2, for the positive contingencies, the participants judged sufficient and necessary causes differently; they also showed judgment deviations from the real contingencies. The $\Delta$P rule could not account for these data. An alternative weighted $\Delta$P rule was proposed and, along with the Rescorla-Wagner model, it successfully explained these results. In Experiment 3, negative contingencies were included. The pattern of judgements among the negative sufficient and necessary causes mirrored that of the positive contingencies but did not reach statistical significance. The $\Delta$P rule could not account for the judgments in Experiment 3, the adjusted $\Delta$P rule did not either. However, the Rescorla-Wagner model accounted for the results very well. In Experiment 4, the predictive power of these different models was compared. In general, the Rescorla-Wagner model remains the best descriptive model for explaining and predicting the patterns of contingency judgments.