The less-is-more effect: Predictions and tests

In inductive inference, a strong prediction is the less-is-more effect: Less information can lead to more accuracy. For the task of inferring which one of two objects has a higher value on a numerical criterion, there exist necessary and sufficient conditions under which the effect is predicted, assuming that recognition memory is perfect. Based on a simple model of imperfect recognition memory, I derive a more general characterization of the less-is-more effect, which shows the important role of the probabilities of hits and false alarms for predicting the effect. From this characterization, it follows that the less-is-more effect can be predicted even if heuristics (enabled when little information is available) have relatively low accuracy; this result contradicts current explanations of the effect. A new effect, the below-chance less-is-more effect, is also predicted. Even though the less-is-more effect is predicted to occur frequently, its average magnitude is predicted to be small, as has been found empirically. Finally, I show that current empirical tests of less-is-more-effect predictions have methodological problems and propose a new method. I conclude by examining the assumptions of the imperfect-recognition-memory model used here and of other models in the literature, and by speculating about future research.