Extracting Models From Data Sets: An Experiment

About this event

We experimentally study the types of mental models people form by learning from a set of observations. Specifically, we study the kinds of inferences individuals make regarding the statistical relationships among variables from examining data. Remarkably, participants identify the correct correlation structure 61% of the time, despite having almost no information about the data-generating process. The remaining 39% of the cases correspond to two recurring errors: The first error, observed 13% of the time, involves misidentifying the underlying correlation structure and occurs most often in the presence of confounding variables, rising to 28% in such cases. The second error, the most common overall (observed 26% of the time), involves a failure to identify any correlations. On the one hand, the first error results in small losses in the most common occurrences; on the other hand, the second error is associated with randomizing behavior, creates substantial losses, and indicates significant difficulties in learning from the data. Importantly, participants display a high degree of consistency in the types of mistakes they make. Hence, different subjects draw opposing conclusions on what constitutes optimal behavior when presented with identical information and those differences are predictable.

This event will take place in MAR 6.33, 6th floor, The Marshall Building, 44 Lincoln's Inn Fields, London WC2A 3LY.

The building is labelled MAR on the map.