Predicting Risk Perception: New Insights from Data Science

Abstract

We outline computational techniques for predicting perceptions of risk. Our approach uses the structure of word distribution in natural language data to uncover rich representations for a very large set of naturalistic risk sources. With the application of standard machine learning techniques, we are able to accurately map these representations onto participant risk ratings. Unlike existing methods in risk perception research, our approach does not require any specialized participant data and is capable of generalizing its learned mappings to make quantitative predictions for novel (out-of-sample) risks. Our approach is also able to quantify the strength of association between risk sources and a very large set of words and concepts and, thus, can be used to identify the cognitive and affective factors with the strongest relationship with risk perception and behavior. This paper was accepted by Elke Weber, accounting.