Bayesian hierarchical modeling of people’s decision-making during an extreme weather event

Abstract

Investigating effects of a alarm type on people’s level of fear and protective actions during a tornado event can result in a more effective warning system. After collecting data on people’s responses to hypothetical tornado displays of the Probabilistic Hazard Information (PHI) warning system, we studied effects of providing some uncertainty information about the tornado threat on people’s levels of fear and protective action. There is an ongoing debate, however, on the appropriate method for analyzing these ordinal and Likert-type data. We adopted a Bayesian multivariate approach with non-informative priors to analyze the ordinal data. It has been demonstrated that the common practice of treating ordinal data as metric can lead to false conclusions, and that the Bayesian approach is more powerful than nonparametric tests for handling this kind of data. Our method finds that providing some uncertainty information about any particular tornado occurrence through PHI significantly increases people’s levels of fear and severity of the protective actions. The results also show that as the distance between the tornado and the weather warning recipient decreases, people’s levels of fear and severity of protective actions increase and vice versa.