Preference Learning for Policy Analysis

Abstract

In policy analysis, individual preferences are used to measure welfare across the population. Nonetheless, individuals are heterogeneous in both what they want or their preference content, and whether they know what they want or their preference structure. Prior work typically restricts preference heterogeneity analysis to differences in preference content. This dissertation explores the intersection of public policy analysis with preference heterogeneity along these two dimensions. We present a general framework for analyzing and discoveringpreference content and structure from choice data. Our framework extends welfare measurementto fully account for preference heterogeneity and can help to better understand the welfare impacts of new policies for sub-populations. As heterogeneity in preference canbe related to judgment structure, we first study how heterogeneity in preference content is related to heuristic judgment. We establish the relationship between judgment and choice for cumulative flood risks. Second, we propose a model that can directly determine differencesin both preference content and structure for individual decision makers empirically using graph matching methods. Finally, we measure heterogeneity in preference content across sub-populations. We develop and test a method to uncover relevant sub-populations in a choice model automatically using machine learning tools. We illustrate the approach discovering relevant socioeconomic covariates in a recent and real decision facing the Chilean government about the environmental impacts of electricity generation. Our framework can help to design policy interventions tailored for the heterogeneous preferences of the public.