UCLA

This dissertation studies the effects and elicitation of beliefs. In the first chapter we estimate the effects of macroeconomic expectations on consumer decisions. We examine this question using an experiment with 2,872 credit card customers at a large commercial bank. We provide participants with expert forecasts of inflation and the nominal exchange rate and measure the consumption response to this information using detailed data on individual credit card transactions. We find that forecasts strongly affect inflation and exchange rate expectations, but do not change spending or self-reported consumption plans as predicted by standard models of intertemporal choice. Results from a supplementary survey experiment suggest that consumers are sophisticated enough to anticipate nominal rigidities that lower expected real income and reduce spending on durables for precautionary reasons, counteracting the effects predicted by standard models of intertemporal optimization. The absence of a link between consumer expectations and behavior has potentially important implications for macroeconomic policies such as forward guidance.

The counter-intuitive results motivate development of more flexible method of belief elicitation, which could facilitate understanding of the subjects' decision making. While the elicitation of numerical variables, such as inflation, is well understood, certain beliefs, such as action plans during an inflation hike, cannot be represented numerically and require verbal elicitation. Verbal elicitation forces the researcher either to use open-ended questions or know the most important answers in advance. We propose a method to crowdsource potential answers to open-ended questions. This ensures that the survey is adaptive and does not miss important answers while maintaining a low-cost, multiple-choice format. We propose two measures of information loss to evaluate the quality of multiple-choice questions: the semantic similarity of selected answers to the open-ended answers to the same question and the probability of selecting any answer from the list. We conduct an experiment to examine the impact of monetary incentives and characteristics of respondents on the quality of crowdsourced answers. Our findings show that incentives can increase quality and effort, but the effects are relatively small compared to the variation across respondents. We find that option authors’ similarity to the respondents in beliefs, political views, and demographics, explains a large fraction of the variation in quality. These results imply that sample selection is likely to be more important for crowdsourcing hypotheses than incentive design.

The third chapter charts a research agenda extending the work on the verbal elicitation techniques. In particular, we focus on the potential of large language models (LLMs) to improve belief elicitation techniques with natural language. We review the current state of belief elicitation in economics. Next, we introduce the main architectures and relevant fine-tuning techniques for LLMs. Lastly, we discuss the potential applications of LLMs to belief elicitation and the use of beliefs in empirical work. This includes using LLMs for representing beliefs numerically, interacting with subjects during belief elicitation, and generating hypotheses regarding how beliefs affect actions.