Collaborative Research: Causal Structures: Experiments and Machine Learning

New York University, New York NY

Investigators

Abstract

To make decisions, people must rely on their understanding of the relevant environment: what are the causes and outcomes of the various forces at play. In other words, in many settings, including economic ones, people rely on subjective causal models (or narratives) to understand the world. Such models help agents organize and interpret information, allowing them to make forecasts about the future, and providing them with a way to evaluate counterfactuals. The main goal of this research is to take a first step towards understanding how economic agents come to adopt (possibly incorrect) models and how this depends on the information available to them. The researchers will approach this topic from two different perspectives. The first involves a series of experiments that aim to understand how people’s subjective models arise from patterns they identify in data. Some experiments will be conducted in an abstract setting, while others involve natural context. Natural context can trigger preconceptions about how different variables are associated with each other that may help or hinder people from correctly identifying actual patterns in a set of observations. The second approach aims to better understand whether news media plays a role in heterogeneous subjective models. The goal is to study whether different news outlets organize and explain the same outcomes using different causal models. A growing literature in economic theory studies ramifications of adopting possibly incorrect subjective models, referring to economic agents relying on such models as ‘misspecified.’ But, for the most part, the literature is silent on how a person comes to adopt a subjective model to begin with, how such a subjective model may depend on the setting, and how it may be shaped by the person’s experiences. In addition, it is an open question under what conditions people adopt subjective models that are consistent with the true data generating process. The goal of this research is to take a first step towards understanding how such misspecifications may arise and how they depend on features of the data-generating process. The researchers will approach the topic from two different perspectives. A first approach involves a series of laboratory experiments to understand how people extract patterns from their observations. The novel experimental design asks subjects to organize different sets of observations (data) with the goal of making predictions in similar situations. The experimental data will let the researchers understand whether the predictions subjects make in each environment are consistent with them using some model that posits specific statistical relationships between different variables. Complemented with ancillary non-choice data that emerges as a by-product of the experimental design, the results will provide insights into how people form models of the world by studying data and how they use these models to make predictions. Experiments will be conducted both with an abstract setting and with context. Understanding how people come to adopt (possibly incorrect) models and how this is impacted by the information available to them is important to determine in what situations they are more vulnerable to being manipulated. Furthermore, it can help us design policies that are effective in correcting beliefs and inducing optimal behavior. The second approach aims to better understand whether news media plays a role in shaping heterogeneous subjective models. The goal is to study whether different news outlets organize and explain the same outcomes using different causal models. To do so, the researchers will use an end-to-end trained Machine Learning pipeline that will take text (news articles) as input and identify the main causal statements advanced in this text as output. Documenting the heterogeneous causal models propagated by news outlets is important for understanding why voters with different political affiliation disagree on the optimal response to problems that are accepted by both sides. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.