Research

Microeconomic Theory

A Welfare Analysis of a Steady-State Model of Observational Learning with Margaret Meyer (working paper)

Abstract It is now well understood that when individuals learn not only from private sources of information but also by observing the choices of previous decision-makers, information may fail to be aggregated, and incorrect herds may result. Extrapolating from such findings, the conventional wisdom seems to be that, relative to equilibrium outcomes, efficiency would be improved if individuals relied more on their private information and less on the choices of predecessors. The canonical models of observational learning are, however, not particularly well suited to welfare analysis. In this paper, we develop a flexible, steady-state framework for characterizing equilibrium outcomes under a variety of observation structures, and we use it to examine the existence and form of welfare-improving interventions for a social planner. The steady-state approach transforms the planner’s problem from a dynamic one to an effectively static one. We allow the planner to intervene only by altering the decision rules used by individuals, not by directly altering the information structure. Our key results are as follows: We show that whether there exist welfare-improving adjustments for the planner and what form they take depends very much on the observation structure. In particular, we prove that there is no scope for the planner to improve welfare in a symmetric environment where each individual observes (with noise) the choice of only a single predecessor. For more general observation structures, we identify two distinct forces which can generate inefficiencies; we term these the "dispersion effect" and the "mean effect". We demonstrate that both forces can, for some observed data, make it optimal for individuals to place less rather than more weight on their private information. 

Analysing electoral reform: the case of Hungary (2017 version)

Abstract While accusations of gerrymandering are common, it is usually hard to establish robust empirical evidence. In this paper, I use data on the Hungarian electoral reform of 2011 to look for gerrymandering by Fidesz, the incumbent party. I develop an empirical test that draws on the Hungarian context and gerrymandering strategies of ‘cracking’, ‘packing’ and ‘slicing’ suggested by the theoretical literature. First, my analysis gives clear empirical confirmation of strategically manipulated district boundaries. Second, I find evidence of so-called `cracking' (Owen and Grofman 1988), with Fidesz moving its supporters away from districts where it already enjoyed large majorities, but not that of packing, which points to suboptimal partisan redistricting. Suboptimality might be explained by cognitive and practical constraints prohibiting the implementation of more involved redistricting strategies.

Artificial Intelligence

Can Artificial Intelligence Solve Strategic Decision Problems? with Peter Eso, Edmund Kelly and James Tilley

Abstract Whether, when, and how Artificial Intelligence (AI) will substitute for human labor are all active debates. This paper evaluates the ability of several leading Large Language Models (LLMs) to solve strategic decision-making problems from the social sciences. On average, we find that LLMs perform similarly to highly competent humans, but the type of problem affects solution rates. LLMs perform better when faced with problems that require numerical calculations of expected values. They perform worse when given tasks that are: a) not based on well-known ‘textbook’ problems, b) involve complex equilibrium reasoning (such commitment problems) or c) require novel economic insights.