Behavioral Causal Inference

TL;DR

This paper models how decision makers infer causal effects from observational data, highlighting the impact of control variable choices on inference accuracy and welfare loss, with bounds depending on decision maker types.

Contribution

It introduces a formal model of lay decision makers learning causal effects from data, analyzing welfare loss bounds based on control variable choices and type structures.

Findings

Bounds on welfare loss due to incorrect causal inference

Control variable selection critically affects inference accuracy

Ordered type structures reduce costs of wrong inferences

Abstract

When inferring the causal effect of one variable on another from correlational data, a common practice by professional researchers as well as lay decision makers is to control for some set of exogenous confounding variables. Choosing an inappropriate set of control variables can lead to erroneous causal inferences. This paper presents a model of lay decision makers who use long-run observational data to learn the causal effect of their actions on a payoff-relevant outcome. Different types of decision makers use different sets of control variables. I obtain upper bounds on the equilibrium welfare loss due to wrong causal inferences, for various families of data-generating processes. The bounds depend on the structure of the type space. When types are "ordered" in a certain sense, the equilibrium condition greatly reduces the cost of wrong causal inference due to poor controls.