Agent‐based Models and Causal Inference

Abstract

This chapter explains the ideal conditions under which agent-based models (ABM) can contribute to causal inference from a production point of view on causality and a vertical perspective on mechanism; describes the practical limitations hampering the use of ABMs for causal inference on this mechanistic ground. It discusses from-within-the-method solutions to circumvent such limitations. "Theoretically informed" ABMs constitute a first improvement. The "input realism" of an ABM can be increased through "empirical calibration". The major difficulty with empirical calibration and validation of an ABM comes from data availability. The ABM's granularity and the requirement of empirical embeddedness to increase ABMs' input and output realism push in opposite directions. Sensitivity analysis can be performed in different ways. If sensitivity analysis focuses on the values of an ABM's parameters, robustness analysis is directed to the ABM's "internal" cogs and wheels and aims at assessing how varying these cogs and wheels impacts on the ABM's outcomes.