Causal Inference in Experimental and Observational Methods

Abstract

This chapter aims to demonstrate that observational and experimental methods, contrary to widespread perceptions well illustrated by Diez Roux's statement, similarly to agent-based models (ABMs), have to face equally demanding challenges when employed to establish "horizontal" causal claims, i.e. claims about robust probabilistic or counterfactual dependences. It argues that, in fact, the use of statistical models for causal inference is itself more controversial than it may seem at first. The chapter systematically scrutinizes randomized control trials, instrumental variables, and directed acyclic graphs more specifically with respect to assumptions that cannot be tested in practice because of data availability, assumptions that are untestable in principle because they would require infinite knowledge capabilities, and how reliability is established in practice within these methods. It discusses putative fundamental differences between ABMs and experimental/observational methods.