Observational Causality Testing

TL;DR

This paper advances observational causality testing by developing empirical bounds, adapting methods for finite populations, and integrating covariate data, thereby enhancing practical applicability and robustness.

Contribution

It introduces new methods for estimating randomness bounds, extends causality testing to finite populations, and incorporates covariate data, along with a novel covariate selection criterion.

Findings

Empirical bounds for randomness can be estimated from twin studies.

Methodology is adaptable to finite populations with corrections.

Incorporating covariates improves causal inference accuracy.

Abstract

In prior work we have introduced an asymptotic threshold of sufficient randomness for causal inference from observational data. In this paper we extend that prior work in three main ways. First, we show how to empirically estimate a lower bound for the randomness from measures of concordance transported from studies of monozygotic twins. Second, we generalize our methodology for application on a finite population and we introduce methods to implement finite population corrections. Third, we generalize our methodology in another direction by incorporating measured covariate data into the analysis. The first extension represents a proof of concept that observational causality testing is possible. The second and third extensions help to make observational causality testing more practical. As a theoretical and indirect consequence of the third extension we formulate and introduce a novel criterion for covariate selection. We demonstrate our proposed methodology for observational causality testing with numerous example applications.