When can relative risks provide causal estimates?

TL;DR

This paper discusses when relative risks from epidemiological studies, especially with rare diseases, can be interpreted causally using causal inference criteria like backdoor and frontdoor, highlighting conditions for valid causal estimates.

Contribution

It applies causal inference criteria to epidemiological models, clarifying when and how relative risks can be used for causal inference in rare disease studies.

Findings

Conventional proportional hazards models can often estimate causal associations for rare diseases.

Backdoor criteria help identify when causal estimates are valid.

Frontdoor criteria can be used even with unmeasured confounders, yielding similar results to measured confounder analyses.

Abstract

It is emphasised that for epidemiological studies where disease incidence is rare, results from conventional proportional hazards models can often correctly estimate causal associations. The well-known "backdoor criteria" from causal-inference is applied to the common epidemiological study of rare diseases with a proportional hazards model, providing an example of when and how estimates from conventional proportional hazards studies can be used. A similar study with the "frontdoor criteria", that allows studies with unmeasured confounders, finds similar results to conventional mediation analysis with measured confounders. Reasons for this are discussed.