Causal interpretation of the sibling comparison and its relation to the cross-over design

TL;DR

This paper critically examines the causal interpretation of sibling comparison designs, linking them to cross-over designs, and proposes weighting methods to address confounding and interaction issues for valid causal inference.

Contribution

It provides a formal characterization of sibling comparisons as cross-over designs and introduces weighting techniques to improve causal effect estimation under certain assumptions.

Findings

Sibling comparison can be viewed as a cross-over design.

Weighting methods help address carry-over effects and interactions.

Simulations and examples demonstrate the approach's potential benefits.

Abstract

The intuitive motivation for employing a sibling comparison design is to adjust for confounding that is constant within families. Such confounding can be caused by variables that otherwise might prove difficult to measure, for example factors relating to genetics, environment, and upbringing. Recent methodological investigations have shown that despite its intuitive appeal, the conventionally employed analysis does not relate to a well-defined causal target, even in the case of constant confounding. A main challenge is that the analysis will target the subpopulation of exposure discordant pairs. In the presence of an effect of the cosibling's exposure on the sibling's outcome, there is a second challenge, namely that the effect corresponds to an intervention that always exposes the cosibling to the opposite exposure from the sibling. We characterise the sibling comparison in terms of the cross-over design. Estimands of interest are discussed before using this characterisation to establish more natural conditions for targeting an appropriate causal parameter. We cast the above-mentioned challenges of the sibling comparison in terms of those facing the cross-over trialist: in order to target an appropriate estimand one must be able to argue the absence of a certain type of carry-over effect as well as absence of trial-by-treatment interaction, thus establishing that the former study design emulates the latter warts and all. We explore weighting to counter the effects of such interactions and to target other estimands. The weights rely on estimates of the unobserved confounding structure. Through simulations and an example analysis, we illustrate its potential usefulness to assess the validity of the assumptions of the matched analysis. We briefly discuss an extension of the weighting procedure to remove selection bias based on data from a population-level reference sample.