Co-Twin Control Models: Assessing Bias from Measured and Unmeasured Confounders

Between-within models are increasingly used to add causal evidence to observational data. In particular, co-twin control models, a subset of between-within models, make use of discordant twin pairs to evaluate whether the... [ view full abstract ]

Between-within models are increasingly used to add causal evidence to observational data. In particular, co-twin control models, a subset of between-within models, make use of discordant twin pairs to evaluate whether the relationship between an exposure and outcome is consistent with a causal interpretation. While the popularity of these models is increasing, they have not been fully explored methodologically. Prior research has shown that bias can be introduced into model estimates through unmeasured confounders that are not shared within a twin pair¹. The amount of bias is a function of the correlation in exposure and the correlation in the unmeasured confounder within twin pairs. The current study makes use of simulation and analytic proofs to explore whether this bias can be reduced by the inclusion of a measured confounder. It can be shown that both the crude and within-pair estimates in a model including a measured confounder do indeed reduce the bias introduced by a non-shared confounder within twin pairs. The reduction in bias is a function of the individual-level correlation between the unmeasured and measured confounder. Interpretation of the within-pair effect, as an estimate of the causal effect of exposure on outcome, in co-twin control models is discussed.

¹Frisell T, Öberg S, Kuja-Halkola R, & Sjölander A. (2012). Sibling comparison designs: Bias from non-shared confounders and measurement error. Epidemiology, 23, 713–720.