Statistical methods for twin and sibling designs

Abstract: Twin and sibling studies are valuable in that they allow adjustment for potential confounding factors that are impossible or hard to measure. By measuring associations ‘within-cluster’ it is possible to adjust for many factors that are shared between individuals in the same cluster. Using Swedish national registers, it is possible to obtain information about a large number of potential confounders. While this gives medical researchers great opportunities to control for confounding, it also increases the risk of model misspecification leading to biased estimates. One strategy to reduce the risk of such bias is to use doubly robust(DR) estimation. In DR estimation two working models are combined in such a way that the resulting estimate will remain asymptotically unbiased when one of the models is misspecified. In study I, we implement existing DR estimators for parameters in linear, log-linear and logistic regression models in the R package drgee. In study II, we propose a new class of DR estimators for ‘within-cluster’ association measures in linear and log-linear regression models. In study III we propose a DR estimator for the ‘within-cluster’ log odds ratio parameter in logistic regression models. The estimators proposed in studies II and III are also implemented in the R package drgee. In study IV, we discuss what shared factors the ‘within-cluster’ association actually is adjusted for. Using the formal theory of causal diagrams we demonstrate that the standard methods for estimating ‘within-cluster’ association parameters implicitly adjust for shared confounders, shared mediators, but not shared colliders. Therefore, the estimated parameter may have a causal interpretation as a direct effect, i.e. as the part of the causal effect that is not mediated through shared factors.