Method developments for the attributable fraction in causal inference

Abstract

In public health and policy making, understanding the overall impact of an intervention is of essential importance. A way to quantify the disease burden due to some risk factor is by the attributable fraction (AF). The AF is a measure of the proportion of some disease that could be prevented if all would have been unexposed to the risk factor of interest. From the definition of the AF, it is a causal parameter and in order to achieve a causal interpretation of the AF estimate, we have to tackle the challenges of estimating causal effects in observational data. One of them is the problem of confounding, which may cause the researcher to confuse a spurious correlation with a causal effect.

In this work, we stress the importance of using model-based adjustment to estimate the AF and develop novel methods for AF estimation. In project I we implemented methods for AF estimation for cross-sectional, case-control (matched and unmatched) and cohort study designs in the statistical software R by the package AF. The package serves as a platform for the novel methods of AF estimation developed in project II-IV. While project I focuses on estimation methods for the AF that rely on the fact that all confounders, sufficient for confounding control, are measured, researchers often face the problem with unmeasured confounding. In some situations, we may have access to clusters that share these unmeasured confounders. Thus, clustered data can be used to adjust for cluster-shared unmeasured confounding. In project II we develop a method that enables estimation of the AF, as a function of time, and adjusts for cluster-shared unmeasured confounders. In practice, confounders may be unmeasured, but not shared within clusters, or we may lack access to clustered data. One remedy is to use an instrumental variable to mimic a randomized controlled trial and estimate the causal effect. In project IV, we developed a method for AF estimation based on instrumental variable analysis.

Genetics play an important role in the disease development and the concept of heritability, i.e. the variation in a trait explained by genetic factors, is often used to quantify the role of genetics. However, heritability does not convey any information on the population impact of some disease due to genetics. In project III we show how the AF can be conceptualized for complex traits, with the overall genetic risk as the exposure, and how heritability and the AF are formally related.