Study Design
Causal diagrams of epidemiological study designs conceptualized as instrumental variable analyses Guoyi Yang* Guoyi Yang Abigail Ada Rath Heidi Jones C Mary Schooling
Background: Most observational studies evaluate causation using confounder-control methods. Directed acyclic graphs (DAGs) are useful in depicting causal structures and have been increasingly used to identify confounders in observational studies. However, the structures of alternative study designs that do not rely on confounder adjustment have not been explained systematically.
Methods: We propose that epidemiological study designs that do not rely on confounder adjustment can be conceptualized as instrumental variable (IV) analyses. We use DAGs to illustrate the causal structures of randomized controlled trials, quasi-experimental studies, including Mendelian randomization, regression discontinuity designs, and interrupted time series, self-controlled case series, and ecological studies. We use a selection diagram for difference-in-differences analysis.
Results: We argue that these seemly heterogenous study designs share a common causal structure as IV analysis. Specifically, IVs are random allocation in randomized controlled trials, genetic allocation in Mendelian randomization, a continuous variable in regression discontinuity designs, time in interrupted time series and self-controlled case series, place in ecological studies, and the interaction between time and place in difference-in-differences analysis. We further explain how IV assumptions are applied to these study designs and provide solutions to address biases due to violation of IV assumptions.
Conclusions: Conceptualizing different epidemiological study designs as IV analyses provides a unified approach to understanding their causal structures and recognizing threats to study validity.
