A critical look at observational studies

Observational studies serve as a critical alternative when randomized trials are precluded by ethical concerns, high costs, or the need for rapid evidence-based hypothesis generation. Increasing reliance on routinely collected observational data (electronic health records, registries, and claims) has been accompanied by an increase in the risk of systematic errors that might threaten the validity. In this review, we focus on a selected set of common and crucial issues, confounding, and the use of directed acyclic graphs to describe and analyze causal relationships. We also highlight selection processes that can induce collider bias and create spurious associations between exposure and outcome. Time-varying confounding, where prior treatment affects future confounders, necessitates the use of specific estimation methods. We discuss measurement error and misclassification in routinely collected data and time-to-event pitfalls, the handling of competing events, and missing data.