Abstract
Real-world data from sources, such as patient registries and electronic health records, can complement randomized controlled trials by providing timely, generalizable insights that better reflect routine clinical practice. However, the absence of randomization can introduce bias, particularly when treatment switching-defined as deviation from or discontinuation of the initial treatment-is influenced by time-varying confounders, that is, variables that are associated with both treatment decisions and outcomes over time. This study presents a comprehensive overview of statistical methods used to adjust for treatment switching in real-world studies to improve causal inference. We systematically searched MEDLINE and Embase for studies comparing at least two statistical methods for adjusting for treatment switching, from inception to December 2024. Forty-five studies were included, identifying four main categories of methods: (1) traditional approaches (intention-to-treat, per-protocol, as-treated, repeated measures); (2) propensity score-based methods (adjustment, matching, marginal structural models); (3) g-methods other than marginal structural models (g-computation, structural nested models, longitudinal targeted maximum likelihood estimation); (4) methods addressing unmeasured confounding (regression calibration, instrumental variables). Traditional methods are straightforward, but often yield biased estimates in the presence of treatment switching. Advanced methods, such as g-methods, are designed to adjust for time-varying confounding and can produce less biased estimates, though they require complex modeling. Instrumental variables and regression calibration relax the assumption of no unmeasured confounding, but rely on strong, often untestable conditions. By evaluating each method's assumptions, strengths, and limitations, we support applied researchers in selecting appropriate methods to strengthen causal inference in real-world studies.