Weighting methods to solve selection bias in clinical epidemiology

Weights are crucial to draw real-world evidence, especially when studies diverge from ideal conditions due to practical or ethical constraints. In this work, we formulated the deviation from an ideal, hypothetical study as a basic principle for understanding weights. The deviations in a given study are phrased as selected data challenge leading to bias. This work presented three cases from clinical epidemiology, illustrating the application of weighting methods to reduce bias: improving comparability between groups (case A), transfer of study results to a different population using either target population characteristics (case B), or sampling information of the study design (case C). Case A was motived by the insufficient evidence on the impact of new imaging technique used for radiotherapy guidance. We artificially created a study using entropy balancing and inverse probability of censoring weights to enable a fair comparison between an observational cohort of patients, and a randomized controlled trial. In this case the unmeasured confounding assumption remained questionable, due to the more stringent exclusion of metastatic patients in the treatment group at baseline and missing time-varying covariate information. Randomized controlled trials are needed to justify guided radiotherapy usage with larger evidence. Case B aimed to conduct a representative analysis of the burden of ventilator-associated pneumonia. This was achieved by prospectively collecting data from 11 European countries, as by using inverse probability of inclusion weights. Further strengths of the design and analysis were a 90-day follow-up period, modeling pneumonia as time-varying, and using multi-state models to account for competing risk bias. Study limitations included high proportion of missing values and the simplified modelling of the hierarchical data structure. Case C demonstrated how sampling information of the study design can be used to correctly estimate absolute risks when patient groups are oversampled in order to improve precision in resource limited settings. For three study designs, we explained how unequal selection strategies introduce bias and constructed inverse probability of inclusion weights. When combined with multistate models and cause-specific proportional hazard models that account for competing risks, these studies can contribute to unbiased evidence in resource-limited settings. In conclusion, methods like inverse probability weighting offer significant advantages in non-ideal study designs. They reduce bias, enable the investigation of effects in constrained settings, and support clinical study results with real-world data. However, they involve untestable assumptions, reduced precision, and increased complexity, which can affect the perceived validity. Extensions such as double-robust methods combined with ensemble learning, alongside high-quality prospective data, may help address some limitations and improve the robustness of findings.