Causal Inference
Benchmarking observational analysis against registry-based trials: estimands, assumptions and estimators Camila Olarte Parra* Camila Olarte Parra Anita Berglund Issa Dahabreh
To increase our confidence in observational analyses, we can compare real trials with observational emulations of target trials designed to be similar. Obtaining equivalent results could make observational analyses more trustworthy to evaluate other (long-term) outcomes outside the scope of the real trial and to assess treatment heterogeneity. Here, we outline a framework for formal benchmarking in an ideal setting of a registry-based trial, where the real trial is embedded in a nationwide registry. We describe the conditions required for valid comparisons and propose methods to evaluate the implications of these conditions on the observed data. We evaluate the proposed methods through plasmode simulations which allow us to generate datasets that mimic the features of real-world observational data while embedding known causal effects as found in trials. These simulations are designed to resemble a large cardiovascular trial nested in a large Swedish registry, where we have follow-up data for patients who did not enrol in the trial. Well-performing methods in simulations are subsequently applied to the real data from the Swedish registry, exploiting the rich variables available from the linkage of different population registers that contain clinical and socio-economic variables.