Joint incorporation of randomised and observational evidence in estimating treatment effects

• Published in: Statistical methods in medical research 2019-01, Vol.28 (1), p.235-247
• Main Authors: Ferguson, John, Alvarez-Iglesias, Alberto, Newell, John, Hinde, John, O’ Donnell, Martin
• Format: Article
• Language: English
• Subjects: Bias; Confidence Intervals; Data Accuracy; Data Interpretation, Statistical; Discounts; Estimation; Evidence-based medicine; Humans; Meta-analysis; Models, Statistical; Observational studies; Observational Studies as Topic; Probability; Randomization; Randomized Controlled Trials as Topic; Simulation; Standard error; Treatment Outcome
• ISSN: 0962-2802; 1477-0334
• DOI: 10.1177/0962280217720854

In evidence-based medicine, randomised trials are regarded as a gold standard in estimating relative treatment effects. Nevertheless, a potential gain in precision is forfeited by ignoring observational evidence. We describe a simple estimator that combines treatment estimates from randomised and observational data and investigate its properties by simulation. We show that a substantial gain in estimation accuracy, compared with the estimator based solely on the randomised trial, is possible when the observational evidence has low bias and standard error. In the contrasting scenario where the observational evidence is inaccurate, the estimator automatically discounts its contribution to the estimated treatment effect. Meta-analysis extensions, combining estimators from multiple observational studies and randomised trials, are also explored.