Statistical methods to analyse registry data in a comparative setting

In rare diseases, randomised controlled trials are not always feasible for ethical reasons or because the required number of patients is too large. In this case, the use of single arm trials (all patients in the treatment group) is preferred, and the control group may be taken from historical data (e.g., registries). Patients from registries are selected so that they respect the same inclusion/exclusion criteria present in the clinical trial. However, in the absence of randomisation, treatment and control groups may still have differences in baseline characteristics. It is necessary to take into account these differences in order to avoid, or limit, a bias in the treatment effect estimation. There are several statistical methods to achieve this purpose, and we focus here on propensity score (PS) methods. The PS is a score summarising patients' baseline characteristics, and it can be used to select similar patients belonging to treatment and control arms. We describe the following PS methods: matching, inverse probability weighting and stratification. An application is presented using a fictive, but realistic, example where the response variable is a time-to-event endpoint. On this example, the results from the PS methods are compared to those obtained with a naïve analysis (without any adjustment for baseline characteristics) and to those obtained with a covariate adjustment method, in which baseline covariates are simply included in the survival analysis. The naïve approach results in a biased treatment effect estimation, compared to covariate adjustment and "doubly robust" (i.e., relevant covariates are included in the survival analysis) PS methods that provide almost the same, unbiased, results. However, in general, relying on only one single method of analysis can be too misleading and several approaches should be used as sensitivity analyses to check the robustness of the results.