All Times EDT
Keywords: Many-to-1 Matching, Propensity Scores, Natural History Controls, Single Arm Trials, Natural History Augmented Randomized Trials, Tipping Point Sensitivity Analysis
Comparison of efficacy in patients receiving a new treatment on a clinical trial with efficacy in patients from natural history studies has substantial challenges. When patients in natural history studies are followed over different courses of the disease progression, it is essential that only data that match the disease progression of clinical trial patients are extracted for analysis.
The many-to-1 matching method developed by the FDA for Brineura in patients with CLN2 provides matching by an index time that characterizes the time course of the disease progression and then by static and time-dependent risk factors for each patient receiving the new treatment using all contributing patients from the natural history study. The many-to-1 matched data were used for summary statistics and graphic presentation in Brineura statistical review. The data can be used for N-of-1 analysis of treatment effect for each patient of the clinical trial. We show that the many-to-1 matched data can also be used for between group comparisons, and therefore, we are able to develop methods for statistical inference, including exact p-values, point estimates, confidence intervals and Bayesian probabilities. We also develop a sensitivity analysis to assess the robustness of the result against various potential sources of bias and uncertainties in statistical modeling. The methodology applies to single arm trials with natural history controls or randomized control trials utilizing both concurrent and natural history controls.
The propensity score matching method lacks the critical step of index time matching to define on both the comparative data and time-dependent risk factors, and therefore, time-dependent risk factors are either ignored or their values at different time are used in propensity score calculation. This leads to violation the required exchangeability assumption, resulting in systematic bias with substantially inflated type 1 error rates and interpretable results.