Dynamic Historical Data Borrowing Using Weighted Average

Abstract

In many clinical trials, especially trials in rare diseases or a certain population like paediatric, it is of great interest to incorporate historical data to increase power of evaluating the treatment effect of an experimental drug. In practice, historical data and current data may not be congruent, and borrowing historical data is often associated with bias and Type-1 error rate inflation. It remains a challenge for historical data borrowing methods to control Type-1 error rate inflation at an adequate level and maintain sufficient power at the same time. To address this issue, dynamic historical borrowing methods can borrow historical data more when historical data are similar to current data and less otherwise. This paper proposed to use a weighted average of historical and current control data, with the weight being set as an approximation to the optimal weight that minimizes the mean-squared errors in the treatment effect estimation.Comparing to selected existing methods, the proposed method showed reduced bias, robust gain in power and better control in Type-1 error rate inflation through simulation studies. The proposed method enables the utilization of all possible historical data in the public domain and is readily used by skipping the need for external expert input in some existing approaches. K E Y W O R D S clinical trial, historical data borrowing, rare disease, study design | 1261 CHU and YI variance 2H . The null hypothesis under consideration is H 0 : μ T = μ C , with two-sided alternatives H a : μ T ≠ μ C . | Test then poolThis frequentist approach first performs a congruence test between Y H• and Y C• at a prespecified significance level. If there is no significant difference between Y H• and Y C• , then Y H• will be fully borrowed, that is, Y H• and Y C• will be pooled as the control group for the treatment effect evaluation. If there is significant difference between Y H• and Y C• , then Y H• will be discarded and only Y C• will be used as the control group for treatment effect evaluation. Note that the selection of the significance level for congruence test is subjective. Larger significance level means more chance for historical data to be borrowed. More details can be found in Viele et al. (2014).With additional assumptions, certain frequentist approaches, such as matching, can be more efficient. In this article, we focus on test then pool (TTP) because it requires less assumptions and can be implemented without information of individual patient data. C | Y H• ∝ L( C | Y H• ) 0 ( C ), where 0 ≤ ≤ 1. = (1 + exp(a + blog(S))) − 1