Relaxed MEMs: Enhancing Borrowing in Multisource Exchangeability Models for Small Sample Applications
Information sharing in clinical trials provides a promising way to increase trial efficiency, specifically in rare disease settings where sample sizes are small. Multisource Exchangeability Modeling (MEM) is a method that incorporates historical data from potentially multiple data sources, such as past trials, by exploring all possible configurations of exchangeability and provides a weight estimate for each configuration, allowing the user to quantify how much is being borrowed from each source. While MEMs have previously been shown to be asymptotically consistent, small sample sizes introduce uncertainty that naturally limits the amount of borrowing MEMs can facilitate. The proposed ``relaxed MEMs'' provide a mathematical framework for relaxing or removing those limits and allowing greater borrowing even in small-sample settings. Several relaxed MEM versions are proposed, and all strategies provide enhanced borrowing with minimal sacrifice in the bias and variance of the mean estimates. Specifically, the maximum absolute mean bias was 3.5\% (0.035) of the true mean in the two source continuous outcome settings and 6.4\% (0.032) of the true proportion in the two source binary outcome settings (absolute mean bias was 1.4\% (0.014) and 3.6\% (0.018), respectively, for traditional MEMs) and variance decreased relative to traditional MEMs across all exchangeable and small difference scenarios. The methods were then demonstrated using two endpoints from a series of phase II trials in cystic fibrosis.