Part III ♥ 180 PROLOGUE Borrowing of historical data is a statistical practice that has gained prominence in recent years.1–3 This approach involves the use of previously collected data for new research purposes, and it was originally proposed by Pocock in 1976.4 Pocock’s primary work was devoted to borrowing information from historical control groups of randomized trials, to potentially augment the current control group of a randomized trial. Pocock proposed strict criteria under which he deemed borrowing acceptable. Obviously, the type of control treatment should be the same. More stringently, information could only be borrowed when research was carried out in the same research centers, preferably by the same researchers. Patient characteristics should be similar, the method of treatment evaluation should be comparable, and there must be no external factors (unmeasured confounding) to believe that the results would differ. In practice, Pocock’s criteria are almost never satisfied.5 The advantages and drawbacks of borrowing In the ideal scenario, borrowing historical data offers several advantages. Firstly, it can save time and resources needed to conduct a trial. Rather than collecting data from scratch, researchers can use already existing datasets to answer current research questions. Secondly, borrowing data can be particularly useful when studying rare or difficult-to-find populations, where new data collection may be challenging or impossible.6 Lastly, it could also help address issues with patient reluctance to be randomized to control groups, although we have witnessed in Part I that patients often benefit from being randomized to a control, rather than a treatment group.Apart from practical advantages, borrowing of information potentially leads to an increase in statistical power and precision by combining data from multiple studies or time periods.2,7 But is all that glitters really gold? There are several potential drawbacks to the use of borrowing information across data sets.8 The primary concern is the introduction of bias and confounding.9,10 Therefore, researchers must carefully evaluate the quality and comparability of historical datasets before attempting to integrate the data sets. This is the exact reason why Pocock invented his criteria. The incorporation of (biased) datasets will undoubtedly increase the type I error, and without properly accounting for this introduced confounding, the chance of making erroneous decisions will inevitably increase.11 For regulatory agencies, this is a critical concern, and it has so far decelerated the implementation of Bayesian borrowing methods in regulatory analyses. In recent years, such methods are now explicitly mentioned in the Guidance for the Use of Bayesian Statistics in Medical Device Clinical Trials by the FDA.12 Practical borrowing examples exist where safety information obtained in adult populations is used to extrapolate safety to pediatric populations.13 Nonetheless, methods of (dynamic) borrowing are not used at a large scale.
RkJQdWJsaXNoZXIy MTk4NDMw