Subgroup analyses in randomized clinical trials - methodological steps and pitfalls towards personalized medicine

Individualized or personalized medicine has become a buzzword in the academic as well as public debate surrounding health care. The word personalized is appealing and transports the message of a new medicine evoking hopes for patients and physicians. However, personalized medicine is not necessarily about persons - it’s about subgroups and the more refined nosology of modern medicine which is based on much more profound knowledge on the pathological processes. To identify benefits and harms for these subgroups implicates several methodological issues, which I investigated in my PhD thesis.
Subgroup analyses in randomized clinical trials (RCT) can have important impact on patient care if their results are true. However, most subgroup analyses have been shown to be false with detrimental effects on patients’ health. To investigate the planning of subgroup analyses in protocols of RCTs and the agreement with corresponding full journal publications, we established a cohort of RCT protocols and subsequent full journal publications. Protocols were approved between 2000 and 2003 by six research ethics committees in Switzerland, Germany, and Canada. We included 894 protocols of RCTs involving patients; 515 subsequent full journal publications were identified. About a third of protocols planned one or more subgroup analyses, but of those, only a small fraction (< 10%) provided a clear hypothesis for at least one subgroup analysis, and only a third planned an appropriate statistical test for interaction. 515 of 894 (58%) studies were published as journal article; of those, almost 50% reported at least one subgroup analysis. In 33% of all publications reporting subgroup analyses, authors stated that subgroup analyses were pre-specified but this was not supported by a third of the corresponding protocols. Furthermore, in those 86 publications in which authors claimed a subgroup effect, only 42% corresponding protocols reported a planned subgroup analysis. More than one third of statements in RCT publications about subgroup pre-specification had no documentation in the corresponding protocols. We conclude that subgroup analyses are insufficiently described in RCT protocols and investigators rarely specify the anticipated direction of subgroup effects. Credibility of claimed subgroup effects cannot be judged without access to RCT protocols.
In statistical analysis, categorizing an inherently continuous predictor (e.g. age) raises several critical methodological issues. This problem also applies to investigation of interaction between e.g. treatment assignment and a continuous predictor in RCTs – e.g. do older patients benefit from a certain therapy compared to younger patients? We applied the new multivariable fractional polynomial interaction (MFPI) approach to investigate interaction between continuous patient baseline characteristics and the allocated treatment in an individual patient data meta-analysis of 3 RCTs (N=2299) from the intensive care field. In all included RCTs, patients requiring mechanical ventilation were randomized into two treatment groups: higher versus lower positive end expiratory pressure (PEEP) ventilation strategy. For each study, we used MFPI to calculate a continuous treatment effect function for four baseline characteristics and 3 outcomes. These functions were plotted with a 95% point wise confidence interval: 1. For each study separately, 2. For all studies combined (averaged function using a fixed effect model). This novel approach allows assessing whether treatment effects interact with continuous baseline patient characteristics and avoids categorisation-based subgroup analyses. These interaction analyses are exploratory in nature. However, they may help to foster future research using the MFPI approach to improve interaction analyses of continuous predictors in randomized trials and individual patient data meta-analyses.