The power to identify patient subgroup effects with meta-analyses of randomized controlled trials
*Stephanie Ann Kovalchik, National Cancer Institute 

Keywords: Comparative effectiveness research, Meta-analysis, Subgroup analyses

Context: Meta-analyses of randomized controlled trials (RCTs) could be an important tool for comparative effectiveness research (CER) by using subgroup analyses to identify modifiers of treatment effect.

Objective: To compare the power of meta-analysis of individual patient data (IPD) to meta-analysis of aggregate data in detecting a clinically significant patient-level effect modifier according to current levels of data sharing in the medical sciences.

Methods: The subgroup effect power (SEP) to detect a clinically significant covariate-treatment interaction was measured in a sample of 500 meta-analyses of RCTs. A clinically significant effect was defined as an effect indicative of a subgroup of treatment nonresponders. An exact and estimated SEP was computed for meta-regression and IPD meta-analysis, respectively, at empirically-determined levels of covariate heterogeneity and participant data sharing. Patterns in seeking and providing clinical trial data were determined from a survey of the sample authors.

Results: 28% of the 377 survey respondents (Response rate, 75.4%) sought participant data. 42.1% of the remaining authors reported that they either had not considered or did not see any statistical advantage to obtaining IPD; 30.8% did not believe their efforts would be successful. Out of a median of 4 (IQR, (2, 7)) primary research investigators contacted, patient data could be obtained for 2 trials (IQR, (0, 5)), on average.

When covariate heterogeneity was between 0 and 30%, the mean SEP for meta-regression ranged from 5.3% 95% CI (5.2, 5.3)% to 22.6% (20.4, 24.9)%. With data from 2 trials, the corresponding mean power gain with an IPD meta-analysis was 23.6% (25.9, 30.9)% to 11.1% (4.0, 8.6)%; with 4 trials, 38.2% (40.2, 45.6)% to 25.6% (18.2, 23.6)%, and with all trials 54% (55.0, 60.8)% and 40.6% (33.8, 39.6)%. Of the 500 meta-analyses of the study sample, 60.6% performed subgroup analyses but only 3.2% included participant data.

Conclusion: Analyses with patient data can substantially increase the chance of discovering clinically relevant modifiers of treatment effect. Although the majority of current meta-analyses in the medical sciences are performing subgroup analyses, few are based on participant data. Present norms in the seeking and sharing of clinical trial data will have to change if meta-analysis is to help achieve the aims of CER.