Subgroup analyses

BMJ 2012; 344 doi: https://doi.org/10.1136/bmj.e2022 (Published 15 March 2012) Cite this as: BMJ 2012;344:e2022
  1. Andrew D Oxman, researcher
  1. 1Global Health Unit, Norwegian Knowledge Centre for the Health Services, N-0130 Oslo, Norway
  1. oxman{at}online.no

The devil is in the interpretation

Sun and colleagues found, in the linked systematic review (doi:10.1136/bmj.e1553), that about a third of a representative sample of recent randomised trials published in core clinical journals report subgroup analyses.1 After judging these reports of subgroup analyses using 10 carefully developed predefined criteria, the authors conclude that only in very few instances can we be confident that subgroup analyses provide a better estimate of effect than the overall results of trials, and they describe in detail the reasons why. These findings are consistent with what could be expected on the basis of previous reviews and the play of chance.2 3 4 5 6 7

Previous reviews of published trials and protocols have found that subgroup analyses are commonly reported (38-87% of the time),2 3 4 5 6 7 and that appropriate statistical analyses (tests of interaction) are not used 38-91% of the time.2 3 4 5 In addition, planned subgroup analyses are commonly not reported (48-69% of the time) and 43-91% of randomised trials report subgroup analyses that were not planned.6 7 When subgroup analyses are reported, authors claim differences in 25-69% of cases, and these results are commonly featured prominently …

View Full Text

Sign in

Log in through your institution