HONcode Certified

This website is certified by Health On the Net Foundation. Click to verify.

This site complies with the HONcode standard for trustworthy health information: verify here.

Reporting of subgroup analyses: look for the test of interaction

Carl Heneghan
Last edited 28th March 2011

Today’s BMJ looks at the ‘influence of study characteristics on reporting of subgroup analyses in randomised controlled trials.’

The objective was to investigate what industry funding does to the reporting of subgroup analyses. The ‘so what’ factor, of why is this important in the first place, probably comes to mind.

Subgroups are common, and in modern trial publication appear pretty much all the time. They are used to try and determine if certain baseline characteristics may affect the treatment outcome. For example, women with heart disease have very different outcomes compared to men and often this results in less aggressive treatment.

Yet, subgroups are open to all sorts of misuse: if they aren’t predefined they should be treated with extreme caution. A number of subgroups have subsequently been shown to be false: for example In 1988, the early breast triallists collaborative group showed in tamoxifen trials, there was a clear reduction in mortality only among women 50 or older. Yet, later reviews show benefits were irrespective of age.

Xin Sun and colleagues selected RCTs published in 118 core clinical journals in 2007, including 469 RCTs. What they found was that the high impact journals, non-surgical trials and larger sample size led to more reporting of subgroups. Of interest, and why this study is worth reading, is when the primary outcome was not significant, pharma trials were more likely to report subgroup analyses than non-pharma trials.

The one learning point that readers of trials should stick to from now on, and why this is important, is pharma trials used a test for interaction less for subgroups than non-pharma trials. To compare treatment effects in subgroups in a RCT, such as by sex, a test of interaction should be used. Even if it looks to the reader the two treatment effects look very different, and the P value looks very different, the test of interaction may not be significant.

So, if the interaction test isn't significant, there is no observable subgroup effect. And, if it isn't reported at all then you should ignore the result.

Twitter TrustTheEvidence.net


Search the TRIP Database

TRIP Database


Recent Comments