Rapid Responses to:

PRACTICE: John Fletcher Subgroup analyses: how to avoid being misled BMJ 2007; 335: 96-97 [Full text]

Rapid Responses: Submit a response to this article

Rapid Responses published:

A refutation

GH Hall   (18 July 2007)

A refutation 18 July 2007
GH Hall, Retired physician EX1 2HW Send response to journal: Re: A refutation	Fletcher’s objection to the inclusion of subsets in analysis if they hadn’t been specified in the research protocol is easily refuted by noting that the relevant data had indeed been identified as worth while being collected. This can be construed as regarding the factor in the subset as a possible confounder, which would undoubtedly have been chosen were a regression analysis to be done. In a sense, therefore, there is no cherry picking taking place. If the data set were studied again for new material then bias would be possible. “Data mining” is often regarded by statistical purists with much suspicion. Personally I find nothing wrong with, say, looking at the features of those well over or under the general mean to gain hints as to the reasons for these deviations. After all, these were the events that Claude Bernard found most interesting. Competing interests: None declared