Choice of summary statistics: relative and absolute measuresBMJ 2013; 346 doi: http://dx.doi.org/10.1136/bmj.f1092 (Published 26 February 2013) Cite this as: BMJ 2013;346:f1092
- 1Department of Psychiatry and Center for Tobacco Control Research and Education, University of California, San Francisco, CA 94143-0984, USA
- 2Department of Epidemiology and Biostatistics, University of California, San Francisco, CA, USA
Our 2012 meta-analysis of cardiovascular related serious adverse events associated with varenicline use has had more than 12 000 online full text viewings.1 In this article we drew attention to an earlier widely publicised meta-analysis,2 which differed from ours in its outcome definition and choice of summary statistic. We restricted outcomes to treatment emergent adverse events in active and placebo arms, whereas the earlier study allowed differential outcome follow-up by arm,2 which extended beyond the treatment emergent period, especially in the varenicline arm. In addition, that study used the Peto odds ratio (OR), whereas we summarised the study effects with the risk difference (RD), an estimate of absolute effect that usefully conveys the clinical effect of a treatment. The risk difference is particularly appropriate in examining rare event data because it accommodates trials with zero events.3 4 5 Thus, our meta-analysis of 22 trials included eight zero event trials that were excluded from the earlier analysis.
The literature on selecting the optimal summary statistic for fixed effects meta-analyses in the setting of rare events is limited. Rucker and colleagues recommended that “if the observation times, and thus the numbers of events, vary considerably between studies, relative measures such as the RR [relative risk] and the OR are preferable.”5 They also proposed that “the RD should be presented alongside the OR to put it in perspective, and thus allow ‘zero’ trials to contribute to the overall evidence and its interpretation.”5 In our analysis, there was no advantage to using a relative measure because the observation periods and the number of events were comparable for all but two of the 22 trials. Consistent with Rucker and colleagues’ guidance,5 we reported Peto and Mantel-Haenszel ORs, the relative risk, and the risk difference.
Our methods should not be applied to all reviews. We encourage further investigations of the best choice of relative and absolute measures in a wide range of event rate and allocation ratio settings.
Cite this as: BMJ 2013;346:f1092
Competing interests: Detailed disclosures are included in our BMJ publication,1 including grant funding from NIDA, NIMH, TRDRP, FAMRI, and a single investigator initiated research award from Pfizer (WS981308 registered online with www.clinicaltrials.gov, title is Varenicline Inpatient Study). The funding from Pfizer did not support the work of this meta-analysis.