incorrectly lumping data
We thank Dr. Marcia Valenstein from the University of Michigan for
quickly and kindly alerting us to a possible error in the odds ratio
estimate for fatal attempts in the SSRI versus TCA analysis. Indeed, this
was an undetected error in manuscript preparation and we let BMJ know
immediately. We regret and take full responsibility for the error.
Fortunately, it does not change our primary conclusions.
As for our analyses being “miscalculated dramatically” or “blatantly
misleading” as suggested, respectively, by Mitchell and Sakinofsky and
Streiner in their BMJ rapid responses, we strongly disagree. Ignoring
individual trials and calculating summary statistics from lumped data, as
they have done, is an incorrect approach to meta-analysis.
Nowhere in our Methods section do we state that we calculated odds
ratios from the aggregated data. Instead, we conducted meta-analyses of
randomized controlled trials using Peto’s method for calculating odds
ratios of rare event data. Failing to preserve the randomization of
subjects at the trial level introduces bias and confounding.(1)
Furthermore, simply calculating odds ratios from aggregated data fails to
account for the influence of chance in small trials. By weighting each
study by its precision, smaller trials have less influence than larger
(1) Egger M, Davey Smith G, Altman D. Systematic reviews in health
care. Meta-analysis in context. London: BMJ Publishing, 2001.
Competing interests: No competing interests