Confidence intervals for the number needed to treat

BMJ 1999;318:1764 ( 26 June )

Letters

Confidence intervals for the number needed to treat

Pooling numbers needed to treat may not be reliable

Absolute risk reduction is less likely to be misunderstood

Pooling numbers needed to treat may not be reliable

The first 150 words of the full text of this article appear below.

EDITOR $---$ The number needed to treat has become a popular summary statistic for the results of randomised controlled trials becauseit combines the treatment effect with the background level ofrisk in the population studied. Patients in a single trial arerandomised for both of these factors, and a confidence intervalcan be calculated which estimates the statistical uncertaintyof the number needed to treat in this particular population.¹

Problems arise when comparisons are made between numbers needed to treat from different randomised trials, or when the numbersneeded to treat from several trials are combined in a meta-analysis.Often the background level of risk varies between trials in anon-random fashion, depending on the entry criteria in each trial.If the relative benefit of the treatment is constant across thesebackground levels of risk then the number needed to treat in eachtrial will decrease as the severity . . . [Full text of this article]