Limits of confidence in pooled NNT results
Number Needed to Treat (NNT) has become a popular summary statistic for
the results of randomised controlled trials because it combines the
treatment effect with the background level of risk in the population
studied. Patients in a single trial are randomised for both of these
factors and a confidence interval can be calculated which estimates the
statistical uncertainty of the NNT in this particular population.
Problems arise when comparisons are made between NNTs from different
randomised trials, or when NNTs from several trials are combined in a meta
-analysis. Often the background level of risk varies between trials in a
non-random fashion, depending on the entry criterion of the patients in
each trial. If the relative benefit of the treatment is constant across
these background levels of risk, the NNT in each trial will decrease as
the severity of the patients included in the trial rises.
Pooling of NNT results may not give a reliable answer under these
circumstances, as the entry criteria of each trial will confound the
treatment effect. Moreover the meaning of a confidence limit around a
pooled NNT poses difficulties when the background risk level between
trials varies widely.
For this reason I would support the suggestion of Egger et al (1)
that the pooled results of Meta-analyses are reported in terms of a
summary statistic which describes the relative benefit of a treatment
(such as Relative Risk). If the pooled Relative Risk is reported with its
confidence interval both can be applied to any chosen control group event
In the example quoted by Professor Altman in Figure 3 of his paper
the pooled Relative Risk is 0.62 (95% CI 0.52 to 0.74). When there is a
background rate of angina in the PTCA group of 28% (such as found in the
GABI trial which included patients with more severe angina) the NNT(B) for
CABG would be 8.67 (95% CI 6.87 to 12.67). If there is a lower background
rate of angina of 16% in the PTCA group (such as found in the CABRI
trial), then the NNT(B) would be 16.85 (95% CI 13.34 to 24.63).
Finally I would suggest that NNT results are always accompanied by
the control group event rate to which they apply, and the Relative Risk
and confidence interval from which they are derived.
1. Egger M, Davey Smith G, Phillips A. Meta-analysis. Principles
and procedures. BMJ 1997;315:1533-7
Competing interests: No competing interests