In a recent study, Hellquist et al.1 estimated the mortality reduction attributable to mammographic screening among Swedish women aged 40-49 years to be 29% (95% confidence interval [CI], 20%-38%). The study is of high quality, and is a landmark study on an important topic. However, we suggest another way of calculating the estimated mortality reduction that would reduce the estimated screening effect on breast cancer mortality somewhat.
Hellquist et al. divided Swedish counties into those that have or have not introduced screening for women 40-49 years of age. They first tested for any statistically significant prescreening differences in mortality between the later screening and nonscreening counties, concluding that there were no differences in mortality. Later, the postscreening differences in mortality between screening and nonscreening counties were used as an estimate of the screening-related mortality reduction, assuming that the groups were equal with the exception of screening. Although this may seem like a reasonable approach, it is important to remember that a nonsignificant P value does not imply that there is no difference between 2 groups, only that it is not possible to separate a difference from random variation. Prior to screening, the mortality was 6% lower in the screening counties, and the authors argue that the regional differences may have become smaller over time. However, because we do not know for certain whether this is the case, we propose an alternative effect measure. We would prefer a measure equivalent to that of randomized trials, estimating the mortality ratio between the screened and nonscreened groups taking into account the prescreening regional differences. For the given data, this yields an estimated mortality reduction of 0.71/0.94 ≊ 0.76 among screened women.
As for the statistical uncertainty, it is larger when taking the prescreening mortality into account. Without access to the original data, we cannot calculate exact CIs, but we can make an approximation by back-calculating from the given CIs. Using this approximation, the standard error is (1.05-0.85)/(2*1.96) ≊ 0.05 for the prescreening mortality ratio and (0.80-0.62)/(2*1.96) ≊ 0.05 for the postscreening mortality ratio. After performing 1 million bias-corrected bootstrap replications using the R software package,2 we found a 95% CI for the estimated mortality reduction of 0.62-0.87. Hence, we conclude that with our preferred approach, the screening-related mortality reduction is 24%, with a 95% CI of 13%-38%.