Re: All-cause mortality in randomized trials of cancer screening.

To support their argument that results of cancer screening trials based on disease-specific mortality are unreliable, Black et al. (1) compared the treatment effect measured by disease-specific mortality with the effect measured by all-cause mortality in 12 such trials. They found “major inconsistencies” between the two measures. We disagree with their interpretation and illustrate our reasons with results from the Minnesota study of fecal occult blood testing (2). After 13 years, that study found a 33% lower colorectal cancer mortality in the annually screened group than in the control group. (Note that the number of colorectal cancer deaths per 10 000 person-years in the annual group given in Table 1 of Black et al. (1) is in error. The figure should be 82/18.4160 4.5, not 5.4.) Because colorectal cancer deaths constituted only 3% of deaths from all causes, the expected reduction in all-cause mortality is only 1%, i.e., 3% of 33%. The reduction in all-cause mortality actually observed in the study was 0.0 per 10 000 person-years, with a 95% confidence interval (CI) of –7.6 to 7.6. The expected 1% reduction in the all-cause mortality rate, corresponding to a decrease of 1.8 per 10 000 personyears, is consistent with this interval. Black et al. considered the treatment effect of 1.2 for disease-specific mortality inconsistent with the 0.0 for all-cause mortality but, in fact, the result 1.2 falls well within the 95% CI of –7.6 to 7.6 for the difference in all-cause mortality. Similar consistencies can be shown for most of the studies cited by Black et al. that were designed for disease-specific outcomes and, as indicated by the large CIs, are underpowered for all-cause analysis. We agree that, in some cancer screening trials, all-cause mortality may provide assurance against the biases that Black et al. identified. However, a problem with the design of studies using an all-cause mortality end point is the enormous sample size required. For example, with all-cause mortality as the outcome, the aforementioned Minnesota trial (2) with 15 000 subjects per group would have required 20 times as many subjects, or 300 000 per group. Lung cancer trials would have to be about 10 times as large. According to the Nordic Cochrane Centre, a breast cancer screening trial would need 1.2 million women in each group (3), 25–60 times the size of some previous diseasespecific studies. Given that the concerns raised may involve only certain cancers in specific populations, rejecting findings across the board based on the possibility of bias is premature.