Purpose:Modeling life expectancy (LE) gains from disease interventions often require consideration of mortality rates from other causes, for which all-cause mortality is often used as a proxy. We examine the bias from double-counting disease-specific mortality in computing competing-cause mortality estimates, identifying when it may influence decision making.

Methods:Using a life table approach, we computed undiscounted and discounted life expectancy under different assumptions about competing-cause mortality: 1) all-cause mortality with disease-specific mortality removed; 2) unadjusted all-cause mortality. US and Brazilian data were used to construct disease-specific mortality rates, which varied in magnitude and age pattern.

Results:In the US, undiscounted LE was consistently biased lower when disease-specific mortality was not removed from all-cause mortality. The magnitude of the bias was 3.5, 2.7, 0.8 and 0.4 years for cardiovascular disease (CVD), cancer, respiratory diseases, and infectious diseases, respectively. When discounting 3% annually, the bias was 0.4, 0.3, 0.09 and 0.08 years. For lymphoma and all cancer in Brazil, the bias was 0.06 and 2.13 undiscounted years, and 0.01 and 0.29 discounted years lower. In all scenarios, as interventions that reduced disease-specific mortality became more effective, the bias was increasingly magnified. In the US, an intervention that reduced CVD mortality by 50% compared to current levels produced a LE gain of 1.58 vs. 2.38 years (biased lower by 0.8 years). If CVD could be eliminated, LE gains would be 3.49 vs. 5.47 years (biased 1.98 years lower). With discounting, results were biased lower by 0.06 (50% CVD mortality reduction) and 0.17 years (CVD elimination). In Brazil, a 50% reduction in cancer-specific mortality produced a gain of 1.03 vs. 1.17 years (biased lower by 0.14 years), and for a 100% reduction, 2.13 vs. 2.41 years (biased lower 0.28 years).

Conclusions:For models that incorporate competing causes of mortality, the discounted and undiscounted LE will be underestimated unless the disease-specific rate is removed from the all-cause rate. For example, for CVD interventions we calculated differences in LE gain between the two methods of 2 years, whereas published estimates of LE gains from interventions range from 6-13 months, such that the bias exceeds the actual gain. Whether this underestimation will consistently alter policy recommendations depends on the disease, interventions available, and the underlying risk of death from other causes.