Estimates of burden from air pollution may be severely biased: a methodological request

Lelieveld et al.¹ offer extensive work on cardiovascular disease burden from ambient air pollution. We wish to complement this study with methodological considerations and we have a request.

This procedure is ‘used to estimate how many deaths could be avoided per year if the population were exposed to a lower counterfactual level than current, ambient concentrations of air pollution’.¹ Yet, this AF method cannot identify numbers of deaths due to exposure. The reason is straightforward: even though the date of death of an exposed person may be observed, the counterfactual date of death given a reduced exposure cannot be determined.² The resulting bias of this AF method may be substantial.² Thus, the authors’ main statement that the number of deaths ‘from all ambient air pollution is estimated at 8.8 (7.11–10.41) million/year’¹ is not scientifically justified. We emphasize that the reported confidence interval reflects imprecision in the estimation of R(z), which is distinct from the inability to identify the number of deaths caused by exposure, even if R(z) is known precisely: the bias outlined above is not quantified by typical confidence or uncertainty intervals. As a consequence, the published death counts are not a reliable basis to compare or rank smoking and air pollution as health risk factors: ‘For comparison, the WHO estimates the global mortality from tobacco smoking (active and passive) at about 7.2 million per year’.¹

A subsequent pitfall is to break down years of life lost (YLL) from exposure by the cause of death/disease to calculate a specific mean loss of life expectancy (LLE): ‘Figure 3 presents the relative contributions of disease categories to LLE’.¹ Regarding lower respiratory infections (LRIs), the authors state: ‘LRI … make up 21.4% of LLE, influenced by childhood mortality’.¹ This statement and Figure 3 are not scientifically justified: the disease-specific YLL due to exposure cannot be identified from epidemiological data.² This is because, as stated already above, the total number of deaths from exposure cannot be identified and, consequently, the specific number of deaths by cause cannot be identified either. But the latter is assumed to calculate specific YLL and LLE.

Leading methodologists identified these biases three decades ago.^3,4 That several scientists use this AF methodology is not sufficient evidence that this practice is right. Hammitt et al.² demonstrate, explain, and discuss the resulting distortions of burden of disease estimates.

May we request to prove that Greenland’s and Robins’ mathematics and reasoning are wrong? If so, this challenge of methodology would simply go away. Yet without such proof, should scientists not abandon methodological approaches that produce uninterpretable outcomes in important research fields such as ambient air pollution—and beyond?

Conflict of interest: none declared.

References