Editor—Liu et al used the term “proportional mortality study” to describe their method of comparing the smoking habits of 0.7 million adults who died of neoplastic, respiratory, or vascular causes with those of a reference group of 0.2 million who died of other causes in China.1 The term can be confusing as it is used only for proportional mortality ratio analysis in standard epidemiology textbooks.2 We suggest that the study can be more easily understood if it is described as a case-control mortality study.
An important assumption in such analyses is that the other causes of death should be unrelated to the exposure “not only in the sense of causation but also in terms of ‘self-selection’ for the exposure and the diagnosis and certification of the underlying cause of death.”3 Liu et al validated this assumption by showing that the smoking rates of the male and female reference groups were only slightly higher than those of the surviving spouses of the people who had died. However, they did not elaborate whether this similarity was true for each city or rural area in China, and, if it was not, why.
Could this similarity be a feature of populations in which the tobacco epidemic is at an early stage? The authors’ assumption may not be valid in other studies (such as our Hong Kong study4) or future studies that use a similar design. One potential confounding factor is social class, which is often associated with both smoking and mortality, and it may lead to an association between smoking and other causes of death. Studies elsewhere have observed some association between smoking and other causes of death (for example, in the American Cancer Society’s cohort the mean annual mortality from other medical causes was 39/100 000 men in never smokers and 81/100 000 in current smokers)5; choosing such other causes as referents would underestimate the risks from smoking.
It is fairly easy to define a priori which are the other causes of death for smokers as relations between smoking and many diseases are known, but it is difficult to define them when other risk factors (such as alcohol consumption) are studied in relation to mortality. Information on smoking (and confounders and other risk factors) in another control group randomly selected from the surviving population should be collected for validation; if the results do not support the assumption, classical case-control analysis comparing the dead and the living is necessary.