Methodologies used to estimate tobacco-attributable mortality: a review

Since the association between tobacco and mortality was first discovered [1, 2], the task of attributing a given number deaths to smoking has been and continues to be a controversial process, beset by limitations and questioned from different quarters, including the powerful tobacco industry. With the appearance of the successive revisions of the International Classification of Diseases (ICD), there has been considerable progress in the process of categorizing mortality, but little in methods for attributing mortality to risk factors such as tobacco. Obtaining reliable estimates of the impact of tobacco on mortality would facilitate to have a clearer picture of the problem caused by smoking and would be of help in the planning of health policy.

The task of quantifying smoking-attributable mortality has been performed mainly through indirect methods. This review sought to list and to describe the different methods of estimating mortality attributed to tobacco use, to indicate the principal methodological differences existing among them, and to identify the possible sources of variability in the results.

In order to obtain papers that addressed the methodology employed for attributing mortality to tobacco use, a search was made in the Medline database until March 2005, using the terms, mortality, attribut,*method* and tobacco or smok*. The search was completed with a manual review of the bibliographic references cited by the papers retrieved and of other publications, such as the monographs published by the Centers for Disease Control and Prevention (CDC). The main inclusion criteria was the use of an epidemiological method to estimate attributable mortality. Papers describing mortality, such as cohort follow-up or mortality studies were excluded unless an epidemiological analysis had been used. Animal studies and communications presented at congresses were also excluded from the search.

The estimation of attributable mortality is also applied to other risk factors in addition to tobacco, such as alcohol consumption or obesity. In order to avoid the exclusion of valid methodologies the search was repeated without restricting it to tobacco or smoke.

The search yielded a total of 372 papers. Of these, 74 were finally included, as the rest did not apply mortality attribution methods. Some papers included more than one method. The unrestricted search, without the terms tobacco or smoke, did not furnish any new alternative methodology.

Revision of the 74 papers enabled us to identified 2 types of mortality attribution procedures for the specific case of tobacco. The first one is based on individual analysis of deaths to ascertain if tobacco use had any role in mortality. Only three studies applied this procedure [3‐5]. The second is based on the application of indirect methods and constitutes the most commonly used methodology for attributing mortality. The total number of papers that employed this indirect methodology was 73, with 61 of these being yielded by the automatic and 12 by the manual search.

Seven indirect methods for estimating tobacco-attributable mortality were identified. The applied methodology in these 7 methods can be classified under four categories: Prevalence-based analysis (Prevalence-based analysis in cohort studies, prevalence-based analysis in case-control studies and the basic method), Peto and colleagues' method, methodologies based on the calculation of excess mortality (Garfinkel's and Roger's method) and predictive models (Prevent). The methods differ in terms of calculation processes, information requirement, data sources and assumptions required for their application. A summary of these methods is showed in Table 1. The main characteristics of the different indirect methods are described below.

This paper constitutes, to our knowledge the first methodological review of procedures for estimating smoking-attributable mortality. In the context of decision-making it is essential to know, albeit approximately, the impact that a given risk factor has on the mortality of a population. Estimation of tobacco-related mortality is not confined to one procedure alone, inasmuch as any of the different methods outlined above can be used for the purpose.

Despite the fact that different methodologies have been found, the foundations of more of them are the same and only few differences arise in the calculation procedures (Table 2). Data availability has been taken into account when choosing a method and also methodology limitations and assumptions have to be considered. Some of them are described below.

The first limitation affecting intercomparison of methods and studies stems from the absence of a universal definition of the categorization of tobacco use. The publications analyzed furnish different definitions of "smoker", "non-smoker" and "ex-smoker" [6, 92‐95], something that inevitably determines the result of the estimation [96].

To view smokers as a single entity could lead to a distorted mortality estimate, since failure to take account of the number of cigarettes smoked, age at initiation, years of smoking and other variables that could modify risks values can occur. It would thus be interesting to explore tobacco use in the studied populations [88]. A correct classification of ex-smokers is very important for estimating and predicting mortality attributable to tobacco use. To avoid overestimation of attributed mortality, Anthonisen [97] proposed that account must be taken of the decrease in risk that takes place at 15 years after quitting the habit. But this decrease is also determined by the subject's age at cessation [98, 99], the duration of smoking [100] and the cause studied. The fact that this information was not expressly gathered in the majority of surveys means that mortality among ex-smokers may be overestimated. This problem is solved, at least in part, by ex-smokers reclassifying themselves as non-smokers after the elapse of a long time without smoking [101].

The second limitation, present mainly in the proportional method, resides in their reliance on current smoking prevalences to reflect mortality occasioned by tobacco use in previous years. Knowing current smoking prevalence could be a great help when it comes to predicting future mortality, but not present[102]: indeed, knowing the prevalence of tobacco use in any given year could help predict lung cancer mortality in 20 years' time [103]. As yet, this problem has no easy solution, due to the absence of historical series of smoking prevalence in most countries. Moreover, even if such series were to exist, lack of knowledge of the latency and induction times for each of the tobacco-related causes of death would constitute another problem. The use of current prevalence may overestimate or underestimate the attributable mortality. In countries where the prevalence is decreasing, as U.S.A. or some European nations, the use of current prevalence is conservative in the proportional attribution method. The opposite occurs in countries where prevalence is increasing. Given the unavailability or inaccuracy of prevalence data, and emphasizing that current prevalence is a poor proxy for cumulative hazards of smoking, the knowledge of the period of time from tobacco consumption until mortality related to this use it is necessary.

Ascertaining the induction period might be feasible if only one specific component cause was active in triggering the disease. However, if one allows for the presence of more than one component cause, then each may have its own induction time; furthermore, the action of effect modifiers could alter the induction period [104]. It would therefore seem that ascertainment of the induction period is complicated; nevertheless, ascertaining the latency period is no easy matter either, since it varies according to the diagnostic methods. What should be clear, however, is that an induction time is needed for tobacco to cause harm, and it is for this reason that the age ranges between 30–35 years are considered the time to begin measuring the effects of exposure. Measuring such effects without taking into account an induction time could lead to overestimated mortality results. On the other hand, some authors [88] feel that ignoring mortality under the age of 35 years may give rise to underestimates of mortality figures, due to the existence of individuals who started smoking at early ages.

Peto et al. avoided the problem entailed in prevalence-dependent methods of attribution. For the application of their estimation procedure, lack of knowledge of the tobacco consumption or latency and induction periods are no a limitation. But this method has not been exempt from criticism [25, 105‐110] directed, mainly, at the calculation of summarized prevalence. Some of these critics were supported by the tobacco industry, which tried to undermine the studies focused on estimations of mortality attributable to tobacco consumption. Peto and colleagues defined synthetic prevalence as an indicator that summarizes a population's smoking history, and calculate it by assuming CPS II data on lung cancer mortality rates among smokers and non-smokers to be valid. The use of these 2 sets of data gave rise to numerous criticisms that highlighted the low population representativeness of the CPS II [25, 107, 111, 112]. Most of the population included in this cohort study was middle class, which may result in lung cancer mortality in non-smokers being underestimated [88] leading, in turn, to an overestimation of lung cancer mortality attributable to tobacco use and, by extension, to an overestimation of the summarized prevalence [25]. To justify their validity and universality, these data were compared with those yielded by the study that targeted British physicians [93]; despite the fact that the results obtained were similar, no conclusion could be drawn, since the representativeness of this latter study was also limited. The only thing that could be said was that the lung cancer mortality rate among non-smokers had not varied over the years[111]. Nonetheless, in countries where the use of coal is widespread, lung cancer mortality among non-smokers is higher, and thus the data, rather than being drawn from the CPS II, have been drawn from a local study [72].

The third limitation centers on the absence of world-wide risk indicators that would reflect the degree of association between tobacco and smoking related-causes of mortality. The most widely used effect measure is RR, and a sensitivity analysis has shown that changes in its value lead to a greater impact on the estimation of mortality than do changes in prevalence [102]. Although drawn from different sources, the RRs used in the various studies mainly came from the CPS II [55, 56]. Applying these risks to populations other than that of the USA aroused criticism because, inter alia, of their only being adjusted for age and sex, and because of the difficulty inherent in assimilating identical tobacco consumption and genetic variability patterns, or the same influence of confounding factors or effect modifiers. A solution to these problems was sought through a re-analysis of the data [9, 113‐115], and the RRs were shown robust. Notwithstanding this, the criticisms continued unabated [116].

The risks obtained from the CPS II are plausible in the light of current knowledge [25] and have been extrapolated [117] to different EU countries, in absence of other high quality indicators. Nevertheless, other authors have chosen to apply RRs which are drawn from studies with less robust designs or possibly inconsistent with present knowledge.

A fourth limitation of the attribution methodology is the uncertainty present in the relationship between exposure, tobacco use, and different causes of death. While lung cancer was the first disease to be causally associated with tobacco use, many studies have observed more causal associations. The latest report of the Surgeon General [56] has added 2 further causes of mortality that had not been considered to date, i.e., stomach cancer and acute myeloid leukemia and excludes hypertension.

Some methods have been compared by applying them in the same population. Published comparisons are the individual analysis and SAMMEC [3, 4], Peto and Prevent methods [83], Peto and proportional attribution method [11], and Garfinkel's and proportional attribution method [85]. The results obtained in all of these comparisons have proved to be similar estimations, thereby conferring validity on the respective methodologies. Observational epidemiology and, despite their limitations, the use of the above-described calculation procedures offer a good approximation of the impact of tobacco on the mortality of a population [4].

Prior to conducting a study on estimation of tobacco-attributable mortality, it is essential to assess which method is best suited to the type and quality of the available information.

When the mortality estimation objective is going to be the knowledge of tobacco impact on a population, it is important to take into account all the diseases related with consumption. For this reason, the applications of methodologies that involve all the causes of disease are important. These methodologies are: Prevalence-based analysis in cohorts and in case-control studies, Peto et al. and Roger's methodology. All of them supply accurate and reliable estimations of mortality attributed to tobacco consumption.

The absence of a simulation study involving and comparing all calculations procedures do not allow us to recommend a method over other one.

These types of methods furnish estimates that constitute valuable information and help forming a more accurate picture of the problem that smoking poses to world health.