A consortium of six research groups used data from common sources to recreate detailed cigarette smoking histories under three distinct tobacco control scenarios as inputs for mathematical models to quantify the impact of changing smoking behavior on lung cancer mortality rates in the United States during 1975–2000. We used a comparative modeling approach to address this complex problem; comparative modeling produces a range of results across models but, when these are reasonably consistent, enhances their credibility. During the period 1975–2000, approximately 2
042 lung cancer deaths among men and women combined could have been averted had tobacco control efforts been completely effective in eliminating smoking as of 1965; of these, we estimate that approximately 795 851 lung cancer deaths were averted or about one-third of what was possible. During the period 1991–2000, we estimate that approximately 345
000 lung cancer deaths among US men and 175
000 deaths among US women were averted due to changes in smoking behaviors starting in the mid-1950s. These estimates of reduced lung cancer mortality associated with reduced tobacco use are much larger than an estimate from demographic projections that 146
000 lung cancer deaths among men were averted in 1991–2003 (5
). We estimate that in the year 2000 alone, approximately 44
000 deaths were averted among US men and 26
000 deaths among US women.
It is not surprising that the various models used in this article yielded a range of estimates of the fraction of lung cancer deaths averted by the tobacco control efforts in the United States. This range of results represents the uncertainty associated with model choice. First, some of these models were calibrated against US mortality data, and, as a consequence, these models describe the lung cancer mortality trends in the United States very well under the ATC scenario. Second, although five of the six groups used the TSCE version of multistage models (8
) as the dose–response module, the estimated parameters were different because they were estimated by their fit to different cohorts. In addition to the TSCE model, the Yale group also used the models developed by Knoke et al. (9
) and Flanders et al. (26
) and obtained similar estimates of the relative effect of tobacco control. It is well known that the risks of tobacco smoking have changed over time; moreover, they could be modified by other factors such as diet that are not accounted for in any of the models. Despite these limitations, the estimated numbers of deaths averted and deaths that could have been averted under the assumption of CTC were reasonably consistent across models (). The main message of these analyses is clear. Tobacco control strategies implemented mid-century have averted hundreds of thousands of lung cancer deaths in the United States during the period 1975–2000, but these are only approximately 30% of the lung cancer deaths that could have been averted had all cigarette smoking ended in 1965.
The FHCRC, MGH-HMS, and Yale groups calibrated their models to US mortality during 1975–2000 using birth cohort and period effects. These calibrations are necessary to describe lung cancer mortality rates and trends in the United States and indicate that the lung cancer mortality experience of the entire population cannot be adequately described by extrapolating from the SEER registry in one decade, or from various cohort and case–control studies of smoking and lung cancer (please see the supplementary material
, available online, for the datasets used by each of the groups for parameter estimation). Particularly among men, US lung cancer mortality is considerably higher than would be expected from the cohort studies against which the dose–response modules were calibrated. In addition, models from cohort studies and available population smoking histories cannot adequately describe temporal components of trend, that is, the effects of age, period, and birth cohort.
There could be several reasons why the models were poor at predicting population lung cancer rates without additional calibrations. First, the datasets used for estimating the parameters of the dose–response modules were almost certainly not representative of the US population. Second, the smoking history generator was based on smoking histories for birth cohorts in the general population that were inferred from simulations using cross-sectional histories that often relied on subjects’ recall of events that occurred several years earlier. Third, potentially important covariates (eg, diet, air pollution, and radon exposure) and occupational exposures (including asbestos and ionizing radiation) were not available for the overall population, and different exposure distributions could contribute to rate discrepancies. Fourth, although the models discussed assume a consistent effect of exposure on lung cancer mortality, temporal changes in the manufacture of cigarettes and smoking behaviors could explain some of the discrepancies in trend, and data on changes in cigarette manufacturing and composition are not readily available. Changes in tobacco or cigarette composition, which were not explicitly addressed in these analyses, could be important contributors to population trends in lung cancer mortality. However, one would expect changes in tobacco or cigarette composition to manifest themselves as period effects, whereas models that used age–period–cohort calibrations find that trends are dominated by birth cohort effects. Finally, uncertainty remains with respect to the models themselves.
In particular, our estimates of the lung cancer rates in the US population under the CTC scenario appear to be higher than would have been expected on the basis of recent work on lung cancer rates among never smokers (27
). However, for the reasons given above (cohorts not representative of the general population, omission of important covariates), the never-smoker rates reported by Thun et al. (27
) may not reflect the never-smoker rates in the general population. Some confidence in the lung cancer rates under the CTC scenario estimated from the models in this article can be derived from the fact that the dose–response modules describe lung cancer rates among former smokers well (9
One limitation of the calibrations is that the same period and cohort parameters are applied to current smokers, former smokers, and never smokers. Factors, such as diet, that could affect trends in lung cancer rates might be expected to have different effects among current smokers, former smokers, and never smokers. However, different cohort and period effects could not be estimated in these subgroups because of identifiability issues. The FHCRC group did fit period and cohort effects to never smokers alone and to current smokers and former smokers separately, but the original model in which these effects are applied equally to all groups described the data better as judged by the Akaike Information Criterion Statistics
Overall, our study shows that changes in smoking behaviors led to a substantial reduction in the lung cancer mortality that would have been expected had the smoking trends in the 1950s continued into the future. Our analysis was conducted through to the year 2000, the latest year for which we were able to obtain sufficiently detailed data when this project was initiated. Consistent with trends for continued gains due to past tobacco control policies, smoking prevalence continued to fall from 23.2% in 2000 to 20.6% in 2008. Much of this decrease can be attributed to tobacco control policies, especially the cigarette price increases in 1998–1999 (29
There are also other limitations to our study. We did not quantitatively assess the relative contributions made by changing patterns of smoking initiation and cessation to decreases in lung cancer mortality. It is clear, however, that most of the benefits of tobacco control policies during the period 1975–2000 have accrued from smoking cessation because changing patterns of smoking initiation would have impacted only individuals who were aged 55 or younger in 2000 and thus younger than the age at which lung cancer mortality begins to increase rapidly. Also, our numbers are likely to greatly underestimate the overall health impact of tobacco control efforts because they neither consider the substantial impact of non-cigarette forms of tobacco use (eg, cigars and pipes) nor the impact of tobacco smoking behaviors on diseases other than lung cancer. Smoking-associated diseases other than lung cancer, such as cardiovascular disease, were outside the scope of this work.
The results of this article show the dramatic impact of the reduction in smoking associated with tobacco control efforts in the second half of the 20th century on lung cancer mortality during the period 1975–2000. Even though other factors, including genetic polymorphisms (27
), contribute to lung cancer risk, the vast majority of lung cancer cases could be eliminated by eliminating smoking. Our results indicate that only approximately 30% of the total lung cancer deaths that could have been averted had tobacco control been complete were actually averted. This is because smoking rates took time to decline after the first Surgeon General's Report in 1965; smokers’ risk of lung cancer remains elevated for many years after smoking cessation and a sizable fraction of the population continued to smoke.
Clearly, further reductions in smoking rates will be required to reduce lung cancer incidence and mortality rates substantially. The recently reported 20% reduction in lung cancer mortality (30
) as a result of early detection using low-dose spiral CT suggests that screening of high-risk individuals may play a role in reducing mortality from this disease. Because risk of lung cancer remains elevated for a long time among smokers who quit, effective screening techniques may have a role in reducing lung cancer mortality among ex-smokers. However, continued implementation of evidence-based tobacco control policies, programs, and services remains the most promising approach to reducing the burden of lung cancer.