We used a discrete dynamical systems population model, as described by Vugrin et al.15 The model is initiated with a starting population, which is divided into subgroups that are defined according to age, sex, and tobacco-use status. The analysis projects population changes in 1-year increments while accounting for births, net migration (including both immigration and emigration), and deaths (a function of age, sex, and tobacco-use status). Members of each subpopulation have a specified probability of dying and of transitioning from one tobacco-use status to another. Using baseline inputs from the year 2000, we have found that projections of smoking prevalence based on a previous version of our model were aligned closely with National Health Interview Survey (NHIS) estimates through 2012 and that projections of the total U.S. population and annual mortality were similar to those of the Census Bureau.15 In this analysis, we updated the baseline to 2015 to account for recent changes in smoking initiation and cessation.
Using a projection period from 2016 through 2100, we simulated a baseline scenario that predicted the future use of cigarettes and noncombusted tobacco products (including smokeless tobacco and e-cigarettes) and then compared the baseline scenario with the policy scenario described below. Although longer-term projections are subject to increased uncertainty, this time period was chosen to account for the potential effects of reduced initiation of smoking on tobacco-related mortality, since such effects would not be observed until many decades into the future. The model, which was implemented with the use of MATLAB software, version R2017a (MathWorks), projects the effect of the policy on the prevalence of the use of cigarettes and noncombusted tobacco products and the effect on tobacco-related mortality and life-years gained.
The model accounts for initiation, cessation, and dual use of tobacco products, along with switching between two products: cigarettes (including the very-low-nicotine cigarettes introduced in the policy scenario) and noncombusted tobacco products. Inputs are summarized below and described in detail in the Supplementary Appendix, available with the full text of this article at NEJM.org.
Initial Population, Births, and Migration Inputs
We used 2015 Census estimates to determine the population distribution according to age and sex16 and age distribution among immigrants.17 We used the 2015 NHIS to determine the prevalence of tobacco use among adults (accounting for all combinations of current, former, and never use for cigarettes and noncombusted tobacco products) according to age, sex, and time since cessation, with the last variable evaluated for cigarettes only.18 We used the 2015 National Youth Tobacco Survey to determine the prevalence of current use and nonuse of cigarettes and noncombusted tobacco products among children and teenagers under the age of 18 years, according to age and sex19 (see the Supplementary Appendix). We derived inputs for annual births and net migration20 according to sex from Census projections for 2015 through 2060; we projected inputs for births and migration for the period from 2061 through 2100 using a state space model for exponential smoothing.21,22 We used NHIS data from 2011 through 2015 to estimate the prevalence of smoking among immigrants according to sex.18 In these analyses, we assumed no consumption of noncombusted tobacco products by immigrants.
We applied the rates of death among persons who had never smoked that were derived from the NHIS Linked Mortality Files15 to persons who had never used tobacco in the model. We adjusted the rates for low mortality in the civilian noninstitutionalized population in the NHIS, which were projected to account for expected improvements in life expectancy with the use of the Lee–Carter mortality forecasting method (see the Supplementary Appendix) and were converted into probabilities of death with the use of standard demographic methods.23,24
We used the NHIS Linked Mortality Files to estimate hazard ratios for death among cigarette smokers, as compared with persons who had never smoked, according to age and sex, using age as the time scale and adjusting for race or ethnic group, education, alcohol consumption, and body-mass index.15 We also estimated hazard ratios among former smokers according to age at cessation. For some age groups, hazard ratios among persons who had recently quit smoking were greater than among current smokers, as has been observed previously,25,26 since some persons quit smoking because of smoking-related illnesses. We used the observed hazard ratios in the baseline scenario, given that the increased risks among persons who had recently quit smoking were attributable to smoking. In the policy scenario, we capped hazard ratios for former smokers at the levels for current smokers of the same age group and sex, since an increased rate of smoking cessation in this scenario would be due to the policy rather than the cessation of smoking because of illness. To be conservative, we excluded the first 3 years after the implementation of the policy from our cumulative estimates of tobacco-related deaths averted and life-years gained.
We applied estimates of the risk of death for smokeless tobacco use from the Cancer Prevention Study II (CPS-II) to estimate mortality in our model among users of noncombusted tobacco products. In the CPS-II, current users of chewing tobacco or snuff at baseline had a higher risk of death than did persons who had never used such tobacco products (hazard ratio, 1.18), whereas no increased risk was observed among former tobacco users.27 In an analysis involving persons who had switched from cigarette smoking to the use of smokeless tobacco, the risk of death was significantly higher than that among those who had quit smoking entirely (hazard ratio, 1.08).28
Given the limited data on long-term health risks of e-cigarettes, the model applies the risks of using traditional smokeless tobacco to e-cigarette users. Although we recognize that e-cigarettes may vary widely in their attributes and the potential to expose users to harmful and potentially harmful constituents, implicit in our assumption about risk is the fact that since the FDA is responsible for premarket approval of new tobacco products, including e-cigarettes, over time the market would come to be dominated by the least harmful of these products.
Among adults who are 35 years of age or older, the model applies a relative risk of 1.18 for current users of noncombusted tobacco products, as compared with those who had never used tobacco products, and a relative risk of 1.08 for former cigarette smokers who subsequently use noncombusted tobacco, as compared with former smokers who did not use such tobacco products. In sensitivity analyses, we applied relative risks up to 1.50 for current users of noncombusted tobacco products, as compared with persons who had never used tobacco products, and relative risks up to 1.30 for former cigarette smokers who currently used noncombusted tobacco products, as compared with former smokers who did not use such products. We assumed that dual users of cigarettes and noncombusted tobacco maintained the same risk as
cigarette smokers who did not use noncombusted tobacco products.
Inputs Regarding Tobacco-Use Behavior
Annual rates of smoking initiation and cessation were derived by Cancer Intervention and Surveillance Modeling Network (CISNET) researchers on the basis of analyses of NHIS data from 1965 through 2015.29 We generated sex- and age-specific initiation rates for exclusive cigarette use, exclusive use of noncombusted tobacco products, and dual use by scaling the 2015 rates according to the prevalence estimates for current use of cigarettes, smokeless tobacco, and e-cigarettes from the 2015 National Youth Tobacco Survey19 (see the Supplementary Appendix). We used smoking-cessation rates from 2015 for cessation of both cigarettes and noncombusted tobacco products. In the baseline scenario, age-specific initiation and cessation rates were assumed to remain constant in all years, with no new product initiation (and therefore no new switching between products or new dual use) after the age of 30 years. This assumption was relaxed in the policy scenario, which allowed for uptake of noncombusted tobacco among smokers at any age, either as dual users or product switchers. During model development, we conducted sensitivity analyses that allowed product switching in the baseline scenario, a variable that did not materially affect the results. We also conducted sensitivity analyses in which we assumed that baseline rates of smoking initiation in the future would be 20% higher and 20% lower than those estimated for 2015.
Policy Scenario Inputs
We obtained data inputs for the policy scenario from a formal expert elicitation, which is a systematic process of formalizing and quantifying judgments about uncertain quantities. This process is typically conducted with subject-matter experts who provide subjective probability distributions for questions of interest. A contractor selected experts on the basis of mutually agreed upon, prespecified criteria that identified authors with extensive publication records on relevant topics. Candidates were required to certify that they had no actual, apparent, or potential conflict of interest in any tobacco-related business or any nicotine- or tobacco-related pharmaceutical products.
Eight experts were asked to provide estimates of the anticipated effects of a hypothetical policy that would require the reduction of nicotine in cigarettes to minimally addictive levels. This reduction would be achieved through setting a maximum limit on the amount of nicotine in cigarette tobacco filler and, therefore, the amount that could be extracted by the user. Experts were asked to assume that combusted tobacco products that are highly likely to serve as substitutes for traditional cigarettes (e.g., roll-your-own tobacco, pipe tobacco, and nonpremium cigars) would be included in the policy, whereas other tobacco products (e.g., premium cigars, water pipe or hookah, e-cigarettes, and smokeless tobacco) would be excluded. Experts estimated the effect of the policy on rates of cigarette-smoking cessation, switching from cigarette smoking to products excluded from the policy, dual use, cigarette-smoking initiation, and initiation of products excluded from the policy. For this model, we made the simplifying assumption that switching to and initiation of tobacco products that were excluded from the policy would be restricted to noncombusted tobacco products. This assumption was largely consistent with the views of the experts.
Experts were asked to provide their best estimate of the true value of each variable, minimum and maximum plausible values, and the 5th, 25th, 75th, and 95th percentile values. They were also asked to estimate the effects of the policy for the year immediately after implementation and in subsequent years. Experts could provide separate estimates for men and women. To account for uncertainty in responses to the policy, we used the distributions of the experts’ estimates in a Monte Carlo simulation. There were 20 distribution responses associated with each expert; these captured each expert’s response for each of the five questions (related to cessation, product switching, dual use, cigarette initiation, and initiation of other products), including differences according to sex and year (first year after implementation vs. subsequent years).
The distributions of the responses from the eight experts varied widely (Table 1). For example, the experts’ median estimate of the percentage of smokers who would quit smoking in the first year after the introduction of the policy ranged from 4.5 to 55.0%, and estimates for subsequent years ranged from 4.5 to 80.0% (Tables F4 and F5 in the Supplementary Appendix). Estimates of the percent change in annual rates of cigarette-smoking initiation that resulted from the policy in its first year were similarly variable, ranging from −21 to −70% for the median estimate; for subsequent years, median estimates ranged from −21 to −75% (Tables F14 and F15 in the Supplementary Appendix). For each expert’s distributions, a Latin Hypercube sampling with 1000 sample values was performed, resulting in a total of 8000 simulations. In the simulation, the policy scenario is introduced in 2020. We ran the model using each of the 8000 sample parameters, and results were aggregated into one set of output distributions. We report median estimates from the output distributions, with ranges that represent 5th and 95th percentile estimates. (Additional details regarding the expert elicitation and statistical methods are provided in the Supplementary Appendix.)