Simulation model

We employed a microsimulation model, ModelHealth^TM: Tobacco, developed in Java. Previous iterations of the model were developed for the US as a whole¹⁴ and Minnesota^15,16. The adaptation of the model to Tobacco Nation states has been previously described¹⁷ and is detailed in Supplementary file Section 1. Here, we briefly describe aspects of the microsimulation model pertinent to the current analysis.

The model simulates annual changes in smoking status and the health and economic harms of smoking by poverty status and race. For the current analysis, we simulated two different age-adjusted populations to facilitate comparisons among population groups within states: one in which the two poverty status groups within each state have the same age distribution as that of the state as-a-whole, and a second in which the race-ethnicity groups have the same age-distribution of the state¹⁸. We measured household poverty status as a dichotomous indicator of living at or below the 138% of the FPL based on household size and total household income. Race and ethnicity in the simulation model are represented by four broad categories (Hispanic of any race, non-Hispanic Black, non-Hispanic White, and all other races) that are available in the demographic, tobacco use, and disease burden data sets used to parameterize the model.

Within the simulation model, individuals’ poverty status does not change over time. However, keeping poverty status for individuals static as they age does not meaningfully limit population-level estimates of policy effects because the portion of the population living at less than 138% of the federal poverty level is relatively stable across age groups. The age, sex, race and ethnicity, and poverty status composition of the population does change over time in the model due to differences in population composition, as new individuals age into the model and other individuals exit due to death. Individuals who are born into the model as the simulation progresses are assigned the race-ethnicity probabilities of baseline 12-year-olds.

We estimated the probability of youth smoking and net initiation by state of residence, age, sex and race-ethnicity from 2019 Youth Risk Behavior Surveillance System (YRBSS) and we estimated adult cigarette smoking status and cessation probabilities by the same characteristics plus education level and poverty status using combined 2016 to 2019 BRFSS surveys^19,20. We estimated long-term relapse probabilities from literature as described in Supplementary file Section 1.

The model simulates smoking-attributable (SA) diseases identified in Smoking-Attributable Mortality, Morbidity, and Economic Costs (SAMMEC) as updated in 2014¹⁰. We obtained incidence and deaths from SA cancers by state, sex, and race/ethnicity from data that inform the US Cancer Statistics Data Visualizations Tool²¹. We obtained deaths for other SA conditions from Detailed Mortality Data²². We used hospitalizations to measure annual cardiometabolic and respiratory disease events. We assigned hospitalization rates for each state’s Census Division as tabulated from the 2018 National Inpatient Sample using the Healthcare Cost and Utilization Project (HCUP)²³. As described in Supplementary file Section 1, we disaggregated cancer incidence and hospitalization rates by smoking status using standard attributable-risk calculations¹⁷ and the relative risks of mortality of current and former smokers relative to never smokers from the 2014 Surgeon General’s Report¹⁰.

Simulation of tax increase scenario

We obtained the average per-pack price of cigarettes by state in 2021 from The Campaign for Tobacco Free Kids (CTFK)²⁴. In the tax increase policy scenario, we assume the $1.50 tax adds $1.50 to the point-of-sale price in the first year. We determined the difference in the risk of smoking initiation among youth aged 12–24 years and of smoking cessation among adults over 24 years, in response to a price change based on price elasticity estimates obtained from structured literature reviews (described and referenced in Supplementary file Section 2). Based on the literature review, we used a price elasticity for those at or below 138% of the FPL that is 2.56 greater than the elasticity for those above that threshold. We obtained elasticities for young adults below and above 138% of the poverty level of -0.516 and -0.201, respectively, and -0.295 and -0.115 for adults aged ≥25 years.

We assumed the tax will have a one-time effect on smoking cessation in the year of the increase and an ongoing effect on smoking initiation because new youth and young adult cohorts will be exposed to higher prices as they age into years of risk for starting smoking.

Simulation of simultaneous tax and tobacco control expenditure increase scenario

A second policy scenario simulates a simultaneous increase in cigarette taxes by $1.50 per pack and an increase in state-level TCE to the CDC-recommended level for each state. Following the literature, the simulation uses state-level TCE as a measure of the intensity of comprehensive tobacco control. We conducted a structured literature review for expenditure elasticities and chose estimates from studies among youth, young adults and other adults which employed the same core statistical analysis for each age group (Supplementary file Section 2)^25-27. In the simulation, changes from baseline inflation-adjusted per capita TCE results in new smoking initiation and cessation probabilities each simulated year. The literature review did not reveal any consistent modification of the benefit of expenditure increases by socio-economic status or race.

Conducting the simulation and sensitivity analysis

We compared outcomes in three scenarios: tax increase, simultaneous tax increase and TCE increase, and no policy change (i.e. static policy) scenarios. For every state, we conducted 30 sets of simulations for 1 million individuals, for each scenario with a different random number sequence for each of the 30 sets. To ensure that the difference between the scenarios was attributable only to policy effects, the model used the same random number sequence in each scenario within a set.

We used each state’s 2021 population²⁸ to compute weighted-average Tobacco Nation effects for the simulation results of individual states. We then computed a weighted-average for non-Tobacco Nation states as the difference between the US results and the Tobacco Nation average.

Internal validation using model-testing sensitivity analyses ensured that differences in model outputs were consistent with changes to model inputs and with differences in inputs among states and demographic groups. External model validation ensured that the model produced expected population demographics, smoking status, and disease rates consistent with baseline summary statistics from the model input databases and that the outputs of future years were not out-of-bounds of reported effects from similar simulation models. Validation was conducted primarily to ensure that the extensive input tables were correctly populated (providing a second, indirect check on the SAS code that produced the input tables) and that the model’s Java code read the input tables and performed tabulations correctly. No calibration was conducted for model validation or other reasons.

Sensitivity analyses

We previously reported that results for the overall population are most sensitive to different estimates of TCE elasticities, baseline smoking cessation probabilities, the baseline incidence of SA disease, and price elasticities²⁹. The sensitivity analysis reported in Table 5 and Supplementary file Tables B–E for Tobacco Nation states indicate, that changes to these variables have little influence on the ratio of effects between population groups when the changes are applied equally to all population groups. In other scenarios, even when the price elasticity is set equal for those above and below the 138% of the poverty threshold, those living below this threshold still benefit more from a tax increase alone (Table 5).

Our literature review did not indicate that historically disadvantaged individuals were more likely than other individuals to modify smoking behavior with changes in TCE. However, we constructed a scenario to examine what might happen if programs could be targeted such that adults in households below 138% of the federal poverty level were 50% less likely to initiate smoking and 50% more likely to quit if they are smoking, similar to the relative effect of tax increases. The results of this hypothetical scenario (Supplementary file Table C, see 'Expenditure elasticity 50% higher below poverty threshold than above poverty threshold') indicate that simultaneous price and expenditure policies would reduce smoking about 20% more for those below 138% of the poverty threshold compared to the base case (9.5% vs 7.8%).

Limitations

The model used in this study is limited by model inputs and assumptions as previously described and evaluated in sensitivity analysis¹¹. Notably, our sensitivity analyses did not indicate that changes to model parameters altered the conclusions about relative magnitude of benefits of tax and expenditure policies by population groups. Nevertheless, readers should keep in mind that residual confounding may bias our estimates of some model inputs, including our indirectly tabulated estimates of the burden of tobacco-attributable illness (Supplementary file Section 1) and policy effectiveness derived from literature based, by necessity, on observational studies.

It is important to consider that while we incorporated differences in the risk of smoking-attributable mortality by race and ethnicity, mortality data do not record measures of SES. As a result, any impact of poverty status on smoking-attributable mortality independent of smoking status, age, sex, and race and ethnicity is not reflected in the simulation results. We chose not to incorporate from secondary data any differential mortality from non-smoking-attributable causes of death by poverty status, race or ethnicity. We made this choice to avoid incorporating societal, health system, or environmental bias in a manner that quantifies lower gains in life years and QALYs (quality-adjusted life years) of preventing a smoking-attributable death in disadvantaged individuals³⁴. The trade-off is that our results may overstate the actual gains of tobacco control policy for disadvantaged individuals given persistent disparities in life-expectancy.

Additionally, the simulation model was constructed and validated using data from late 2018 to 2021, and therefore we describe our results as covering the 20 years starting in 2021. Smoking rates have continued to fall since 2021 (as predicted by the model), the e-cigarette market and the market for other novel tobacco products continues to evolve, and wage growth for some lower income workers have outpaced inflation in recent years. Therefore, while it is reasonable to expect the absolute benefits predicted for 20 years starting today would be smaller due to lower cigarette smoking prevalence, it is less clear whether the percent reduction in tobacco harms by population groups will be meaningfully impacted by current and future changes to the tobacco use environment.