^{1}

^{2}

^{2}

^{3}

^{4}

The authors have declared that no competing interests exist.

Conceived and designed the experiments: JL SAG. Analyzed the data: JL SAG. Wrote the first draft of the manuscript: JL. Contributed to the writing of the manuscript: JL SAG. Agree with the manuscript’s results and conclusions: JL SAG. All authors have read, and confirm that they meet, ICMJE criteria for authorship.

Reductions in smoking in Arizona and California have been shown to be associated with reduced per capita healthcare expenditures in these states compared to control populations in the rest of the US. This paper extends that analysis to all states and estimates changes in healthcare expenditure attributable to changes in aggregate measures of smoking behavior in all states.

State per capita healthcare expenditure is modeled as a function of current smoking prevalence, mean cigarette consumption per smoker, other demographic and economic factors, and cross-sectional time trends using a fixed effects panel data regression on annual time series data for each the 50 states and the District of Columbia for the years 1992 through 2009. We found that 1% relative reductions in current smoking prevalence and mean packs smoked per current smoker are associated with 0.118% (standard error [SE] 0.0259%,

Changes in healthcare expenditure appear quickly after changes in smoking behavior. A 10% relative drop in smoking in every state is predicted to be followed by an expected $63 billion reduction (in 2012 US dollars) in healthcare expenditure the next year. State and national policies that reduce smoking should be part of short term healthcare cost containment.

James Lightwood and Stanton Glantz report on a modeling study in which health spending across US States during 1992–2009 was studied as a function of smoking prevalence, indicating that reduced smoking could yield substantial early benefits in healthcare expenditure as well as long-term improvements in health.

There have been many estimates of the medical costs of smoking at both the national and state levels, but these estimates do not capture the changes in health care expenditure over time that are associated with changes in smoking behavior and the effects of tobacco control programs.

Estimates from California and Arizona have shown that medical savings accrue quickly as the prevalence and intensity of smoking decreases, when adjusted for the history of smoking reduction and tobacco control program activity in the United States.

This study examined the year-to-year relationship between changes in smoking and changes in medical costs for the entire United States, taking into account differences between different states and historical national trends in smoking behavior and healthcare expenditures.

The study found that 1% relative reductions in current smoking prevalence and mean packs smoked per current smoker are associated with 0.118% and 0.108% reductions, respectively, in per capita healthcare expenditure (elasticities).

Historical regional variations in smoking behavior (including those due to the effects of state tobacco control programs, smoking restrictions, and differences in cigarette taxation rates) are associated with substantial differences in per capita healthcare expenditures across the United States.

A 10% relative drop in smoking in every state is predicted to be followed by a $63 billion reduction (in 2012 US dollars) in healthcare expenditure the next year.

Changes in healthcare costs appear quickly after changes in smoking behavior.

State and national policies that reduce smoking should be part of short term healthcare cost containment.

Smoking causes a wide range of diseases, including cardiovascular and pulmonary disease, complications of pregnancy, and cancers [

Previous research found that increases in per capita funding for population-based tobacco control programs in California [

This paper estimates how much on average a 1% relative reduction in smoking prevalence in a US state reduces health costs in that state a year later. The analysis estimates this association (elasticity) while controlling for the effects of a variety of other differences between states that may produce a spurious association between reduction in smoking prevalence and reduced health expenditure, e.g., changes in population composition and other health behaviors that may also reduce health expenditure. To obtain this estimate for each state, we use a regression approach, with various refinements that take account of correlated time series. In the main and supplemental sensitivity analysis, we control—as much as possible when using state aggregated data—for the effects of other variables that may influence health care expenditure at the state level in addition to smoking (e.g., demographic factors, such as population age composition and ethnic composition; other health risk behaviors in the population, such as alcohol use; and obesity). We also control for the possible effects of unmeasured variables (e.g., cross-state cigarette purchases) on the validity of the measure of cigarette consumption per smoker in each state.

The dependent variable in the regression model (

Because available data on mean consumption per smoker may be contaminated with measurement error that increases over the sample period due to increasing interstate tax differentials, the individual state cigarette tax rates are included to adjust for the effects of this possible measurement error. Other state-specific control variables that might affect per capita healthcare expenditure are included. To account for long run trends in healthcare expenditure that are correlated with the observed state-specific explanatory variables as well other correlated but unobserved trends, the national averages of the dependent and explanatory variables are included in the regression. Finally, state-specific intercepts are included in the regression to model regional and state-specific factors that may affect state healthcare expenditure and that remain constant over the sample period. All the independent (explanatory) variables are lagged by 1 y.

Measures of smoking behavior, the other population factors we are considering, and healthcare costs change over time unpredictably because of changes in technology, access to care, and the nature of the population itself. From a statistical perspective, that means that the underlying process is nonstationary, and we need to account for this in the analysis. To do so, we also include the national cross-sectional averages of the dependent and independent variables as independent variables in the regression equation to account for their long run trends and trends in other correlated but unobservable variables associated with per capita healthcare expenditure that vary over the sample period [

There is also a possibility that the reported cigarette sales in a state (which we used to estimate annual per smoker cigarette consumption) might not be equal to the numbers of cigarettes smoked in a state. To adjust for possible measurement error in mean cigarette consumption per smoker, state-specific cigarette tax rates are also included in the regression model (

The independent variables are taken from the year before the healthcare expenditure data (i.e., lagged by 1 y), to allow for time for the independent variables to affect healthcare expenditure.

The estimated effects of smoking on healthcare costs are based on cross-sectional time series (panel) data on smoking, healthcare costs, and demographics for the 50 states and the District of Columbia (considered and referred to hereafter as 51 “states”) for the years 1992 through 2009.

The main results use the Centers for Medicare and Medicaid Services (CMS) estimates of total (public and private payer) healthcare expenditure by state of residence [

Prevalence of current smoking and state and federal cigarette tax data were from the Behavioral Risk Factor Surveillance System (BRFSS) provided by the Centers for Disease Control and Prevention (CDC) State Tobacco Activities Tracking and Evaluation (STATE) System [

Total state resident population data and the proportion of state resident population age 65 y or older were from the US Census Bureau [

All monetary values are expressed in year 2010 US dollars using the regional medical care (for healthcare expenditures) and regional all-item (for cigarette taxes and personal income) Consumer Price Index for All Urban Consumers (CPI-U) [

There were up to 18 annual observations for the individual 51 states, making 918 data points. There are only 27 missing data points (2.9%) because of individual states not participating in the BRFSS in some years. All but three missing observations are due to delayed entry of 11 states into the BRFSS or a BRFSS component. Fisher’s exact test and continuity-corrected Spearman’s and Kendall’s tau-a correlation coefficients were used to evaluate the association between the presence and length of lagged state entry into BRFSS and each state’s smoking behavior and socio-demographics used in the analysis, state population, and geographic region. No statistically significant geographical or socio-demographic or economic relationships were found to explain the patterns of delayed entry among the states, so we consider the missing observations to be missing completely at random.

The regression model explains state per capita healthcare expenditure as a function of state per capita income, population age structure (proportion of the population that is elderly), proportion of the population that is African-American, proportion of the population that is Hispanic, and additional control variables that describe national trends in health care expenditure, such as changes in medical technology and the market for health care. Other variables that may affect the results were missing for some years and states, such as prevalence of insurance coverage and prevalence of other health risks (e.g., obesity and high blood pressure). A sensitivity analysis (detailed in

Previous research compared smoking behaviors and per capita healthcare expenditures in California [

The model used for these national estimates has two parts (

The natural logarithm of state per capita healthcare expenditure in each state is explained using the lagged natural logarithms of state smoking prevalence, mean cigarette consumption per smoker, per capita income, and several demographic variables and the lagged natural logarithms of their associated national averages across all the states. Using logarithms in this way yields regression coefficients that are interpreted as elasticities, which are dimensionless constants that give the percent change in the dependent variable associated with a 1% (relative) change in each explanatory variable. The logarithmic transformation produced better behaved residuals for individual state data than the linear specifications used in earlier work [

The second part of the model adds an adjustment for possible measurement error in individual state observations of mean cigarette consumption per smoker due to untaxed cigarette consumption induced by differences in state cigarette taxes. A state-specific model for this type of measurement error (that would use different coefficients for each of the 51 states) led to severe multicollinearity and model specification problems, so the eight BEA economic regions were chosen as the most appropriate grouping for modeling variations in the effect of the individual state-specific cigarette tax rates over time. In particular, we retained information on individual state variation in cigarette tax rates while restricting the associated coefficients’ values regionally. The BEA regions were chosen for the regional pattern of cigarette tax adjustment effects because the BEA regions reflect economically homogenous groups of states [

Several sensitivity analyses were conducted to check the possibility that the estimates that attribute changes in population health to smoking are related to other risk factors than smoking (and secondhand smoke exposure). The results of these sensitivity analyses are summarized below. Detailed results appear in

The prevalence of other health risk factors were measured in the BRFSS surveys (prevalence of high blood pressure and high cholesterol among respondentswho had those checked, prevalence of abusive drinking, no insurance coverage, no regular exercise, diabetes, and obesity), and these prevalence estimates were all added to the final model (

Description of Variable | Variable | Coefficient (Elasticity) | Standard Error | |
---|---|---|---|---|

Prevalence of smoking | ln(_{i, t−1}) |
0.118 | 0.0259 | <0.001 |

Cigarette consumption per smoker | ln(cps_{m, i, t−1}) |
0.108 | 0.0253 | <0.001 |

Per capita personal income | ln(_{i, t−1}) |
0.224 | 0.0674 | 0.001 |

Percent of population age ≥ 65 y | ln(_{i, t−1}) |
0.530 | 0.0936 | <0.001 |

Percent of population Hispanic | ln(hs_{i, t−1}) |
0.0108 | 0.00763 | 0.156 |

Percent of population African-American | ln(_{i, t−1}) |
0.0130 | 0.00632 | 0.039 |

Cigarette tax, New England | ln(tx_{i, NE, t−1}) |
0.0477 | 0.0103 | <0.001 |

Cigarette tax, Mideast | ln(tx_{i, ME, t−1}) |
0.0203 | 0.0106 | 0.056 |

Cigarette tax, Great Lakes | ln(tx_{i, GL, t−1}) |
−0.00662 | 0.0151 | 0.660 |

Cigarette tax, Plains | ln(tx_{i, PL, t−1}) |
0.0358 | 0.0179 | 0.045 |

Cigarette tax, Southeast | ln(tx_{i, SE, t−1}) |
0.0190 | 0.0229 | 0.418 |

Cigarette tax, Southwest | ln(tx_{i, SW, t−1}) |
5.45 × 10^{−7} |
0.0248 | 1.00 |

Cigarette tax, Rocky Mountain | ln(tx_{i, RM, t−1}) |
−0.0108 | 0.0131 | 0.409 |

Cigarette tax, Far West | ln(tx_{i, FW, t−1}) |
0.0178 | 0.0312 | 0.568 |

National average per capita healthcare expenditure | ln(hr_{ue, t−1}) |
0.864 | 0.0959 | <0.001 |

Principal component term |
pc3_{ue, t−1} |
−0.564 | 0.132 | <0.001 |

* The “principal component term” is the third principal component of the cross-sectional average terms other than per capita healthcare expenditure. It was the only principal component that entered the regression at the 5% significance level.

Changes in smoking behavior may be correlated to other public health measures and general population awareness of healthy lifestyles, environmental health, and public policies that affect access to care. A sensitivity analysis of possible confounding by these factors was conducted by adding available time series variables that would be correlated with these factors, in the same way as was done for other health risks (

Consistent time series are not available for other factors that may be correlated with unmeasured changes in health risks or public health programs and policies. Perhaps the most prominent such variable is educational attainment in the population. A robustness check of the omission of this variable was conducted by studying the stability of relative state levels of educational attainment across time. Another robustness check was conducted by estimating the correlation over time between state educational attainment and a variable that should be highly correlated: state real per capita personal income.

Additional sensitivity analyses were conducted to evaluate the results of instrumental variable estimation for cigarette consumption per smoker by including instruments for the variables mean consumption per smoker, prevalence of cigarette smoking, per capita income, and proportion of the population age 65 y or older (

The estimated elasticities in

Deviations in per capita healthcare expenditures from the average national level (savings below or excess expenditures above) were calculated for each state in four steps, and then aggregated to the BEA regional level. First, for each state, the arc elasticity estimate of the deviation in state healthcare expenditure attributable to the two smoking behavior variables were calculated by multiplying the estimated elasticities of per capita healthcare for prevalence of current smoking and measured mean cigarette consumption per smoker by the average percent difference between the respective individual state and national averages of the smoking behavior variables over the sample period. The elasticities estimated in the coefficients are valid for modeling the effect of infinitesimal changes in the explanatory variables; the arc elasticity is an adjustment to account for finite differences in the data. Second, the adjustments to per capita healthcare expenditures due to state tax differentials were calculated in the same way: arc elasticities for the tax rates were calculated by multiplying the estimated elasticities of healthcare expenditure by the average percent difference between the respective individual state and national averages of the state cigarette tax variables over the sample period. Third, the net regional healthcare expenditure attributable to smoking adjusted for mismeasurement was calculated for each state by subtracting the results of the second step from the results of the first step, by state. Fourth, the excess per capita expenditures for each BEA region were calculated by taking the simple arithmetic average of each state in each respective region. Total aggregate values for each state and region were calculated by multiplying the state or regional per capita estimates by the state or regional residential populations.

As a check on the reasonableness of the results, the proportion of measured cigarette consumption per smoker due to estimated untaxed consumption was calculated. The calculation was done by dividing the healthcare expenditure due to tax differentials—and therefore attributable to mismeasurement of cigarette consumption (found in step two above)—by the average regional price of cigarettes to calculate the estimated unmeasured consumption in packs of cigarettes per capita. Estimated unmeasured consumption in packs of cigarettes per capita was then divided by the prevalence of current smokers to calculate the estimated unmeasured consumption in terms of packs per smoker. Then the estimated unmeasured consumption in terms of packs per smoker was divided by the measured mean cigarette consumption per current smoker to estimate the estimated unmeasured consumption as a proportion of measured consumption. This estimate gives the proportion of measured cigarette consumption in each region, which can be compared to survey estimates of the proportion of untaxed cigarettes consumed in the United States [

Interval estimates for the excess expenditures and proportion of measured cigarette consumption that is untaxed were calculated using the covariance matrix of the elasticities (which for the logarithmic transformation is the same as the covariance matrix of coefficient matrix of the regression coefficients). The distributions of excess expenditures and proportion of unmeasured cigarette consumption were normally distributed, so formulas for the variances of functions of normal distributions were used to calculate standard errors (SEs).

Because we used the estimated elasticities to calculate the healthcare expenditure attributable to differences in smoking behavior, the estimates are independent of the sample distributions of the other variables in the model. The results can be thought of as quantifying the effects of changes in smoking behavior while holding all the other variables, such as per capita personal income and age distribution of the population, constant.

The elasticities of healthcare expenditure with respect to smoking prevalence and measured mean cigarette consumption per smoker are 0.118 (SE 0.0259, ^{2} statistics indicate that the regression has good explanatory power, particularly for describing variations in per capita healthcare expenditure within each state over time (

^{2} |
Error Structure | ||
---|---|---|---|

Source | Value | Statistics for Regression Residuals | Value |

Within | 0.914 | ρ | 0.940 |

Between | 0.258 | corr(_{i}, Xb) |
−0.291 |

Total | 0.495 | RMSE | 0.0295 |

ρ, proportion of regression error variance due to cross-sectional state-specific constants; corr (_{i}, Xb), correlation between linear state-specific intercept and linear score; RMSE, root-mean-square error.

These estimates of decline in per capita healthcare expenditure associated with changes in smoking behavior are counterfactual predictions that assume that all other factors other than smoking behavior remain constant. The actual observed changes in healthcare expenditure in future years will also depend on additional state-specific variables such as per capita income and age structure of the population, in addition to their evolution via common trends across states.

None of the sensitivity analyses for omitted variables produced a statistically significant or even barely noticeable change in the regression coefficients of the estimated model (

The results of the sensitivity analysis on instrumental variables did not produce evidence of serious bias produced by problems with the instruments used for cigarette consumption per smoker, except for proportion of the population age 65 y or over (

Without adjustment for mismeasurement of cigarette consumption per smoker, the Far West region has the largest estimated savings in annual per capita healthcare expenditure associated with departures of its smoking behavior from the national average: $210 (SE $45.5); the Southeast region has the largest excess expenditure: $154 (SE $30.7) (

Average Excess Expenditure | BEA Region | |||||||
---|---|---|---|---|---|---|---|---|

New England | Mideast | Great Lakes | Plains | Southeast | Southwest | Rocky Mountain | Far West | |

Mean | −37.0 | −34.8 | 62.5 | −21.7 | 66.4 | −6.54 | −119 | −34.5 |

SE | 6.80 | 7.65 | 13.8 | 4.76 | 14.6 | 1.45 | 26.1 | 7.62 |

Mean | 42.1 | −68.6 | −19.1 | 10.9 | 87.8 | −134 | −16.7 | −175 |

SE | 9.86 | 16.0 | 4.50 | 2.55 | 20.5 | 31.4 | 3.90 | 41.1 |

Mean | 5.30 | −103 | 43.4 | −10.7 | 154 | −141 | −135 | −210 |

SE | 9.00 | 21.0 | 12.1 | 4.09 | 30.7 | 32.1 | 28.3 | 45.5 |

Mean | 98.5 | 30.0 | −2.65 | −34.0 | −59.9 | 0.00104 | 14.6 | 28.0 |

SE | 21.5 | 15.8 | 6.01 | 17.0 | 74.2 | 6.29 | 17.8 | 49.6 |

Mean | 0.416 | 0.163 | −0.0165 | −0.141 | −0.236 | 0.00000317 | 0.0791 | 0.164 |

SE | 0.0906 | 0.0860 | 0.0374 | 0.0704 | 0.292 | 0.0192 | 0.0962 | 0.290 |

Mean | 104 | −73.4 | 40.7 | −44.8 | 94.4 | −141 | −121 | −182 |

SE | 25.4 | 25.4 | 11.5 | 17.5 | 90.2 | 34.0 | 32.7 | 51.7 |

Mean | 1,500 | −3,530 | 1,890 | −910 | 7,330 | −5,210 | −1,310 | −9,470 |

SE | 370 | 1,220 | 367 | 356 | 7,010 | 1,260 | 355 | 2,690 |

Data are given as 2010 US dollars per capita unless otherwise indicated. Negative dollar amounts indicate savings compared to national average smoking behavior; positive dollar amounts indicate excess expenditures compared to national average smoking behavior. Negative proportions indicate that estimated true consumption is less than measured consumption; positive proportions indicate that estimated true consumption is less than measured consumption.

After adjustment for state tax differentials, the Far West still has the largest total estimated annual per capita savings, $182 (SE $51.7), but the New England region now has the largest excess per capita expenditure, $104 (SE $25.4); the Southeast has the next largest, $94.4 (SE $90.2) (

The difference between measured and estimated true cigarette consumption per smoker was less than 20% for all BEA regions except the Southeast, where estimated true consumption was 23.6% (SE 29.2%) less than measured consumption, and New England, where estimated true consumption was 41.6% (SE 9.06%) higher than measured (

Survey Estimates [ |
Model Estimates | |||||
---|---|---|---|---|---|---|

Metropolitan Area | Range | Area | Point Estimate | 95% Confidence Interval | ||

Low | High | Low | High | |||

New York City | 47.9% | 49.9% | New York State | 20.1% | 8.02% | 32.2% |

Boston | 36.8% | 38.4% | Massachusetts | 34.2% | 27.5% | 40.9% |

Providence | 29.6% | 55.4% | Rhode Island | 35.3% | 28.1% | 41.9% |

Philadelphia | 1.2% | 1.3% | Pennsylvania | 4.9% | 2.8% | 7.0% |

District of Columbia | 29.0% | 59.9% | District of Columbia | 13.1% | 4.7% | 21.5% |

Survey estimates provide ranges based on modeling assumptions, rather than 95% confidence intervals.

Our estimates provide strong evidence that reducing smoking prevalence and cigarette consumption per smoker are rapidly followed by lower healthcare expenditure. The model is dynamic and predicts per capita healthcare expenditures in the current year as a function of smoking behavior in the previous year. For example, 1% relative reductions in current smoking prevalence and mean cigarette consumption per smoker in one year are associated with a reduction in per capita healthcare expenditure in the next year of 0.118% + 0.108% = 0.226% (SE 0.0363%), with all other factors including common trends held equal. In 2012, total healthcare expenditures in the US were $2.8 trillion [

These are short run 1- to 2-y predictions, and while they indicate that the effects of changes in smoking on healthcare expenditure begin to appear quickly, they do not imply that all changes in the costs and savings of smoking in the population are immediate. If all states reduce their prevalence of smoking and cigarette consumption per smoker, then the corresponding common trends will gradually change over time. The elasticity of the common trend for the prevalence of smoking (from the model estimated with all cross-sectional averages entered as separate variables, rather than using principal components) is relatively small and not statistically significant (−0.0545, SE 0.0581,

These estimates are consistent with previous research on healthcare expenditures attributable to cigarette smoking in California [

This analysis uses aggregate measures of population characteristics to estimate the relationships between smoking behavior variables and per capita healthcare expenditures. The elasticity estimates are not directly comparable to estimates of the economic burden of cigarette smoking using cross-sectional data on individuals in national health surveys [

Our estimates do avoid some problems of estimates based on cross-sectional data. An example is the “quitting sick” effect, which imputes expenditure savings to smokers who quit smoking after being diagnosed with a serious chronic tobacco-related disease, such as lung cancer or cardiovascular disease. The expected expenditure savings from quitting by a smoker who remains well will not be realized in those who quit sick because expensive and irreversible health effects of smoking have already occurred. The quitting sick effect is a consequence of incorrectly imputing missing information (the unobservable health status of the smoker at the time of cessation) that is not present in cross-sectional data. This study uses longitudinal data on measures of smoking behavior and healthcare expenditures on large populations and therefore is not subject to quitting sick effects because the excess health care costs of those who quit sick will be included in a state’s total aggregate healthcare expenditure data along with the reduction in prevalence that occurs when the reduction in smoking of comparable people is recorded in surveys that represent the population of that state. It should be noted that some estimates of the health burden of cigarette smoking that account for quitting sick and other problems with estimates based on cross-sectional data find a higher burden of smoking-related disease and therefore higher smoking-attributable expenditures than most published cross-sectional estimates [

The estimates presented here cannot be used to reliably estimate the change in healthcare expenditure associated with complete elimination of cigarette consumption because the estimated elasticities apply only to modest variation around the status quo, but they do capture expenditures attributable to cigarette smoking in a large population that are difficult to measure from national health surveys (such as the effects of second- and third-hand smoke exposure, and long term effects of developmental problems from premature birth and low birth weight or asthma contracted during childhood, attributable to parental cigarette smoking).

Our methods may suffer from spurious regressions and attribute non-smoking public health factors that are correlated with smoking behavior to the smoking behavior. Specifically, this research does not estimate a smoking attributable fraction of healthcare costs for each state that corresponds to a measure that can be derived from individual survey data. Rather, it estimates the average national effect of variations in aggregate-level state-specific smoking behavior variables around the national trend in those variables on variations in state-specific real per capita healthcare expenditure around its national trend.

The results of this study are subject to the limitations of analysis of aggregate observational data. A study of this nature that uses aggregate data and a relatively small sample size cannot, by itself, establish a causal connection between smoking behavior and healthcare costs, and that is not the goal of this study. Rather, this study should be evaluated in the context of the existing body of research that has already established that the relationship between smoking behavior and healthcare costs is causal using a variety of study designs [

These estimates do not address the issue of whether, over the whole life cycle, a population without any cigarette smoking would have higher healthcare expenditures due to longer lived non-smokers. Forecasting the very long run effects of reductions in smoking over the life cycle in a US population would require the construction of a model to forecast the eventual changes in the age structure of the population and resulting changes in per capita healthcare expenditures as a function of smoking behavior.

Lower smoking prevalence and cigarette consumption per smoker are associated with lower per capita healthcare expenditures. Historical regional variations in smoking behavior (including those due to the effects of state tobacco control programs, smoking restrictions, and differences in taxation) are associated with substantial differences in per capita healthcare expenditures across the United States. Those regions (and the states in them) that have implemented public policies to reduce smoking have substantially lower medical costs. Likewise, those that have failed to implement tobacco control policies have higher medical costs. Changes in healthcare costs begin to be observed quickly after changes in smoking behavior. State and national policies that reduce smoking should be part of short term healthcare cost containment.

(PDF)

US Bureau of Economic Analysis

Behavioral Risk Factor Surveillance System

common correlated effects

Centers for Medicare and Medicaid Services

standard error