We used the 2001 and 2002 MEPS to develop a model that predicts smoking-attributable medical expenditures for the Medicaid population. MEPS is a nationally representative survey of the civilian, noninstitutionalized population that quantifies each participant's total annual medical spending, including expenditures from public- and private-sector health insurers and out-of-pocket payments. The data also include information about each participant's source of health insurance (eg, any evidence of Medicaid coverage during the year) and sociodemographic characteristics (such as race/ethnicity, sex, and education). Information about MEPS is available at www.meps.ahrq.gov/mepsweb/.

The MEPS sampling frame is drawn from participants in the National Health Interview Survey (NHIS). NHIS is a nationally representative survey that collects data on selected health topics. Although MEPS does not capture information on smoking, self-reported smoking variables are available for a subset of adult NHIS participants (the Adult Sample File) and can be merged with MEPS data. We used responses to the question "Have you smoked at least 100 cigarettes in your entire life?" to differentiate between ever smokers and nonsmokers. We excluded from the analysis sample respondents with missing data on smoking variables (≈1% of respondents aged ≥18 years and all respondents aged <18 at the time of the NHIS interview) and those who did not receive Medicaid coverage. Our final MEPS-NHIS population included 1,588 adults with weighting variables that allowed us to generate nationally representative estimates of the adult, civilian, noninstitutionalized Medicaid population (Table 1).

Before constructing our national model, we used the Medical Care component of the Consumer Price Index to inflate all MEPS annual medical spending data to 2004 dollars.

The BRFSS is a state-based telephone survey of the adult (aged ≥18), noninstitutionalized population that tracks health risks in the United States. The most recent BRFSS surveys do not allow for stratifying participants by type of health insurance. This information was, however, available before 2001. Therefore, we used 1998-2000 BRFSS data to predict state-level medical expenditures for the Medicaid population. Information about BRFSS is available at www.cdc.gov/BRFSS/. Although BRFSS does not collect medical expenditure data, it includes information about each participant's smoking status, insurance status (before 2001), and sociodemographic characteristics (such as race/ethnicity, sex, and education). Because these variables match those from MEPS-NHIS, we were able to construct an expenditure prediction model with MEPS-NHIS data and use the results to generate expenditure estimates for smokers and nonsmokers on the basis of state-representative population characteristics of BRFSS participants.

As we did with our MEPS-NHIS restrictions, we excluded those with missing smoking data (≈1%) and those who did not receive Medicaid coverage. Our final BRFSS population included 16,201 adults with weighting variables that allowed us to generate state-representative estimates of the adult, noninstitutionalized Medicaid population (Table 1).

Estimating state-specific smoking-attributable medical expenditures for the Medicaid population involved 3 steps. First, we used MEPS-NHIS data to create a model that predicts annual medical expenditures for Medicaid recipients as a function of smoking status, body weight, and sociodemographic characteristics. Second, we used state-representative BRFSS data and results from our MEPS-NHIS national model to estimate the fraction of medical expenditures for Medicaid recipients that was attributable to smoking for each state. Third, we multiplied these fractions by previously published estimates of state-specific Medicaid expenditures to compute smoking-attributable Medicaid expenditures for each state. These steps are described in detail below.

We used a 4-part regression model to predict annual medical expenditures for each MEPS-NHIS Medicaid recipient. The 4-part regression approach was pioneered by authors of the RAND Health Insurance Experiment to control for several unique characteristics of the medical expenditures distribution and is now commonly applied to medical expenditures data (12,13). The model estimates predicted expenditures by using the following functional form: EXP = Pr(C × EXP_IP + [1 − C]EXP_NIP), where EXP represents predicted annual expenditures; Pr represents the predicted probability of positive medical expenditures during the year and is estimated with a logistic regression model; C represents the conditional probability of positive inpatient expenditures, given positive expenditures, and is estimated with a logistic regression model; EXP_IP represents ordinary least squares (OLS)-predicted medical expenditures, given positive inpatient expenditures during the year; and EXP_NIP represents OLS-predicted medical expenditures, given positive expenditures but no inpatient expenditures.

All OLS regression models are estimated on the logged expenditure variable to adjust for the skewness in annual expenditures (mean annual expenditures are significantly greater than the median). Logged expenditures are converted back to expenditures by using the homoscedastic smearing factor (14).

Including dummy variables that indicate smoking status (ever smoked set equal to 1 and the referent group, never smoked, set equal to 0) in each regression model allowed us to quantify the effect of smoking on annual medical expenditures. In addition to smoking status, all regressions controlled for other variables assumed to influence annual medical expenditures, including self-reported body weight, sex, race/ethnicity, age, region of residence, education, and marital status. Regression models were estimated by using SUDAAN version 8 (RTI International, Research Triangle Park, North Carolina) to control for the complex survey design used in MEPS-NHIS. Table 2 presents results from the 4-part regression model.

We used the coefficient estimates from the MEPS-NHIS models to predict annual medical expenditures for each BRFSS Medicaid recipient. To do this, we multiplied each person's characteristics (the independent variables) by his respective coefficients generated from the 4 MEPS-NHIS regression models and combined the results with the equation above. Using the BRFSS weighting variables and each person's predicted medical expenditures, we computed total predicted medical expenditures for each state's Medicaid population.

We estimated smoking-attributable medical expenditures as the difference between predicted expenditures for ever smokers and predicted expenditures for nonsmokers, leaving all other variables unchanged. This method allowed us to isolate the effect of smoking while maintaining any other population characteristics that may contribute to higher annual medical expenditures among smokers.

For the Medicaid population in each state, the percentage of aggregate medical expenditures attributable to smoking was calculated by dividing aggregate predicted expenditures attributable to smoking by total predicted expenditures for adult Medicaid recipients in each state. Because BRFSS is limited to adults, our results should be interpreted as the fraction of adult medical expenditures that are attributable to smoking among adults in each state.

For a variety of reasons, including the lack of data on institutionalized populations, MEPS national spending estimates (and state-level spending estimates based on MEPS) underestimate actual US health care spending (15). Therefore, to quantify annual adult smoking-attributable medical expenditures for each state, we multiplied our state-by-state smoking-attributable fractions by published estimates of 2001 state-specific Medicaid expenditures, available from the Centers for Medicare and Medicaid Services (16). We used 2001 because it is the most recent year that annual, state-specific Medicaid expenditure estimates are available. To match our regression population, we limited Medicaid expenditures to those accrued by adult recipients (≥18 years). We then inflated medical expenditure estimates to 2004 by using a national adjustment factor (1.31). This adjustment factor, calculated as the ratio of 2004 projected expenditures (actual expenditures not yet available) to 2001 actual expenditures, was based on data from Centers for Medicare and Medicaid Services National Health Expenditure Accounts, generally considered the standard for measuring annual health care spending (17).

The MEPS-NHIS national model that was used to calculate our state-level estimates is an improvement on a previous study that used data from the 1987 National Medical Expenditure Survey (NMES) to estimate smoking-attributable Medicaid expenditures (11). Results from the 1987 NMES are dated, and unlike NHIS, many of the key smoking variables that NMES used were imputed (25). Using recent data and actual, as opposed to imputed, smoking information in our calculations provides states with updated and accurate information that may better inform policy decisions. In addition, these differences may, in part, explain why the nationwide Medicaid smoking-attributable fraction of 11.0% is more conservative than the previous estimate of 14.5% generated for 1993 (11). Other changes that may account for the difference in our estimated smoking-attributable fraction include potential changes in the number of people treated for smoking-related illness from 1993 to 2002 and any change in treatment disposition from inpatient to outpatient care. Finally, our estimates differ from previous estimates, and probably understate Medicaid smoking-attributable expenditures, because they exclude expenditures for nursing home care, which are not available in the MEPS-NHIS model.

Despite these strengths, our study has several limitations. First, both the MEPS-NHIS and BRFSS are limited to noninstitutionalized populations, but we apply the resulting smoking-attributable fractions to expenditure estimates that include both institutionalized and noninstitutionalized populations. If these fractions are different for the institutionalized population, our expenditure estimates would be biased. Second, data limitations precluded us from quantifying smoking-attributable medical expenditures for smokers younger than 18 years and nonsmokers exposed to secondhand smoke. The effects of secondhand smoke on children's health are considerable (7). Secondhand smoke exposure can lead to acute lower respiratory infections, such as bronchitis and pneumonia in infants and young children, and can cause children who already have asthma to experience more frequent and severe attacks (26). Although health care expenditures attributable to secondhand smoke exposure may be high, quantifying these expenditures is difficult. As a consequence, our estimates understate smoking-attributable expenditures. Third, our analysis is limited to health care expenditures and therefore does not address other expenses (eg, disability, decreased productivity, absenteeism) that result from smoking (7). Finally, because our focus was not to test statistically whether smoking-attributable expenditures were larger in some states than others, we did not calculate standard errors at the state level.

An estimated 443,000 Americans die prematurely each year as a result of smoking or exposure to secondhand smoke (27). Medicaid recipients are disproportionately affected by tobacco-related disease because their smoking prevalence is approximately 53% greater than that of the overall US adult population (10). In addition to the individual health toll, the disproportionately higher smoking prevalence among Medicaid recipients imposes substantial costs on society. We estimate that smoking accounts for approximately 11% of Medicaid program expenditures. To improve the health of Medicaid recipients and potentially reduce the growth rate of expenditures, Medicaid programs in all 50 states and the District of Columbia are encouraged to follow the 2000 PHS guidelines and cover all recommended tobacco-dependence treatments and approved medications (18). The cost-effectiveness of these programs, combined with the high cost of smoking, suggests that such coverage may provide cost savings to the financially strapped Medicaid programs.