does spending more on tobacco control programs make economic sense? an incremental benefit-cost...

DOES SPENDING MORE ON TOBACCO CONTROL PROGRAMS MAKEECONOMIC SENSE? AN INCREMENTAL BENEFIT-COST ANALYSIS

USING PANEL DATA

SUDIP CHATTOPADHYAY and DAVID R. PIEPER∗

This paper presents a benefit-cost analysis of the ongoing, state-level tobaccoprevention and control programs in the United States. Using state-level panel datafor the years 1991–2007, the study applies several variants of econometric modelingapproaches to estimate the state-level tobacco demand. The paper finds a statisticallysignificant evidence of a sustained and steadily increasing long-run impact of thetobacco control program spending on cigarette demand in states. The study also showsthat, if individual states follow the Best Practices funding guidelines, potential futureannual benefits of the tobacco control program can be as high as 14–20 times the costof program implementation. (JEL C2, H5, I1)

I. INTRODUCTION

State-level tobacco control programs gath-ered momentum in the United States duringthe early 1990s and have since been continuingwith a varying degree of success in loweringsmoking prevalence. Recently, Centers for Dis-ease Control and Prevention (CDC) modified its1999 guidelines with a new set of guidelineson recommended funding levels for comprehen-sive tobacco control programs for each state(Best Practices 2007). While states decide theannual funding levels (costs) for the control pro-grams, assessment of the economic benefits ofthe programs, based on their past performance,is rarely undertaken. Unless benefits of the rec-ommended funding levels are shown to clearlyoutweigh the costs of the programs, states maydeviate from the CDC-recommended fundinglevels and divert the tobacco-related revenuesaway from these programs. This is, in fact, arecent trend in many states and is a concernoften expressed in policy debates (American

∗The authors would like to thank the two anonymousreviewers for their helpful comments, which have signifi-cantly improved the paper from its earlier version. Data forthis research are all from published sources and the findingsof this research have no potential conflict with any agency.Chattopadhyay: Department of Economics, San Francisco

State University, 1600 Holloway Avenue, San Francisco,CA 94132. Phone +1 415 338 1447, Fax +1 415 3381057, E-mail [email protected]

Pieper: Department of Geography, University of California,Berkeley, Berkeley, CA 94720. E-mail [email protected]

Lung Association 2009). This paper carries out abenefit-cost analysis of the ongoing, state-leveltobacco control programs for the first time inlight of the CDC’s Best Practices (2007) guide-lines. The analysis is based on several vari-ants of econometric approaches for modelingstate-level tobacco demand, using panel data ontobacco sales from all 50 states for the period1991–2007.

There is a large body of research thatshows that state-level tobacco control programsalong with federally and privately funded initia-tives have played a significant role in tobaccoprevention and control during the last twodecades (Farrelly, Pechacek, and Chaloupka

ABBREVIATIONS

ASSIST: American Stop Smoking Intervention StudyCDC: Centers for Disease Control and PreventionFE: Fixed EffectsIMPACT: Initiatives to Mobilize for the Prevention

and Control of Tobacco UseMSA: Master Settlement AgreementMTF: Monitoring the FutureNCI: National Cancer InstituteNTCP: National Tobacco Control ProgramOLS: Ordinary Least SquaresOSH: Office of Smoking and HealthRE: Random EffectsRWJF: Robert Wood Johnson FoundationTUS-CPS: Tobacco Use Supplement of the Current

Population SurveysVIF: Variance Inflation Factor

1

Contemporary Economic Policy (ISSN 1465-7287)doi:10.1111/j.1465-7287.2011.00302.x© 2011 Western Economic Association International

2 CONTEMPORARY ECONOMIC POLICY

2003; Farrelly et al. 2008; Hu, Sung, and Keeler1995b; Tauras et al. 2005). Farrelly, Pechacek,and Chaloupka (2003) investigate the jointeffects of the funding level and the durationof the control program on tobacco demand andreport that the longer states invest in comprehen-sive tobacco prevention and control programsthe greater and faster the reduction in smoking.A significant amount of additional annual rev-enue being generated as a result of the 1998Tobacco Master Settlement Agreement (MSA)has also opened up further opportunities to helpstates continue to invest in these programs.Unfortunately, in recent years funding levels forsuch programs have shown a steady downturn.For example, total state spending on tobaccocontrol at 2008 prices has dropped from a highof $930.8 million in 2002 to $694.3 million in2007. Moreover, as states face record budgetdeficits due to recession, many are turning tocigarette taxes to raise revenue for spending inother programs. In 2009, only North Dakotafunded tobacco prevention and control programsat the level recommended by the Best Practices(2007) guidelines. Forty states and the Districtof Columbia spent less than 50% of the BestPractices recommended level in 2009 despitereceiving millions of dollars from tobacco taxrevenues and the MSA funds. Funding in thestate of Washington that has been able to bringdown adult smoking rate by 30% and youthsmoking rate by 50% since 2002, has beenreduced by half in the fiscal years 2010 and 2011(American Lung Association 2009). This trendmay adversely affect the long-run sustainabil-ity of the tobacco control efforts and the CDC’slong-term target to reduce adult smoking preva-lence rate down to 10% by 2025 (Best Practices2007).

A benefit-cost analysis of the comprehensivetobacco control programs in light of the CDC’sBest Practices (2007) guidelines involves, as afirst step, a systematic assessment of the per-formance of the existing programs over a longtime horizon. Specifically, we ask the followingquestions in the present research: (1) Does state-level spending on the tobacco control programsduring the past two decades show any indepen-dent impact on tobacco demand? (2) If so, doesthis impact grow over time? The second ques-tion is more important, as it takes a long time toquit smoking and, as such, an immediate impactmay not be discernable (Chaloupka and Warner2000; Gruber 2001). Moreover, continued fund-ing can lead to better infrastructure over time,

which, in turn, can lead to increased efficiency inprogram delivery and evidence-based interven-tions, increased awareness among existing andprospective smokers. Our approach in answer-ing the two aforementioned questions is a dis-tinct improvement over the work of Farrelly,Pechacek, and Chaloupka (2003) in that (1) wetake full advantage of the panel data structureof the state-level data by adopting the fixed-and random-effects models; (2) we model thetobacco demand recognizing that the level ofcontrol funding is endogenous in the model;(3) we model the long-run effect of the tobaccocontrol funding as a function of elapsed time,which is the time since the initial control fundingduring the study period; and (4) we control forthe substitution possibilities of cigarette demandby explicitly considering cigarettes available inthe bordering states as a substitute good.

We extend our analysis of the two aforemen-tioned issues further to include a benefit-costanalysis of the ongoing, state-level compre-hensive tobacco control program for the firsttime. A benefit-cost analysis involves takinginto consideration all direct and indirect costsassociated with smoking. CDC (2006) reportssmoking-attributable costs under three cate-gories, namely medical costs, productivity costs,and Medicaid cost. These costs are based onnumerous epidemiological and actuarial studieslinking health consequences of smoking behav-ior. For example, CDC (2006) reports that in2004 tobacco addiction cost the nation morethan $96 billion per year in direct medicalexpenses as well as more than $97 billion annu-ally in lost productivity. In 2005, the Societyof Actuaries estimated that the effects of expo-sure to secondhand smoke cost the United States$10 billion per year. Thus, any reduction intobacco demand can result in benefits to societyin terms of costs avoided. This research drawson the reported costs in the CDC (2006) reportand links it to the econometric model of tobaccodemand, estimated in the paper, to further esti-mate the annual benefits associated with costsavings at the state level under various scenarios.The study finds a clear evidence of a sustainedand steadily increasing long-run impact of thecontrol program spending on cigarette demand.Moreover, the paper shows that the possiblefuture benefits of the program spending alongthe lines of Best Practices (2007) guidelines faroutweigh the cost.

The remainder of this paper is organized asfollows. Section II presents a brief history of the

CHATTOPADHYAY & PIEPER: BENEFIT-COST ANALYSIS OF TOBACCO CONTROL PROGRAMS 3

state-level tobacco control programs. Section IIIsummarizes the existing literature on this issue.Section IV provides a description of the data.In Section V, we present the tobacco demandmodels. In Section VI, we report and analyze theempirical findings. The benefit-cost analysis ispresented in Section VII. Section VIII providesthe summary and conclusion.

II. A BRIEF HISTORY OF TOBACCO CONTROLPROGRAMS

There have been three major sources of rev-enue for the state-level tobacco control programsduring the last two decades, namely the tobaccotax-generated revenue, a number of federallyor privately funded initiatives, and the fundsreceived by states through the MSA. While thefunding during the 1990s was dominated by thefirst two sources, the MSA funds have provideda significant additional source of revenue duringthe 2000s.

The first major state-level initiative in com-prehensive tobacco control program started inCalifornia in 1988. Massachusetts followed nextin 1992. Each of these programs was fundedthrough tax revenues generated by a sharp $0.25per pack increase in the cigarette tax. Manyother states followed the California and Mas-sachusetts examples by raising tobacco taxes tofund similar programs in the 1990s. During the2000s, the state of Washington raised the stateexcise tax by $0.60 followed by other states.More recently, the American Lung Association(2009) reports that, in 2009 14 states raised taxesto record high levels. As of August 3, 2010, theaverage state cigarette tax is at $1.45 a pack withNew York being the highest cigarette tax state inthe nation ($4.35 per pack) and Missouri beingthe state with the lowest tax (17 cents per pack).

Federally or privately funded initiativesstarted in 1991 with the American Stop SmokingIntervention Study (ASSIST) program. This wasa joint initiative by the National Cancer Insti-tute (NCI) and the American Cancer Society thatsupported 17 states during 1991–1998. In 1993,CDC’s Office of Smoking and Health (OSH)started supporting tobacco control activities in32 of the remaining 33 states (excluding Cal-ifornia) through the Initiatives to Mobilize forthe Prevention and Control of Tobacco Use(IMPACT) program. In 1999, the NationalTobacco Control Program (NTCP) waslaunched, combining the NCI and CDC ini-tiatives into one coordinated national program

funded and managed by the CDC. The CDCfunding is designed to support and leveragestate funding for evidence-based interventionsand to help states evaluate their program efforts.Starting 1994, Robert Wood Johnson Foundation(RWJF) funded many states through its Smoke-Less States program. The program, intendedto strengthen state-level tobacco control poli-cies, initially supported 19 states but later onextended the support to 42 states.

In November 1998, the tobacco industryreached an agreement (MSA) with 46 statesthat sued the industry to recoup funding forMedicaid expenses for treatment of tobacco-related illnesses. Under this settlement agree-ment, the industry would pay $206 billion tostates over a period of 20 years. The remainingfour states, namely Florida, Minnesota, Missis-sippi, and Texas had settled with the industryseparately.

The revenues from the cigarette tax and thefunds from the MSA taken together have placeda substantial amount of funds at the disposal ofthe states, annually. In addition to MSA funds,states, such as Florida, Minnesota, Mississippi,and Texas, also have funds from previous set-tlements.1 Unfortunately, the funding levels arenegligible at best compared to the total tobaccorevenues. Figure 1 depicts the status of the aver-age state tobacco control funding in 2008 dollarsvis-a-vis total revenues from taxes and MSArevenues during the period under study, sepa-rately. It is disheartening to note that despiteproven successes of the programs in reducingsmoking prevalence, states, on average, havespent only between 0.09% (1992) and 4.5%(2002) during the study period.

In order to facilitate and strengthen the com-prehensive tobacco control programs in states,CDC came up with the Best Practices guide-lines for the first time in 1999 and thenagain in 2007 (Best Practices 2007). Best Prac-tices (2007) is a refinement of the set ofits 1999 guidelines in that it is based onevidence-based analyses of tobacco tax-fundedand settlement-funded programs and publishedevidence of effective tobacco control strate-gies. Under the new guidelines, recommendedper-capita annual funding ranges from $9.23

1. MSA funds were given to states as reimbursement forincreased medical expenses from smoking-related diseasesand were not intended to be used solely for tobacco controlprograms. However, access to these funds has enabled statesto mobilize its revenues from other sources to tobaccocontrol programs.


FIGURE 1Control Funding as Opposed to Revenue (in Million Dollars)

(Utah) to $18.02 (DC) with the recommendednational average being $12.34. Recommendedtotal funding ranges from $9 million (Wyoming)to $441 million (California). Unfortunately, fun-ding by states is fast declining with average statespending being 19.4% and 17.0% in 2009 and2010, respectively, of the CDC recommendation(Campaign for Tobacco-Free Kids 2010).

III. REVIEW OF THE CURRENT LITERATURE

The impact of spending by individual stateson tobacco prevention programs has been widelystudied. Interesting studies include Bauer et al.(2000); Biener, Harris, and Hamilton (2000);CDC (1999); Connolly and Robbins (1998); Hu,Sung, and Keeler (1995a, 1995b); Lightwood,Dinno, and Glantz (2008); and Pierce et al.(1998). Manley et al. (1997) provide specificfocus to the ASSIST program by comparing theASSIST states with the non-ASSIST states.

To date the only comprehensive state-levelanalysis that focuses on the effect of tobaccocontrol expenditure on cigarette demand inall 50 states is by Farrelly, Pechacek, andChaloupka (2003). The authors use annual datafor the period 1981–2000 to estimate per-capita cigarette demand and adopt six econo-metric demand specifications, including a laggedspecification for the per-capita control expen-diture and a specification with cumulative per-capita control expenditure. They find statisticallysignificant effects of state-level control fundingon cigarette demand in five of the six models.

The authors conduct a simulation on their esti-mated models and find that in the cumulativeper-capita expenditure model a one-time per-capita funding increase of $6 in 1994 wouldlead to a 1.2% decline in per-capita cigarettesales in 2000 and a permanent increase of per-capita funding by $6 every year starting from1994 would lead to an 8.7% decline in per-capitasales in 2000. This finding has an interestingpolicy implication in that sustained investmentsin comprehensive statewide programs may leadto faster and larger declines in smoking rates.

Tauras et al. (2005) use individual levelannual survey data on youth smoking collectedby the University of Michigan’s Monitoring theFuture (MTF) project for the period 1991–1999to evaluate the success of the tobacco controlprograms. The authors use a two-part (probitand ordinary least squares [OLS]) model andfind that not only funding for tobacco con-trol decreased youth smoking prevalence in the1990s, but increased tobacco control fundingin line with the CDC recommendation wouldhave a substantial impact on youth smokingprevalence. Farrelly et al. (2008) links individ-ual survey response information, obtained fromthe supplements to the Current Population Sur-vey by the Bureau of Census, to the annualcontrol expenditure by states in a logit modelto study the effect of state-level tobacco pro-grams on smoking prevalence during the period1985–2003. They conduct a simulation similarto Farrelly, Pechacek, and Chaloupka (2003)and find that an increase in cumulative control


TABLE 1Data Sources

Variable Description Source

Quantity of cigarettesales

State tax-paid cigarette sales, in millions ofpacks

Tax Burden on Tobacco (2008)

Price Average price of cigarettes in dollars perpack in a state, inclusive of all taxes


Price: bordering states Average price of all the bordering states Based on USA map and Tax Burden onTobacco (2008)

Total tax Average state excise tax and federal tax (indollars) for the year per cigarette pack.


Total tax: borderingstates

Average total tax of all the bordering states Based on USA map and Tax Burden onTobacco (2008)

Personal disposableincome

Total state annual personal disposableincome, in thousands of dollars

Bureau of Labor Statistics: www.bea.gov/regional/spi/default.cfm?satable=summary—Disposable Personal Income

Total state tobaccocontrol funding

Total amount of state tobacco controlprogram funding in million dollars peryear

ImpacTEEN.orgwww.impacteen.org/tobaccodata.htm

Percent college graduates Percentage of population 25 and older witha bachelor’s degree or more

U.S. Census Bureau: www.census.gov/population/www/socdemo/educ-attn.html—detailed Table 13 for each state

Population aged 15–24 Estimate for state population 15–24 yearsof age as of July 1st of each year

U.S. Census Bureau: www.census.gov/popest/states/asrh/index.html—Selected Age Groupsby States

Population aged 25 yearsor more

Estimate for state populations aged25 years or more

Same as above

Unemployment rate Annual unemployment rate by state Bureau of Labor Statistics: www.bls.gov/data/home.htm—Labor Force Statistics from theCPS

Population State population estimates as of July 1st ofeach year

U.S. Census Bureau: www.census.gov/popest/states/asrh/index.html

Consumer price index National consumer price indices convertedto 2008 as base year

Bureau of Labor Statistics: www.bls.gov/data/home.htm—CPI-All Urban Consumers(Current Series)

Smoke-free air law score Total score representing number of laws inplace in a given year


Total Alciati score Measures the extensiveness of state tobaccocontrol youth access laws for a givenstate and year


program expenditure is associated with a signif-icant reduction in smoking prevalence. Marlow(2006) carries out a state-level analysis, sepa-rately for cigarette sales and for youth smokingrates across 45 states and finds little or no evi-dence of the effect of the tobacco control spend-ing on cigarette sales or youth smoking. How-ever, the Marlow study uses only 2 years of data(2001, 2002) and probably missed a lot of vari-ability in program spending by not using a suf-ficiently longer time horizon and thus, the resultof the study should be treated with caution.

IV. DATA

Data for this analysis come from the U.S.Bureau of the Census, the U.S. Bureau of LaborStatistics, the Tax Burden on Tobacco study

published by Orzechowski and Walker (2008),ImpacTEEN.org (http://www.impacteen.org/tobaccodata.htm), and the Tobacco Use Supple-ment of the Current Population Surveys (TUS-CPS). Table 1 provides the details of the datasources. The data are by state for each of the 50states and for each year from 1991 through 2007.This provides a rectangular panel of 50 × 17 =850 observations. For the purpose of analysis,all price, tax, and income data are convertedfrom nominal to real values using constant 2008dollars. Per-capita cigarette sales, per-capita dis-posable income, proportion of the population inthe 15–24 age group, 25 and above age group,and the percentage of college graduates are cal-culated by dividing by the total state popula-tion for each state. Descriptive statistics for key


TABLE 2Descriptive Statistics at the State Level

Variable Observations Mean Std. Dev. Min Max

Cigarette salesPacks sold per adult per year (q) 850 85.5 28.26 31.4 185.7Prices, taxes, incomes, and expendituresPrice per pack in 2008 dollars (p) 850 3.42 0.96 1.94 6.45Average price of all bordering states (psub) 850 3.38 0.89 2.10 6.24Total tax (state + federal) per pack in 2008

dollars (tax )850 0.96 0.47 0.32 3.19

Average total tax of all bordering states (taxsub) 850 0.93 0.38 0.41 2.48Per-capita disposable income in 2008 dollars (pdi ) 850 30,465 4,705 19,875 47,203Total tobacco control program funding in millions of

2008 dollars (totfund )850 9.87 21.78 0 213.73

DemographicsPercent college graduates (pcgr) 850 24.29 5.01 11.4 40.4Percent age of population aged 15–24 (pc1524 ) 850 14.3 1.2 11.7 19.9Percentage of population aged 25 or above (pc25 ) 850 64.61 2.97 52.35 71.20Unemployment rate (unemp) 850 5.1 1.4 2.3 11.3Population (million) (popul ) 850 5.51 6.06 0.46 36.38

variables are summarized in Table 2. The aver-age price and average state tax in borderingstates of a given state are calculated for eachgiven state in the following way: first the groupof all states with physical borders with the givenstate are identified; then a simple average of theprices and a simple average of the total (stateand federal) taxes, averaged over each memberstate in the group, are computed.

V. TOBACCO DEMAND MODELS

A. Modeling Issues

Modeling aggregate demand for cigaretteshas a long history. Interesting research inthis area includes Baltagi and Levin (1985);Chaloupka and Saffer (1992); Becker, Gross-man, and Murphy (1994); Sung, Hu, andKeeler (1994); Evans, Ringel, and Stech (1999);and Farrelly, Pechacek, and Chaloupka (2003).Chaloupka and Warner (2000) provide synthesisof many such papers. In the present research,we take into consideration a number of impor-tant factors discussed in the previous literaturebut not effectively addressed to model tobaccocontrol funding, before specifying a tobaccodemand model.

First, we apply the fixed effects (FE) and ran-dom effects (RE) estimation approaches to con-trol for unobserved, state-specific heterogeneity.There can be numerous unobserved state-level,time-constant effects that cannot be accountedfor in a standard OLS regression analysis, which

make estimation of the effect of the time-varyingcontrol program expenditure biased and incon-sistent.2 There may be other factors in addi-tion to the factors discussed in footnote 2. Forexample, it may often be the case that high taxstates are clustered, which may affect cigarettesales in these states. As the clustering phe-nomenon does not usually change over time,the FE and the RE models can eliminate theclustering effects (Wooldridge 2002). In short,panel data analysis with fixed or random effectscan consistently estimate the effect of controlprogram expenditure by capturing and eliminat-ing the unobserved state-specific effects that canpotentially weaken a standard OLS regressionestimation.

Second, we use state-level total annualexpenditure as an explanatory variable, unlike

2. There are a number of reasons for unobserved hetero-geneity. Different states may have different attitudes towardtobacco usage. For example, states with a strong economicdependence on the tobacco industry (such as North Car-olina, Virginia, and Kentucky) may show different cigarettesales trends. In addition, states have different laws restrictingtobacco usage and different programs to encourage and sup-port smokers trying to quit. Other state-level heterogeneity iscaused by geographic location, including proximity to stateswith lower tax rates and access to Native American reserva-tions and military bases where state cigarette taxes are notassessed. These factors vary among the states but tend tochange slowly, if at all, over time. The models presentedin the Farrelly, Pechacek, and Chaloupka (2003) study haveused 50 state-specific dummies to control for fixed effects.However, in order to effectively eliminate state-level unob-served heterogeneity, in addition to the state dummies, fixed-or random-effect approaches must be used for the purpose.


some earlier studies (Farrelly, Pechacek, andChaloupka, 2003; Farrelly et al. 2008; Hu, Sung,and Keeler 1995a) that use per-capita expendi-ture as the explanatory variable to analyze theimpact of tobacco control funding. Per-capitaexpenditure attaches equal weight to each popu-lation group within a state. However, populationat risk differs from state to state and variouspopulation groups within a state are at varyinglevel of risk depending on a state’s demographyand prevalence of tobacco use. This means thatthe effect size of the control program expendi-ture differs across population groups within astate. Thus, a reliable estimation of the effectsize requires appropriate weighting adjustmentfor calculation of per-capita funding level. Inthe absence of such weighting information itis appropriate to use total state-level fundinginstead of per-capita funding. Also, noting thatthe effect of control funding may vary from stateto state depending on the size of the state, weinclude states’ aggregate population separatelyas control.3

Third, recognizing control funding as astock variable, previous studies model cigarettedemand as a function of cumulative fundinglevel adjusted based on an arbitrary discount-ing factor (Farrelly, Pechacek, and Chaloupka2003; Farrelly et al. 2008; Hu, Sung, and Keeler1995a). We deviate from this approach andhypothesize that control program expenditurehas two independent effects, namely contem-poraneous effect and elapsed time effect, thelatter signifying the time since the initial con-trol funding during the study period. Whilethe magnitude and direction of the former oncigarette demand is unknown and can only beassessed empirically, the latter is expected tohave an adverse effect on cigarette demand and

3. As states do not allocate tobacco control fundingbased on its population size, controlling for population sizein a state is necessary even if one uses per-capita tobaccocontrol expenditure. This has been overlooked in some ofthe earlier studies cited before. This is an issue researchersand policy makers often bring out when comparing the per-formance of the tobacco control program in Massachusettsrelative to California during the nineties (see, e.g., CDC2005). According to them Massachusetts being a smallerstate requires a higher level of per-capita tobacco controlexpenditure compared to California to achieve the samelevel of per-capita reduction in cigarette sales. As we men-tion in the next section, when we control for the size ofthe state with state’s population, the coefficient associatedwith per-capita tobacco control expenditure turns out statis-tically insignificant at any acceptable level. However, in ourproposed specification, with total control funding expendi-ture the coefficient turns out to be significant at 1% level orbetter.

its magnitude is expected to increase over time.The reason being that infrastructure, efficiencyin program delivery, quality of interventions,effectiveness of media campaigns are importantfeatures of a comprehensive tobacco control pro-gram that can only improve over time and thus,the longer states invest in these, the greater theimpact.

There are two advantages of the “elapsedtime effect” formulation over the “cumulativefunding effect” formulation. The first relatesto the fact that in the present study we areinterested in the long-run impact associated witha change in the level of control funding ina given year. For example, consider the twosimple formulations below:

Elapsed time effect formulation:

yit = α0 + α1xit + α2t∗xit + v1t

Cumulative effect formulation:

yit = γ0 + γ1zit + v2t

where yit is tobacco demand, xit represents thelevel of control funding, t denotes time elapsedin years, zit represents cumulative funding upto year t (ignoring any discounting), and vrepresents random error.

In the case of the former formulation, theeffect associated with a change in the controlfunding in a given year t reflects a long-runeffect (contemporaneous effect: α1 and elapsedtime effect: α2t). In the case of the latter,however, the effect (γ1) of a change in thecumulative funding in a given year is constantregardless of any year. In other words, the effect(γ1) reflects the long-run effect that is averagedover the entire duration of the study period,and thus, in itself it does not account for thefact that longer investment will have greatereffects. Another advantage of the “elapsed timeeffect” formulation, presented below, is that itavoids assumption of an arbitrary discount rate.The “elapsed time effect” and the “cumulativefunding effect” formulation are discussed inmore detail in the empirical section of the paper.

Fourth, controlling for the effect of cross-border sales has been an important exercise un-dertaken by Farrelly, Pechacek, and Chaloupka(2003) and Lovenheim (2008). However, thesestudies control for the incentives for makingcross-border purchases/sales based on interstateprice or tax differentials and population distri-bution. In the present research, we control forprices of cigarettes in the bordering states by


recognizing that cigarettes available in the bor-dering states serve as a substitute good.

Fifth, because of the oligopolistic structure ofthe tobacco market in the United States cigaretteprices can be endogenous (Gruber 2001).4 Endo-geneity is also an issue because prices areoften mis-measured (Farrelly, Pechacek, andChaloupka 2003). Gruber (2001) notes thatstate-specific taxes, which is exogenous in ademand model, can capture about 80% of thevariation in cigarette prices. Farrelly, Pechacek,and Chaloupka (2003) use tax instead of pricein the tobacco demand model to overcome theprice endogeneity problem. We specify twotypes of the demand—one with price as the keyindependent variable, which does not address theprice endogeneity and the other with tax as thekey independent variable, which overcomes theprice endogeneity problem.5

Finally, in a state-level tobacco demand anal-ysis the level of tobacco control funding candepend on the effectiveness of the control pro-gram in the past as well as the numerous unob-served determinants of tobacco consumption.This makes funding level endogenous in thedemand model. In order to effectively addressthis endogeneity issue, we identify a set ofappropriate instruments for the funding level andcarry out a 2SLS analysis, while utilizing thepanel data structure of the model.

B. Model Specification

The dependent variable in the regressions isthe natural log of the quantity of per-capita tax-paid cigarette sales for each state (q). The inde-pendent variables include the natural log of thereal price of cigarettes per pack in dollars (p),the natural log of the real (average) price perpack of cigarette in the bordering states (psub),natural log of total (state + federal) real tax (tax )per pack, natural log of real (average) total taxper pack in the bordering states (taxsub), nat-ural log of real per-capita personal disposableincome in thousands of dollars (pdi ), natural logof aggregate population (lpopul ), and the realtotal annual tobacco control funding in millions

4. Keeler et al. (1996) find that cigarette firms pricediscriminately across states and that the regulatory anti-smoking efforts are countered by cigarette firms by loweringcigarette prices. This behavior of firms can also make theprice endogenous.

5. If state taxes are dependent on some unobserveddeterminants of the level of tobacco consumption, then taxcan be endogenous. In our model, however, we assume taxto be exogenous as is done in previous research.

of dollars (totfund ).6 As mentioned before, alldollar figures are inflation-adjusted based on2008 prices. The socioeconomic variables thatare measured in percentages are not logged.These include the percentage of college grad-uates (pcgr), the percentage of the populationbetween ages 15 and 24 (pc1524 ), percentageof population aged 25 and above (pc25 ), andthe unemployment rate (unemp). Finally, as iscommon in panel data analyses, each specifi-cation includes 16 year dummies to account forthe year fixed effects. With 17 years in the panelfrom 50 states there are 850 data points for esti-mation. The econometric specifications undereach estimation strategy are presented below:

2SLSspecifications of the pooled OLS model:Price-based:

lqst =β0 +16 year dummies+β1 lpst +β2 lpsubst

+ β3 lpdist + (β41 + β42∗t) totf undst

+ β5 lpopulst + β6 pcgrst +β7 pc1524st

+ β8 pc25st + β9 unempst + ust

Tax-based:

lqst = β0 + 16 year dummies + β1 ltaxst

+ β2 ltaxsubst +β3 lpdist + (β41 +β42∗t)

× totf undst + β5 lpopulst

+ β6 pcgrst + β7 pc1524st + β8 pc25st

+ β9 unempst + ust

2SLS specifications of the FE or RE Model:Price-based:

lqst = β0 + 16 year dummies + β1lpst

+ β2 lpsubst + β3 lpdist + (β41 + β42∗t)

× totf undst + β5 lpopul + β6 pcgrst

+ β7 pc1524st + β8 pc25 + β9 unempst

+ as + ust

Tax-based:

lqst = β0 +16 year dummies+β1 ltaxst

+ β2 ltaxsubst +β3 lpdist + (β41 +β42∗t)

× totf undst + β5 lpopulst + β6 pcgrst

+ β7 pc1524st +β8 pc25st + β9 unempst

+ as + ust

6. We did not use natural log of totfund, since thevariable has zero values for many states in the years1991–1993.


In each of the above models, the two suspectedendogenous variables representing funding levelare totfund and t*totfund. For each of thetwo variables, the instruments used for 2SLSestimation are: (1) “Total Smoke-Free Air Law”(airscore); (2) “Total Alciati Score” (alciati-score); and (3) funding level in the immediatelypreceding year (totfund-1 ).7 A measure of thestate budget condition could be an additionalinstrument for the two endogenous variables, asit has been driving funding decisions in nearlyall states over the past decade. However, suchan instrument is not considered because of theabsence of an appropriate measure of the budgetcondition and the lack of availability of budget-related information from states.

The airscore and alciatiscore variables arescores provided exogenously based on the num-ber of policies in place and, as such, canbe treated as truly exogenous. The details ofairscore and alciatiscore are provided in the lasttwo rows of Table 1. Further details are avail-able in the ImpactTEEN.org website.

Each variable in the above models is observedfor each state (s) and each year (elapsed timet = 0, 1, . . . , 16). The term ust represents i.i.d.random error. The term as represents unob-served state-specific fixed effects in the FEmodel that do not change over time. This unob-served fixed-effects term is assumed to be cor-related with the observed explanatory variables.The term as in the case of the RE model repre-sents a collection of unobservable, state-specificeffects that is randomly distributed and is uncor-related with the observed explanatory variables.

7. American Lung Association publishes on an annualbasis states’ performances in anti-tobacco campaign (Ameri-can Lung Association 2007). Four factors by which a state’sefficacy of the control program is judged are: “tobacco pre-vention and control spending,” “tax rate per pack,” “smoke-free air laws,” and “youth access laws.” Grades of eachstate are available in this report only for the years 2003,2004, 2006, and 2007. As exogenous instruments for currentyear’s control funding, we compile similar information on“smoke-free air laws” and “youth access laws” that are avail-able for all the years in the study period and for each statefrom the ImpacTEEN.org website. Also, as tobacco preven-tion and control spending for the current year is suspectedas endogenous in the demand model, we consider fundingin the immediately preceding year as the third exogenousinstrument explaining the efficacy of the control program.The use of lagged values of an endogenous variable asinstrument in a panel model is common in the literature(see, e.g., Ziliak 1997 for an application and Cameron andTrivedi 2008 for a theoretical discussion). Finally, we didnot use tax rate per pack in the immediately prior year asan exogenous instrument, since total tax in the current year(ltax ) is already an included exogenous variable in the threetax-based regression specifications.

In the price-based specifications we assumethat cigarettes in the bordering states are sub-stitutes for the cigarettes sold within the statesand as such, the aggregate demand for cigarettesnot only depends on the price of the cigarette inthe state, it also depends on the average price inthe bordering states. Similarly, in the tax-basedspecifications we assume that tax differential inthe bordering states has an important effect onthe state aggregate cigarette demand. Accord-ingly, we specify the demand for cigarette as afunction of the total tax imposed within a stateas well as the average tax charged in the bor-dering states. Both specifications can effectivelycontrol for the degree of cross-border sales in astate that are discussed in previous studies (Far-relly, Pechacek, and Chaloupka 2003; Loven-heim 2008).

Notice that the above formulation includesboth contemporaneous effect (β41) and theelapsed time effect (β42t) of control funding.The elapsed time effect is important because,tobacco being an addictive good, behaviorchange is a gradual process consisting of precon-templation, contemplation, action, maintenance,and relapse (Farrelly, Pechacek, and Chaloupka2003; Prochaska et al. 1988). Moreover, effi-ciency in program delivery, quality of interven-tions, effectiveness of media campaigns, etc., areimportant features of a comprehensive tobaccocontrol program that can only improve overtime.

VI. EMPIRICAL RESULTS

A. Endogeneity Issue

For the price-based pooled model, we carryout Hausman’s test for endogeneity of the twofunding variable, namely totfund and t*totfund.The Hausman’s test suggests that the twovariables are jointly endogenous (F(2, 820) =13.96; p value <.001) and thus, demands appli-cation of 2SLS. Sargan’s score test for overi-dentifying restriction in the price-based pooledmodel fails to reject the null hypothesis of exo-geneity of the set of three excluded instruments(χ2

1 = .20; p value = .65), implying that theinstruments are truly exogenous.8

8. Reduced-form regression of totfund on airscore, alci-atiscore, and totfund-1 yields an R2 of 0.77 and the regres-sion of t* totfund on airscore, alciatiscore, and totfund-1yields an R2 of 0.64. Also, in the reduced form regressionof totfund, the only variable that is significant is totfund-1 (p value <.001), where as in reduced form regressionof t* totfund all the variables are very highly significant(p value <.001).


B. Issue of Spatial Autocorrelation

States differ with respect to the propor-tion of dollars spent on various componentsof the tobacco control program, which mayresult in spatial heterogeneity across states. Forexample, a high proportion spent on mediaadvertising (particularly TV) may lead to sub-stantial cross-border influences where mediamarkets stretch across state borders. A spa-tial econometric analysis is needed to correctfor such spatial heterogeneity. Unfortunately,detailed data on this are unavailable and thus,such cross-border effects remain unobserved,leading to spatially autocorrelated errors in thetobacco demand model. Although, under spa-tial heterogeneity parameters are still consis-tent, ignoring such heterogeneity may resultin inefficient estimates of the demand param-eters (Anselin et al. 1996; Anselin and Bera1998).

We conduct Moran’s I test to test for thepresence of spatial autocorrelation for the years1991, 1999, and 2007, separately. The yearsselected are the first year, middle year, andthe last year of the study period. Moran’s Itest in our analysis involves, as a first step,finding the coordinates (latitude and longitude)of each of the 50 states. This information isavailable at http://www.xfront.com/us_states/. Inorder to facilitate the analysis, we use thecoordinate of each state capital to represent thecoordinate of the corresponding state. In thenext step, we create a Euclidian distance fromeach coordinate to all the other coordinates.The third step involves creating a 50 × 50weight matrix for the 50 states of which theoff-diagonal entry (i, j) in the matrix is equalto 1/(distance between point i and point j).As the first quartile Euclidian distance for theU.S. states is found to be 10.0, we attach zeroweight for the states that are more than 10.0Euclidian distance away from a given state.Finally, we compute Moran’s I statistics, usingthe following formula:

Moran’s I statistic I

=⎡⎣n

∑i

∑j

wij (qi − q)(qj − q)

⎤⎦ /

×⎡⎣

⎛⎝∑

i

∑j

wij

⎞⎠ ∑

k

(qk − q)2

⎤⎦

where wij are the weights, and qi and q arethe per-capita cigarette demand in the statei in a given year and average demand overall the n = 50 states in a given year, respec-tively. If there is no spatial autocorrelation,I statistic should be close to zero. A valueof I statistic close to 1 indicates cluster-ing. The third and the final steps are car-ried out using the STATA software specificallydesigned to test for spatial autocorrelation. TheMoran’s I statistics for the 3 years are asfollows:

Variable I E(I ) SD(I ) Z-score p value n

1991: q 0.229 −0.020 0.062 4.027 .000 50

1999: q 0.080 −0.020 0.062 1.611 .054 50

2007: q 0.044 −0.020 0.061 1.055 .146 50

The above tests suggests that in 1991, thenull hypothesis of “no spatial autocorrelation”is rejected at any significance level; in 1999it is rejected at 10% level; and in 2007,data do not provide any evidence of spatialautocorrelation.

Although Moran’s I tests suggests thatthe severity of spatial autocorrelation is nota concern, there is statistical evidence of itand thus, in order to ensure parameter effi-ciency our estimation must be robust to spa-tial autocorrelation. In the following subsec-tion, we report Driscoll and Kraay (1998) stan-dard errors for coefficients estimates for thepooled and FE regressions. The Driscoll andKraay (1998) approach is designed for spa-tially dependent panel data and the standarderrors obtained using this approach are robustto general forms of cross-sectional (spatial) andtemporal dependence when the time dimensionbecomes large, which is the case in the presentanalysis.

C. Estimation Results

The estimation results from the pooled, FE,and RE models are reported in Table 3, underthe price-based and tax-based specifications. Asdiscussed above, the standard errors, presentedwithin parentheses, are all corrected for het-eroscedasticity and spatial autocorrelation inthe case of the pooled OLS and FE models.However, since the Driscoll and Kraay (1998)approach is not capable of correcting for spatialautocorrelation in the case of the RE model, we


TABLE 3Coefficient Estimates for Log Cigarette Packs Sold by States

Price-Based Specification (2SLS) Tax-Based Specification (2SLS)

VariablePooledModel

FixedEffects

RandomEffects

PooledModel

FixedEffects

RandomEffects

lp −1.53∗∗∗(0.10)

−0.91∗∗∗(0.16)

−0.91∗∗∗(0.08)

lpsub 0.30∗∗∗(0.13)

0.24∗∗∗(0.05)

0.25∗∗∗(0.10)

ltax −0.57∗∗∗(0.02)

−0.42∗∗∗(0.03)

−0.42∗∗∗(0.02)

ltaxsub 0.17∗∗∗(0.04)

0.14∗∗∗(0.02)

0.14∗∗(0.04)

lpopul −0.05∗∗∗(0.007)

−0.43∗∗∗(0.07)

−0.07∗∗(0.03)

−0.04∗∗∗(0.009)

−0.36∗∗∗(0.04)

−0.07∗∗∗(0.03)

lpdi 0.37∗∗∗(0.05)

−0.16(0.12)

−0.15(0.12)

0.24∗∗∗(0.06)

−0.12(0.11)

−0.11(0.12)

totfund 0.0021(0.0015)

0.004∗∗∗(0.001)

0.0036∗∗∗(0.0008)

0.00027(0.0006)

0.0026∗∗∗(0.0009)

0.0025∗∗∗(0.0006)

t*totfund −0.00037∗∗∗(0.0001)

−0.00044∗∗∗(0.00007)

−0.0004∗∗∗(0.00007)

−0.00026∗∗∗(0.00004)

−0.0004∗∗∗(0.00004)

−0.0003∗∗∗(0.00006)

pcgr −0.021∗∗∗(0.001)

−0.006(0.004)

−0.0065∗∗∗(0.002)

−0.022∗∗∗(0.001)

−0.0067(0.0042)

−0.007∗∗∗(0.002)

pc1524 −0.034∗∗(0.016)

−0.016(0.011)

−0.017∗∗∗(0.006)

−0.036∗∗∗(0.019)

−0.011(0.013)

−0.012∗∗(0.006)

pc25 0.027∗∗∗(0.004)

0.006∗∗∗(0.002)

0.012∗∗∗(0.001)

0.029∗∗∗(0.006)

0.009∗∗∗(0.002)

0.014∗∗∗(0.001)

unemp −0.013(0.012)

−0.004(0.005)

−0.003(0.006)

0.004(0.013)

−0.0075(0.0045)

−0.007(0.005)

Sample size 850 850 850 850 850 850R2 0.70 0.78 0.77 0.72 0.81 0.81

Notes: Each equation also includes an intercept term and 16 year dummies. Numbers in parentheses are standarderrors. For pooled and FE models Driscoll and Kraay (1998) standard errors with lag (4) are reported. For the RE model,heteroscedasticity-robust standard errors are reported.

∗∗Significant at 5%; ∗∗∗significant at 1%.

correct for only heteroscedasticity in the case ofthe RE model.9,10

We note that in the present sample theestimated values of the error serial correla-tion σ2

a/(σ2a + σ2

u) are: price-based models =0.96 (FE) = 0.76 (RE); tax-based models =0.96 (FE) = 0.81 (RE). The numerator in theexpression for serial correlation represents thevariation in the state-specific unobserved effectsand the denominator is the total variation in theerror term. This signifies that the state-specific

9. We perform the White (1980) test for heterscedastic-ity on the demeaned model based on which FE estimationis carried out. The test shows the presence of heteroscedas-ticity.

10. To ensure that multicollinearity is not a problem inthe set of explanatory variables, we computed the VarianceInflation Factor (VIF) in the pooled model specifications.The mean VIF are 3.85 and 2.77 in the price-based and thetax-based specifications, respectively, which are much lessthan 10. Moreover, none of the variables shows a VIF valuemore than 7 in either of the specifications.

unobserved heterogeneity accounts for a sub-stantial portion of the variation in the residuals.Since the pooled model cannot control for theunobserved, state-specific heterogeneity, in thepresence of such a high magnitude of serialcorrelation the results from the pooled modelshould be treated with caution.11

As discussed in the previous section, wepresent both the price-based model that does

11. In the presence of heteroscedasticity, Hausman’s(1978) specification test statistic to compare the FE and REmodels has a non-standard limiting distribution (Wooldridge2002) and, as such, the test result can be misleading.Despite this limitation, we carry out the specification test tocompare our policy parameters associated with the variablet*totfund in the FE and RE models. The test turns outto be inconclusive at any acceptable significance level inboth price-based and tax-based specifications, indicatingthat either model can be used for analysis. A high serialcorrelation in our sample also points to the possibilitythat the FE parameter estimates could be close to the REparameter estimates, which is indeed the case in the presentsample, except for the control variable lpopul.


not address price endogeneity and the tax-basedmodel, which is a better model because itovercomes the problem of price endogeneity.A comparison of the parameter estimates ofthe three econometric approaches based on theprice- and tax-based models reveals that theestimates are somewhat similar in magnitudeto their respective counterpart. This signifiesthat addressing the price endogeneity problemin the price-based regression may not noticeablyaffect the parameter estimates. The coefficienton the unemployment rate variable (unemp) isinsignificant in all the models. The coefficientsassociated with percent college graduates (pcgr)and percent population in 15–24 age group(pc1524 ) are insignificant in the FE modelin both price-based regression as well as tax-based regression; the coefficient associated withpersonal disposable income per-capita (lpdi ) isinsignificant in both FE and RE models; andfinally, the coefficient associated with total state-level tobacco control program funding (totfund )is insignificant in both price-based and tax-based pooled model. All other coefficients aresignificant at 10% level or better in all themodels.

The variable representing percent collegegraduates (pcgr) proxies for the level of edu-cation in a state and the significant negativesign on its coefficient shows that the more edu-cated the state population the less the demandfor cigarette. The negative sign on the coeffi-cient of percent population in 15–24 age group(pc1524 ) may be counter-intuitive. However, ifone believes that during the reference period thetobacco control and prevention initiatives couldtarget the 15–24 age group more effectively rel-ative to other population groups in a state thenthis result may not be surprising. In fact, ifmarginal smokers in a state are assumed to bein the 15–24 age category and the populationof these marginal smokers is proportional to thetotal population in this age group, one wouldexpect that the benefits to society of deterringthe marginal smoker are greater, on average,compared to other smokers. A similar interpre-tation would apply in the case of percent collegegraduates (pcgr), if one assumes that educatedsmokers are influenced on the margin. Statisti-cal tests, however, suggest that the effect sizesassociated with pcgr and pc1524 are statisticallyequal, signifying that whichever of the two cat-egories is influenced on the margin, benefits tosociety of deterring the marginal smokers in thetwo categories would be statistically the same.

Positive and significant coefficient on the vari-able representing percent population 25 or above(pc25 ) captures adult smokers as a subset. Thus,if the size of the adult smoking population variesdirectly with the total adult population and ifone believes that it is harder for adult smokersto quit smoking, a positive coefficient should beexpected.

The price elasticity in the price-based specifi-cation reveals some very interesting results. Forexample, own price elasticities show very highmagnitudes ranging from −0.91 (FE and RE)to −1.53 (pooled model). However, own priceelasticity can be grossly misleading as an elas-ticity indicator, as elasticity of cigarette demandshould be based on the broadly defined mar-ket (for example, in state and out of states) notjust a narrowly defined market, such as a state.According to Lovenheim (2008), who developsa model to control for cross-border sales, ignor-ing smuggling can cause an upward bias in priceelasticity of cigarettes in absolute value. Wecontrol for cross-border sales by including aver-age price of cigarettes in the bordering states.Our price-based regression models reveal thisupward bias when one looks only at the own-price elasticity in isolation. The true price elas-ticity is the full price elasticity in our modelsand is the sum of the own- and the cross-priceelasticities. For example, in the case of the price-based FE model for a 1% increase in the pricewithin states as well as the average price of thebordering states quantity demanded decreasesby 0.67% (= −0.91 + 0.24). In the tax-basedFE and the RE models overall tax elastic-ity is −0.28 (= −0.42 + 0.14), which trans-lates to price elasticity of −0.58.12 Lovenheim(2008) finds the smuggling-corrected elasticitiesto range between −0.45 and −0.47, which aresmaller than that obtained in our price-based FEand RE models (−0.67) as well as that obtainedin our tax-based FE and RE models (−0.58).Our elasticity estimates are significantly higherthan −0.32, −0.30, and −0.15 reported by Far-relly, Pechacek, and Chaloupka (2003), Evans,Ringel, and Stech (1999), and Sloan, Smith, andTaylor (2002), respectively.

Significant border price (lpsub) and tax elas-ticities (ltaxsub) in all the models suggestthat consumers react to cigarette price or tax

12. We run a regression of ltax on lp, all the statedummies, and all the year dummies, which yields priceelasticity of tax equal to 2.06. We multiply this number tothe tax elasticity of cigarette demand of −0.28 to arrive atthe price elasticity of cigarette demand of −0.58.


FIGURE 2Control Funding Effect Size Over Time

increases by resorting to purchase in the border-ing states with lower price or tax, respectively.

The variable relating to tobacco controlexpenditure, namely totfund that appears sep-arately to reflect contemporaneous effect andas the interaction term to reflect elapsed time(t*totfund ) is the main policy variable forthe present study.13 While the contemporane-ous effect is highly significant and positivein the FE and RE models, the elapsed timeeffect of the control expenditure is highly sig-nificant and negative in all the models, asexpected, implying that the effect of the currentperiod’s control expenditure increases steadilyover time.14 An analysis of the combined effect

13. A comparative analysis of our model based on totalfunding and the one based on per-capita funding, adoptedby Farrelly, Pechacek, and Chaloupka (2003), would bea worthwhile exercise. In addition to the proposed modelwith totfund and t*totfund as the explanatory variables,we also estimate the model based on per-capita funding.Unfortunately, a comparative analysis of the two modelingapproaches could not be carried out, since, under the latterspecification, the interaction coefficient associated with theper-capita funding variable turned out to be statisticallyinsignificant in all the specifications at any acceptablelevel. As explained before in Section V.A, the per-capitafunding measure may be inappropriate in the absence of anappropriate weighting adjustment, and thus, this finding maynot be surprising.

14. Findings in the previous literature are mixed withregard to contemporaneous effect. Some studies find thatstate-level policies, such as advertising restrictions, healthwarnings, territorial restrictions, limits on youth access totobacco products, counter advertising, etc., have positive

is important to understand the role of the con-trol programs in reducing cigarette demand. InFigure 2, we show the combined effect in reduc-ing cigarette demand. The effects on cigarettedemand reduction are shown in percentagesusing the three econometric modeling strate-gies, applied to the tax-based regression, for theperiod 1991 through 2008.15 The effect size istrending up steadily in all the models with theFE and the RE models showing a high degree ofagreement and the pooled model showing a rel-atively larger effect size. The combined effect isnegative (increase in cigarette demand) up until1998 in the case of the FE and RE models, sig-nifying that it took 7 years after the start of thestudy period to realize the benefits of the controlprograms. This result is not surprising since, asdiscussed earlier, public policies to change peo-ple’s preference for addictive goods can take along time to be effective.

In the tax-based models, the effect sizes(β41 + β42 ∗ t) for the year 1999 are: FE:0.022%, and RE: 0.025% and those for the year2007 are: FE: 0.308%, and RE: 0.302%. In theFE model it means, all else equal, for every one

contemporaneous effect on smoking (Laugesen and Meads1991), while others find that they have adverse effect(Baltagi and Levin 1986; Goel and Morey 1995).

15. As the study period is from 1991 (t = 0) to 2007(t = 16), the last observation (t = 17) represents year 2008,which gives the out-of-sample estimate of the overall effectof control funding in 2008.


million dollar increase in control funding thereis an expected reduction in per-capita cigarettesales of 0.022% and 0.308% in the years 1997and 2007, respectively. These effects are aftercontrolling for the year-specific fixed effect. Inshort, our study clearly shows that, controllingfor the year-specific fixed effect, the same levelof current control expenditure has larger andlarger effect in reducing the per-capita cigarettesales as time passes. Our results thus fully justifyour own hypothesis and reinstate the findingsby Farrelly, Pechacek, and Chaloupka (2003),and the CDC’s position on this important issue,which states that “longer investment will haveeven greater effects” (Best Practices 2007).

As noted before, in our “elapsed time effect”formulation the effect size varies every yeardepending on the time elapsed t , whereas inthe “cumulative funding effect” formulation pre-sented in Farrelly, Pechacek, and Chaloupka(2003) and Farrelly et al. (2008) the effect sizeis fixed regardless of time t , which essentiallyreflects the average effect size over the periodof 17 years. We carry out the Farrelly et al.approach with discount rates of 0%, 5%, 10%,25%, and 50%, using both price-based modelsand tax-based models. We find that for discountrates of 25% and above the control fundingeffect is statistically insignificant. For discountrates 0%, 5%, and 10%, the effect sizes based onthe tax-based FE model are: 0.034%, 0.046%,and 0.055%, respectively. The average controlfunding effect based on our tax-based FE modelis computed as 0.040%. Thus, our estimatesbased on the “elapsed time effect” are com-parable to those obtained from the cumulativefunding approach that assumes between 0% and5% discount.16

We also note that, as discussed before, underthe “elapsed time effect” formulation the effectsize associated with a change in the currentyear’s control funding reflects a long-run effectfor that particular year. In Section VII, we takeadvantage of this formulation and carry out abenefit-cost analysis of the long-run effect forthe year 2008 by estimating the effect sizeunder various hypothetical levels of the controlfunding in the year 2008 (one year after thestudy period). Under the “cumulative fundingeffect” formulation, however, the effect of achange in the level of the year 2008 controlfunding reflects only an average effect over the

16. The results based on the discount rate approach areavailable with the correspondent author upon request.

17 years of the study period. Thus, in itself itdoes not take into account the fact that longerinvestment will have greater effects.17

Figure 3 shows the time path of per-capitacigarette demand for the actual data as wellas for the predicted values obtained from thepooled, FE, and RE models, under the tax-based specification. The time paths for FE andRE models are indistinguishable and the plotclearly shows that the predicted time paths aregenerally in agreement among themselves andshow similar movements like the actual timepath, although they are slightly but consistentlylower than the actual time path.

VII. A BENEFIT-COST ANALYSIS OF THECDC-RECOMMENDED FUNDING

We estimate the benefits (cost avoided) ofsmoking reduction keeping 2007 (last year ofthe study period) as the benchmark year and2008 as the study year, that is the year when ahypothetical change in the control funding willbe instituted and its long-run impact assessed.18

We use both the tax-based as well as theprice-based specifications. In Table 4, we reportthe detailed results only for the FE model,under the tax-based specification. Column 1represents various levels of additional controlfunding over and above the average year-2007spending of $13.89 million (in 2008 dollars).The last value in this column represents the CDCrecommended Best Practices (2007) averagefunding of $73.722 million (in 2008 dollars),which is $59.832 million over and above theaverage year-2007 spending of $13.89 million.Column 2 reports reduction in the predicted per-capita number of packs in 2008 as a resultof additional funding, keeping all the otherexplanatory variables at their 2007 averages,including the elapsed time t , which takes the

17. An advantage of the “cumulative funding effect”formulation is that it follows from a relatively straightfor-ward consumer demand model. If a new cohort of potentialsmokers enters every year and is influenced by antismokingfunding, in aggregate a cumulative formulation may bettercapture the stock–flow nature of the population of smokers.

18. The “elapsed time effect” formulation proposed inthe study is capable of assessing the effect of a hypotheticalchange in the control funding in the future. Also, since thepast costs are sunk or unrecoverable, the theory of benefit-cost analysis dictates that only prospective (future) costsshould be considered as relevant to an investment decision.We take advantage of the strength of our model in assessingthe future and follow what the benefit-cost analysis theorydictates and carry out such an analysis for the future (year2008).


FIGURE 3Time Path of Per-Capita Cigarette Demand

2007 value of 16. Column 3 reports aggregateaverage pack reduction in a given state, obtainedby multiplying column 2 by the average statepopulation as of July 31, 2008.

The last four columns are based on Table 4of the CDC (2006) report, which gives vari-ous costs associated with smoking, namely themedical cost per pack ($5.31), productivity costper pack ($5.16), and Medicaid cost per pack($1.63) in 2004 dollars.19 We first convert thesefigures into equivalent of 2008 dollars. We thenmultiply the converted figures with the aggre-gate average pack reduction in a given state,obtained in column 3, to arrive at the averagebenefits (cost avoided) of the tobacco controlprogram in a given state.

In Table 5, we present the summary resultsof the aggregate benefits (cost avoided), undervarious levels of funding for each model.In particular, the three benefit (cost avoided)columns in the case of the tax-based speci-fication show that if the CDC-recommendedaverage amount of $73.722 million (which is

19. In response to an email inquiry, Office of Smokingand Health, Centers for Disease Control and Preventionconfirmed that the three smoking-related costs are mutuallyexclusive.

$59.832 million over and above the averagespending of $13.89 million in 2007) was spentin a given state in 2008, the estimated bene-fits would have ranged from $853 million (RE)to $1.0559 billion (pooled model), resulting ina net benefit in the range of, approximately,$779 million to $981 million. The same esti-mated net benefits figures in the case of price-based specification range from $848 million(RE) and $1.005 billion (pooled model). It isalso interesting to note from the last threecolumns of this table that the benefit-cost ratiosrange from 17 to 20 in the case of the pooledmodel and from 14 to 17 in the FE and REmodels, which means that benefits far outweighthe costs. A recent study by Lightwood, Dinno,and Glantz (2008) on California’s tobacco pre-vention program found that for every dollar thestate spent on its tobacco control program from1989 to 2004, the state received tens of dol-lars in savings in the form of sharp reductionsto total healthcare costs in the state. Our studyshows similar results based on the average ofall the 50 states. However, our empirical modelcan also be used to analyze the impact of actualCDC-recommended funding for each state. Inshort, our empirical results show that as the


TABLE 4Total Costs and Benefits under Various Levels of Control Funding (Fixed Effects Model)a

AdditionalFunding in aState in 2008(Million Dollar)

PredictedPer-Capita Packs

Reduction in a Statein 2008

Average PackReduction in aState in 2008

(Million)

Medical CostAvoided (Million

Dollar)

ProductivityCost Avoided

(Million Dollar)

Medicaid CostAvoided (Million

Dollar)

Total CostAvoided(MillionDollars)

1 0.19 1.4 7.0 6.8 2.1 15.910 1.90 14.0 68.8 66.8 21.1 15720 3.75 27.5 135 132 42 30950 8.97 65.1 324 314 99 73759.832 10.57 76.5 382 371 117 869

aThe results are based on the tax-based specifications.

TABLE 5Summary of Aggregate Benefits in a State and the Benefit-Cost Ratios

Total Cost Avoided (Million Dollars) Benefit-Cost Ratios

Additional Funding in 2008(Million Dollar)

PooledModel

FixedEffects

RandomEffects

PooledModel

FixedEffects

RandomEffects

Tax-based specification1 19.7 15.9 15.6 19.7 15.9 15.610 194 157 154 19.4 15.7 15.420 380 309 303 19.0 15.4 15.150 898 737 724 17.9 14.7 14.559.832 1,055 869 853 17.6 14.5 14.3Price-based specification1 20.2 17.1 16.9 20.2 17.2 16.910 199 169 167 19.9 16.9 16.720 389 333 329 19.5 16.6 16.450 919 792 783 18.4 15.8 15.759.832 1,079 932 922 18.0 15.6 15.4

tobacco control funding effect gathers momen-tum over time, benefits of a sustained level ofCDC-recommended funding will grow over theyears and thus, our estimates of the year 2008benefits may well be the lower bound of thepotential benefits to be accrued in all the futureyears.

A major limitation of the above benefit-cost analysis is that the benefits (costs avoided)associated with the smoking figures reported inthe CDC (2006) report are based on studiesthat compare experiences of smokers as opposedto nonsmokers. It is quite likely that many ofthose who quit smoking as a result of successfultobacco control programs might have had otherhealth or productivity impairments. This maypose the problem of selectivity in arriving at thereliable benefits (costs avoided) estimates. Giventhe absence of any other reliable benefits dataat the national level, the benefit-cost analysispresented in this research can be viewed asmerely an attempt to provide a perspective

on the effectiveness of the tobacco controlprograms at the state level.

VIII. SUMMARY AND CONCLUSION

In recent years, state-level tobacco controland prevention funding in the United States hasshown a declining trend. This paper presents aneconomic benefit-cost analysis of the ongoing,state-level tobacco control programs for the firsttime, which suggests that such a move by thepublic authorities may be short-sighted. Build-ing on the current literature on the effect oftobacco control funding on state-level cigarettesales the study provides several methodologi-cal extensions. Using the fixed- and random-effects techniques as applied to panel data andaccounting for endogeneity of the tobacco con-trol funding level, the study models the effect ofthe current tobacco control funding on cigarettesales as a function of the elapsed time, which isthe time since the initial control funding during


the study period. On the basis of the state-level panel data for 1991–2007, compiled frommultiple sources, the study finds a statisticallysignificant evidence of a sustained and steadilyincreasing long-run impact of the control pro-gram spending on cigarette demand.

Moreover, the study reveals that if the stateswere to follow the CDC-recommended BestPractices (2007) funding guidelines in year2008, the aggregate benefits of the control pro-grams, which is the sum of the cost savingsassociated with the medical cost, productivitycost, and Medicaid cost, could range between14 and 20 times the average cost of implement-ing the tobacco control program in a state. Inshort, the benefits of the control programs faroutweigh the costs.

REFERENCES

American Lung Association. “State of Tobacco Control.Report Card: Does Your State Make the Grade?”www.lungusa.org, 2007.

. “State of Tobacco Control. Report Card: How DoesYour State Measure Up?” www.lungusa.org, 2009.

Anselin, L., and A. Bera. “Spatial Dependence in LinearRegression Models with an Introduction to SpatialEconometrics,” in Handbook of Applied EconomicStatistics, edited by D. Giles and A. Ullah. New York:Marcel Dekker, 1998.

Anselin, L., A. Bera, R. Florax, and M. Yoon. “SimpleDiagnostic Tests for Spatial Dependence.” RegionalScience and Urban Economics, 26, 1996, 77–104.

Baltagi, B. H., and D. Levin. “Estimating Dynamic Demandfor Cigarettes Using Panel Data: The Effects of Boot-legging, Taxation and Advertising Reconsidered.” TheReview of Economics and Statistics, 68, 1986, 148–55.

Bauer, U. E., T. M. Johnson, R. S. Hopkins, and R. G.Brooks. “Changes in Youth Cigarette Use and Inten-tions Following Implementation of a Tobacco ControlProgram: Findings from the Florida Youth TobaccoSurvey, 1998–2000.” Journal of the American MedicalAssociation, 284(6), 2000, 723–28.

Becker, G. S., M. Grossman, and K. M. Murphy. “AnEmpirical Analysis of Cigarette Addiction.” The Amer-ican Economic Review, 84(3), 1994, 396–417.

Best Practices for Comprehensive Tobacco Control Pro-grams. U.S. Department of Health and Human Ser-vices, Atlanta, GA: Centers for Disease Control andPrevention, National Center for Chronic Disease Pre-vention and Health Promotion, Office on Smoking andHealth, 2007.

Biener, L., J. E. Harris, and W. Hamilton. “Impact of theMassachusetts Tobacco Control Program: PopulationBased Trend Analysis.” British Medical Journal, 343,2000, 1772–77.

Cameron, A. C., and P. K. Trivedi. Microeconometrics:Methods and Applications. Cambridge: CambridgeUniversityPress, 2008.

Campaign for Tobacco-Free Kids. American Heart Asso-ciation, American Cancer Society, American LungAssociation. “History of Spending for State TobaccoPrevention Programs FY2006–FY2010.” Washington,DC: Campaign for Tobacco-Free Kids. 2010.

Centers for Disease Control and Prevention (CDC). “Declinein Cigarette Consumption Following Implementationof a Comprehensive Tobacco Prevention and Educa-tion Program—Oregon, 1996–1998.” Morbidity andMortality Weekly Report, 48(7), 1999, 140–43.

. “A Broken Promise to Our Children: The 1998 StateTobacco Settlement, Seven Years Later.” AmericanHeart Association, American Cancer Society, Cam-paign for Tobacco-Free Kids, American Lung Asso-ciation, 2005.

. “Sustaining State Programs for Tobacco Control:Data Highlights 2006.” Atlanta, GA: U.S. Departmentof Health and Human Services, Centers for DiseaseControl and Prevention, National Center for ChronicDisease Prevention and Health Promotion, Office onSmoking and Health, 2006.

Chaloupka, F. J., and H. Saffer. “Clean Indoor Air Lawsand the Demand for Cigarettes.” Contemporary PolicyIssues, 10(20), 1992, 72–83.

Chaloupka, F. J., and K. E. Warner. “The Economics ofSmoking,” in The Handbook of Health Economics,The National Bureau of Economic Research WorkingPaper Series, No. 7047, edited by J. P. Newhouse andA. J. Cuyler. New York: North-Holland, 2000.

Connolly, G., and H. Robbins. “Designing an EffectiveStatewide Tobacco Control Program—Massachusetts.”Cancer Supplement, 12(83), 1998, 2722–27.

Driscoll, J. C., and A. C. Kraay. “Consistent CovarianceMatrix Estimation with Spatially Dependent PanelData.” Review of Economics and Statistics, 80, 1998,549–60.

Evans, W. N., J. Ringel, and D. Stech. “Tobacco Taxes andPublic Policy to Discourage Smoking,” in Tax Policyand the Economy, vol. 13, edited by James Poterba.Cambridge, MA: National Bureau of Economic Re-search, 1999.

Farrelly, M. C., T. Pechacek, and F. Chaloupka. “TheImpact of Tobacco Control Program Expenditures onAggregate Cigarette Sales: 1981–2000.” Journal ofHealth Economics, 22, 2003, 843–59.

Farrelly, M. C., T. F. Pachacek, K. Y. Thomas, and D.Nelson. “The Impact of Tobacco Control Programs onAdult Smoking.” American Journal of Public Health,98(2), 2008, 304–09.

Goel, R. K., and M. J. Morey. “The Interdependence ofCigarette and Liquor Demand.” Southern EconomicJournal, 62, 1995, 451–59.

Gruber, J. “Tobacco at the Crossroads: The Past and Futureof Smoking Regulation in the United States.” The Jour-nal of Economic Perspectives, 15(2), 2001, 193–212.

Hausman, J. A. “Specification Tests in Econometrics.”Econometrica, 46, 1978, 1251–71.

Hu, T.-W., H.-Y. Sung, and T. E. Keeler. “The State Anti-smoking Campaign and the Industry Response: TheEffects of Advertising on Cigarette Consumption inCalifornia.” American Economics Association Papersand Proceedings, 85(2), 1995a, 85–90.

. “Reducing Cigarette Consumption in California:Tobacco Taxes vs. an Anti-Smoking Media Cam-paign.” American Journal of Public Health, 85(9),1995b, 1218–22.

ImpacTEEN.org. Institute for Health Research and Policy,National Program Office, University of Illinois atChicago.

Keeler, T. E., T.-W. Hu, P. G. Barnett, W. G. Manning,and H.-Y. Sung. “Do Cigarette Producers Price-Discriminate by State? An Empirical Analysis of LocalCigarette Pricing and Taxation.” Journal of HealthEconomics, 15, 1996, 499–512.

Laugesen, M., and C. Meads. “Tobacco Advertising Restric-tions, Price, Income, and Tobacco Consumption in


OECD Countries. 1960–86.” British Journal of Addic-tion, 86, 1991, 1343–54.

Lightwood, J. M., A. Dinno, and S. A. Glantz. “Effect ofthe California Tobacco Control Program on PersonalHealth Care Expenditures.” PLOS Medicine, 5(8),2008, 1214–22.

Lovenheim, M. F. “How Far to the Border? The Extent andImpact of Cross-Border Casual Cigarette Smuggling.”National Tax Journal, 61(1), 2008, 7–33.

Manley, M. W., J. P. Pierce, E. A. Gilpin, B. Rosbrook, C.Berry, and L.-M. Wun. “Impact of the American StopSmoking Intervention Study on Cigarette Consump-tion.” Tobacco Control, 6(Suppl. 2), 1997, S12–16.

Marlow, M. L. “Tobacco Control Program and TobaccoConsumption.” Cato Journal, 26(3), 2006, 573–91.

Orzechowski and Walker Consulting Firm. The Tax Burdenon Tobacco, Historical Compilation, Volume 43, Wash-ington, DC: The Tobacco Institute, 2008.

Pierce, J. P., E. A. Gilpin, S. L. Emery, M. M. White, M. S.Rosbrook, and C. C. Berry. “Has the California To-bacco Control Program Reduced Smoking?” Journalof the American Medical Association, 280(10), 1998,893–99.

Prochaska, J. O., W. F. Velicer, C. C. DiClemente, andJ. Fava. “Measuring Processes of Change: Application

to the Cessation of Smoking.” Journal of Consultingand Clinical Psychology, 56, 1988, 520–28.

Sloan, F. A., V. K. Smith, and D. H. Taylor, Jr. “Information,Addiction, and ‘Bad Choices’: Lessons from a Centuryof Cigarettes.” Economic Letters, 77(2), 2002, 147–55.

Sung, H. Y., T. W. Hu, and T. E. Keeler. “Cigarette Taxa-tion and Demand—An Empirical Analysis.” Contem-porary Economic Policy, 12(3), 1994, 91–100.

Tauras, J. A., F. J. Chaloupka, M. C. Farrelly, G. A. Giovino,M. Wakefield, L. Johnston, P. M. O’Malley, D. D.Kloska, and T. F. Pechacek. “State Tobacco ControlSpending and Youth Smoking.” American Journal ofPublic Health, 95(2), 2005, 338–44.

White, H. “A Heteroscedasticity-Consistent CovarianceMatrix Estimator and a Direct Test for Heteroscedas-ticity.” Econometrica, 48, 1980, 817–38.

Wooldridge, J. M. Econometric Analysis of Cross Sectionand Panel Data. Cambridge, MA: The MIT Press,2002.

Ziliak, J. P. “Efficient Estimation with Panel Data WhenInstruments Are Predetermined: An EmpiricalComparison of Moment-Condition Estimators.” Jour-nal of Business Economics and Statistics, 15, 1997,419–31.

does spending more on tobacco control programs make economic sense? an incremental benefit-cost...

Documents