J Fam Econ Iss (2007) 28:4967 DOI 10.

1007/s10834-006-9045-4 ORIGINAL PAPER

Does smoking have a causal effect on weight reduction?

Zhuo Chen Steven T. Yen David B. Eastwood

Published online: 15 December 2006 Springer Science+Business Media, LLC 2006

Abstract This study examines the relationship between smoking and body mass index (BMI) with a simultaneous equations system allowing for censoring and endogeneity of the number of cigarettes smoked, which alleviates simultaneity bias caused by unobserved heterogeneity and expansion bias by censoring in the regressor. The results suggest smoking may not have a strong long-term causal effect on body weight after controlling for the endogeneity. The negative relationship between smoking and BMI reported in the literature is potentially attributable to the aforementioned biases and should be interpreted with caution. The statistical procedure developed can be useful in other applications with a censored endogenous regressor. Keywords Body mass index Censored regressor Overweight Simultaneous equations system Smoking Introduction Obesity and smoking are two major public-health concerns in the US. The prevalence of obesity has reached an alarming level in the country. Hedley et al. (2004) combined data from the 19992000 and 20012002 National Health and Nutrition Examination Surveys (NHANES) and found that 65.1% of the adults aged 21 and over in 19992000 were either overweight or obese, 30.4% were obese, and 4.9% were extremely obese. Brown, Mishra, Kenardy, and Dobson (2000), Kuchler and
Z. Chen (&) The Chicago Center of Excellence in Health Promotion Economics, The University of Chicago, Chicago, IL 60637, USA e-mail: zchen1@cdc.gov S. T. Yen D. B. Eastwood Department of Agricultural Economics, The University of Tennessee, Knoxville, TN 37996-4518, USA e-mail: syen@utk.edu D. B. Eastwood davideastwood@verizon.net

123

50

J Fam Econ Iss (2007) 28:4967

Variyam (2003), and Mokdad et al. (2003) assessed the deleterious effects of an overweight population. Paeratakul, Lovejoy, Ryan, and Bray (2002) examined the obesity-related chronic diseases in US adult population according to gender, race and socioeconomic status (SES) and concluded that the disease burden associated with obesity in the population is substantial. Meanwhile, tobacco smoking, accountable for more than 400,000 deaths annually, remains the leading preventable cause of mortality in the US (Perkins, Hickcox, and Grobe, 2000). Dardis and Keane (1995) examined the policy implications using a risk-benet analysis of cigarette smoking decisions. Researchers have recently paid increased attention to the relationship between obesity and smoking (e.g., Brown et al., 2000; Lin, Huang, & French, 2004; Wilson, Habib, & Philpot, 2002). They have found an inverse relationship between smoking and body weight. Kuchler and Variyam (2003) concluded that nonsmokers were more likely to be obese than smokers. Lin et al. (2004) found a strong association between smoking and body mass index (BMI).1 Nayga (2000) reported similar ndings after controlling for the effect of health knowledge. Chou, Grossman, and Saffer (2004) suggested that the anti-smoking campaign, especially the state and Federal excise tax hikes and the settlement of state Medicaid lawsuits, had contributed to the recent upward trend in obesity. Although smoking may be associated with lower body weight, it has some undesirable health related effects. Jee, Lee, Nam, Kim, and Kim (2002) found that smoking increases the likelihood of an individual having a low BMI but a high waistto-hip ratio, which is a high-risk factor for diabetes mellitus. Furthermore, the weight-reducing effect of smoking in some studies may have been overestimated. Based on a medical experiment, Audrain, Klesges, and Klesges (1995) found that smoking increased resting energy expenditure in both normal-weight and obese smokers, but the metabolic effect is larger and lasts longer in normal-weight smokers. These ndings have potential implications for discouraging obese persons from taking up smoking and intervention among those who already smoke. The objective of this paper is to examine the relationship between smoking and BMI with a simultaneous equations system. It is, to our knowledge, the rst effort to correct for the simultaneity bias and expansion bias caused by endogeneity and censoring of the smoking variable in examining the relationship between cigarette smoking and body weight. The rest of the paper is organized as follows. The next section reviews relevant literature. A simultaneous equations model is outlined in the third section. The fourth section describes the dataset and variables used. Results are presented in the fth section, which is followed by the conclusion.

Literature Lifestyle and SES variables are important determinants of body weight (e.g., Chen, Yen, and Eastwood, 2005; Chou et al., 2004; Lin et al., 2004; Sun, 2003). Smoking of cigarettes is also related to lifestyle and SES; see, e.g., Yen (1999, 2005) for determinants of cigarette smoking, and Abdel-Ghany and Wang (2003) and Sharpe, Abdel-Ghany, Kim, and Hong (2001) for the role of SES on substance use. Lifestyle
1 BMI is calculated as weight in kilograms divided by height in meters squared; also see Table 1. In this study, both weight and height are self-reported.

123

J Fam Econ Iss (2007) 28:4967

51

variables include food/nutrition intakes and energy expenditures. Socioeconomic factors include household and individual characteristics.2 Lifestyle, SES and Bodyweight Lifestyle is considered an important factor contributing to overweight and obesity. Using data from the 199496 Continuing Survey of Food Intakes by Individuals (CSFII), Lin et al. (2004) examined the relationships among eating behaviors, dietary intake, physical activity, attitudes toward diet and health, sociodemographic variables and BMI among women and children in different income groups. Signicant correlations between BMI and age, race, dietary patterns, TV watching, and smoking were found for women. Kennedy, Bowman, Spence, Freedman, and King (2001) concluded that BMIs were signicantly lower for men and women on high carbohydrate diets, and the highest BMIs were noted for those on low carbohydrate diets. These ndings are consistent with the fact that diets high in carbohydrates and low/moderate in fats are usually lower in energy. Macdiarmid, Vail, Cade, and Blundell (1998) suggested that among women consumption of high-fat and sweet products contributes to obesity, and the altered representation of the data created by low-energy reporters appeared to distort the relationship. Field, Gillman, Rosner, Rockett, and Colditz (2003) found that fruit and vegetable consumption was not signicantly related to weight. Socioeconomic variables such as education, employment, income and personal rrez-Fisac, Rodr guez characteristics have been used to explain body weight. Gutie Artalejo, Guallar-Castillon, Banegas, and de Rey Calero (1999) identied illiteracy, sedentary lifestyle and energy intake as the determinants of geographic variations in BMI and obesity in Spain. In a study of 1966 northern Finland birth cohort, Laitinen, rvelin (2002) found that unemployment was associated with Power, Ek, Sovio, and Ja an increased risk of obesity among women after controlling for other SES variables. Subjects with low school performance and women with long unemployment histories were also found to be at increased risk of being obese. Examining CSFII 199496 survey data, Townsend, Peerson, Love, Achterberg, and Murphy (2001) found an unexpected and puzzling link between food insecurity and overweight status among women who were overweight and having a greater potential for increased incidence of obesity-related chronic disease. Kan and Tsai (2004) found a relationship between individuals knowledge concerning the health risks of obesity and their tendency to be obese. Nayga (2000) found a statistically signicant effect of health knowledge on obesity after controlling for education among US adults using the Diet and Health Knowledge Survey (DHKS) component of the 1994 CSFII. Based on the NHANES III survey for 1988 94, Kuchler and Variyam (2003) argued that information programs linking overweight and obesity with health risks might fail to induce diet and lifestyle changes if individuals do not recognize that they are overweight or obese. They suggested that correcting such misperceptions may increase the size of the population susceptible to a weight-health risk information program.
2

We note that the genetic factors are important in the study of obesity but unfortunately not available for the adult individuals in the CSFII survey and probably unavailable in most other datasets. It is hoped that the racial information and some of the social and food intake behavior we consider capture some genetic effects.

123

52

J Fam Econ Iss (2007) 28:4967

Cigarette consumption and bodyweight In general, as previously mentioned, researchers have found a negative association between cigarette smoking and body weight. Appendix Table 1 presents a review of studies that examined this relationship. However, smoking is likely an endogenous decision, and thus simultaneity between smoking and body weight equations may exist. Conventional statistical procedures undertaken in previous studies may be susceptible to at least two types of statistical bias. First, unobserved heterogeneity may affect the number of cigarettes smoked and body weight simultaneously. Perkins et al. (2000) found substitution effects between food intake and nicotine reinforcement. Yen (1999, 2005) and Yen and Jones (1996) found that SES variables are signicantly related to cigarette consumption. The set of determinants of cigarette consumption, therefore, is likely to overlap with that of body weight. Although we could control for certain types of food intake, the regressors are unlikely to include every aspect of smoking and body weight. Stress and many psychological attributes, for instance, are usually unavailable. Hence, treating the smoking decision as exogenous will produce biased estimates of the causal effects if, in fact, the decision is endogenous (e.g., Wooldridge, 2002b, p. 86 and p. 253). The second type of bias is the expansion bias discussed by Rigobon and Stoker (2003). The bias arises because the existence of a censored regressor could lead to a higher (in absolute value) coefcient estimate for the regressor. The combined effects of simultaneity bias and expansion bias may explain why researchers have obtained strong negative effects of smoking on BMI. Unfortunately, their conclusions may lead to unexpected and undesirable consequences. Klesges and Klesges (1988) noted that about one-third of smoking participants in a university sample reported using smoking to control weight, which would not have been recommended by any health professional. The potential endogeneity and simultaneity raise interesting questions. What exactly is the causal relationship between smoking and overweight/obesity or weight lossdoes smoking contribute to weight reduction after controlling for individual characteristics? Is smoking a potential means of weight control? Does model misspecication in previous studies lead to overestimation of the weight-reducing effect of smoking? What are the policy implications of an improved estimate of the relationship between smoking and being overweight? The following section develops a simultaneous equations model to address these issues.

Linear regression model with a censored endogenous regressor Our interest centers on the effects of smoking on BMI. Let C be the number of cigarettes smoked. Many previous studies treat C as exogenous, which likely leads to the regression of BMI (denoted y) on C and other explanatory variables (x). This model also ignores the expansion bias caused by a censored regressor and is formulated as y x0 b c C e ; 1

where b is a vector of parameters, c is a scalar parameter, and e is a random error. Quitting smoking is known to have at least a short-term effect on body weight

123

J Fam Econ Iss (2007) 28:4967

53

(e.g., Perkins et al., 2000). A binary variable indicating whether the individual had quit smoking is constructed from the survey responses and is included in the regression equation as well. Another approach is to treat the consumption of cigarettes as binary, i.e., replace C with smoker, which is (1) for smokers and (0) for non-smokers; see, e.g., Brown et al. (2000) and Lin et al. (2004). The binary variable approach does not permit estimating the effects of the extent of addiction, i.e., it could not distinguish a heavy smoker from an occasional smoker. More importantly, the endogeneity of cigarette smoking is still ignored in these approaches. Treating the binary variable smoker as endogenous results in a treatment effects model as illustrated in Barnow, Cain, and Goldberger (1981). We treat the endogenous smoking variable as censored, versus binary, from an information preserving point-of-view as the number of cigarettes smoked is available. In order to account for the simultaneity between smoking (C) and BMI (y) and to accommodate zero consumption of smoking (and thus to minimize the simultaneity bias and the expansion bias), a reduced form simultaneous equations model is proposed: C max0; w0 a u1 y x0 b cC u2 ;

where w is a vector of exogenous variables, a is a conformable parameter vector, and the error terms u1 and u2 are distributed as bivariate normal with zero means, standard deviations r1 and r2, and correlation q. The rst Eqn. (2) is a censored regression model which accounts for non-smokers in the sample. Derivation of the sample likelihood function is presented in Appendix A. The parameter estimates can be obtained using the method of maximum likelihood (ML). It is important to recognize that, due to the endogeneity and censoring of C, the parameter estimates by themselves may not reect the change in the dependent variable (y) resulting from a change in an explanatory variable. The effects of explanatory variables are more appropriately explored from the relevant conditional and unconditional means of the dependent variable. The expected values of y conditional on smoking, non-smoking and unconditional are, respectively, EyjC[0 x0 b cw0 a cr1 qr2 ! /w0 a=r1 Uw0 a=r1 3

EyjC

0 x0 b qr2

/w0 a=r1 Uw0 a=r1

Ey x0 b cr1 /w0 a=r1 cw0 aUw0 a=r1 ;

where /() and F() are the probability density function (pdf) and cumulative distribution function (cdf), respectively, of the univariate standard normal distribution. The marginal effects are derived by differentiating Eqns. (35). Detailed derivations of these equations and their marginal effects are presented in Appendix B. The marginal effects of each discrete variable are calculated as changes in the means of y

123

54 Table 1 Variable denitions and sample statistics Variable Denition

J Fam Econ Iss (2007) 28:4967

Mean

Standard Deviation

Continuous variables BMI Body mass index: (weight in kg.) (height in meters)2 C Average number of cigarettes per day Income Per-capita annual income (thousand) Age Age in years Education Highest grade completed (years) TV hours Average hours of TV/video game per day Beer Beer consumed per day (100 grams) Wine Wine consumed per day (100 grams) Bev. ratio % of carbonated beverages consumed among allnon-alcoholic beverages HH size Household size Fruits Total fruits consumed per day (100 grams) Grains Total grains consumed per day (100 grams) Sugar Total sugar and sweets consumed per day (100 grams) Milk Total milk and milk products consumed per day (100 grams) Eggs Total eggs consumed per day (100 grams) Meats Total meat, poultry, sh consumed per day (100 grams) Legumes Total legumes consumed per day (100 grams) Beverage Total non-alcoholic beverage consumed per day (100 grams) Discrete variables (1 = yes; 0 otherwise) Smoker (if the individual) smokes cigarettes City Resides in central city Suburban Resides in suburban area Rural Resides in rural area (reference) South Resides in the South Northeast Resides in the Northeast Midwest Resides in the Midwest West Resides in the West (reference) Food stamps Resides in a household authorized to receive food stamps Male Gender is male Hispanic Of Hispanic origin Other race Race is not Black or White Black Race is Black White Race is White (reference) Employed Employed full- or part-time Active work Occupation needs to be physically active Exercise Exercises at least once a week Low-fat milk Chooses to use skim or 1% milk Vegetarian A vegetarian Vitamins Takes any vitamin supplement(s) Alcohol Had used alcohol during the year Homeowner A homeowner Quit smoking Had quit smoking Source: Compiled from the CSFII 199496 (USDA-ARS, 2000)

26.40 4.94 16.85 49.66 12.68 2.65 0.10 0.13 0.22 2.81 1.57 2.89 0.22 2.12 0.20 2.13 0.25 9.72 0.26 0.29 0.45 0.26 0.37 0.18 0.25 0.20 0.07 0.52 0.04 0.07 0.11 0.82 0.57 0.66 0.48 0.30 0.03 0.47 0.64 0.71 0.27

5.26 10.47 12.90 17.10 3.14 2.16 0.39 0.57 0.31 1.47 2.08 2.05 0.40 2.35 0.38 1.65 0.70 7.56

123

J Fam Econ Iss (2007) 28:4967

55

in Eqs. (35) due to a nite change (i.e., from 0 to 1) in the variable, holding all other variables constant. The model used in this paper is the rst approach of handing censored regressors as in Rigobon and Stoker (2003). It is analogous to the treatment effects model, in which the endogenous variable is a binary indicator. There are two methods of estimation. The two-stage method uses the generalized residual from the rst stage as an instrument. The ML approach species a bivariate normal distribution and maximizes the sample likelihood. The two-stage method produces consistent estimates whereas the ML procedure produces efcient estimates when the error distribution is correctly specied. We use the latter.

Data Data from the CSFII 199496, conducted by the US Department of Agriculture (USDA-ARS, 2000), are used in this study. CSFII 199496 is the most recent survey of similar purpose and scale carried out by the USDA. The CSFII 1998 module is not used as it covers only children under 10 years old. The CSFII dataset is a multistage stratied area probability sample of non-institutionalized individuals in the US. Demographic characteristics of each individual, as well as self-reported height and weight index and food and nutrient intake information, are included. Table 1 presents the names and labels of these variables. The dependent variable in the weight equation is BMI. The explanatory variables include regions (South, Northeast, Midwest; reference = West), urbanization (city, suburban; reference = rural), food stamp recipient status, household income, age, education, gender (male), race (Black, other race; reference = White), ethnicity (Hispanic), employment status (employed), time spent watching TV or playing video games (TV hours), whether the individual exercised frequently (exercise) and whether work activities involved active physical efforts (active work). Also included in the BMI equation are food/nutrition intake and dietary pattern variables. They are intakes of fruits, grains, sugar, milk, eggs, meats, and legumes, consumption of beer, wine, and total non-alcohol beverages, and dummy variables indicating use of skim or 1% fat milk instead of 2% fat or whole milk (low-fat milk), whether the respondent was a vegetarian, and was taking vitamin supplement(s) (vitamins). The number of cigarettes smoked is included in the BMI regression equation to reect the effect of smoking on overweight. Also included is the dummy variable indicating whether the individual had quit smoking. To account for endogeneity and non-smokers in the sample, cigarette consumption is modeled as a Tobit model, and the observed consumption (C) is also included as an explanatory variable in the BMI Eqn. (2). All exogenous variables used in the BMI regression are included in the cigarette equation except the quitting indicator (quit smoking). Because only individuals over 21 can legally drink alcohol and smoke cigarettes, respondents under 21 years old were excluded. The nal sample included 7,876 observations (see Table 1), of which 52% were male. The average age was 49.7 years old, and the average year of education was 12.7 years. About 26% of the sample reported smoking of cigarettes on an average day. The average time spent watching TV or playing a video game was 2.65 h per day. Nearly half of the sample reported exercising at least once a week, and about 65% reported their health conditions to

123

56

J Fam Econ Iss (2007) 28:4967

be fair or better. Average per capita income was $15,850, and 7% of the individuals resided in households were authorized to receive food stamps. Average intakes of fruits, grains sugar, milk, eggs, meats, and legumes were 157, 289, 22, 212, 20, 213, 25 grams, respectively.

Estimation results Before estimating the system of equations by the ML procedure, BMI is regressed with four specications using OLS, and the results are presented in Table 2. In the rst and third models, the continuous variable C is used while the dummy variable smoker is used in the other models. The last two linear models include an additional variable quit smoking. Both measures of smoking have strong negative and statistically signicant effects on BMI without controlling for endogeneity. Effects of smoking cessation on BMI are also found, which had been ignored in many previous cross-sectional studies (e.g., Lin et al., 2004). The adjusted R2 ranges from 0.098 to 0.106. Inclusion of the smoking cessation variable gives the last two models slightly more explanatory power. These OLS estimates do not provide sufcient evidence to infer the effects of smoking on bodyweight because of the potential simultaneity bias and the expansion bias of including a censored regressor as discussed in Rigobon and Stoker (2003). ML estimation of the simultaneous equations system is carried out using the maxlik optimization routines in Gauss. Asymptotic standard errors of the parameter estimates are derived from Whites (1982) robust covariance matrix. The parameter estimates are presented in Table 3. The endogeneity test is equivalent to testing the null hypothesis that there is no correlation between the error terms of the two equations; see, e.g., Hamilton, Merrigan, and Dufresne (1997) for a similar endogeneity test. The likelihood ratio test (LR = 23.47, P < .001) leads to rejection of the null hypothesis that there is no correlation between the two error terms, suggesting endogeneity of cigarette smoking and justifying the joint estimation of the two equations. The negative value of q indicates that the unobserved heterogeneity affects body weight and the number of cigarettes smoked in different directions. The pseudo R2 (Wooldridge, 2002a, p. 465) is small (0.032), which is typical for crosssectional analysis. The estimated coefcient of C is negative and signicant according to the OLS estimates but insignicant according to the ML estimates. These conicting results suggest that the negative effects of cigarette smoking in the OLS estimates likely resulted from failure to accommodate endogeneity of cigarette smoking. Previous ndings in the literature that smoking helps weight control therefore may have resulted from simultaneity and expansion biases. Individuals in medical experiments smoke a number of cigarettes preset by the researchers, e.g., in Klesges and Klesges (1988). However, cigarette consumption is not exogenous in the long term because individuals choose their preferred level of cigarette consumption themselves. Quitting smoking has a statistically signicant and positive effect on BMI according to both the OLS and simultaneous-equations estimates. This agrees with previous ndings and suggests that an individuals weight might rebound to a higher level following smoking cessation, which questions the effect of smoking as a means of short-term weight control.

123

Table 2 OLS estimates (dependent variable = BMI) Model 2 Estimates 19.894a 0.689 0.162 0.145 0.166 0.189 0.176 0.250 0.006 0.046 0.021 0.020 0.023 0.129 0.342 0.257 0.200 0.160 0.158 0.118 0.028 0.364 0.142 0.118 0.201 0.008 0.127 0.132 0.102 0.587a 0.324 0.031a 0.723a 0.687a 0.398a 0.336a 0.325a 0.088a 0.405a 0.691b 0.702a 1.377a 0.077 0.075 0.447a 0.182a 1.566a 0.306b 0.021 0.020 0.023 0.131 0.343 0.259 0.201 0.160 0.159 0.119 0.028 0.366 0.142 0.119 0.202 0.008 0.128 0.133 0.103 0.343a 0.332a 0.096a 0.411a 0.661c 0.651b 1.482a 0.084 0.118 0.454a 0.187a 1.547a 0.328b 0.584a 0.302 0.033a 0.699a 0.584a 0.384a 0.481a 0.285b 0.025 0.031 0.409b 1.401a 0.007 0.002 0.163 0.145 0.167 0.189 0.177 0.251 0.006 0.047 0.454a 0.282c 0.053 0.042 0.395b 1.415a 0.008 0.007 0.693 0.450a 0.279c 0.055 0.050 0.393b 1.418a 0.008 0.008 0.346a 0.334a 0.097a 0.435a 0.652c 0.665a 1.475a 0.079 0.123 0.458a 0.188a 1.543a 0.323b 0.579a 0.292 0.034a 0.695a 0.567a 0.377a 19.743a 19.936a SE Estimates SE Estimates SE 0.690 0.162 0.145 0.166 0.189 0.176 0.250 0.006 0.046 0.021 0.020 0.023 0.130 0.342 0.257 0.200 0.160 0.158 0.118 0.028 0.364 0.142 0.118 0.202 0.008 0.127 0.133 0.103 Model 3 Model 4

Variable

Model 1

Estimates

SE

J Fam Econ Iss (2007) 28:4967

0.692

0.163 0.145 0.167 0.189 0.177 0.251 0.006 0.047

Constant 19.617a Household characteristics City 0.474a Suburban 0.279c South 0.028 Northeast 0.047 Midwest 0.406b Food stamps 1.407a Income 0.007 HH size 0.001 Personal characteristics Age 0.342a Age2/100 0.329a Education 0.088a Male 0.463a Hispanic 0.670c Other race 0.743a Black 1.348a Employed 0.064 Active work 0.085 Exercise 0.456a TV hours 0.185a Vegetarian 1.558a Home owner 0.294b Food/nutrient intakes Vitamins 0.576a Bev. ratio 0.300 Beverages 0.032a Low-fat milk 0.715a Alcohol 0.656a Wine 0.382a

0.021 0.020 0.023 0.129 0.343 0.258 0.201 0.160 0.159 0.119 0.028 0.366 0.142

0.119 0.202 0.008 0.128 0.132 0.103

57

123

Table 2 continued Model 2 Estimates 0.061b 0.020 0.111 0.006 0.369b 0.103a 0.218a 0.029 0.029 0.142 0.025 0.151 0.037 0.082 0.044 0.016 0.103 0.004 0.344b 0.104a 0.224a 0.445a 0.047a 4.994 F(35,7840)=25.12 0.099 1.646a 4.974 F(35,7840)=27.05 0.106 0.029 0.029 0.142 0.025 0.152 0.037 0.083 0.140 0.006 0.059b 0.020 0.112 0.006 0.364b 0.103a 0.220a 0.198 SE Estimates SE Estimates SE 0.029 0.029 0.142 0.025 0.151 0.037 0.082 0.143 0.151 Model 3 Model 4

58

Variable

Model 1

123
0.140 1.723a 4.975 F(35,7840)= 27.77 0.106
a

Estimates

SE

0.046 0.017 0.098 0.004 0.357b 0.105a 0.222a

0.029 0.029 0.142 0.025 0.152 0.037 0.083

Fruits Grains Sugar Milk Eggs Meats Legumes Quit smoking C Smoker Standard error F-statistic Adjusted R2 = 1%,
b

0.053a

0.006

4.997 F(35,7840) = 25.52 0.098 = 5%, and c = 10%

Note: Variables indicate levels of statistically signicance:

J Fam Econ Iss (2007) 28:4967

J Fam Econ Iss (2007) 28:4967 Table 3 Maximum-likelihood estimates of simultaneous equations of BMI and cigarettes Variable Cigarettes Estimates Household characteristics City Suburban South Northeast Midwest Food stamps Income HH Size Personal characteristics Age Age2/100 Education Male Hispanic Other race Black Employed Active work Exercise TV hours Vegetarian Home owner Food/nutrient intakes Vitamins Bev. Ratio Beverages Low-fat milk Alcohol Wine Fruits Grains Sugar Milk Eggs Meats Legumes Quit smoking C Constant Error std. dev. (r) R Log-likelihood SE BMI Estimates 0.435b 0.245c 0.058 0.066 0.403b 1.324a 0.007 0.016 0.310a 0.297a 0.066a 0.365a 0.865b 0.582b 1.481a 0.054 0.030 0.370a 0.163a 1.548a 0.174 0.550a 0.363 0.011 0.791a 0.805a 0.371a 0.015 0.000 0.115 0.010 0.309b 0.099a 0.215a 0.335b 0.015 19.769a 5.018a

59

SE

0.845 1.324 0.791 1.634 0.056 2.124 0.045 1.158a 1.825a 2.048a 1.400a 1.930b 10.080a 3.646b 1.967 0.840 6.117a 3.993a 0.878a 2.475 6.330a 1.489c 2.093 0.739a 4.146a 9.020a 1.201c 2.580a 0.878a 0.536 0.212 2.102b 0.206 0.207 36.146a 25.220a 0.177a 35,339.218

1.076 0.991 1.077 1.246 0.998 1.483 0.043 0.313 0.159 0.162 0.152 0.877 2.382 1.704 1.275 1.029 1.056 0.794 0.177 2.608 0.935 0.805 1.334 0.061 0.895 0.935 0.737 0.342 0.212 1.087 0.181 0.945 0.233 0.598

0.170 0.150 0.168 0.188 0.179 0.340 0.005 0.057 0.023 0.023 0.025 0.133 0.350 0.279 0.234 0.165 0.144 0.120 0.035 0.304 0.156 0.120 0.224 0.011 0.134 0.140 0.085 0.028 0.030 0.147 0.024 0.155 0.037 0.079 0.143 0.019 0.703 0.078

4.760 0.421 0.049

a

Note: Variables indicate levels of statistically signicance:

= 1%,

= 5%, and c = 10%

123

60

J Fam Econ Iss (2007) 28:4967

Most of the signs for the other coefcients are as anticipated and consistent across models. Among the household characteristics in the BMI equation, people who resided in the Midwest and those with household members in the food stamp program are more likely to have a high BMI than the reference group, which includes the residents of the Western states. Compared to individuals from rural areas, individuals residing in the cities are less likely to be overweight. Individual characteristics also have strong effects on BMI. Age has a nonlinear effect. Ethnicity and race also play a role, with Hispanic and Black respondents being more likely to have high BMI. Consistent with common belief about lifestyle, exercise has a negative effect whereas time watching TV or playing video games has a positive effect on BMI. Vegetarians have lower BMI, which may reect higher levels of self-control. Men are more likely to have higher BMI than women, while respondents with higher level of education tend to have lower BMI. This is likely related to the well-known bias of self-reported weight and height. Men tend to over-report their weight while women may under-report.3 Among the food/nutrient intake variables, the preference for low-fat milk has a positive effect on the BMI, which echoes the puzzle reported by Lin et al. (2004). Our conjecture is that either the use of low-fat milk increases the total amount of milk consumed or overweight respondents are more concerned with the fat content. Those who take vitamin supplements regularly and those who had used alcohol in the past year are more likely to have lower levels of BMI. More intakes of wine and legumes are associated with lower BMI while intakes of beverages, eggs, and meats are associated with higher levels of BMI. This sheds light on what diet pattern might be preferred for the purpose of weight control. Also reported in Table 3 are estimates of the cigarette equation. Respondents who are members of larger households smoke less than those from smaller households. Age has a nonlinear effect on the number of cigarettes smoked, as evidenced by signicance of age and its squared term. Education reduces the number of cigarettes smoked. Men smoke more than women. Respondents of different race or gender groups do display signicant variations in cigarette consumption. Hispanics smoke less than non-Hispanics, as do Black and individuals of other races than White. Work that involves heavy physical labor increases the number of cigarettes smoked. Time watching TV or playing video games is associated with increased smoking, as are use of beverages and alcohol and consumption of fruits, grains, and wine. Home ownership, active exercises, consumption of low-fat milk and frequent use of vitamin supplements are associated with reduced smoking. The marginal effects (changes in BMI with respect to changes in the explanatory variables) of the variables are presented in Table 4. The unconditional marginal effects are evaluated at the means of the variables for the whole sample. The marginal effects conditional on smoking and non-smoking are evaluated at the means of explanatory variables for the corresponding samples. The unconditional marginal effects are not signicantly different from the coefcient estimates in the BMI equation. In addition, there are notable differences between the unconditional and conditional effects, i.e., the effect of being Hispanic is insignicant for the smoker sample while signicant for the non-smokers and the whole sample.4
3 4

We thank an anonymous referee for suggesting the gender-related self-reporting biases.

We thank an anonymous reviewer for pointing out that it is not appropriate to compute the marginal effects merely at the full-sample means.

123

J Fam Econ Iss (2007) 28:4967 Table 4 Marginal effects of explanatory variables Variable Unconditional Estimates Continuous variables Income 0.007 HH size 0.012 Age 0.149a Education 0.070a TV hours 0.166a Bev. Ratio 0.356 Beverages 0.014 Wine 0.375a Fruits 0.023 Grains 0.003 Sugar 0.113 Milk 0.009 Eggs 0.316b Meats 0.100a Legumes 0.215a Discrete variables City 0.437b Suburban 0.249c South 0.056 Northeast 0.061 Midwest 0.402b Food stamps 1.331a Male 0.371a Hispanic 0.839b Other race 0.594b Black 1.475a Employed 0.057 Active work 0.011 Exercise 0.383a Vegetarian 1.555a Homeowner 0.197 Vitamins 0.555a Low-fat milk 0.778a Alcohol 0.777a SE Conditional on Smoking Estimates SE

61

Conditional on Nonsmoking Estimates SE

0.005 0.056 0.011 0.024 0.035 0.223 0.009 0.085 0.027 0.029 0.146 0.024 0.154 0.037 0.079 0.170 0.149 0.168 0.187 0.178 0.339 0.132 0.349 0.275 0.233 0.165 0.143 0.118 0.302 0.150 0.119 0.131 0.136

0.008 0.018 0.195a 0.106a 0.189a 0.302 0.033a 0.406a 0.090a 0.025 0.100 0.003 0.370b 0.105a 0.221a 0.459a 0.283c 0.035 0.018 0.401b 1.385a 0.421a 0.566 0.689b 1.424a 0.079 0.149 0.486a 1.620a 0.357b 0.593a 0.670a 0.540a

0.005 0.056 0.013 0.027 0.035 0.224 0.009 0.087 0.035 0.030 0.145 0.025 0.154 0.037 0.080 0.169 0.149 0.167 0.185 0.178 0.335 0.131 0.373 0.272 0.230 0.163 0.154 0.120 0.298 0.155 0.118 0.133 0.158

0.006 0.023 0.134a 0.057b 0.158a 0.376c 0.007 0.363a 0.002 0.006 0.119 0.011 0.296c 0.098a 0.213a 0.447a 0.264c 0.047 0.043 0.402b 1.355a 0.392a 0.740b 0.632b 1.454a 0.066 0.056 0.427a 1.582a 0.268c 0.571a 0.733a 0.680a

0.005 0.057 0.011 0.026 0.036 0.225 0.011 0.085 0.030 0.030 0.148 0.024 0.156 0.037 0.080 0.169 0.148 0.167 0.185 0.178 0.337 0.131 0.352 0.275 0.231 0.163 0.143 0.117 0.298 0.151 0.119 0.131 0.137

Note: Marginal effects are evaluated at the sample means of all explanatory variables for the full sample, smokers, and non-smokers, respectively. Variables indicate levels of statistically signicance: a = 1%, b = 5%, and c = 10%

Implications A brief comparison of our results with previous studies on the smokingoverweight relationship is made in Appendix Table 1. Except for Ruidavets, Bongard, Bataille, ` res (2002), most previous studies concluded there was a negative Gourdy, and Ferrie relationship. The inverse relationship between smoking and BMI observed in the cross-sectional literature might be falsely induced by model misspecication. There are two potential sources of model misspecication, i.e., ignoring the simultaneity and expansion biases. First, smoking is an endogenous decision that relates to

123

62

J Fam Econ Iss (2007) 28:4967

individual heterogeneity (personal characteristics, physical conditions, etc.). Regressing BMI on the number of cigarettes smoked does not allow for the simultaneity of cigarette smoking and obesity, and thus the estimates of coefcients are likely to be biased and inefcient. Second, most of the literature either ignored the zero-consumption problem or did not account for the effects of increased cigarette consumption. Chou et al. (2004) used cigarette price in their study of the effect of an anti-smoking campaign, which might induce the zero consumption problem. Those who do not smoke because the cigarette market price is higher than their reservation prices will continue to abstain from smoking. A small portion of the smokers will stop smoking after the cigarette price hike. Studies that treat smoking as a binary variable neglect the quantitative information of cigarettes smoked. The present study addresses these problems by treating the number of cigarettes smoked as an endogenous decision and by accommodating censoring in the variable as suggested in Rigobon and Stoker (2003). The implication of this study is that the negative relationship between smoking and BMI obtained through OLS should be interpreted with caution. The relationship between anti-smoking campaigns and the prevalence of obesity may be spurious due to the endogeneity of smoking. Most of the medical experiments which had shown that smoking has a temporary weight-reducing effect only considered the short-term change of weight or energy consumption. Smoking decisions in these studies were made to be exogenous while in the long run it is not so. The result of this study suggests cigarette consumption may not have a direct long-run effect on body weight after controlling for endogeneity. There is no support for using cigarette smoking as a means of weight control.

Appendix A Derivation of the likelihood function Denote z1 C w0 a=r1 and z2 y x0 b cC=r2 , the likelihood contribution for the regime with C = 0 is gu2 Z
w 0 a

hu1 ju2 du1

1

1 r 2 /z2 U

! z1 qz2 p 1 q2

The integral in the above follows from the bivariate normality of u1 and u2 and properties of the conditional pdf of u1 given u2. The likelihood contribution for the regime C > 0 is simply the bivariate normal pdf. Therefore, the sample likelihood function is L Y
C 0 1 r 2 /z2 U

Y
C[0

1 1 r 1 r2

! 1 1 2 2 p exp 2 q z z z z 1 2 2 : 21 q2 1 2p 1 q2

z1 qz2 p 1 q2

123

J Fam Econ Iss (2007) 28:4967

63

Appendix B Conditional Means of the Dependent Variable and the Marginal Effects of Explanatory Variables Taking the expectation of y in equation (2) conditional on smoking (C > 0), we have that EyjC[0 x0 b c ECjC[0 Eu2 ju1 [ w0 a x0 b c w0 a c Eu1 ju1 [ w0 a Eu2 ju1 [ w0 a ! ! /w0 a=r1 r2 /w0 a=r1 q x0 b c w0 a c r1 r 1 Uw0 a=r1 Uw0 a=r1 r1 ! 0 /w a=r1 : x0 b c w0 a c r1 qr2 Uw0 a=r1

B:1

In (B.1), the third equality follows from properties (truncated moments) of the univariate and bivariate normal distributions. Following a similar procedure, the mean of y conditional on non-smoking (C=0) is EyjC 0 x0 b c0 Eu2 ju1 w0 a x0 b qr2 ! /w0 a=r1 : Uw0 a=r1 B:2

Using equations (B.1) and (B.2), and recognizing that PrC[0 Uw0 a=r1 ; the unconditional mean of y (though conditional on exogenous variables) is Ey x0 b c r1 /w0 a=r1 c w0 a Uw0 a=r1 : B:3

Differentiating the expectations in equations (B.1), (B.2) and (B.3) with respect to a common element of x and w (say xj = wj), we have   @ EyjC[0 r2 r1 bj caj c q @ xj r1 "    #    aj /w0 a=r1 2 aj /w0 a=r1 w0 a B:4 Uw0 a=r1 Uw0 a=r1 r1 r1 r1 ( ) ! !    r2 /w0 a=r1 2 /w0 a=r1 w0 a bj caj aj c q Uw0 a=r1 Uw0 a=r1 r1 r1 @ EyjC @ xj 0 ( !   r2 /w0 a=r1 2 aj bj q r1 Uw0 a=r1 r1 r1  0  ' ! 0 /w a=r1 w a aj Uw0 a=r1 r1 r1 ( !2 !  0 ) 0 r2 /w a=r1 /w0 a=r1 wa bj q aj Uw0 a=r1 Uw0 a=r1 r1 r1

B.5

 0    0   0  0  aj 0 @ Ey wa wa w a w a aj zacU aj c r1 / bj c / @ xj r1 r1 r1 r1 r1 r1 bj c Uw0 a=r1 aj : B:6

123

64

123
Method OLS Result Implication Simultaneous equations system Statistically signicant negative coefcient of cigarette consumption. Statistically insignicant positive coefcient of cigarette consumption. Statistically signicant positive elasticity of BMI with respect to cigarette price. OLS Endogeneity contributed to the negative relationship between cigarette consumption and BMI estimated in previous studies. Smoking should not be suggested as a means of weight control. Anti-smoking campaigns contributed to the upward trend in obesity. Smokers tend to be thinner than non-smokers. Weighted Multivariate Regression of BMI Survey with six month followup, logistic regression Negative and statistically signicant coefcient of smoker dummy variable. Smokers had a lower BMI on admission. Weight gain occurred at a slower rate in smokers compared with non-smokers. For both men and women, the risk for paradox A (low BMI and high waist-hip ratio) was higher among smokers than non-smokers. Logistic regression Smoking cessation be encouraged as an adjunct to nutritional intervention in nursing home residents with nicotine dependence and weight problems. Smoking may increase the risk for paradox A; smokers are thus more likely to have diabetes mellitus.

Appendix Table 1 Comparison to previous smoking-overweight studies

Study

Data

This study

CSFII 1994-96, 7,876 adults

Chou et al. (2004)

Lin et al. (2004)

Wilson et al. (2002)

BRFSS 19841999 (the Behavioral Risk Factor Surveillance System), 1,111,074 adults CSFII 19941996, 2,419 adult women & 1,651 school-age children 88 nursing home residents in St. Louis

J Fam Econ Iss (2007) 28:4967

Jee et al. (2002)

Korean Nationwide Health Examination Survey, 3,450 men and 4,250 women

Appendix Table 1 continued Method Multiple linear regression Positive coefcients for past/current smoking dummy variables. Result Implication

Study

Data

Ruidavets et al. (2002)

330 free-living French men (4564y)

J Fam Econ Iss (2007) 28:4967

Brown et al. (2000)

Summary statistics

Greater representation of smokers in the two lowest BMI categories. Smoking increased resting energy expenditure in both weight categories but the extent differed.

Men who had never smoked had a lower waist hip ratio or BMI than current smokers or former smokers. Smoking is associated with lower BMI.

Audrain et al. (1995)

14,779 women (18 23 years) in the Australian Longitudinal Study survey in 1996 40 female smokers (20 normal weight and 20 obese) Medical experiment

Obese persons should be discouraged from taking up smoking to control weight.

65

123

66

J Fam Econ Iss (2007) 28:4967

References
Abdel-Ghany, M., & Wang, M. Q. (2003). Contemporaneous and intertemporal relationship between the consumption of licit and illicit substances by youth. Journal of Family and Economics Issues, 24, 281289. Audrain, J. E., Klesges, R. C., & Klesges, L. M. (1995). Relationship between obesity and the metabolic effects of smoking in women. Health Psychology, 14, 116123. Barnow, B., Cain, G., & Goldberger, A. (1981). Issues in the analysis of selectivity bias. Evaluation Studies Review Annual, 5, 4359. Brown, W. J., Mishra, C., Kenardy, J., & Dobson, A. (2000). Relationships between body mass index and well-being in young Australian women. International Journal of Obesity, 24, 13601368. Chen, Z., Yen S. T., & Eastwood D. B. (2005). Effects of food stamp participation on body weight and obesity. American Journal of Agricultural Economics, 87, 11671173 . Chou, S., Grossman, M., & Saffer, H. (2004). An economic analysis of adult obesity: Results from the behavior risk factor surveillance system. Journal of Health Economics, 23, 565587. Dardis, R., & Keane, T. (1995). Risk-benet analysis of cigarette smoking: Public policy implications. Journal of Consumer Affairs, 29, 351367. Field, A. E., Gillman, M. W., Rosner, B., Rockett, H. R., & Colditz, G. A. (2003). Association between fruit and vegetable intake and change in body mass index among a large sample of children and adolescents in the United States. International Journal of Obesity, 2, 821826. rrez-Fisac, J. L., Rodr guez Artalejo, F., Guallar-Castillon, P., Banegas, J. R., & de Rey Calero, Gutie J. (1999). Determinants of geographical variations in body mass index (BMI) and obesity in Spain. International Journal of Obesity, 23, 342347 . Hamilton, V. H., Merrigan, P., & Dufresne, E. (1997). Down and out: Estimating the relationship between mental health and unemployment. Health Economics, 6, 397406. Hedley, A. A., Ogden, C. L., Johnson, C. L., Carroll, M. D., Curtin, L. R., & Flegal, K. M. (2004). Prevalence of overweight and obesity among U.S. children, adolescents, and adults, 19992002. Journal of American Medical Association, 29, 28472850. Jee, S. H., Lee, S. Y., Nam, C. M., Kim, S. Y., & Kim, M. T. (2002). Effect of smoking on the paradox of high waist-to-hip ratio and low body mass index. Obesity Research, 10, 891895. Kan, K., & Tsai, W. (2004). Obesity and risk knowledge. Journal of Health Economics, 23, 907934. Kennedy, E. T., Bowman, S. A., Spence, J. T., Freedman, M., & King, J. (2001). Popular diets: Correlation to health, nutrition, and obesity. Journal of the American Dietetic Association, 101, 411420. Klesges, R. C., & Klesges, L. M. (1988). The use of cigarette smoking as a weight loss strategy in a university population. International Journal of Eating Disorders, 7, 413419. Kuchler, F., & Variyam, J. N. (2003). Mistakes were made: Misperception as a barrier to reducing overweight. International Journal of Obesity, 27, 856861. rvelin, M. R. (2002). Unemployment and obesity Laitinen, J., Power, C., Ek, E., Sovio, U., & Ja among young adults in a northern Finland 1966 birth cohort. International Journal of Obesity, 26, 13291338. Lin, B., Huang, C. L., & French, S. A. (2004). Factors associated with womens and childrens body mass indices by income status. International Journal of Obesity, 28, 536542. Macdiarmid, J. I., Vail, A., Cade, J.E., & Blundell, J. E. (1998). The sugar-fat relationship revisited: Differences in consumption between men and women of varying BMI. International Journal of Obesity, 22, 10531061. Mokdad, A. H., Ford, E. S., Bowman, B. A., Dietz, W. H., Vinicor, F., Bales, V. S., & Marks, J. S. (2003). Prevalence of obesity, diabetes, and obesity-related health risk factors, 2001. Journal of American Medical Association, 289, 7679. Nayga Jr., R. M. (2000). Schooling, health knowledge and obesity. Applied Economics, 32, 815822. Paeratakul, S., Lovejoy, J. C., Ryan, D. H., & Bray, G. A. (2002). The relation of gender, race and socioeconomic status to obesity and obesity comorbidities in a sample of U.S. adults. International Journal of Obesity, 26, 12051210. Perkins, K. A., Hickcox, M. E., & Grobe, J. E. (2000). Behavioral economics of tobacco smoking. In W. K. Bickel, & R. E. Vuchinich (Eds.), Reframing health behavior change with behavioral economics (chap. 10pp. 265292). Mahwah, NJ: Lawrence Erlbaum. Rigobon, R., & Stoker, T. M. (2003). Censored regressors and expansion bias. (Working Paper 445103) Cambridge: Massachusetts Institute of Technology,,Sloan School of Management. ` res, J. (2002). Eating frequency and Ruidavets, J. B., Bongard, V., Bataille, V., Gourdy, P., & Ferrie body fatness in middle-aged men. International Journal of Obesity, 26, 13291338.

123

J Fam Econ Iss (2007) 28:4967

67

Sharpe, D. L., Abdel-Ghany, M., Kim, H. Y., Hong, G. S. (2001). Alcohol consumption decision in Korea. Journal of Family and Economics Issues, 22, 724. Sun, H. (2003). Physical and mental health of contemporary Chinese children. Journal of Family and Economics Issues, 24, 355364. Townsend, M. S., Peerson, J., Love, B., Achterberg, A., & Murphy, S. P. (2001). Food insecurity is positively related to overweight in women. Journal of Nutrition, 131, 17381745. US Department of Agriculture, Agricultural Research Service (USDA-ARS). (2000). Continuing Survey of Food Intakes by Individuals 19941996, CD-ROM, Washington, DC. White, H. (1982). Maximum likelihood estimation of misspecied models. Econometrica, 50, 126. Wilson, M. M. G., Habib, S., & Philpot, C. (2002). Cigarette smoking and weight loss in nursing home residents. The International Journal of MedicineMedicine On-Line. URL http:// www.priory.com/med/cigsmoking.htm (Accessed 1 November 2004). Wooldridge, J. M. (2002a). Econometric analysis of cross section and panel data. Cambridge, MA: MIT Press. Wooldridge, J. M. (2002b). Introductory Econometrics. Mason, OH: Thomson SouthWestern. Yen, S. T. (1999). Gaussian versus count-data hurdle models: Cigarette consumption by women in the US. Applied Economics Letters, 6, 7376. Yen, S. T. (2005). A multivariate sample-selection model: Estimating cigarette and alcohol demands with zero observations. American Journal of Agricultural Economics, 87, 453466. Yen, S. T., & Jones, A. M. (1996). Individual cigarette consumption and addiction: A exible limited dependent variable approach. Health Economics, 5, 105117.

123

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.