Background
Smoking and obesity are two of the greatest health hazards in the world, so a large body of research has investigated the causes and consequences of smoking and obesity. Recently, attention has been paid to the connection between smoking and obesity [
1]. It has been argued that smoking increases physical metabolism and reduces the consumption of sweet food; hence, smoking is believed to be related to lower body weight and reduced obesity [
2,
3]. Although there is much evidence to support the negative relationship between smoking and weight-related outcomes, almost all studies are based on findings of correlation rather than causation. In contrast to these studies, some health economists have debated the
causal effects of smoking on weight-related outcomes [
4‐
9]. They typically exploited instrumental variables (IVs), that is, variables that are correlated with smoking but uncorrelated with weight-related outcomes (referred to as exclusion restrictions). Examples include state cigarette costs in the US. Medical researchers generally do not use IVs, but IVs provide an effective means to tease out causality when double-blind, randomized, placebo-controlled trials are unavailable, as in this case. Therefore, health economists are keen to use IVs, whenever plausible IVs are available. Nevertheless, the small number of studies in health economics demonstrated that it was difficult to find plausible IVs, and the results were sensitive to the inclusion and construction of some covariates and the stratification of the samples.
Despite the growing interest in the issue, relatively little attention has been paid to it in the developing world. Considering the sheer number of and the rate of increase in the number of smokers there, however, the developing world deserves more attention. For example, in 1980, there were 441 million smokers (both sexes combined) in developing countries and 280 million smokers in developed countries. In 2012, the corresponding figures changed to 726 and 242 million [
10]. In addition, the number of deaths attributable to tobacco use was estimated to increase from 3.4 to 6.8 million deaths in low- and middle-income countries between 2002 and 2030 but to decrease in high-income countries during the same period [
11]. Meanwhile, in 1997, the World Health Organization (WHO) declared that obesity was a global epidemic [
12]. With obesity at a global scale, the developing world, thought to be unaffected by obesity, also witnessed an emerging epidemic. The proportion of overweight or obese (body mass index, BMI ≥25) adult women aged 18–49 grew by about 0.7 percentage points per year in 42 developing countries [
13].
Furthermore, a developing country offers an interesting opportunity for smoking-related issues because smoking is little discouraged there [
14]. Public awareness of the adverse effects of smoking is limited; anti-smoking policies are either absent or poorly enforced, and the advertisement, promotion, and sale of tobacco take place openly and widely, even to minors. Empirical results under these circumstances could provide insight into a situation where restrictions on smoking in the developed world are relaxed. If the relationship between smoking and weight is entirely based on a biological mechanism, these environmental differences between the developed and developing worlds would be inconsequential, and separate studies for developing countries would be redundant. It is, however, possible that some environmental factors, such as the health-consciousness of a population, influence smoking and weight-related outcomes. It is thus of great interest to separately consider developing countries. We aimed to estimate the approximately causal effects of smoking on weight, BMI, and obesity in Indonesia, while controlling for bias stemming from time invariant individual heterogeneity.
Methods
In 1993, the IFLS started to follow over 22,000 people in 7224 households in 13 provinces (IFLS1), which is representative of 83 % of the 1993 Indonesian population. The IFLS sampling scheme stratified by provinces and then randomly selected 321 enumeration areas (EAs) within each of the 13 provinces and then households within a selected EA. For each household selected, a representative member provided household-level demographic and economic information, and several household members were randomly selected and asked to provide detailed individual information.
Four follow-ups ensued in 1997 (IFLS2), 1998 (IFLS2+), 2000 (IFLS3), and 2007 (IFLS4). IFLS2+ is not publicly available and covers only a quarter of the original respondents for some ad hoc purposes. We thus excluded IFLS2+ and used all the other surveys. Most data were collected by interview, but some anthropometrics (e.g., height and weight) were measured in every survey year by two specially trained nurses. The user’s manual does not mention informed consent or institutional review board approval.
One questionnaire module concerned smoking, from which we extracted information on current smoking status and smoking intensity. Smoking status was determined by the answer to the following question: “Do you still have the habit [smoking] or have you totally quit?” If the answer was “still have,” we considered the respondent a current smoker, and if the answer was “quit” or he had not smoked before, a nonsmoker. We measured smoking intensity by the number of cigarettes smoked per day. For a small number of men who chewed tobacco or smoked a pipe, 1 g was assumed to be equal to one cigarette; the ratio of 0.8 g to one cigarette did not affect the following results (not shown). When we analyzed smoking intensity, we treated nonsmokers in two ways. First, we assigned zero to the number of cigarettes per day and created a dummy variable indicating a non-smoker. Second, we excluded the dummy of non-smoking and used only the continuous variable (i.e., number of cigarettes per day). In these two cases, we grouped smoking intensity as follows to allow nonlinearity in the effects of smoking on weight-related outcomes: 0, 1–9, 10–19 and 20+ cigarettes consumed per day. Smoking is an exclusively male habit in Indonesia: in the raw data of IFLS4, only 2.0 % of women aged 15+ had ever smoked, and 1.6 % were currently smoking. Consequently, we considered only men aged 15–55 years in IFLS1, each of whom was observed in all the survey years.
The longitudinal scheme allowed us to use fixed effects models. Specifically, we employed the following specification to estimate the effects of smoking on weight-related outcomes:
$$ {w}_{it}={\beta}_1{S}_{it}+{X}_{it}{\beta}_2+{u}_i+{\varepsilon}_{it}, $$
(1)
where
w
it
refers to individual
i’s weight-related outcome in year
t,
S to current smoking status or smoking intensity,
X to a vector of time-varying covariates,
β
1 and
β
2 to coefficients,
u to any individual characteristics that do not vary over time (i.e., time invariant individual heterogeneity), and
ɛ to the random error term. Potential correlation of
u and
S threatens against arguing
β
1 to be causal effects. For example, impatience in
u may lead to both smoking and low weight. Then,
β
1 would be biased upward (in absolute values) because
u and
S are positively correlated. On the other hand, health-consciousness in
u could lead to nonsmoking and low weight. In this case,
β
1 would be biased downward. Moreover, heredity contained in
u is believed to affect both
w and
S [
15‐
17]. It is thus empirically compelling to control for
u, which is a critical confounder; almost all studies in the literature neglected this point. For comparison purposes, we presented results derived from ordinary least squares (OLS) with standard errors clustered at the individual level. Comparisons of the results estimated by fixed effects models and OLS demonstrated the degree of bias in
β
1.
One concern for fixed effects models is reverse causality, but reverse causality is likely to be small. This is because although absence of evidence is not the same as evidence of absence, information on smoking as a weight control tool is lacking in Indonesia. Even if such information were as common in Indonesia as in the US, this would not make reverse causality serious because even in the US, where such information abounds, reverse causality occurred only for women [
18,
19]. Furthermore, because women are more sensitive to weight in general, some Indonesian women might smoke if they strongly believe that smoking reduces weight. The fact that they rarely smoke suggests that reverse causality is likely to be negligible. Another concern for fixed effects models is bias stemming from omitted time-varying covariates. For example, people care more about health over time, so they smoke less and lose weight during the same period. In this case, time-varying health consciousness drives changes in both smoking status and weight, thereby biasing
β
1. As shown in the next section, however, comparisons of results estimated by fixed effects models and OLS suggest that bias in
β
1 was upward, and even potentially overestimated
β
1 in the fixed effects models was small. Therefore, our main argument that
β
1 was small remains valid.
We considered weight, BMI, and obesity for
w. BMI is not an ideal measure for assessing obesity because it cannot distinguish fat from muscle, and among fat, total fat from abdominal fat, while obesity is concerning mainly because of abdominal fat. Nevertheless, BMI is still widely used, particularly in the social sciences because it is readily available and cost-effective in a large survey. In addition, particularly for Indonesia, BMI is a better predictor of obesity-related diseases than are other anthropometrics such as waist circumference, waist-to-height ratio, and waist-to-hip ratio [
20]. Hence, we relied on BMI for the assessment of obesity and used the cutoff for obesity not 30 but 25, following the suggestion by the WHO for Asians [
21]. This cutoff is reasonable because only 1.5 % of the sample exceeded a BMI of 30.
X included age and its squared term, height, marital status, urban (vs. rural) residence, and earnings. Age, marital status, and earnings were self-reported, height was measured, and urban residence was determined based on administrative information. We controlled for age because it is a basic demographic factor related to both smoking and weight-related outcomes; its squared term was intended to capture a potentially nonlinear relationship between the two. We controlled for height because tall people are generally heavier and height reflects early life conditions [
22‐
29]. We entered marital status because it is another basic demographic factor, but we found that it contributed little to explaining
w. Earnings refer to salaries or wages for paid employees or net profits for the self-employed, earned during the last month [
30‐
32]. We created a dummy indicating men with no earnings and assigned zero to their natural log of earnings. The idea was that earnings affect weight through smoking and including earnings-related variables would tease out the pure effects of smoking on weight.
In all specifications, we performed Hausman’s specification tests to check whether u was correlated with S or X, that is, whether fixed effects models were more appropriate than random effects models. In all cases, the test rejected the null hypothesis that both models would yield the same coefficients, suggesting that fixed effects models were more appropriate (not shown). The sample size varied depending on the covariates controlled for, but it exceeded 9000. We applied longitudinal weights to make the sample representative. Because the data are publically available, no ethical approval was required.
Discussion
We acknowledge that the IFLS does not contain information on fat and fat distribution. This information is important because even if smokers lose weight, they may gain abdominal fat, which is a more accurate measure of harm to health. That said, we argue that a lack of this information does not dramatically change our main argument because the effects of smoking on weight-related outcomes were consistently small for all the outcomes. Even if changes in smoking behavior change the distribution of fat, the small effects of smoking suggest that the change in abdominal fat is unlikely to be dramatic enough to pose a serious health threat.
Researchers have typically correlated smoking and weight-related outcomes, thereby failing to estimate the causal effects of smoking on weight-related outcomes. Recently, some health economists tried to address this issue by using IVs, but their results were not robust to specification changes. As a powerful alternative method, we employed fixed effects models. The most important feature of this method for our purposes is to control for time invariant individual heterogeneity. Methods based on correlation, notably OLS, cannot address this concern. We demonstrated the importance of controlling for it by comparing results derived from both methods. Although fixed effects models cannot address reverse causality and bias stemming from time-varying individual heterogeneity, we argue that they were probably negligible. Therefore, a fixed effects model provides a valuable alternative that estimates the effects of smoking on weight-related outcomes—possibly close to causality.
Our main finding is that estimation precision notwithstanding, the influence was very small: a quitter would gain weight by at most 1 kg. The results were similar for BMI and obesity. These small sizes of potentially overestimated effects strengthen our argument that the causal effects of smoking on weight are small. At the same time, our results demonstrate that ignoring time invariant individual heterogeneity results in great overestimates of the effects. Therefore, some previous findings of the large effects of smoking on weight-related outcomes might result from the failure of controlling for individual heterogeneity.
One may wonder why the effects of smoking on weight-related outcomes are so small. Given that we detected a nonlinear relationship between smoking intensity and weight-related outcomes in Table
6, it could be that smoking generates two contrasting effects. One is that smoking increases physical metabolism and reduces consumption of sweet food; the other is that chronic inflammation caused by smoking increases weight. The negative relationship is well recognized, but the positive one is not. Moreover, the fact that no study has found a positive relationship between smoking and weight suggests that the negative relationship is stronger than the positive one, but, according to our results, only slightly so. Based on this speculation, future research can determine the exact mechanisms, while focusing on the positive relationship as well.
Our results are based on data derived from a developing country but in line with the current knowledge in the US. Chou et al. initiated this line of research in health economics, using repeated cross-sections from Behavioral Risk Factor Surveillance System data, augmented by other data [
4]. However, they focused on the causal effects of cigarette prices (not smoking per se) on weight in a reduced form model. Analyzing the same data, Gruber and Frakes used IVs and found that people who quit smoking were 56 %
less likely to be obese; as they admitted, however, the size was implausibly large [
6]. At the same time, they argued that the sizes of Chou et al.’s estimates were similar to theirs with the opposite sign, casting doubt on the results of both studies. Chou et al. responded that the sizes of their estimates were reasonable, but this rejoinder still focused on the effects of cigarette prices on weight loss [
5].
Nonnemaker et al. reconciled these contrasting results [
7]. They pointed out that the key lay in controlling for state-specific time trends by demonstrating that the trends were correlated both with state cigarette costs (prices or taxes) and obesity and this led to a spurious relationship between cigarette costs and obesity; the real effects identified in this study were very small. Using the same idea but different data (the National Longitudinal Study of Youth 1979) and methods, Baum agreed with Nonnemaker et al. [
8]. Specifically, he controlled for the trends by creating comparison and treatment groups; the control group was not affected by state cigarette costs, whereas the treatment group was affected by the costs. The results suggested that types of cigarette taxes were immaterial once state-specific time trends were controlled for; the magnitudes of cigarette costs were also small.
Fang et al.’s study is of interest because they examined a developing country, namely, China [
9]. Their estimation strategy for causality was based on IVs, which also partially relied on cigarette costs. Their results suggested that if an average male smoker quit smoking (a reduction of 16.8 cigarettes per day), his BMI would increase by about two. The size of the effect was considerably larger than that suggested by medical research. Note that it took almost two decades even during the “obesity boom” in the US in 1976–1994 for BMI to increase by 1.5 [
36]. The large magnitude argued by Fang et al. could be attributed to the IVs: the coefficient on the number of cigarettes per day with the IVs was more than six times as great as that without the IVs. Although they argued that the IVs corrected for endogeneity, the difference seems too large, casting doubt on the exclusion restrictions of their IVs. This implausibility supports our empirical strategies.
Competing interests
I declare that I have no competing interests.