BMJ Open. How Smoking Affects the Proportion of Deaths Attributable to Obesity

Transcription

1 How Smoking Affects the Proportion of Deaths Attributable to Obesity Journal: BMJ Open Manuscript ID: bmjopen-0-00 Article Type: Research Date Submitted by the Author: -Jun-0 Complete List of Authors: Stokes, Andrew; Boston University, Global Health Preston, Samuel; University of Pennsylvania, Population Studies Center <b>primary Subject Heading</b>: Epidemiology Secondary Subject Heading: Epidemiology Keywords: EPIDEMIOLOGY, PUBLIC HEALTH, STATISTICS & RESEARCH METHODS BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

2 Page of BMJ Open How Smoking Affects the Proportion of Deaths Attributable to Obesity Andrew Stokes, PhD and Samuel H. Preston, PhD Department of Global Health and Center for Global Health and Development, Boston University School of Public Health, Boston, MA, USA; Department of Sociology and Population Studies Center, University of Pennsylvania, Philadelphia, PA, USA Keywords: body mass index, obesity, smoking, mortality, confounding Address for Correspondence: Andrew Stokes, PhD Boston University School of Public Health 0 Massachusetts Ave. rd Floor, Boston, MA 0 Tel: -- Fax: -- acstokes@bu.edu Word Count:, BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

3 Page of Strengths and Limitations Smoking and obesity are leading causes of premature mortality in the United States. This study considers in greater detail than previous studies how these two exposures interact. Using high quality nationally representative data, we show that smokers have a lower average body mass index than non-smokers and a lower relative risk of death associated with obesity. We demonstrate that the attributable fraction for obesity estimated among never-smokers is biased as an estimate for the entire population. We use a novel indicator of adiposity, the maximum body mass index an individual has attained in the course of life. A distinct advantage of this indicator is that it is less affected by reverse causality than body mass index at time of survey. A limitation is that it may be subject to measurement error. BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

4 Page of BMJ Open ABSTRACT Background Although ever-smokers make up the majority of the older adult population in the US, they are often excluded from studies examining the impact of obesity on mortality. Understanding how smoking and obesity interact is critical to assessing the proportion of deaths attributable to obesity. Methods We estimated the mortality risks of obesity by smoking status in US adults using data from the National Health and Nutrition Examination Survey. Analyses were carried out using Cox models and maximum body mass index (BMI) was specified as the key exposure variable. We also calculated population attributable fractions (PAFs) by smoking status and investigated differences in PAFs in a decomposition analysis. Results The hazard ratio associated with a one-unit increment in BMI beyond.0 kg/m was.0 for never-smokers (% confidence interval (CI).0-.0; p<0.00),.0 for former smokers (% CI.0-.0; p<0.0), and.0 for current smokers (% CI 0.-.0). We estimated that.% of deaths were attributable to excess weight. The PAFs were., 0. and. for never, former and current smokers respectively. The difference in PAFs between never and current smokers was almost entirely explained by the difference in hazard ratios. Conclusions The proportion of deaths attributable to obesity is nearly three times as high among neversmokers compared to current smokers. This finding is consistent with the fact that smokers are BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

5 Page of subject to significant competing risks. Analyses that exclude smokers are likely to substantially overestimate the proportion of deaths attributable to obesity in the US. BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

6 Page of BMJ Open INTRODUCTION Smoking and obesity are leading causes of premature mortality in the United States.[] How these risk factors interact has not been thoroughly investigated. In this paper, we focus on the impact of smoking on the proportion of deaths attributable to obesity in the contemporary United States. The mortality risks of obesity are often estimated after eliminating ever-smokers and people with chronic conditions from the sample and assuming that the estimated risks for never-smokers and healthy people apply to the entire population.[ ] These restrictions generally strengthen associations between obesity and mortality, in some cases greatly. However, the restrictions can exclude up to 0% of deaths, leading some researchers to question the external validity of the results.[] Proponents of strict exclusion criteria argue that such measures are necessary for obtaining valid estimates of the mortality risks of obesity.[] Smoking is thought to be so strongly related to obesity and mortality that it is difficult to avoid residual confounding even when using typical adjustments for smoking status and intensity. However, by removing smokers, a major source of risk that competes with obesity is removed, leading to the false impression that obesity is associated with a larger relative burden than it is.[,] Smokers may have a lower proportion of deaths attributable to obesity than non-smokers for two reasons. First, smoking changes the body mass index (BMI) distribution of the population. Many smokers lose weight, or fail to gain weight, because of smoking. Second, smoking may change BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

7 Page of the relative risk of death associated with obesity. With respect to the effect of smoking on the BMI distribution, many smokers lose weight, or fail to gain weight, because of smoking. A review of epidemiologic and biomedical studies of the effect of smoking on weight suggests that U.S. smokers weigh, on average, kg less than nonsmokers.[] When smokers quit, they gain, on average,. kg within months after quitting, and their weight returns to the same weight trajectory over age as that observed in nonsmokers. Smoking increases -hour energy expenditure by about 0%. Nicotine s effects on the brain also leads to suppression of appetite, and smoking per se can serve as a behavioral alternative to eating. A second reason why smokers may have a lower fraction of deaths attributable to obesity than non-smokers is that the relative risks of death associated with obesity may be lower among smokers. When one major exposure is added to the environment in which another exposure is operative, a high fraction of deaths may be caused by the additional exposure, reducing the fraction of deaths that remain to be caused by the original exposure. Whether the mortality risks associated with obesity are different for smokers and non-smokers has been investigated in four large American cohort studies ranging from,000 to. million participants. All found that the risk of death associated with obesity was greater among nonsmokers or never-smokers than among current smokers or ever smokers.[0 ] The Prospective Studies Collaboration pooled data on the mortality risks of obesity from studies including,000 participants.[] This study concluded that the excess risks for BMI and smoking were roughly additive rather than multiplicative. The implication is that the death rate, when graphed as a function of linear BMI, was displaced upwards by a nearly constant amount for BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

8 Page of BMJ Open smokers relative to non-smokers above the minimum BMI interval of. to kg/m. Such a displacement implies a lower relative risk of death associated with obesity for smokers than for non-smokers (see the Appendix, which develops an additive model of relations between two exposures). This paper builds on this literature by providing estimates of the proportion of deaths attributable to obesity for the entire population of the United States and for population sub-groups distinguished by smoking status. We use a summary measure of weight history, the maximum weight an individual has achieved, in constructing our exposure variable.[] [] Relative to baseline BMI, historical measures have the advantage of being less affected by reverse causality, the downward bias in the estimated mortality effects of obesity that results from weight loss among those who are seriously ill.[,] METHODS The data for this analysis were drawn from the National Health and Nutrition Examination Survey. We combined the NHANES III (-) and continuous NHANES cohorts (- 00) and linked these to mortality status in the National Death Index through 00.[] NHANES is a nationally representative survey of the non-institutionalized population of the United States that combines interviews and clinical measurement. A unique feature of NHANES is that it asks questions about weight histories, including an individual s maximum weight (exclusive of weight during pregnancy). This information, along with height measured at the time of survey, was used to construct the key exposure variable in our analysis, maximum BMI. This variable was chosen to represent the effects of BMI as a cumulative process and to BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

9 Page of minimize the effect of reverse causation, which biases downwards the estimated relative risk of death from obesity.[,] This effect is expected to be especially powerful among smokers because of the high incidence of illness and death in that group. We used Cox models to investigate the mortality risks of obesity among individuals aged 0-. We began with a relatively older age in order that individuals have accumulated substantial weight trajectories that can be suitably summarized by maximum BMI. Observations after individuals reach age were censored. We estimated hazards models separately for the three smoking groups. We distinguished those who never smoked (defined as having smoked fewer than 00 cigarettes in one s lifetime), those who currently smoke, and those who formerly smoked but who have quit. The key independent variable in the analysis, maximum BMI, was specified as units of BMI above. Using this variable linearly in a hazard model implies that risks increase exponentially above a BMI of (i.e., they increase linearly in the log of the hazard). Strong empirical support for such a shape emerged from the Prospective Studies Collaboration.[] Those in the BMI range of. to.0 were assigned a value of zero on this measure and those with BMI s below. were dropped from the analysis ( observations). In a preliminary analysis, we tested quadratic models to investigate non-linearities in the relation between BMI and the log of mortality. Coefficients on the quadratic terms were insignificant and thus dropped in subsequent modeling. BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

10 Page of BMJ Open The Cox models were adjusted for sex, age, race/ethnicity (Hispanics, Black Non-Hispanic and Other), and educational attainment (less than high school graduate, high school graduate, and more than high school). For current and former smokers, we additionally adjusted for smoking intensity using the following categories: less than pack/day, to packs/day, or more packs/day. Smoking intensity for former smokers was determined based on a question that asked about the number of cigarettes smoked per day at the time of quitting. The proportion of deaths attributable to obesity (population attributable fraction (PAF)) was calculated as the weighted average of PAFs in different strata, with the appropriate weights being the number of cases or deaths in the various strata.[] The formula used to assess the PAF within each smoking category is the following: = where pd i refers to the proportion of decedents in BMI category i and HR i refers to the hazard ratio with respect to mortality for an individual in category i. Individuals in the normal weight category were assigned a hazard ratio of.0. We used a two-step process to attribute the difference in PAF s between current smokers and never-smokers to differences in BMI distributions versus differences in the mortality risks of obesity. The approach is based upon equation (). To estimate how much of the difference in PAF s between current smokers and never-smokers is caused by differences in their BMI distributions, we re-calculated the PAF combining the hazard ratios of never-smokers with the BMI distribution of current smokers to produce a hypothetical distribution of deaths by BMI () BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

11 Page 0 of status. This process produces an estimate of what the PAF would be for never-smokers if they maintained their BMI-specific death rates but had the BMI distribution of smokers. Second, we used an analogous process of substitution to calculate what the PAF would be for never-smokers if they had the hazard ratios of current smokers while maintaining their own BMI distribution. We adjusted for unequal probabilities of selection and non-response using sample weights and accounted for the complex survey design of NHANES. Analyses were performed using STATA (StataCorp, Texas, USA) and variances were estimated with the SVY routine, which uses Taylor series linearization. This study was based on anonymous secondary data and thus did not require approval from an ethics committee. RESULTS Table describes characteristics of never-smokers, former smokers, and current smokers. Although current smokers were numerically the smallest of the three smoking groups, they experienced the largest number of deaths over the follow-up period. Current smokers were less likely to have been obese in the course of life than the other groups and more likely to have had a maximum weight in the normal BMI range of. to.0 (also see Figure A). On the other hand, former smokers were slightly more likely to have been overweight or obese in the course of life than never-smokers. (Table & Figure here) Table and Figure B present the estimated hazard ratios among the three smoking groups. The hazard ratio associated with a one-unit increment in BMI beyond.0 was.0 (% confidence interval (CI).0-.0; p<0.00) for never-smokers,.0 (% CI.0-.0; 0 BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

12 Page of BMJ Open p<0.0) for former smokers, and.0 (% CI 0.-.0) for current smokers. Thus, higher levels of BMI are a survival threat among all smoking groups, but the threat is more than twice as great among never-smokers as current smokers, with former smokers intermediate. The predicted relative risk associated with a BMI of 0 was.0 =.0 among never-smokers and.0 =. among current smokers. (Table here) Figure shows estimates of the population attributable fraction for obesity among the three smoking groups based on equation. It uses the hazard ratios associated with obesity that are shown in Table, in combination with the actual distribution of deaths by BMI in -unit wide categories of BMI. Among never-smokers,.% of deaths in this cohort were attributable to high BMI. Among current smokers, the proportion was only.%. Former smokers were located about halfway between the other two groups at 0.%. The PAF for the entire population is the death-weighted mean of these figures, or.%. (Figure here) Why is the fraction of deaths attributable to obesity lower among current smokers? One reason is that the BMI distribution of current smokers is shifted to the left relative to that of neversmokers. If the BMI distribution of current smokers were combined with the relative risks of never-smokers, Table shows that the PAF for never-smokers would be.%, compared to its actual value of.%. Thus, of the original difference in PAF s of 0. percentage points between never-smokers and current smokers, approximately.0 percentage points, or 0%, would be eliminated if never-smokers had the same BMI distribution as current smokers. If the relative risks for current smokers were combined with the BMI distribution of never-smokers, the PAF would be.0%, a reduction of.0 percentage points compared to the original PAF of BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

13 Page of never smokers of.%. This reduction represents % of the original difference in PAF s between the groups (note that the two hypothetical changes add to % rather than 00% of the original difference because of interactions between the two factors). Thus, these two exercises are consistent in showing that the low relative risks associated with BMI among current smokers are the dominant reason why their PAF is far below that of never smokers. (Table here) DISCUSSION It is clear that current smokers have a much lower fraction of deaths attributable to obesity than those who never smoked. It is also clear that the principal source of this difference is that the relative risks associated with obesity are much lower among current smokers than among neversmokers. Is the disparity between their PAF s consistent with the mortality risk from smoking itself? Two studies based on data from the National Health Interview Survey and the NHANES estimated mortality risks of smoking in the range of.-. compared to never-smokers.[0,] If the death rate of smokers were only one-third as great had they never smoked, then approximately two-thirds of deaths among smokers would be attributable to smoking itself. That leaves only the remaining one-third of deaths to be attributable to obesity and other factors. As a result, the PAF for obesity would be approximately one-third as great for smokers as for neversmokers. In fact, the PAF s for the two groups of 0. and 0. correspond closely to that expectation. The Appendix includes a more precise illustration of how the proportion of deaths attributable to one exposure should be expected to change when a new exposure is added. A similar correspondence also prevails among former smokers. The risk of death among former smokers relative to never-smokers is approximately..[,] So about one-third of the deaths BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

14 Page of BMJ Open among former smokers are attributable to smoking. If the remaining two-thirds had the same PAF as never-smokers, then the PAF among former smokers would be about 0.*0.= 0., which is its actual value. The implication of these findings is that when the population attributable fraction for obesity is calculated by excluding groups that are subject to other very significant health risks not only smokers but also people with diagnosed illnesses or health impairments and very old people then the attributable fraction will increase. When such exclusions are extensive, they can raise the estimated attributable fraction for obesity well above that pertaining to the population as a whole. We believe that the strategy pursued in this paper provides a useful approach to dealing with this dilemma. We exclude no one from the attributable risk estimation, so that it pertains to the population as a whole, including sick people and smokers. On the other hand, we use an indicator of obesity, an individual s maximum BMI, which pertains to the life cycle rather than simply to baseline circumstances. Health problems that may have reduced weight at baseline should have a smaller impact on an individual s maximum lifetime BMI. As a result, the use of maximum BMI should produce less bias in estimating the mortality effects of obesity. A potential weakness of the paper is the possibility of error in self-reported maximum weight. A Prior studies report a correlation between measured and recalled weight in the neighborhood of 0..[ ] However, maximum weight may be identified more accurately than weight at a BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

15 Page of specified earlier age since it does not require establishing a correspondence between one s age and one s weight trajectory. If we had used baseline BMI rather than lifetime maximum BMI in our analysis, the hazard per one-unit increment in BMI would be.0 (vs..0) for never-smokers,.00 (vs..0) for former smokers, and 0. (vs..0) for current smokers. Thus, among current smokers, higher BMI s would appear to be protective and the attributable risk of obesity among smokers would be negative. This implausible result is an indication of the strong biases that result from using baseline BMI in such calculations, especially among smokers. This is one of many examples of an obesity paradox that emerges when individuals are located in a highly diseased, high mortality environment. Reverse causation becomes a much more serious threat in such an environment[] and poses a more serious risk of bias. The fact that the obesity paradox is eliminated among current smokers by the use of lifetime maximum BMI is an indication that weight loss associated with illness has affected the distribution of baseline BMI is such a way as to produce an obesity paradox. The estimated fraction of deaths attributable to obesity is higher than most previous estimates in the United States.[,] Prior studies suggest that the association between obesity and mortality declines with age,[] thus exclusion of older individuals from the sample may be partly responsible for these findings. A second reason is that we use maximum BMI rather than baseline BMI as our indicator of adiposity. As just noted, the relative risks associated with maximum BMI are substantially greater than those associated with baseline BMI. We believe that the relative risks that we use are a more accurate representation of the hazards of adiposity BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

16 Page of BMJ Open than those associated with baseline BMI because estimates of the latter are biased downwards by confounding. The higher fraction of deaths attributable to obesity among never-smokers has implications for the set of future death risks facing Americans. Declines in smoking that have already occurred and that may keep occurring are likely to increase the proportion of deaths attributable to obesity. In the extreme, if no one ever smoked we could anticipate that the proportion of death attributable to obesity would rise from.%, the value for the contemporary US population aged 0-, to something closer to.%, the PAF for contemporary never-smokers. In conclusion, the proportion of deaths attributable to obesity among US adults ages 0- is nearly three times as high among never-smokers as among current smokers. The principal reason for this discrepancy is that current-smokers have a lower risk of death associated with obesity than non-smokers. Such a reduction is consistent with the fact that smokers are subject to a major risk that is competing with obesity and that is itself responsible for many deaths. Former smokers have a PAF that is roughly halfway between that of current smokers and that of never-smokers. BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

17 Page of Acknowledgements This project was supported by the National Institute on Aging [R0AG00]. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes on Aging or the National Institutes of Health. We thank Irma T. Elo, Doug Ewbank, Neil K. Mehta and Lucia Tiererova for their comments and suggestions. Competing Interests We declare no conflicts of interest Author Contributions Both authors contributed to all aspects of this article, including conceptualization, data analysis and writing. BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

18 Page of BMJ Open REFERENCES Murray CJL, Abraham J, Ali MK, et al. The state of US health, 0-00: Burden of diseases, injuries, and risk factors. JAMA 0;0:-0. Allison D, Fontaine K, Manson J, et al. Annual deaths attributable to obesity in the United States. JAMA ;:0-. Calle E, Rodriguez C, Walker-Thurmond K, et al. Overweight, obesity, and mortality from cancer in a prospectively studied cohort of US adults. N Engl J Med 00;:. Mokdad AH, Marks JS, Stroup DF, et al. Actual causes of death in the United States, 000. JAMA 00;:. Flegal KM, Graubard BI, Williamson DF, et al. Impact of smoking and preexisting illness on estimates of the fractions of deaths associated with underweight, overweight, and obesity in the US population. Am J Epidemiol 00;:. Berrington de Gonzalez A, Hartge P, Cerhan JR, et al. Body-mass index and mortality among. million white adults. N Engl J Med 00;:. Flegal KM. Bias in calculation of attributable fractions using relative risks from nonsmokers only. Epidemiology 0;:. Stokes A. Body-mass index and mortality among white adults [Letter]. N Engl J Med 0;:. Audrain-McGovern J, Benowitz N. Cigarette smoking, nicotine, and body weight. Clin Pharmacol Ther 0;0:. 0 Calle E, Thun M, Petrelli J, et al. Body-mass index and mortality in a prospective cohort of U.S. adults. N Engl J Med ;:0 0. Koster A, Leitzmann M. The combined relations of adiposity and smoking on mortality. Am J Clin Nutr 00;:0. Ma J, Jemal A, Flanders WD, et al. Joint association of adiposity and smoking with mortality among U.S. adults. Prev Med 0;:. Van Dam R, Li T, Spiegelman D, et al. Combined impact of lifestyle factors on mortality: prospective cohort study in US women. BMJ 00;:a0 a0. Prospective Studies Collaboration. Body-mass index and cause-specific mortality in 00,000 adults : Collaborative analyses of prospective studies. Lancet 00;:0. BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

19 Page of Stokes A. Using maximum weight to redefine body mass index categories in studies of the mortality risks of obesity. Popul Health Metr 0;:. Mehta NK, Stenholm S, Elo IT, et al. Weight histories and mortality among Finnish adults: The role of duration and peak body mass index. Epidemiology 0;:0 0. Joshy G, Korda RJ, Bauman A, et al. Investigation of methodological factors potentially underlying the apparently paradoxical findings on body mass index and all-cause mortality. PLoS One 0;:e. National Center for Health Statistics. Office of Analysis and Epidemiology. The Third National Health and Nutrition Examination Survey (NHANES III) Linked Mortality File, mortality follow-up through 00: Matching Methodology. Hyattsville, MD: 00. Benichou J. A review of adjusted estimators of attributable risk. Stat Methods Med Res 00;0:. 0 Jha P, Ramasundarahettige C, Landsman V, et al. st-century hazards of smoking and benefits of cessation in the United States. N Engl J Med 0;: 0. Thun MJ, Carter BD, Feskanich D, et al. 0-year trends in smoking-related mortality in the United States. N Engl J Med 0;:. Mehta N, Preston S. Continued increases in the relative risk of death from smoking. AJPH 0;0:. Bayomi DJ, Tate RB. Ability and accuracy of long-term weight recall by elderly males: the Manitoba follow-up study. Ann Epidemiol 00;:. Dahl AK, Reynolds C. Accuracy of recalled body weight: A study with 0-years of follow-up. Obesity 0;:. Perry G, Byers T, Mokdad A, et al. The validity of self-reports of past body weights by US adults. Epidemiology ;:. Preston SH, Stokes A. Obesity paradox: conditioning on disease enhances biases in estimating the mortality risks of obesity. Epidemiology 0;:. Wang Z. Age and the impact of obesity on mortality. Am J Public Health 0;0:. BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

20 Page of BMJ Open Figure Components of the Population Attributable Fraction: (A) BMI Distribution in the Deaths (%) and (B) Hazard Ratios for Mortality from All Causes (A) Percent (B) Hazard Ratio Never Smokers Former Smokers Current Smokers Body Mass Index (kg/m) Never Smokers Former Smokers Current Smokers Body Mass Index (kg/m) BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

21 Page 0 of Figure Population Attributable Fractions for US Adults Ages 0-, Total and by Smoking Category PAFs are calculated using Equation in the text. 0 BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

22 Page of BMJ Open Table Characteristics of US Adults Ages 0- by Smoking Status Never Smokers (n=,) Former Smokers (n=,0) Current Smokers (n=,0) No. % or mean No. % or mean No. % or mean Deceased Age at survey, years 0... Education Less than high school,.,.0,0. High school or equiv., More than high school,0.0,.. Race/ethnicity Hispanic,0... Non-Hispanic black... Other,0.0,.,0 0. BMI, Maximum (kg/m) Obesity status at maximum Normal... Overweight,.,.. Obese Class I,.,.. Obese Class II... Smoking Intensity (packs/day) 0 to <,..0 to <,. 0. >. 0. missing 0.. on September 0 by guest. Protected by copyright. BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from

23 Page of Categories of BMI are normal weight (.-.0 kg/m ); overweight (.0-. kg/m ); obese class (0.0-. kg/m ); and obese class (.0 kg/m or greater). Never smokers are defined as those having smoked less than 00 cigarettes in their lifetime. The characteristics pertain to persons aged 0-, surveyed in years -00 with mortality follow-up through 00. Estimates of means and percentages incorporate NHANES sample weights. Sources: National Health and Nutrition Examination Survey. BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

24 Page of BMJ Open Table Hazard Ratios for Mortality from All Causes Never Smokers Former Smokers Current Smokers Hazard Hazard Hazard % CI % CI Ratio Ratio Ratio % CI Sex Women Men Race/ethnicity Other Hispanic Non-Hispanic Black. *.0.. ***... **.0. Education Level Less than High School High School More than High School 0. * ** Smoking Intensity 0 to < to < >=. * BMI -.0 *** ** BMI: body mass index. Never smokers are defined as those having smoked less than 00 cigarettes in their lifetime. The sample includes persons ages 0-. Entry years are -00 with mortality follow-up through 00. Hazard ratios are derived from Cox proportional hazards models that adjust for gender, race/ethnicity (non-hispanic black, Hispanic, other) and educational attainment (less than high school, high school, some college or greater). Age is specified as analysis time. BMI- is calculated by subtracting on September 0 by guest. Protected by copyright. BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from

25 Page of from each person s maximum BMI. Individuals with BMI values between. and kg/m constitute the reference category and are assigned a value of 0. All estimates are weighted and account for complex survey design. Sources: National Health and Nutrition Examination Survey. ***p<0.00; **<0.0; *p<0.0 BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

26 Page of BMJ Open Table Decomposition of Difference in Population Attributable Fractions between Never and Current Smokers PAFs Counterfactual PAFs Decomposition Never Smokers Current Smokers Difference in PAFs Never Smokers HR with BMI of Current Smokers Change in PAF produced by BMI Never Smokers BMI with HR of Current Smokers Change in PAF produced by HR Contribution of BMI (%) Contribution of HR (%) PAF: population attributable fraction; BMI: body mass index; NS: never smoker; HR: hazard ratio. on September 0 by guest. Protected by copyright. BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from

27 Page of Appendix Formal Example of How the Population Attributable Fraction for Obesity Varies with the Relative Risk of Death from Smoking In this appendix, we develop an expression for the ratio of population attributable fractions (PAF) for obesity among non-smokers to that among smokers. This expression shows how the ratio depends upon the relative risk of death from smoking and on other parameters. The model from which this expression is derived assumes that the risks of death associated with obesity and with smoking are additive. It also assumes that the prevalence of obesity is the same among smokers and non-smokers. Although this latter assumption is not strictly correct, the empirical results presented in the text suggest that attributable risk fractions are not very sensitive to differences in obesity distributions between smokers and non-smokers. We recognize two categories of BMI, obese and non-obese; and two categories of smoking, smokers and nonsmokers. Notation ON µ = Death rate of non-obese, non-smokers BN µ = Death rate of obese non-smokers OS µ = Death rate of non-obese smokers BS µ = Death rate of obese smokers p= Proportion of each group that is obese Assumptions A) The relative risk of death from obesity among non-smokers is K, so that = K BN ON µ µ B) The relative risk of death from smoking among the non-obese is j, so that = j OS ON µ µ BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

28 Page of BMJ Open C) The risks of death associated with obesity and smoking are additive, so that BS ON ON ON ON ON ON µ = µ + ( Kµ µ ) + ( jµ µ ) = µ ( K+ j ) Dividing all death rates by the death rate of non-obese non-smokers produces the set of relative risks shown in Table A.. Table A. Relative Death Rates Classified by Smoking and Obesity Status Smoke Yes No Obese Yes j+k- K No j Note that if risks of obesity and smoking were multiplicative (i.e, additive in the logs), the value in the upper-left cell would be jk. In this case, it is easy to demonstrate that both the relative risks of death associated with obesity, and the attributable risks, would be the same for smokers and non-smokers. In the case of additive risks, the relative risk of death associated with obesity among smokers is (j+k-)/j, which is less than the relative risk among non-smokers, K, as long as j>. On the other hand, the additional absolute risk of death attributable to obesity is the same among smokers, K, as among non-smokers. The prevalence of obesity among the deaths of non-smokers is ON pµ K pk () = ON ON pµ K+ ( p) µ + p(k ) BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

29 Page of The prevalence of obesity among the deaths of smokers is ON p(j+ K ) µ p(j+ K ) () = ON ON p(j+ K ) µ + ( p) jµ p(k ) + The PAF for obesity among non-smokers, using equation () in the text, is () N pk K p( K ) PAF = = + p( K ) K + p( K ) The PAF for obesity among smokers is, S p(j+ K ) K p( K ) () PAF = = p( K ) + j j K + p(k ) + j The ratio of the PAF for obesity among non-smokers to the PAF for obesity among smokers is N PAF p( K ) + j () = S PAF p( K ) + Equation () is the expression that we seek. It says that the ratio of the PAF among non-smokers to that among smokers depends on the relative risk of death from smoking among the non-obese, j, as well as on the prevalence of obesity, p, and the relative risk of death from obesity among non-smokers, K. As K gets very large, the value approaches.0; smoking makes less and less of a difference for the ratio of PAFs when the relative risk of obesity gets larger and larger. As K is reduced to a value close to.0, the ratio of PAFs approaches j, the relative risk of death from smoking among the non-obese. Assume that the prevalence of obesity is 0., roughly the proportion in Table. Assume further that the relative risk associated with obesity among non-smokers, K, is. (corresponding to the relative risk for never-smokers at a BMI of.0) and that j =. (with j =., the relative risk of death from smoking among the obese would be (j+k-)/k=., so that the weighted BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

30 Page of BMJ Open mean relative risk for the two groups would be.0). When these values are substituted into equation (), the ratio of PAFs becomes., which is the approximate ratio identified in the paper (0.:0.). The model in this section and the parameters that we have chosen are first approximations. But they help to establish the plausibility of the relative PAFs that we have identified for current and never smokers. BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

31 How Smoking Affects the Proportion of Deaths Attributable to Obesity: A Decompositional Analysis Journal: BMJ Open Manuscript ID bmjopen-0-00.r Article Type: Research Date Submitted by the Author: 0-Nov-0 Complete List of Authors: Stokes, Andrew; Boston University, Global Health Preston, Samuel; University of Pennsylvania, Population Studies Center <b>primary Subject Heading</b>: Secondary Subject Heading: Epidemiology Epidemiology, Nutrition and metabolism, Smoking and tobacco, Public health Keywords: EPIDEMIOLOGY, PUBLIC HEALTH, STATISTICS & RESEARCH METHODS BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

32 Page of BMJ Open How Smoking Affects the Proportion of Deaths Attributable to Obesity: A Decompositional Analysis Andrew Stokes, PhD and Samuel H. Preston, PhD Department of Global Health and Center for Global Health and Development, Boston University School of Public Health, Boston, MA, USA; Department of Sociology and Population Studies Center, University of Pennsylvania, Philadelphia, PA, USA Keywords: body mass index, obesity, smoking, mortality, confounding Address for Correspondence: Andrew Stokes, PhD Boston University School of Public Health 0 Massachusetts Ave. rd Floor, Boston, MA 0 Tel: -- Fax: -- acstokes@bu.edu Word Count:, BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

33 Page of ABSTRACT Objective Although ever-smokers make up the majority of the older adult population in the US, they are often excluded from studies examining the impact of obesity on mortality. Understanding how smoking and obesity interact is critical to assessing the proportion of deaths attributable to obesity. Setting Nationally representative sample of the non-institutionalized population of the United States. Baseline data were drawn from the National Health and Nutrition Examination Survey, - and -00. Participants US adults aged 0- (n=,) Primary Outcome Measure We used Cox models to estimate the mortality risks of obesity by smoking status. All-cause mortality was assessed prospectively through December 00 (n=, deaths). Maximum body mass index (BMI) was specified as the key exposure variable. We also calculated population attributable fractions (PAFs) by smoking status and investigated differences in PAFs in a decomposition analysis. Results The hazard ratio associated with a one-unit increment in BMI beyond.0 kg/m was.0 for never-smokers (% confidence interval (CI).0-.0; p<0.00),.0 for former smokers (% CI.0-.0; p<0.0), and.0 for current smokers (% CI 0.-.0). BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

34 Page of BMJ Open We estimated that.% of deaths were attributable to excess weight. The PAFs were., 0. and. for never, former and current smokers respectively. The difference in PAFs between never and current smokers was almost entirely explained by the difference in hazard ratios. Conclusions The proportion of deaths attributable to obesity is nearly three times as high among neversmokers compared to current smokers. This finding is consistent with the fact that smokers are subject to significant competing risks. Analyses that exclude smokers are likely to substantially overestimate the proportion of deaths attributable to obesity in the US. BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

35 Page of STENGTHS AND LIMITATIONS Using high quality nationally representative data, this study considers in greater detail than previous studies how two leading causes of premature mortality in the United States interact. We use a novel indicator of obesity, an individual s maximum body mass index, which pertains to the life cycle rather than simply to baseline circumstances. Compared to body mass index measured at the time of survey, maximum body mass index is less affected by reverse causality a major source of bias in observational studies of the association between obesity and mortality. In contrast to many studies, we do not exclude major subgroups from the attributable risk calculation, so that it pertains to the population as a whole, including sick people and smokers. A limitation of the study is that maximum body mass index is self-reported and may be subject to measurement error. BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

36 Page of BMJ Open INTRODUCTION Smoking and obesity are leading causes of premature mortality in the United States.[] How these risk factors interact has not been thoroughly investigated. In this paper, we focus on the impact of smoking on the proportion of deaths attributable to obesity in the contemporary United States. The mortality risks of obesity are often estimated after eliminating ever-smokers and people with chronic conditions from the sample and assuming that the estimated risks for never-smokers and healthy people apply to the entire population.[ ] These restrictions generally strengthen associations between obesity and mortality, in some cases greatly. However, the restrictions can exclude up to 0% of deaths, leading some researchers to question the external validity of the results.[] Proponents of strict exclusion criteria argue that such measures are necessary for obtaining valid estimates of the mortality risks of obesity.[] Smoking is thought to be so strongly related to obesity and mortality that it is difficult to avoid residual confounding even when using typical adjustments for smoking status and intensity. However, by removing smokers, a major source of risk that competes with obesity is removed, leading to the false impression that obesity is associated with a larger relative burden than it is.[,] Smokers may have a lower proportion of deaths attributable to obesity than non-smokers for two reasons. First, smoking changes the body mass index (BMI) distribution of the population. Many smokers lose weight, or fail to gain weight, because of smoking. A review of epidemiologic and BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

37 Page of biomedical studies of the effect of smoking on weight suggests that U.S. smokers weigh, on average, kg less than nonsmokers.[] When smokers quit, they gain, on average,. kg within months after quitting, and their weight returns to the same weight trajectory over age as that observed in nonsmokers. Smoking increases -hour energy expenditure by about 0%. Nicotine s effects on the brain also leads to suppression of appetite, and smoking per se can serve as a behavioral alternative to eating. A second reason why smokers may have a lower fraction of deaths attributable to obesity than non-smokers is that the relative risks of death associated with obesity may be lower among smokers. When one major exposure is added to the environment in which another exposure is operative, a high fraction of deaths may be caused by the additional exposure, reducing the fraction of deaths that remain to be caused by the original exposure. Whether the mortality risks associated with obesity are different for smokers and non-smokers has been investigated in four large American cohort studies ranging from,000 to. million participants. All found that the risk of death associated with obesity was greater among nonsmokers or never-smokers than among current smokers or ever smokers.[0 ] The Prospective Studies Collaboration pooled data on the mortality risks of obesity from studies including,000 participants.[] This study concluded that the excess risks for BMI and smoking were roughly additive rather than multiplicative. They demonstrate that the death rate, when graphed as a function of linear BMI, was displaced upwards by a nearly constant amount for smokers relative to non-smokers above the minimum-risk BMI interval of. to kg/m. Such a displacement implies a lower relative risk of death associated with obesity for smokers BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

38 Page of BMJ Open than for non-smokers (see the Appendix Table A, which develops an additive model of relations between two exposures). This paper builds on this literature by providing estimates of the proportion of deaths attributable to obesity for the entire population of the United States and for population sub-groups distinguished by smoking status. We use a summary measure of weight history, the maximum weight an individual has achieved, in constructing our exposure variable.[,] Relative to baseline BMI, maximum weight may better capture potential cumulative effects of past obesity status and has the advantage of being less affected by reverse causality, the downward bias in the estimated mortality effects of obesity that results from weight loss among those who are seriously ill. METHODS The data for this analysis were drawn from the National Health and Nutrition Examination Survey. We combined the NHANES III (-)[] and continuous NHANES cohorts (-00)[] and linked these to mortality status in the National Death Index through 00. [] NHANES is a nationally representative survey of the non-institutionalized population of the United States that combines interviews and clinical measurement. A unique feature of NHANES is that it asks questions about weight histories, including an individual s maximum weight (exclusive of weight during pregnancy). This information, along with height measured at the time of survey, was used to construct the key exposure variable in our analysis, maximum BMI. This variable was chosen to represent the effects of BMI as a cumulative process[0,] and to minimize the effect of reverse causation, which biases downwards the estimated relative risk of BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.

39 Page of death from obesity.[,] This effect is expected to be especially powerful among smokers because of the high incidence of illness and death in that group. Cox models were used to investigate the mortality risks of obesity among individuals aged 0-. We began with a relatively older age in order that individuals have accumulated substantial weight trajectories that can be suitably summarized by maximum BMI. Observations after individuals reach age were censored. Hazards models were estimated separately for the three smoking groups. We distinguished those who never smoked (defined as having smoked fewer than 00 cigarettes in one s lifetime), those who currently smoke, and those who formerly smoked but who have quit. The key independent variable in the analysis, maximum BMI, was specified as units of BMI above. Using this variable linearly in a hazard model implies that risks increase exponentially above a BMI of (i.e., they increase linearly in the log of the hazard). Strong empirical support for such a shape emerged from the Prospective Studies Collaboration.[] Those in the BMI range of. to.0 were assigned a value of zero on this measure and those with BMI s below. were dropped from the analysis ( observations). In a preliminary analysis, we tested quadratic models to investigate non-linearities in the relation between BMI and the log of mortality. Coefficients on the quadratic terms were insignificant and thus dropped in subsequent modeling. The Cox models were adjusted for sex, age, race/ethnicity (Hispanics, Black Non-Hispanic and Other), and educational attainment (less than high school graduate, high school graduate, and BMJ Open: first published as 0./bmjopen-0-00 on February 0. Downloaded from on September 0 by guest. Protected by copyright.