Smoking vs Other Risk Factors as the Cause of Smoking-Attributable Deaths: Title and subTitle BreakConfounding in the Courtroom FREE

The US surgeon general attributes to cigarette smoking approximately 400,000 deaths annually, approximately one fifth of all deaths that occur in the United States.¹ Tobacco industry defendants criticize these estimates,^{2 - 3} arguing that the Centers for Disease Control and Prevention's (CDC's) method of estimating "smoking-attributable" deaths^{4 - 5} adjusts for age and sex, but fails to consider the lower socioeconomic and educational status and associated dietary, occupational, and other risk factors of modern smokers,^{6 - 13} thereby exaggerating the burden of mortality attributed to smoking.

We examined the impact of adjusting the relative and attributable risk estimates associated with smoking for other demographic and behavioral correlates of smoking aside from age and sex. We first assessed the effect of multivariate adjustment on relative risk (RR) estimates associated with current and former smoking in a large prospective American Cancer Society study begun in 1982,^{14 - 16} the same cohort that provides the RR estimates used by the CDC to calculate deaths attributable to smoking in the United States.^{4 - 5 ,17} We then used the CDC computer program Smoking-Attributable Mortality, Morbidity, and Economic Costs 3.0 (SAMMEC 3.0)⁴ to measure the extent to which smoking-attributable deaths may be overestimated or underestimated because of factors associated with smoking.

Our analyses also illustrate the concept of confounding, a term derived from the Latin verb confundere, "to mix together." Epidemiologists use the technical term confounding to signify a distortion of the apparent association between a factor and disease risk, brought about by the former being associated with other factors that influence (cause or prevent) the condition being studied.¹⁸ More general definitions of confounding include "To mix up or confuse—to confound one's enemy, to confound truth and lies."¹⁹ The 2 meanings may coexist in legal and public relations battles regarding tobacco.²⁰

The Cancer Prevention Study II (CPS-II) is a nationwide prospective mortality study of nearly 1.2 million US adults, aged 30 years or older, that began in 1982.^{14 - 16} Enrollees were asked by an American Cancer Society volunteer to complete a confidential, 4-page mailed questionnaire on tobacco and alcohol use, diet, and other factors potentially affecting mortality. The present analyses include 974,150 people (370,037 men and 604,113 women) who at enrollment reported either never smoking or currently or formerly smoking cigarettes only. We restricted analyses to persons who provided complete information on smoking and excluded men who had ever smoked pipes or cigars (Table 1). The mean age at the time of enrollment was 57 years. Study participants are more likely than the general US population to be college-educated, married, middle-class, and white (Table 2).¹⁶

Vital Status Follow-up. Deaths were ascertained from the month of enrollment through August 1988 via personal inquiries by the volunteers in September 1984, 1986, and 1988. During these 6 years, 8.6% of the men and 4.7% of the women participants died, and 1.8% and 2.3%, respectively, were lost to follow-up (Table 1). Death certificates were successfully obtained for 97.3% of deceased persons; from these the underlying cause of death was coded according to an abbreviated version of the International Classification of Diseases, Ninth Revision (ICD-9).²¹ We based our analyses on 6 rather than 4 years of follow-up (as used by the CDC)^{4 - 5 ,22} to maximize the number of deaths available for analysis, especially in lifelong nonsmokers, while minimizing misclassification of exposure due to active smokers quitting smoking during more prolonged follow-up.

End Points. Our analyses considered deaths in which the underlying cause was coded to any of the chronic diseases included in the CDC estimates of smoking-attributable mortality.^{4 - 5} Deaths due to cancer included malignancies of the trachea, bronchus, or lung (n = 4847), cancers of the lip, oral cavity, or pharynx (n = 289), esophagus (n = 290), pancreas (n = 1107), larynx (n = 91), uterine cervix (n = 82), urinary bladder (n = 375), and kidney and other urinary tract (n = 484). Because of the small number of cancers of the oropharynx, larynx, and esophagus among lifelong nonsmokers, we combined these sites into upper aerodigestive tract malignancies (n = 670) (Table 1). Circulatory disease deaths included ischemic heart disease (n = 14,685), arterial disease (n = 1223), other heart conditions (n = 5514), and cerebrovascular diseases (n = 3539). Respiratory deaths included chronic obstructive pulmonary disease (COPD) (n = 1884) and other respiratory diseases (n = 1146). We did not consider deaths from smoking-related pediatric diseases, burns, or environmental tobacco smoke. To parallel the RR estimates used in the CDC SAMMEC 3.0 software⁴ we included persons who reported prevalent cancer at the time of enrollment.

Classification of Smoking Behavior. The information on smoking provided in 1982 was used to categorize respondents into: (1) current cigarette smokers who reported smoking cigarettes at least once a day for 1 year or more and had never smoked other tobacco products; (2) former cigarette smokers who met the above criteria but who reported having stopped smoking before 1982; and (3) lifelong nonsmokers who reported never having smoked any tobacco product.¹⁴ No further information on smoking was collected for the overall cohort after 1982. Although the vital status of participants continues to be followed, we based the present analyses on the first 6 years of follow-up to minimize misclassification of current smokers who may have quit during follow-up. A repeat survey of 185,000 CPS-II participants in 1992-1993 found that approximately half of those who smoked when enrolled in 1982 reported no longer smoking 10 years later (data not shown).

Estimation of Relative Risk. Cox proportional hazards methods²³ were used to estimate the hazard ratio (HR) (corresponding to the RR) and 95% confidence intervals associated with current or former smoking in men and women separately, first adjusting only for age, and later for multiple other covariates. All multivariate models adjusted for age at enrollment within 1 year, race (white vs nonwhite), education (3 levels), marital status, "blue collar" employment in most recent or current job, and total weekly consumption of vegetables and citrus fruit.²³ For cardiovascular end points, statistical models also adjusted for any current aspirin use or alcohol consumption, body mass index (calculated as weight in kilograms divided by the square of height in meters), physical activity in work or play (4 levels), and weekly consumption of fatty foods.²⁴ For lung cancer and COPD the analyses controlled for a history of occupational asbestos exposure. Analyses of uterine cervical cancer controlled for age at menarche (4 levels), number of live births (5 levels), oral contraceptive use (yes/no), and menopausal status (6 levels). People with missing data on 5 factors were assigned to the most commonly reported category; thus, an additional 1.3% were classified as high school graduates, 1.5% as moderate exercisers, 0.4% as white, 0.5% as married, and 2.0% of men and 3.7% of women as unemployed. Persons reporting eating 3 or fewer servings of vegetables per week due to incomplete answers were classified as missing data on vegetable consumption. The imputation of missing covariates had negligible effect on the RR estimates.

We assessed the validity of the proportional hazards assumption by graphing Kaplan-Meier curves²⁵ and by testing the statistical significance of an interaction term between smoking and year of follow-up. P<.05 was considered statistically significant.

Estimation of Attributable Risk. Deaths attributable to smoking were estimated using the SAMMEC 3.0 software, which uses RR (or HR) estimates from CPS-II to calculate the attributable fraction of deaths due to smoking among current and former smokers for 19 disease categories.^{4 - 5} These are combined with population-based prevalence data on current, former, and never smoking in the United States in 1990,^{4 ,26} and with the total number of deaths in each of the 19 smoking-related disease categories.²⁷ We assessed the potential impact of confounding on deaths attributed to smoking by entering first the age-adjusted and then the fully adjusted RR estimates from CPS-II.

Men and women who currently smoked cigarettes when enrolled in CPS-II were on average 3 to 4 years younger than lifelong nonsmokers, somewhat less educated, and apt to report lower consumption of vegetables but more regular alcohol consumption (Table 2). The demographic differences between current and never smokers were small. Male current smokers were slightly less likely to be currently employed, more likely to have held a blue collar job and to report ever being exposed to asbestos, more likely to consume a diet higher in fat, not to take vitamin supplements, and to be less active (Table 2). Some of these associations were reversed among female former smokers. Women who had stopped smoking were more educated than the never smokers and more likely to be married and to consume vegetables, alcoholic beverages, and vitamin supplements (Table 2).

Adjusting for the small demographic and behavioral differences between smokers and lifelong nonsmokers produced only minor changes in the HR estimates associated with smoking (Table 3). Adjustment caused the HR estimates associated with ischemic heart disease and lung cancer (which together constitute approximately half of smoking-attributable deaths in the United States)⁵ to decrease by 9% and 5%, respectively, in male current smokers, and by 5% and 0% in female current smokers. Multivariate adjustment had the opposite effect in formerly smoking women, increasing the HR estimates between smoking and ischemic heart disease, COPD, and cancers of the lung, upper aerodigestive tract, bladder, and cervix (Table 3). Adjustment for the alcohol consumption of smokers also increased the HR between smoking and cardiovascular deaths,²⁸ partially offsetting the impact of the lower educational status, less frequent vegetable consumption, and physical inactivity of smokers. The proportional hazards assumption was not violated in any of these statistical models.

Table 4 illustrates the change in the number of deaths attributed to smoking when fully adjusted rather than age-adjusted HR estimates from CPS-II are used. The total estimate of smoking-attributable disease deaths among US adults decreases from 401,109 to 396,741, approximately 1%. The net change is correspondingly small in the 3 subcategories of neoplastic, cardiovascular, and respiratory diseases. Reasons for the negligible change in smoking-attributable mortality when adjusted for multiple correlates of smoking are discussed below.

We also compared the age- and sex-adjusted estimates of smoking-attributable deaths when calculated based on the CPS-II HR estimates from 4 rather than 6 years of follow-up. The estimated total deaths attributable to smoking in adults, again excluding pediatric diseases, deaths from burns, and environmental tobacco smoke, was 421,711 based on the 4-year follow-up and 401,109 based on 6-year follow-up. Most of this 4.9% decline was attributable to lower risk of cardiovascular death among former smokers over the 6-year follow-up.

Our principal finding is that after adjusting for age, further adjustment for behavioral and socioeconomic differences between smokers and nonsmokers minimally affects the RR estimates associated with smoking and has even less impact on the estimates of smoking-attributable deaths. These findings are important because they are based on the same large prospective study^{14 - 16} and attributable risk calculation methods⁴ that the CDC uses to estimate deaths due to smoking in the United States.^{1 ,5 ,17 ,22} Our results suggest that the current federal estimates of smoking-attributable mortality^{5 ,22} are not seriously biased by other demographic or behavioral factors associated with smoking, after controlling for age.

Most previous large prospective studies of smoking conducted between 1950 and 1980 predated the development of computers and statistical software programs that could adjust simultaneously for multiple covariates in very large data sets.^{29 - 33} Only 4 such studies in the United States have, to our knowledge, reported both multivariate- and age-adjusted RR estimates associated with active smoking.^{34 - 37} In none of these did adjustment for factors other than age and sex substantially alter the RR estimates. Our analyses go further by demonstrating that the small changes in the RR after multivariate adjustment result in even smaller changes in smoking-attributable deaths. Our findings complement those of Malarcher et al,³⁷ who reported that the smoking-attributable fraction underlying the CDC estimates was robust compared with alternative approaches for estimating RR.³⁸

Several considerations explain why adjusting for socioeconomic and behavioral factors that have become associated with contemporary smoking only minimally changes the estimates of RR and smoking-attributable deaths. First, the smokers and lifelong nonsmokers are more similar with respect to socioeconomic and educational status in our study than in the contemporary United States. In fact, the cohort is more homogeneous than is the overall US population, reflecting the demographic characteristics of American Cancer Society volunteers at the time. Many CPS-II smokers began tobacco use around World War II, and much of their smoking preceded the era when cigarette smoking became a behavior of the disenfranchised. The socioeconomic homogeneity of the cohort strengthens, rather than weakens, our ability to control for certain extraneous factors that might affect the survival of smokers.

Second, not all of the behaviors with which smoking was associated in our study were detrimentally associated with survival. CPS-II smokers did report less physical activity and lower dietary consumption of vegetables and fruit, both of which were associated with higher cardiovascular death rates in our study. However, smokers also reported more regular consumption of alcohol, which was associated with lower cardiovascular mortality.²⁸ Adjusting for all of these factors in combination produces a smaller net change in the RR estimates for cardiovascular mortality in CPS-II than would be expected if only selected factors were considered separately.

Third, the effect of multivariate adjustment among former smokers, especially in women who have stopped smoking, is to decrease the estimate of smoking-attributable deaths, partly offsetting the increase in the estimate that occurs among male current smokers. The higher educational status and healthy behavior of women who have quit smoking contrast sharply with the characteristics of male current smokers. It also can be argued that some of the factors that we controlled for in these analyses should not have been because they represent adverse effects of smoking and/or intermediates in the causal pathway. For example, both physical inactivity and low body mass index can be caused by smoking or smoking-related diseases. Controlling for these may cause attributable risk calculations to underestimate the true burden of disease caused by smoking.

Adjusting for the lower educational and socioeconomic characteristics of modern-day smokers had little impact on the relative or attributable risks associated with smoking. The absence of evidence of epidemiologic confounding may help resolve at least one aspect of the ongoing tobacco debate.