点击显示 收起
1 From the Department of Social and Preventive Medicine, University at Buffalo, State University of New York, Buffalo, NY (JLF); the Departments of Biostatistics (JR), Nutrition (SAS-W, ER, WCW, and DJH), and Epidemiology (SAS-W, GC, ER, WCW, and DJH), Harvard School of Public Health, Boston, MA; the Cancer Prevention Studies Branch, Division of Clinical Sciences, National Cancer Institute, Bethesda, MD (DA); the Cancer Institute of New Jersey, UMDNJ–Robert Wood Johnson Medical School, New Brunswick, NJ (EVB); the Department of Epidemiology, Maastricht University, Maastricht, Netherlands (PAB); Harvard Center for Cancer Prevention, Boston, MA (GC, WCW, and DJH); the Channing Laboratory, Harvard Medical School/Brigham and Women's Hospital, Boston, MA (GC, DF, WCW, and DJH); the Department of Epidemiology, TNO Nutrition and Food Research Institute, Zeist, Netherlands (RAG); the Division of Epidemiology, School of Public Health, University of Minnesota, Minneapolis, MN (LH); the Department of Public Health Sciences, Faculty of Medicine, University of Toronto, Toronto, Canada (ABM); the Department of Epidemiology and Population Health, Albert Einstein College of Medicine, Bronx, NY (TER); Cancer Prevention and Control, H Lee Moffitt Cancer Center and Research Institute, Tampa, FL (TAS); and the Department of Epidemiology and Health Promotion, National Public Health Institute, Helsinki, Finland (JV)
2 Supported by research grants no. CA55075, CA78548, and AA 09802 from the National Institutes of Health.
3 Reprints not available. Address correspondence to JL Freudenheim, Department of Social and Preventive Medicine, 270 Farber Hall, State University of New York at Buffalo, Buffalo, NY 14214. E-mail: jfreuden{at}buffalo.edu.
See corresponding editorial on page 495.
ABSTRACT
Background: Although smoking is the primary cause of lung cancer, much is unknown about lung cancer etiology, including risk determinants for nonsmokers and modifying factors for smokers.
Objective: We hypothesized that alcohol consumption contributes to lung cancer risk.
Design: We conducted a pooled analysis using standardized exposure and covariate data from 7 prospective studies with 399 767 participants and 3137 lung cancer cases. Study-specific relative risks (RRs) and CIs were estimated and then combined to calculate pooled multivariate RRs by using a random-effects model.
Results: We found a slightly greater risk for the consumption of 30 g alcohol/d than for that of 0 g alcohol/d in men (RR: 1.21; 95% CI: 0.91, 1.61; P for trend = 0.03) and in women (RR: 1.16; 95% CI: 0.94, 1.43; P for trend = 0.03). In male never smokers, the RR for consumption of 15 g alcohol/d rather than 0 g alcohol/d was 6.38 (95% CI: 2.74, 14.9; P for trend < 0.001). In women, there were few never-smoking cases and no evidence of greater risk (RR: 1.35; 95% CI: 0.64, 2.87). Because of possible residual confounding by smoking, we performed sensitivity analyses by reclassifying the never smokers in the highest drinking category as former smokers. Resulting associations for alcohol consumption were somewhat attenuated, but P for trend = 0.05 for men, which was near the original P = 0.03.
Conclusions: A slightly greater risk of lung cancer was associated with the consumption of 30 g alcohol/d than with no alcohol consumption. Alcohol consumption was strongly associated with greater risk in male never smokers. Residual confounding by smoking may explain part of the observed relation.
Key Words: Alcohol consumption • diet • epidemiology • lung neoplasms • meta-analysis
INTRODUCTION
There is clear evidence that smoking causes lung cancer, but much about the etiology of lung cancer is not well understood, including why some nonsmokers develop lung cancer and why only a portion of smokers do so. It has been suggested that some of this variation may be explained by alcohol consumption (1, 2). Alcohol is oxidized to acetaldehyde, a known carcinogen (3). There is evidence that alcohol can act as a prooxidant in tissues, including lung tissue (4–12), and on lipids, including lung membrane lipids (4, 13). Alcohol can induce the expression of enzymes that are related to carcinogen metabolism (14), and compounds other than ethanol that are contained in alcoholic beverages may have carcinogenic effects.
In studies of alcoholics, morbidity and mortality due to lung cancer have been shown to be high (15–21), but the greater risk may be explained, in part or entirely, by the fact that the people in these populations were also more likely to smoke. Most studies of alcoholics have not controlled for the smoking status of participants. In studies measuring both alcohol consumption and smoking in individuals, there is some (although inconsistent) evidence of a modest increase in lung cancer risk in association with alcohol consumption (2, 22). In a meta-analysis, there was evidence of a greater risk of lung cancer associated with heavier drinking in cohort and hospital-based case-control studies but not in population-based case-control studies (22). Many studies examining alcohol and lung cancer have been limited by small sample size, particularly with respect to heavy drinkers. Furthermore, given the correlation between alcohol consumption and smoking, the examination of the effect of alcohol in never smokers is particularly informative. However, most studies have not been able to examine risk in this group because of the small numbers of lung cancer cases identified in never smokers. In the Pooling Project of Prospective Studies of Diet and Cancer (23-25), the primary data from 7 cohort studies of diet and cancer have been reanalyzed by using standardized categories of alcohol consumption and of all potential covariates including the smoking variables. These data allow for the examination of subgroups of interest to address the issue of a possible association of alcohol consumption with lung cancer. We report here pooled results from these cohorts with respect to the association between alcohol consumption and the risk of lung cancer.
SUBJECTS AND METHODS
Methods for the Pooling Project of Prospective Studies of Diet and Cancer have been described previously (23–25). The Pooling Project was originally designed to examine associations between dietary factors and breast cancer risk. It has now been expanded to include analyses related to other cancer sites, and it therefore includes cohorts of men. Inclusion criteria for cohort studies in the pooled analyses of lung cancer were 50 incident cases of lung cancer, an assessment of usual diet, a validation study of the diet instrument or of a closely related instrument, and assessment of smoking status at baseline. For these analyses related to lung cancer, 2 cohorts included previously in the breast cancer analyses [the New York University Women's Health Study (26) and the Swedish Mammography Cohort (27)] were not included because smoking data were not collected at baseline. In addition, as in the previous study of alcohol and breast cancer (24), the Adventist Health Study (28) was not included because of the low prevalence of alcohol consumption in that population. Each of the studies included here had been reviewed and approved by the institutional review board of the institution at which the study was conducted.
As shown in Table 1, included in this report were 5 cohorts with women [the Canadian National Breast Screening Study (29), the Iowa Women's Health Study (1), the Netherlands Cohort Study (30), the New York State Cohort (31), and the Nurses' Health Study (NHS; 32, 33)] and 4 cohorts with men [the -Tocopherol ß-Carotene Cancer Prevention Study (ATBC; 34), the Health Professionals Follow-up Study (32, 33), the Netherlands Cohort Study (30), and the New York State Cohort (31)]. Because there were repeated assessments of smoking and alcohol consumption for the NHS, this cohort has been analyzed in 2 sections: NHS Section A, including the 1980–1986 follow-up, and NHS Section B, including the 1986–1996 follow-up. Incident cases from the NHS were counted in just one of the cohorts (although a woman could contribute person-time as a noncase to NHS Section A and become a case in NHS Section B). Because the person-time in the different time periods was asymptotically uncorrelated even though the data were based on the same participants (35), the use of pooled estimates from the 2 time periods did not differ from the use of estimates from a single time period but had the advantage of using the updated exposure data from 1986.
View this table:
TABLE 1. Characteristics of the cohort studies included in the pooled analysis of alcohol and lung cancer
Lung cancer outcome ascertainment
Each study ascertained incident lung cancers by using follow-up questionnaires with subsequent medical record review (32), linkage with a cancer registry (1, 29–31) or both (34). In addition, some studies used mortality registries or death certificates (1, 29, 31, 32, 34). We categorized lung cancer on the basis of the International Classification of Diseases for Oncology morphology codes (36) or the histologic classification provided by the original study investigators.
Alcohol consumption assessment
For most of the cohorts, alcohol consumption was assessed with queries regarding the consumption of beer, wine, and spirits. Some of the questionnaires included information regarding red and white wine or other specific beverages. For the New York State Cohort, only total alcohol consumption was queried; this cohort was not included in beverage-specific analyses. The format of the questionnaires varied. Some allowed participants to indicate both frequency of drinking and the usual number of drinks on each occasion; others asked participants to choose among categories indicating total usual alcohol consumption. Most of the questionnaires assumed a standard drink size; the Canadian National Breast Screening Study allowed participants to indicate a drink size that differed from the standard indicated, and the ATBC Study allowed participants to choose from 1 of 3 portion-size options. Alcohol consumption in grams per day was calculated for each study by using the reported frequency of consumption, the beverage-specific alcohol content, and the average amount consumed.
Smoking assessment
For each cohort, baseline cigarette smoking status (never, current, or former smoker) was assessed. The ATBC Study (34) was limited by design to men who were currently smokers. For all of the cohorts, data were also obtained regarding duration of smoking in those who had ever smoked. For most studies, the information about the amount smoked was the amount of smoking at baseline. For the New York State Cohort, the usual number of cigarettes smoked daily was ascertained. Because few of the studies included questions regarding pipes, cigars, and other tobacco products, only cigarette smoking was considered here. Only the Netherlands Cohort Study included a detailed assessment of passive smoking.
Statistical analysis
For each dataset, after applying the exclusion criteria used by that study, we further excluded participants if they reported a history of cancer (other than nonmelanoma skin cancer) at baseline, were missing information regarding alcohol consumption, reported energy intakes either > or < 3 SDs from the study-specific loge-transformed mean energy intake of the baseline population, or were lacking information on smoking status, the number of years of smoking by past and current smokers, or the number of cigarettes smoked daily by current smokers.
We examined risk related to alcohol consumption by using the Cox proportional hazards model (37). Relative risks (RRs) were estimated for categories of total alcohol consumption and for consumption of alcohol from beer, wine, and spirits. P for trend across categories was calculated by taking the study-specific median for each category and assigning that value to all the participants in that category. This variable was then entered into the regression model.
Incidence rate ratios for the ATBC Study, the Health Professionals Follow-up Study, the Iowa Women's Health Study, the New York State Cohort, and the NHS were estimated by using SAS PROC PHREG software (version 8.2; SAS Institute, Cary, NC; 38). For the Canadian National Breast Screening Study and the Netherlands Cohort Study, the analysis was of a case-cohort study (39) and used EPICURE software (version 2.11; Hirosoft, Seattle, WA; 40). In all the studies, age at baseline and the year that the baseline questionnaire was returned were included as stratification variables. Person-years of follow-up were calculated from the date the baseline questionnaire was returned until the date of lung cancer diagnosis, the date of death, or the end of follow-up, whichever came first.
We compared the following approaches for controlling for confounding by smoking: adjustment for smoking status, adjustment for pack-years smoked, adjustment for smoking status and duration of smoking, and adjustment for smoking status, duration of smoking, and amount smoked. Of these, the last model (ie, smoking status, duration of smoking, and amount smoked) explained more of the variation in risk than did the others, and it is the one reported here.
Because of differences between men and women in the distribution of alcohol consumption and of beverage-specific use, we conducted sex-specific analyses. For the cohorts that included both men and women, ie, the Netherlands Cohort Study and New York State Cohort, men and women from each cohort were analyzed separately. All RRs were adjusted for smoking status (never, past, or current), smoking duration (for past and current smokers; continuous), amount smoked (for current smokers; continuous), education (<high school graduate, high school graduate, or >high school graduate), body mass index (in kg/m2) (<23, 23 to <25, 25 to <30, or 30), and energy intake (continuous). We had previously found fruit intake and ß-cryptoxanthin to be inversely associated with the risk of lung cancer (25, 41). We examined confounding by both of these factors but found that results were very similar with and without inclusion of either the fruit or ß-cryptoxanthin variable; results reported here are without adjustment for either of these dietary variables. An indicator variable for missing responses for covariates was created, when applicable. There was missing information for 7% of participants for each variable within each study. Two-sided 95% CIs and P values were calculated. We used the random-effects model to combine the loge RRs; study-specific RRs were weighted by the inverse of their variance (42).
Heterogeneity among studies was calculated by using the Q statistic (42, 43). For analysis of associations between specific types of alcoholic beverages and lung cancer, a test for between-study heterogeneity was carried out simultaneously on all alcoholic beverages (beer, wine, and spirits); the results were not statistically significant. A contrast test based on the beverage-specific associations and their estimated covariance matrix was used to test for differences between the associations. This test statistic approximately follows a chi-square distribution with 2 df.
For total alcohol consumption, risk was calculated according to strata of smoking status and several dietary variables. We examined analyses stratified by intake of fruit and of ß-cryptoxanthin because of the previous finding that each was inversely associated with risk of lung cancer in the Pooling Project (25, 41). We examined risks within strata of folate intake because of interactions between folate and alcohol at the level of absorption and metabolism (44) and examined risks within strata of vitamin A intake because of previous findings of differences in risk associated with alcohol depending on vitamin A status (45). We tested whether there were differences in the RRs between the strata by using a meta-regression model (46). We also tested whether associations differed between adenocarcinomas, small cell carcinomas, and squamous cell carcinomas by using a 2-df Wald test statistic (47). Collectively, these 3 histologic types represented 60% of the cases in each study.
RESULTS
This study included 7 cohorts with 399 767 participants and 3137 lung cancer cases (1375 females and 1762 males) (Table 1). Drinking rates varied among the cohorts. Mean drinking was highest in the ATBC Study and lowest in the cohorts of women. The Iowa Women's Health Study had the largest percentage who reported being nondrinkers (55%). Nearly half of the male cases were members of the Netherlands Cohort Study. Among the female cases, the largest number came from the Iowa Women's Health Study and the NHS.
As shown in Table 2, there was a tendency for the prevalence of smoking to increase with greater alcohol consumption; the proportion of never smokers decreased and that of current smokers increased. However, the amount smoked by smokers was similar in each category although somewhat higher for the highest category of alcohol. The one exception was the ATBC Study, which was limited to smokers, such that, at every level of alcohol consumption, 100% of the subjects were smokers. The correlation between alcohol consumption in drinkers and cigarettes smoked daily in current smokers ranged from 0.09 in the females in the New York State Cohort to 0.24 in the subjects (all of whom were male) of the ATBC Study; the median correlation in the cohorts was 0.16, and all correlations were significant (data not shown).
View this table:
TABLE 2. Smoking habits by alcohol consumption category and by study1
Study-specific and pooled multivariate RRs of lung cancer in relation to alcohol consumption are shown in Table 3. In models adjusted only for age, the pooled RRs for consumption of 30 g alcohol/d compared with no alcohol consumption were 1.76 (95% CI: 1.15, 2.67) in men and 3.09 (95% CI: 2.06, 4.63) in women. In multivariate models that included adjustment for smoking, the pooled RRs associated with reported consumption of 30 g alcohol/d compared with no alcohol consumption were 1.21 (95% CI: 0.91, 1.61) in men and 1.16 (95% CI: 0.94, 1.43) in women; neither of these values was statistically significant. For both men and women, the test for trend in this association was statistically significant (P = 0.03). The heterogeneity by sex was not statistically significant (P = 0.72 for between-studies heterogeneity due to sex for 30 g alcohol/d category); the pooled estimate for men and women combined was 1.18 (95% CI: 1.00, 1.39) in a comparison of consumption of 30 and 0 g alcohol/d (data not shown). Among the men, when the analysis excluded those cases diagnosed within 4 y of the baseline data collection, the pooled multivariate RR for the upper category of consumption was 1.48 (95% CI: 0.96, 2.29; P for trend = 0.02). There was significant heterogeneity among the cohorts for this analysis (P = 0.04). In the corresponding analysis in women, the RR was slightly closer to the null than it was when all cases were included; the results were not statistically significant (pooled multivariate RR: 1.06; 95% CI: 0.83, 1.35; P for trend = 0.16).
View this table:
TABLE 3. Alcohol consumption (g/d) and lung cancer: study-specific and pooled multivariate-adjusted relative risks (RR)1
Risks associated with specific beverages are shown in Table 4. For these analyses, because of the small numbers of heavy drinkers of a single beverage, the upper category was collapsed to those drinking 15 g alcohol of that beverage daily. There was significant between-studies heterogeneity for sex in the 15 g/d category for alcohol from beer (P = 0.004) and spirits (P = 0.03) but not from wine (P = 0.24) (data not shown). In men, there was no increase in risk associated with alcohol from beer. For alcohol from wine, there was evidence of decreased risk (pooled multivariate RR: 0.87; 95% CI: 0.55, 1.39; P for trend = 0.04). When men drinking 15 g alcohol/d from spirits were compared with those who drank no spirits, the pooled multivariate RR was1.34 (95% CI: 1.09, 1.66; P for trend = 0.04). Because there were many cases in the consumption category of 15 g alcohol/d, we also examined the risk of lung cancer in men who consumed 15 to <30 (pooled multivariate RR: 1.15; 95% CI: 0.87, 1.51) and 30 (pooled multivariate RR: 1.65; 95% CI: 1.19, 2.27) g alcohol/d from spirits (P for trend = 0.012; data not shown). There was significant heterogeneity among men in the associations of the 3 beverage types with risk of lung cancer (P = 0.03 for consumption of 15 g alcohol/d). In women, among those drinking 15 g alcohol from beer/d, the pooled multivariate RR was 1.88 (95% CI: 1.45, 2.42; P for trend < 0.001). The CIs included the null for the other 2 beverages for the upper category RR estimates. For women, the P value for heterogeneity among the 3 beverages in the upper category of consumption was < 0.01. Separate data on the consumption of red and white wine were available for a subset of studies; these separate results were similar to those shown for total wine consumption by both men and women (data not shown).
View this table:
TABLE 4. Risk of lung cancer associated with alcohol consumption from beer, wine, and spirits by sex: pooled multivariate-adjusted relative risks (RR)1
RRs for lung cancer stratified by smoking status are shown in Table 5. The ATBC Study had no participants who were never or former smokers, and thus it was not included in analyses for those categories. The Canadian National Breast Screening Study and the New York State Cohort were not included in the highest drinking category for never smokers because of insufficient sample size, but they were included in the other categories. Also in these analyses, because of limitations of sample size, the highest category of consumption was 15 g alcohol/d. There was significant (P = 0.01) heterogeneity between studies in the stratum of nonsmokers by sex but not among the current and former smokers (data not shown). In male never smokers, the risk of lung cancer was greater in the consumption category of 15 g alcohol/d than in the category of no alcohol consumption (pooled multivariate RR: 6.38; 95% CI: 2.74, 14.9; P for trend < 0.001). In the same comparison in the women, the increase was more modest and the CI included the null (pooled multivariate RR: 1.35; 95% CI: 0.64, 2.87; P for trend = 0.98), but the estimate was based on only 8 cases in the highest category, and 2 of the cohorts had no cases in that category. In both male and female former smokers, alcohol consumption was not associated with risk of lung cancer.
View this table:
TABLE 5. Alcohol consumption and lung cancer by smoking status: pooled multivariate-adjusted relative risks (RR)
We also examined risk associated with alcohol consumption in former smokers who had quit smoking 10 y before the baseline. There was no association between alcohol consumption and risk in this group (data not shown). For current smokers, upper-category pooled multivariate RRs were 0.94 and 1.10 for men and women, respectively, and CIs included the null (P for trend were = 0.92 and 0.02, respectively). When we examined the association between alcohol consumption and lung cancer risk in current smokers who reported smoking <20 cigarettes/d at baseline, we found no association. The association between alcohol consumption and lung cancer risk was modified by smoking status in men but not in women (P < 0.001 and = 0.76, respectively, for between-studies heterogeneity by smoking status in the upper category of alcohol consumption). We also examined beverage-specific analyses stratified by smoking status; the number of cases in each category was small, and the CIs were wide. Whereas there was evidence of an inverse association with wine consumption in men in the overall group, when results were stratified by smoking status, this inverse association was observed only in former smokers (pooled multivariate RR: 0.54; 95% CI: 0.29, 1.01 for 15 versus 0 g alcohol/d); for never and current smokers, there was a nonsignificantly greater risk in association with wine drinking (data not shown).
We examined risk by tumor histology (Table 6). In the highest category of alcohol consumption, there was significant between-studies heterogeneity by sex for small-cell carcinomas (P = 0.01) but not for adenocarcinomas and squamous cell carcinomas (data not shown). There was some evidence that alcohol consumption was more strongly associated with the risk of adenocarcinomas in both men and women and with the risk of small-cell tumors in men. However, the difference in the RRs was not significant; for men and women in the 30 g alcohol/d category, P for common effects according to cell type was 0.51 and 0.26, respectively.
View this table:
TABLE 6. Alcohol consumption (g/d) and lung cancer by histologic type: pooled multivariate-adjusted relative risks (RR)1
Residual confounding by smoking was a concern in our analyses. We therefore examined the effect on results if all participants coded as never smokers were in fact former smokers. We recalculated RRs after assigning all never smokers the study-specific median duration of smoking reported by former smokers. These RRs were quite similar to our original calculations; CIs overlapped almost completely (data not shown). Because there might be more misclassification in reported smoking status in the heavier drinkers, we also examined a model in which we reclassified as former smokers the never smokers in the upper category of alcohol drinking. For this latter analysis, the RRs were closer to the null than in the original analyses, and the CIs included the null. Nonetheless, the P for trend was 0.05 for the men (data not shown).
We also examined whether there were differences in risk associated with alcohol consumption for dichotomous strata of fruit, ß-cryptoxanthin, folate and vitamin A consumption determined by their respective median values. There was no evidence of heterogeneity in effects across strata (data not shown).
DISCUSSION
In this pooled analysis, there was weak evidence for a positive association between alcohol consumption and lung cancer risk. Trends were significant, although the CIs for the risk estimates in the highest category of alcohol consumption included the null. In men, the consumption of spirits was associated with risk more than was that of other beverage types; in women, risk was associated with the consumption of beer than with that of wine or spirits. There was a strong positive association between alcohol consumption and risk in never smoking men that was not observed in female never smokers. However, the small number of female cases who were never smokers in the highest drinking category limited our ability to examine associations in this group. There was no evidence of an association of alcohol consumption with risk in smokers, who constitute the largest portion of cases in these cohorts.
Several potential mechanisms have been proposed for an effect of alcohol on lung cancer. These include carcinogenesis by the alcohol metabolite acetaldehyde (3), oxidation by alcohol (4–12), and induction by alcohol of cytochrome p450 that affects the metabolism of other procarcinogens (4, 13). Other compounds in alcoholic beverages may also be relevant (2), which could explain the different associations we saw for beer and spirits than for wine. However, we saw differences in beverage-specific associations in both men and women; the observed beverage-specific differences may not be biological. Furthermore, it has been hypothesized that the mechanism of alcohol may be to enhance carcinogenic effects of cigarette smoke on tissues (2); our findings were strongest for nonsmokers, however, which would not be consistent with this explanation.
Many epidemiologic studies have examined the association between alcohol consumption and lung cancer, and the results have been inconsistent (2), perhaps as a result of the strong effect of smoking on lung cancer risk, the likely weak effect (if any) of alcohol, and the different sizes of the studies. In a meta-analysis of studies of alcohol consumption and lung cancer (22), point estimates for combined findings in 13 cohorts [including 2 of the cohorts in our pooled analyses (31, 48)] were similar to those we report here. These investigators also conducted a meta-analysis of 10 case-control studies; there was an increase in lung cancer risk associated with alcohol consumption in hospital-based but not population-based case-control studies (22). In 2 cohort studies not included in our analysis (49, 50) and in several case-control studies (51–53) published since the publication of the meta-analysis, findings were also consistent with ours.
Because smoking is such a strong risk factor for lung cancer and because smoking is correlated with alcohol consumption, the major concern in the examination of an association between alcohol consumption and lung cancer is failure to fully control for confounding by smoking. Measurement error in measured aspects of smoking (ie, smoking status, duration of smoking, and amount smoked) and variations in other unmeasured aspects (ie, depth of inhalation and length of time that smoke is held in the lungs) may have an effect on the estimation of the risk of lung cancer from other factors correlated with smoking, such as alcohol (2, 22, 54). It may be that our findings of differences by beverage type may be the result of uncontrolled confounding. Of interest with respect to residual confounding by smoking is our finding of a large increase in risk of lung cancer in men who never smoked. There have been several other studies of alcohol consumption and lung cancer in nonsmokers (22, 55–58). Most of these studies have found an increased risk associated with alcohol consumption, albeit generally at somewhat lower levels of alcohol consumption.
By contrast with the findings of increased risk in male never smokers, we found little evidence of increased risk in former and current smokers in our study, even in those who had quit smoking 10 y before or for current smokers of <20 cigarettes/d, groups that should be more similar to the never smokers with respect to their lung cancer risk. There was some evidence that female current smokers in the highest category of alcohol consumption were at increased risk. Whereas the CI for the highest category included the null, there was evidence of a significant trend in the data. Other studies have reported increased risk with alcohol consumption in heavy smokers (2, 59, 60).
Because of our finding of increased risk of lung cancer with alcohol consumption in men who reported never smoking, we investigated the possible effect of a misreporting of smoking status on the estimate of risk by recalculating risk estimates after reclassifying never smokers as former smokers. Risk estimates changed only somewhat. In contrast, in simulations regarding misclassification by smoking status, Korte et al (22) found that the RR for their meta-analysis of cohort studies was consistent with no true effect, 10% misclassification of smokers as nonsmokers, and no misclassification of drinking status.
An additional potential source of residual confounding could be the effect of passive smoke exposure. In particular, the group of nonsmokers who were included in the highest category of alcohol consumption might have a heavier exposure to smoke if they drank in smoke-filled environments. We had limited or no information on passive smoke exposure in all but one of the cohorts. However, passive smoke exposure is unlikely to explain an association of the magnitude of that observed in nonsmoking men. Furthermore, there may be some confounding due to the smoking of pipes and cigars. Limited information was available on the smoking of pipes and cigars in the current study.
Misclassification of drinking status—particularly the inclusion of former drinkers in those reporting that they are currently nondrinkers—is another potential source of bias. Misclassification of this nature would attenuate risk estimates if alcohol consumption does increase lung cancer risk. In fact, there was some indication in both men and women that there was a lower risk of cancer in those reporting moderate consumption than in nondrinkers. Our findings were also consistent with an alcohol effect limited to heavier drinking. Because of smaller numbers of heavier drinkers, we were not able to explore in detail the associations with heavy alcohol consumption.
We found some evidence of differences in the association between different beverage types and the risk of lung cancer: a stronger association for spirits in men and for beer in women. Other studies have found greater risk in association with consumption of beer and spirits but not wine (2). These differences may be related to compounds found in beer or spirits. When we looked at more closely defined categories of the consumption of spirits by the men, the increased risk was again primarily confined to those drinking 30 g alcohol/d.
In this study, which entailed pooling data from the cohorts of 7 studies with a total sample of nearly 400 000 participants and >3000 lung cancer cases, we found weak evidence of a positive association between alcohol consumption and lung cancer risk. By pooling data from these large cohort studies, we were able to examine risk in uniformly defined exposure groups and with a consistent group of covariates. The observed risk appeared to be confined to those subjects with consumption of 30 g alcohol/d, which corresponds to 2 drinks/d. However, we could not exclude the possibility, at least in the smokers, that this association could be explained by uncontrolled confounding by smoking. The finding of an increased RR in male never smokers is notable, although the absolute risk of lung cancer in this group is, of course, small. Smoking remains the most important cause of this disease, which has considerable effect on public health.
ACKNOWLEDGMENTS
Pooled data came from existing cohorts. Investigators for each study had contributed to the design of and data collection and analysis of their particular study.
When these analyses were initiated, DJH was the principal investigator for the Pooling Project, leading the overall effort to conduct data analysis on existing data sets. SAS-W (the current principal investigator for the Pooling Project) worked with JLF, DJH, and JR on management of the data and in all phases of data analysis and manuscript writing. The data analysis plan was conceived by JLF. Data analyses were conducted by JR. All other authors contributed to the data analysis plan. The manuscript was written by JLF with input from all of the other authors. None of the authors had a personal or financial conflict of interest.
REFERENCES
Related articles in AJCN: