Document oD6wqr3pD1152MNLXQvYpqQag

-- URES ve, Screen-OrientedUser Interface :xt-Sensitive Help on W S Micros YSIS MODELS !ic Regression 'roporlinal Hazards Regression ?me-DependentCovariates itional Logistic Regression on Regression tic Regressionwith Random Effects ?entia1and Weibull Regression n-Meier Exact Contingency Table Analysis 'HICS* ,e-lndependentGraphics ie or Off-line Plotting 3r Plots in-Meier and Cox Survival Plots 1 Beta and Fitted Value Plots itate Plots with Lines, Text and Legend bc YlIOLOGY orld have contributed niology, edited by edition was published vn attention to some I f epidemiology aring a revision and I volume 138 Number 1 July 1, 1993 American Journal of EPIDEMIOLOGY Copyright Q 1993 by The Johns Hopkins University School of Hygiene and Public Health Sponsored by the Society for Epidemiologic Research L REVIEWS AND COMMENTARY Smoking and Leukemia: Evaluation of a Causal Hypothesis Michael Siegel Leukemia has not traditionally been recognized as a smoking-related malignancy. Until 4years ago, there were no studies that were designed specifically to examine the relation between smoking and leukemia. Of the 21 studies that have considered this relation, 13 have been published within the past 4 years. The recent interest in this research question was stimulated by the review of the evidence by Austin and Cole in 1986 ( I ) , in which they suggested that there was some support for a causal relation between smoking and leukemia and called for more research. Although 14 additional studies have been published since then, there has been no formal reevaluation of the causal hypothesis. Because of the potential public health significance of a causal relation between smoking and leukemia and the availability of much recent data, it seems appropriate to review the studies at this time. This paper evaluates the hypothesis that smoking is a cause of leukemia. First, the Received for publicationAugust 3, 1992, and in final form March 23, 1993. Abbreviation: CI, confidence interval. From the General Preventive Medicine Residency Program, University of California, Berkeley, and University of California, San Francisco,Berkeley, CA. Reprint requests to Dr. Michael Siegel, Office on Smoking and Health, Centersfor Disease Control and Prevention,MS K50,4770 BufordHighway NE,Atlanta GA 30341. findings of the studies, taken as a whole, are evaluated. Second, the validity of the findings is evaluated in terms of the roles of chance, bias, and confounding. Third, the criteria for causal inference are considered. Finally, a conclusion is reached, and the implications for public health and future research are considered. There are 21 published studies that have examined the relation between smoking and leukemia (2-22). Six of these have been updated by subsequent reports on the same populations and will not be considered separately (17-22). Of the remaining 15 (2-16), nine were case-control studies (2-10) and six were cohort studies (1116). The most important features of the methodology of these studies are summarized in table 1. EVALUATION OF FINDINGS OF EPIDEMIOLOGIC STUDIES The findings of the 15 epidemiologic studies are summarized in table 2. With respect to leukemia in general, 10of the studies found an elevated risk estimate for smoking; in eight of these studies, the elevated risk was statistically significant. Five studies found a decreased relative risk for smokers; in one, the decreased risk was statistically significant. An examination of the confidence intervals in table 2 suggests that 1 2 Siege1 TABLE 1. Summary of methodology of studies of smoking and leukemia Study Subjects Case-control studies Williams and Horrn (2) Paffenbarger et al. (3) Kabat et al. (4) Flodin et al. (5) Cartwright et al. (6) Severson et al. (7) Spitz et al. (8) Brownson et al. (9) Brown et al. (10) Third National Cancer survey tumor registry; controls: incident cancer cases not related to smoking Cohort of male Harvard and University of Pennsylvania alumni; controls: random sample of cohort survivors US hospitalized patients; controls: hospitalized noncancer patients and patients with cancer not related to smoking Swedish hospital tumor registries; controls: random population-based sample Hospital hematology clinics, Yorkshire, England; controls: noncancer hospitalized patients Washington State tumor registries, 3 counties; controls: random population-based sample Texas cancer referral center; controls: patients with cancer not related to smoking Missouri Cancer Registry; controls: patients with cancer not related to smoking lowa/Minnesota tumor registries, white males; controls: random state-rnatchied population sample Cohort studies Weir and Dunn (11) Doll and Pet0 (12) McLaughlin et al., (13) Garfinkel and Boffetta (14) Mills et al. (15) Linet et al. (16) * Number of leukemia cases Male California labor union members British male physicians Male US veterans National sample enrolled by American Cancer Society volunteers California Seventh-day Adventists Lutheran Brotherhood policyholders Size* 172 96 539 111 161 106 241 1,131 578 30 43 1,206 1,378 43 74 based on all of the studies, there is an elevated risk of leukemia in smokers, but the table is not entirely convincing. With respect to myeloid leukemia, however, the results are quite consistent. Of the 12 studies that examined myeloid leukemia specifically, eight found an elevated relative risk for smokers. Six of these risk estimates were statistically significant. An elevated relative risk of about 1.5 is consistent with all but two of the studies. Only three of the studies that found an elevated risk for myeloid leukemia considered acute and chronic myeloid leukeda separately (2, 9, 10).Two of these found an elevated risk for both acute and chronic myeloid leukemia (2, 10).The seven risk estimates that are available do not provide evidence of an elevated risk of lymphatic leukemia for smokers. There appears, then, to be an elevated risk of leukemia in smokers that is most striking for myeloid leukemia but is not seen for lymphatic leukemia. An estimate of the magnitude of this risk can be obtained by pooling the study findings. Each individual study can be viewed as a stratum, and a weighted average of the stratum-specific risk estimates can be calculated. Using the Mantel-Haenszel method for pooling of uniform stratum-specificestimates (23), we determined pooled risk estimates for the casecontrol and cohort studies. Confidence intervals were determined by using the testbased method (24). For the cohort studies, the relative risk for all leukemia was 1.34 (95 percent confidence interval (CI) 1.151.56). For the four cohort studies that evaluated myeloid leukemia specifically, the TABLE 2. Summary Study Williams and Horrn Paffenbarger et al. Kabat et al. (4) Flodin et al. (5) Cartwright et al. (6) Severson et al. (7) Spitz et al. (8) Brownson et al. (9) Brown et al. (10) Pooled odds ratio Chi-square, horn( Degrees of freed p value Weir and Dunn (11) Doll and Pet0 (12) McLaughlin et al. (1 Gatfinkel and Boffer Mills et al. (15) Linet et al. (16) pooled relative risk Chi-square, homor Degrees of freedor p value * Risk estimatesare lor r been calculatedfrom availa no results are reported for t t Inthe risk estimatescot @dies. $95% CI, 95% confidenc pooled relative risk 1.31-1.74). For the pooled odds ratio fc (95 percent CI 1.01 leukemia it was 1.2 1.39). While cautio preting the results of nonunifom method a d statistical evider risk estimates amon ies (table 2), the ana; sonable weighted a mates. Moreover, evidence for the ho mates for all leukem kemia among the cc m e n the study findi ~~ ' registry; related to iersity of m sample xpitalized :ancer not controls: England; Its counties; .mple j: patients 3tients with ite males; population Sire* 172 96 539 111 161 106 241 1,131 578 Cancer 30 43 1,206 1,378 43 74 hen, to be an elevated risk Jkers that is most striking :mia but is not seen for iia. An estimate of the risk can be obtained by findings. Each individual wed as a stratum, and a : of the stratum-specific i be calculated. Using the nethod for pooling of uniific estimates (23), we deisk estimates for the case- ,art studies. Confidence ermined by using the test.-). For the cohort studies, 'or all leukemia was 1.34 dence interval (CI) 1.15r cohort studies that evalmkemia specifically, the Smoking and Leukemia 3 TABLE 2. SUmIIUiV Of findings of studies of smoking and leukemia* study All leukemia Risk estimatet 95% 'I* Myeloid leukemia estRimisaktet 95% GI Case-controlstudies Lymphatic leukemia estRimisakte+ 95% CI 1 Williams and Horn (2) Paffenbarger et al. (3) Kabat et al. (4) Flodin et at. (5) Carlwright et at. (6) Severson et ai. (7) Spitz et ai. (8) Brownson et at. (9) Brown et al. (10) Pooled odds ratio Chi-square, homogeneity Degrees of freedom p value 1.89 1.7 0.87 0.71 0.6 2.1 0.78 1.14 1.4 1.36-2.5 1.1-2.7 0.73-1.04 0.4-1.2 0.4-0.96 1.2-3.8 0.55-1.12 1.00-1.30 1.1-1.9 1.09 1.01-1.19 45 8 <0.01 2.09 i . i a 3 . 7 1 2.3 1.1-4.6 0.99 0.76-1.28 0.6 2.1 0.75 1.45 1.3 0.4-0.96 1.2-3.8 0.37-1.54 1.18-1.78 0.9-1.9 i2 3 1. o a i.39 31 7 <0.01 1.3 0.71 0.96 0.98 1.5 0.5-3.2 0.4-1.2 0.5-1.7 0.8-1.2 1.0-2.1 Cohort studies Weir and Dunn (11) Doll and Pet0 (12) McLaughlin et al. (13) Garfinkel and Boffetta (14) Mills et at. (15) Linet et at. (16) 1.32 0.40-4.37 0.73 0.5S1.07 1.28 1.13-1.45 1.41 1.26-1.57 2.00 1.01-3.95 1.1 0.6-1.9 1.51 1.50 2.24 0.8 1.15-1.98 1.25-1.79 0.9-5.53 0.3-1.7 1.09 0.9-1.4 1.4 0.5-3.5 Pooled relative risk Chi-square, homogeneity Degrees of freedom P value 1.34 1.15-1.56 9.15 5 0.11 1.51 1.31-1.74 5.1 3 0.17 * Risk estimatesare for neverleversmokin When relative risks and confidence intervalsare not given in the papers, they have been calculatedfrom available data. For lymptatk leukemia. confidence intervalscould not be calculated for several studies, and no results are reported for these studies. t In the risk estimatescolumn, odds ratios were reportedfor all casecontrolstudies, and relativerisks were reported for all cohort studies. *95% CI.95% confidence interval. pooled relative risk was 1.51 (95 percent CI 1.31-1.74).For the case-control studies, the pooled odds ratio for all leukemia was 1.09 (95 percent CI 1.01-1.19), and for myeloid leukemia it was 1.23 (95 percent CI 1.081.39). While caution must be used in interpreting the results of this analysis because of nonuniform methodology in the studies (25) and statistical evidence for nonuniformity of risk estimates among the case-control studies (table2), the analysis does provide a reasonable weighted average of the risk estimates. Moreover, there is statistical evidence for the homogeneity of risk estimates for all leukemias and for myeloid leukemia among the cohort studies (table 2). When the study findings are pooled, there is a fair degree of confidence that the relative risk of myeloid leukemia in smokers in the 12 studies is in the range of about 1.1-1.7. The pooled relative risk for all leukemias is lower (in the range of about 1.0-1.5). Only nine of the 15 studies included females, and only four of these reported sexspecific results (table 3). The confidence intervals in three of these four studies are wide enough so that the differences in risk between males and females could be explained by chance. THE ROLE OF CHANCE One must consider whether the negative studies are inconsistent with the positive studies. In other words, were these studies I 4 Siege1 TABLE 3. Relation between smoking and myeloid leukemia: sex-specific findings* Study Risk estimate Male 9596 C l t Risk estimate Female 95% CI Williams and Horm (2) Kabat et al. (4) Brownson et al. (9) Garfinkel and Boffetta (14) 1.30 0.61-2.79 0.97 0.6-1.4 1.5 1.1-2.0 1.72 1.26-2.34 2.27 0.93-5.53 0.87 0.6-1.3 1.4 1.0-1.9 0.99 0.75-1.30 Risk estimates are for nevevever smoking When relative risks and confidence intervals are not given in the papers, they havebeen calculated from available data t 95O', CI,95% confidence interval able to accept the null hypothesis of no association, rather than merely failing to reject it? The answer to this question can be obtained by calculating the power of each study to detect a given elevation of relative risk. Table 4 shows this evaluation of power. Calculations were based on formulas outlined by Hennekens and Buring (26) for case-control studies and Hulley and Cummings (27) for cohort studies. Only four of the studies had adequate power (greater than 0.80) to detect a relative risk of 1.5 for myeloid leukemia. Of the three studies that found a nonsignificant decreased myeloid leukemia risk for smokers, the power was 0.81 (4), 0.14 (8), and 0.16 (16). Thus, two of these negative studies do not have the power to accept the null hypothesis of no association be- TABLE 4. Power* of studies to detect relative risk of leukemia for ever smokers ~ ~~~ Study Power* to detect relative risk for. All leukemia Myeloid leukemia Williams and Horm (2) Paffenbarger et al. (3) Kabat et al. (4) Flodin et al. (5) Cartwright et ai. (6) Severson et al. (7) Spitz et al. (8) Brownson et al. (9) Brown ewal. (10) Weir and Dunn (11) Doll and Peto (12) McLaughlin et al. (13) Garfinkel and Boffetta (14) Mills et al. (15) Linet et al. (161 0.64 0.39 0.99 0.26 0.45 0.31 0.53 0.99 0.84 0.11 0.29 0.99 0.99 0.14 0.27 0.32 0.17 0.81 0.45 0.14 0.81 0.58 0.86 0.98 0.01 0.16 * Power to detect relative nsk = 1 5 tween smoking and myeloid leukemia. As a result of low study power, the confidence intervals in these studies are quite wide, and they include a risk estimate of 1.5. The study by Kabat et al. (4) had adequate power to detect a relative risk of 1.5 for myeloid leukemia, and the study by Cartwright et al. (6) found a significant decreased risk of myeloid leukemia in smokers. These two studies, then. are the only ones not consistent with a moderately increased risk of myeloid leukemia for smokers. THE ROLE OF BIAS There are a number of biases that could have affected the risk estimates in the individual studies. This discussion will consider whether these factors could explain a consistently elevated risk estimate for myeloid leukemia. The finding of an elevated odds ratio in two studies that used cancer controls (2, 9) and in three that used a population-based control sample (3,7, 10)precludes the possibility that a single bias might explain the elevated odds ratios. A major methodological problem in cohort studies is that of differential loss to follow-up. However, the possibility of differential losses to follow-up in the cohort studies cannot explain the consistent association between smoking and myeloid leukemia in the case-control studies. Moreover. follow-upwas noted to be at least 95 percent complete in three of the four cohort studies that reported on myel Any differential loss percent of subjects a relative risk estimate the one cohort study elevated risk of myell ers had only 77 percl In general, the co across a variety of re studies makes it diffk bias in explaining thc The studies by 1 Cartwright et al. (6) d( tion because they sut ological weakness nc seven case-control stl sampled from a regis mias. In the study by redefined from a pop1 controls used in a prc resent only a fraction mia cases in these hos Cartwright et al., case. hospital hematology L percent were alive to b tion bias due to the us. viving cases may expl are inconsistent with rates in surviving(prek are likely to be lower t cases. This bias wouI( estimate of the true o( When these two stud the analysis, the poolc eloid leukemia in the c 1.48 (95 percent CI 1. estimates are almost i the cohort studies. Th mpresentative cases i studies may explain ti hwen the pooled risk e control and cohort s t u ~ In addition to bias 1 studies, the possibility tion of positive result must be considered (28 of publication bias w; First, the analy? studies that were not de! dings* Female ?tsk timate 2.27 3.87 1.4 0.99 95% CI 0.93-5.53 0.6-1.3 1 .0-1.9 0.75-1.30 s are not given in the papers, they id myeloid leukemia. As study power, the confi1 these studies are quite wlude a risk estimate of Kabat et al. (4) had adeetect a relative risk of 1.5 kemia, and the study by (6) found a significant deiyeloid leukemia in smok.tudies, then, are the only ent with a moderately inf myeloid leukemia for 3IAS .imber of biases that could :risk estimates in the indihis discussion will consider tctors could explain a cond risk estimate for myeloid )f an elevated odds ratio in used cancer controls (2. 9 ) at used a population-based ( 3 ,7, 10) precludes the posingle bias might explain the atios. hodological problem in cothat of differential loss to vever, the possibility of dif. to follow-up in the cohort explain the consistent adson smoking and myeloid leuse-control studies. Moreover, noted to be at least 95 percent ree of the four cohort studies Smoking and Leukemia 5 that reported on myeloid leukemia (1 1-13). Any differential losses in the remaining 5 percent of subjects are unlikely to alter the relative risk estimates significantly. In fact, the one cohort study that failed to find an elevated risk of myeloid leukemia in smokers had only 77 percent follow-up (16). In general, the consistency of findings across a variety of research designs in the studies makes it difficult to invoke a single bias in explaining the findings. The studies by Kabat et al. (4) and Cartwright et al. (6) deserve individual mention because they suffered from a methodological weakness not present in the other seven case-control studies. Cases were not sampled from a registry of incident leukemias. In the study by Kabat et al., cases were redefined from a population of hospitalized controls used in a previous study and represent only a fraction of the incident leukemia cases in these hospitals. In the study by Cartwright et al., cases were identified from hospital hematology clinics, and only 3 1 percent were alive to be interviewed. Selection bias due to the use of nonincident surviving cases may explain why these studies are inconsistent with the others. Smoking rates in surviving (prevalent)leukemia cases are likely to be lower than those in incident cases. This bias would result in an underestimate of the true odds ratio. When these two studiesare excluded from the analysis, the pooled odds ratio for myeloid leukemia in the case-control studies is 1.48 (95 percent CI 1.26-1.72). These risk estimates are almost identical to those for the cohort studies. Thus, selection of nonrepresentative cases and controls in two studies may explain the inconsistency between the pooled risk estimates for the casecontrol and cohort studies. In addition to bias within the individual studies, the possibility of selective publication of positive results (publication bias) must be considered (28-30). The likelihood of publication bias was evaluated in two ways. First, the analysis was restricted to studiesthat were not designed specificallyto examine the relation between smoking and leukemia. These studies reported on the relation between smoking and multiple causes of mortality (2,ll-14) or between leukemia and a wide range of potential rihk factors (2, 3,6). Publication bias is less likely for nonspecific studies because they examine a wide range of potential associations, and it is less plausible that a negative result for any one specific relation would lead to nonpublication. Of these nonspecific studies, four of five found a positive association between smoking and myeloid leukemia, and five of seven found a positive association between smoking and all leukemia. It is unlikely that publication bias explains the preponderance of positive findings in these studies. Second, graphic analysis of the relation between effect size and study size has been suggested as a method to assess the likelihood of publication bias (30-32). In the absence of publication bias, when effect size is plotted against sample size (or standard error) in a funnel graph (33), a pyramid pattern with the apex at the true effect size should be obtained. Publication bias tends to skew the pyramid by creating a gap in the area of negative studies with small sample size (31, 32).A funnel plot of the 15published studies of smoking and leukemia (figure 1) produced a pyramid shape with apex corresponding to a relative risk of about 1.3 and no gap in the lower left-hand comer of the pyramid. While this analysis does not rule out the possibility of publicationbias, it provides further evidence that it is unlikely to explain the preponderance of positive findings in the 15 studies. THE ROLE OF CONFOUNDING There are several known confounding variables that must be considered in evaluating the relation between smoking and leukemia. These are age, sex, race, socioeconomic status, and exposure to known leukemogens.Age and sex are controlled for in all 15studies.The remaining variables are now considered in terms of the direction and 6 Siege1 20 19 - 18 - 17 - - la 15 - -UI 14 - C 13 - c 0 12 - eL 11 - t 10 - 0 0 0 9- i.aj 8 Ce 7 - .tj 6 - \ 50 4- 3- Do 0 D O 0 00 0 2- 0 1- 0 I III I -OB -0.4 - 0 2 0 02 0.4 0.6 0.8 In (RRI FIGURE 1. Funnel plot of relative risk (on a logarithmic scale) according to the inverse of standard error of lognormal relative risk (In RR) calculated for 15 published studies of smoking and leukemia. magnitude of effect they would have on the estimate of the relative risk of leukemia in risk estimates. smokers. Race Exposure to known leukemogens Only five of the positive studies were controlled for race. Leukemia is more common among whites than among nonwhites (34). Blacks are more likely to be smokers than are whites (35).Thus, confounding by race would result in an underestimate of the relative risk for current smokingand leukemia. Socioeconomic status Education and/or income levels were controlled for in five of the positive studies. There is a slight positive association of leukemia with higher socioeconomic status ( 2 ) . There is a small negative association between both education and income and smoking (35).Thus, confounding by socioeconomic status would result in an under- The major known environmental risk factors for leukemia are exposure to benzene, ionizing radiation, and chemotherapeutic agents (34).Only one of the studies controlled for these variables. Severson et al. (7) found no substantialchange in the odds ratio or dose-responserelation when analysis was restricted to subjects with no reported exposure to these agents. There is certainly a potential for confounding by exposure to known leukemogens, and this confounding would result in an overestimate of the relative risk. First, smokers are at increased risk of cancer at several sites and may be more likely than nonsmokers to be exposed to chemotherapeutic agents. Second, smokers are slightly more likely to report occupational exposure to chemicals. Stellman et al. (36) found that 54 reported occupational compared with 46 pert and 52 percent of ever no difference in expos1 ation. The relative risk. ciated with occupation; have generally been in (34).Using these data greatest potential mag confounding on the rela the cohort studies.To bt assumed that all person cupational chemical er to benzene, and a relatit kemia in benzene-ex1 used. If there were no tween smoking and leuE trol for the effects o f t under these conditions in the detection of a I 1.06. Even under the r tions, the magnitude of posure to known leut small. Confounding by know unlikely to explain the c between smoking and Confounding by an however, could explain risk in the range of 1.5. clinical trial could elim of confounding.Nevertr the likelihood that con known factor explains I must be made. There mogeneity of the subj studies. Paffenbarger e male Harvard and Unik nia alumni. Weir and Di male labor union meml cupations. McLaughlir only male US veterans 3 insurance in 1953.Mil Only California Sevet The homogeneity of within many of the stud] likelihood that confount factor explains the obsc 0 0 0 0 I I1I 0.4 0.6 0.6 the inverse of standard error of md leukemia. .elative risk of leukemia in own leukemogens )wnenvironmental risk faca are exposure to benzene, on, and chemotherapeutic ly one of the studies convlariables. Severson et al. (7) ntial change in the odds ratio 2 relation when analysis was 3jects with no reported exagents. There is certainly a mfounding by exposure to )gens, and this confounding _.an overestimate of the relsmokers are at increased it several sites and may be .I nonsmokers to be exposed eutic agents. Second, smokmore likely to report occuire to chemicals. Stellman et Smoking and Leukemia 7 a]. (36) found that 54 percent of smokers reported occupational chemical exposure compared with 46 percent of nonsmokers and 52 percent of ever smokers. There was no difference in exposure to ionizing radi- ation. The relative risks for leukemia associated with occupational benzene exposure have generally been in the range of 1.5-3.0 (34).Using these data, we estimated the coreatest potential magnitude of effect of confoundingon the relative risk estimatesin the cohort studies.To be conservative,it was assumed that all persons who report any occupational chemical exposure are exposed to benzene, and a relative risk of 3.0 for leukemia in benzene-exposed workers was used. If there were no true association between smoking and leukemia, failure to control for the effects of exposure to benzene under these conditions would have resulted in the detection of a relative risk of only 1.06. Even under the most extreme conditions, the magnitude of confounding by exposure to known leukemogens is quite small. Confoundingby known variables, then, is unlikely to explain the observed association between smoking and myeloid leukemia. Confounding by an unknown factor, however, could explain a consistent relative risk in the range of 1.5. Only a randomized clinical trial could eliminate the possibility of confounding. Nevertheless, ajudgment of the likelihood that confounding by an un- known factor explains the observed relation must be made. There is a remarkable homogeneity of the subjects in many of the studies. Paffenbarger et al. (3) studied only male Harvard and University of Pennsylvania alumni.Weir and Dunn (1 1) studied only male labor union members in high-risk occupations. McLaughlin et a]. (13) studied only male US veterans with government life insurance in 1953. Mills et al. (15) studied only California Seventh-Day Adventists. The homogeneity of cases and controls within many of the studiesargues for a small likelihood that confounding by an unknown factor explains the observed association. CRITERIA FOR CAUSAL INFERENCE Consistency of association There is a good deal of consittency in the epidemiologic studies. In fact, for myeloid leukemia, all but two of the studies are consistent with a relative risk of about 1.5 for ever smokers. Strength of association This criterion is not met by the studies. The relative risk for myeloid leukemia in ever smokers appears to be about 1.5. This is low enough that confoundingby some unrecognized variable could potentially explain the observed association.This does not rule out a causal relation, however, as the true relative risk could be quite small. Dose-response relation Six of the nine studies that looked for a dose-response relation between intensity of smoking and leukemia risk found one (7,9, 13-16). This relation was statistically significant in five of these (7, 9, 13, 15, 16). Decreased risk after removal from exposure The one study that examined this question (7) found a linear trend of decreasing risk with increasing number of years since smoking was stopped. In addition, the elevated risk was present only for smokers who inhaled into the chest. Biologic plausibility There are several lines of evidence for the plausibility of a causal relation between smoking and leukemia. Animal studies. Several studies have demonstrated the induction of hematopoietic tumors in animals by tobacco smoke. Keast et a]. (37) found an elevated incidence of lymphocytic lymphomas in smokeexposed mice. DiPaolo and Levin (38) found a greatly increased incidence of lymphomas in smoke condensate-treated mice. iI i j ! 8 Siegel Cytogenetic studies in humans. Numerous studies have documented an increased frequencyof cytologic changes in peripheral blood lymphocytes of smokers (38). Specifically, an increased frequency of chromosomal aberrations, sister chromatid exchanges, and micronuclei have been found. Most recently, Bridges et al. (39) found an increased frequency of thioguanineresistant mutants in circulating T lymphocytes of smokers. Leukemogens in cigaretie smoke. At least three chemical leukemogens (benzene, urethane, and nitrosamines) are present in tobacco smoke (1). In addition, the concentration of lead-210, a radioactive leukemogen, has been demonstrated to be significantly increased in the bones (40) and soft tissues (41) of smokers. CONCLUSION Based on the existing epidemiologic data; the analysis of the roles of chance, bias, and confounding; and the criteria for causal inference, it appears that smoking causes myeloid leukemia. A causal relation between smoking and leukemia has important public health implications. With a relative risk of 1.5 for myeloid leukemia in ever smokers and the Centers for Disease Control and Prevention estimate of a 55.8percent prevalence of ever smoking ( 3 3 , the population attributable risk is 22 percent. Since the causes of leukemia are largely unknown, this would make smoking the leading known cause of leukemia. Leukemia should be added to the list of smoking-related diseases, and efforts to prevent leukemia should include appropriate attention to the role of smoking. Since there are insufficient data to rule out the qssibility that smoking causes lymphatic leukemia and to separate out the effects of smoking on acute versus chronic myeloid leukemia, future studies should be designed specifically to determine the effects of smoking on cell-specific leukemia risk. ACKNOWLEDGMENTS The author thanks Dr. Warren Winkelstein, Dr. Allan Smith, Dr.Tom Novotny, Dr. Arthur Reingold, Eve Siegel, and David Sklarew for their help in preparing this paper. REFERENCES I . Austin H, Cole P. Cigarette smoking and leukemia. J Chronic Dis 1986;39:417-2 1. 2. Williams RR, Horn JW. Association of cancer sites with tobacco and alcohol consumption and socioeconomic status of patients: interview study from the Third National Cancer Survey. J Natl Cancer Inst 1977;58:52547. 3. Paffenbarger RS, Wing AL, Hyde RT. Characteristics in youth predictive of adult-onset malignant lymphomas, melanomas, and leukemias: brief communication. J Natl Cancer Inst 1978$0: 89-92. 4. Kabat GC, Augustine A, Hebert JR. Smoking and adult leukemia: a case-control study. J Clin Epidemiol 1988;41:907-14. 5 . FIodin U, Fredriksson M, Persson B, et al. Chronic lymphocytic leukemia and engine exhausts, fresh wood, and DDT a case-referent study. Br J Ind Med 1988;45:33-8. 6. Cartwright RA, Darwin C, McKinney PA, et al. Acute myeloid leukemia in adults: a case-control study in Yorkshire. Leukemia 1988;2:687-90. 7. Severson RK, Davis S, Heuser L, et al. Cigarette smoking and acute nonlymphocytic leukemia. Am J Epidemiol 1990;132:418-22. 8. Spitz MR, Fueger JJ, Newell GR, et al. Leukemia and cigarette smoking. Cancer Causes Control 1990;1:195-6. 9. Brownson RC, Chang JC, Davis JR. Cigarette smoking and risk of adult leukemia. Am J Epidemiol 1991;134:938-41. 10. Brown LM, Gibson R, Blair A, et al. Smoking and risk of leukemia. Am J Epidemiol 1992; 135:763-8. 11. Weir JM, Dunn JE. Smoking and mortality: a prospective study. Cancer 1970;25:105-11. 12. Doll R, Pet0 R. Mortality in relation to smoking: 20 years' observations on male British doctors. Br Med J 1976;2:1525-36. 13. McLaughlin JK, Hrubec Z, Linet MS, et al. Cigarette smoking and leukemia. J Natl Cancer Inst 1989;81:1262-3. 14. Garfinkel L, Boffetta P. Association between smoking and leukemia in two American Cancer Society prospective studies. Cancer 1990; 6k2356-60. 15. Mills PK, Newell GR, Beeson WL, et al. History of cigarette smoking and risk of leukemia and myeloma: results from the Adventist health study. J Natl Cancer Inst 1990;82:1832-6. 16. Linet MS. McLaughlin JK, Hsing AW, et al. Cigarette smoking and leukemia-results from the Lutheran brotherhood cohort study. Cancer Causes Control 1991;2:413-17. 17. Brownson RC. Cigare leukemia. J Clin Epide 18. Severson RK. Cigaretrt Cancer 1987:60: I4 1 3 . 19. Harnmond EC. Smokir! rates of one million me cer lnst Monogr 1966: I 20. Kahn HA. The Dorn Sfi tality among U.S. vete one-half years of obsei Monogr 1966:19:1- 125 21. Kinlen LJ. Ropot E. Le[ its among United 1988;297:657-9. 22. Rogot E. Murray JL. death among U.S. vew tion. Public Health Re1 23. Mantel N. Haenszel H' analysis of data from disease. J Natl Cancer 24. Miettinen 0.Estimabili referent studies. An 103:226-35. 25. Greenland S. Quantitati of epidemiologic liic 1987;9: 1-30. 26. Hennekens CH. Burin medicine. Boston. MA 1987. 27. Hulley SB, Cummings cal research. Baltimc Wilkins, 1988. 28. Dickersin. KD. The exi. and risk factors for 1990;263: 1385-9. 29. Rennie D. Flanagin A. ! umph of hope ovt` 1992;267:411-12. 30. Begg CB, Berlin JA. E 'AENTS Lnks Dr. Warren Winkelstein, Dr. Tom Novotny, Dr. Arthur iegel. and David Sklarew for ming this paper. :P. Cigarette smoking and leukeI Dis I986;39:4 17-21. Horm JW. Association of cancer ICCO and alcohol consumption and :status of patients: interview study 3 National Cancer Survey. J Natl )77;58:525-47. <S, Wing AL, Hyde RT. Characterh oredictive of adult-onset m a k nis, melanomas, and leukemias: mication. J Natl Cancer Inst 12. .ugustine A, Hebert JR. Smoking nemia: a case-control study. J Clin 88;41:9O7-14. redriksson M, Persson B, et a]. phocytic leukemia and engine exwood, and DDT: a case-referent id Med 1988;45:33-8. 'A. Darwin C, McKinney PA, et al. 4 leukemia in adults: a case-control shire. Leukemia 1988;2:687-90. Davis S, Heuser L, et al. Cigarette acute nonlymphocytic leukemia. iiol 1990:132:418-22. ieger JJ, Newell GR. et al. Leukeette smoking. Cancer Causes Con95-6. Z. Chang JC, Davis JR. Cigarette risk of adult leukemia. Am J Epi134:938-41. Sibson R, Blair A, et al. Smoking leukemia. Am J Epidemiol 1992; inn JE. Smoking and mortality: a mdy. Cancer 1970;25: 105-11. R. Mortality in relation to smoking: servations on male British doctors. '76;2: 1525-36. JK, Hrubec Z, Linet MS, et al. Cigig and leukemia. J Natl Cancer Inst 1-3. Boffetta P. Association between leukemia in two American Cancer spective studies. Cancer 1990; *wellGR, Beeson WL, et ai. History smoking and risk of leukemia and esults from the Adventist health Cancer Inst 1990;82:1832-6. lchughlin JK, Hsing AW, et al. Cigng and leukemia-results from the rotherhood cohort study. Cancer :rol 1991:2:413-17. Smoking and Leukemia 9 17. Brownson RC. Cigarette smoking and risk of leukemia. J Clin Epidemiof 1989:-42:10-35-6. IX. Severson RK. Cigarette smoking and leukemia. Cancer 1987:60:1414. iy. Hammond EC. Smoking in relation to the death rates of one million men and women. Natl Can- cer Inst Monogr 1966:19:127-204. 20. Kahn HA. The Dorn study of smoking and mortality among U.S. veterans: report on eight and one-half years of observation. Natl Cancer Inst Monogr 1966:19:I-125. 2 1 . Kinlen LJ. Rogot E. Leukemia and smoking habits among United States veterans. BMJ 1988:197:657-9. 3 7 Rogot E, Murray JL. Smoking and causes of death among U S .veterans: 16 years of observation. Public Health Rep 1980:95:2 13-22, 33. Mantel N, Haenszel W. Statistical aspects of the analysis of data from retrospective studies of disease. J Natl Cancer Inst 1959:22:71948. 1-1. Miettinen 0.Estimabilitv and estimation in case- referent studies. Am' J Epidemiol 1976; 1031226-35. 25. Greenland S. Quantitative methods in the review of epidemiologic literature. Epidemiol Rev 1987:9: 1-30. 26. Hennekens CH, Buring JE. Epidemiology in medicine. Boston. MA: Little. Brown, & Co., 1987. 27. Hulley SB, Cummings SR. ed. Designing clini- cal research. Baltimore, MD: Williams & Wilkins, 1988. 28. Dickersin. KD. The existence of publication bias and risk factors for its occurrence. JAMA 1990:263: 1385-9. 39. Rennie D, Flanagin A. Publication bias. The tri- umph of hope over experience. JAMA 1992;267:411-12. 30. Begg CB. Berlin JA. Publication bias and dis- , semination of clinical research. J Natl Cancer Inst 1989:81:107-15. 3 I . Begg CB. Berlin JA. Publication bias: a problem in interpreting medical data. J R Stat Soc '4 1988:151:419-63. 32. Vandenbroucke JP. Passive smo)ing and lung cancer: a publication bias? Br Med J 1988: 296:391-2. 33. Light RJ, Pillemer DB. Summing up: the science of reviewing research. Cambridge. MA: Harvard University Press, 1984. 34. Sandler DP. Collman GW. Cytogenetic and environmental factors in the etiology of the acute leukemias in adults. Am J Epidemiol 1987;1-36:1017-32. 35. Centers for Disease Control and Prevention. Smoking. tobacco. and health: a fact book. Atlanta, GA: Centers for Disease Control and Prevention. 1987. 36. Stellman SD, Boffetta P,Garfinkel L. Smoking habits of 800,000 American men and women in relation to their occupations. Am J Ind Med 1988:1343-58. 37. Keast D. Ayre DJ. Papadimitriou JM. A survey of pathological changes associated with long-term high tar tobacco smoke exposure in a murine model. J Pathol 1981;135:249-57. 38. DiPaolo JA. Levin ML. Tumor incidence in mice after oral painting with cigarette smoke condensate. J Natl Cancer Inst 1965:34:595-600. 39. Bridges BA, Cole J, Green MHL. Smoking. leukemia and lymphocyte mutations. (Letter). Leukemia 1989;3:836. 40. Holtzman RB, Ilcewica FH. Lead-? IO and polonium-210 in tissues of cigarette smokers. Science 1966:153:1259-60. -41. Bfanchard RL. Concentrations of 210Pb and 21OPo in human soft tissues. Health Ph\s 1967;13:625-32.