Tobacco Documents Online

R' Interactions between Smoking and Other Exposures: Occupation and Diet STEVEN D. STELLMAN American Cancer Society New York, New York 10001 INTRODUCTION There are two important reasons for investigating the relationship between tobacco smoking and other factors in the causation of illness. In the first place, smoking is causally related to a very large number of diseases, including those which cause the majority of deaths in our society. Many of these diseases, particularly cancers of the lung and other sites, also have environmental causes in addition to stnuking. Many of these environmental factors increase the qpantitatfive risk of diseases by amounts sintilar to or oftenJess than the risks associated1with smoking; yet many people exposed to these factors also smoke„so that it is often a nrajur methodb- logical problem in epidemiological studies to disentangle the effects due to smoking from those due to other exposures. The second reason for pursuing smoking-environment interactions is that samc substances, notably asbestos, increase the risk of smoking-related disease far above the amount' expected if smoking and asbestos excrtcd their effects indcpcndL•ntly. This effect„often called synergism, has profound implications for predicting future numbers of': cases of diseases (Selikoff 1981), as well' as for developing strategies for prevention. Most multiple-factor studies have centered~ around cigarette smoking and occu- pational exposures. The literature on this subject is now sufficiently abundant that the 1986 Surgeon-General's report on smoking and health is devoted exclusively to occupation. Among the topics treated at length in that report are general work- piace interactions, chronic lung disease, and cancer. Among the exposures con- sidrjred are petrochemicals, aromatic amines, pesticides„asbestos, radon daughters, and cotton dust. Rather than attempt to cover these topics which have already been reviewed in great depth in that report, in detail, this paper will be confined to presenting a superficial summary of the interaction problem, with~ some interesting illustrations from studies of both occupation and nutrition in relation to smoking and cancer. I will mention some of the problems encountered in trying to analyze these situa- tions epidemiologically and present some new American Cancer Society data which may help us as we proceed to investigate the interrelationships between smoking and other exposures. 377 Btrnbury Report 23: Mechanisms in Tobacco Carcinogenesis: Q Cold Spring HarborLahnratorv. 0-87969-223-5-8/86 $1.00 + .00

378 / S. D. StNlman The 1979 Surgeon-General's report lists six ways in which cigarette smoking can interact with the occupational environment to increase risk of illness or injury'(U.S. Dept. Health, Education, and Welfare 1979). (1) A working environment may facilitate body absorption of the toxic com- ponents of cigarcttc smoke. (2) Cigarette smoking can transform workplace chemicals into more toxic sub- stances. (3) A worker can be doubly exposed to the toxic constituents of tobacco smoke and to the same constituents in the workplace. (4) The health effects from environmental exposure can be concurrent with similar health effects from smoking. (5) The synergistic effects of all agents can posc a grave health problem to:workers. (6) Accidents canbc caused by smoking initho industrial cnvironment. During the past few years, an elaborate-and' sometimes controversial-mathe- maticai formalism has been developed for d'escribing and quantifying such inter• actions, particularly as they apply to measuring the contributionrto total risk of disease due to individual exposures (Walker and Rothman 1972; Rothman 1976, 1981; Saracci 1977, 1980; Rothman et al. 1978; Walter and Holford 1981). Formal definitions have been proposed for causal types of concepts, such as interaction and synergy. The main difficulty in this area has not been lack of good statistical ideas so much as lack of good data. As will be shown below, the basic environmental dosage measurements, which are hard enough to obtain reliably for a single ex- posure, become very tenuous when applied two at a time. Nevertheless, there are now available a number of useful examples to illustrate the wide range of inter- actions between these various exposures and smoking. To simplify the discussion and focus attention on ~ the factors themselves, I' willl present data from a number of multfiplc•factor studies in terms of two simple models: additive and multiplicative. The numericali aspects of these models are presented in Table I, in terms applicable to studies in whichieither relative risks or Table 1 Comparison of Additive and Multiplicative Models for Two Simultaneous Exposures Exposure Relative risk Rate Neither 1.0 lo Smoking s sio Additional factor a alo Both Additive modet' (s + a - [i) (s + a - 11) [o Multiplicative model sa sata

Smoking Ihteractions with Occupation and Dia / 379 absolute rates are available. If s is the relative risk conferred by smoking in the absence of the second exposure, and a is the relative risk due to the second' ex- posure in the absence of sluuking; tllen aciording to the :IdditJvl) IIIUdCI the felallve' risk among persons exposed to both shoul& be s+ a - I, whereas according to the multiplicative model it should be sa. Extension, to ntultiple levels of exposure is straightforward. The key to developing both models is that risks due to any combination of exposures must always be measured relative to a common reference point. SMOKING AND OCCUPATION Figure l shows the relative risk for lung cancer in relation to both smoking and shipyard work according to the data of Blot'and'Fraumeni(1981) combine&across four studies. Smoking apart, it is assumed that the excess lung cancer risk among shipyard workers is due mainly to exposure to asbestos, although other contribu- tory factors are certainly possible. Except' for former smokers, the relative risk (RR) among shipyard workers is higher at each level of smoking, compare&to the 24.0 22.0 20.0 1&0 i 1 16.0 14.0 ~ 12.0 10.0 8.0 6.0 4.0 ~ 2.0-i 0.0 LEGENO 9MYUIe WORK M 1:1 NO Y., MaoeliVn MoASI A.M.• ACdiHrr. M.M.:• MWfip/ieulivr 10.7 r1lK 2.2 1.0 6.0 8 6:2 .-.:Y.. 4.9 4 .-A:fa 3.7 F~~.F I ffl ,,.a rY.M 10_2 22.7 ~YY tf.x NON- Ex- 0.50.3-1.3 2+ SMOKERSMOKER PACK PAqCSPACKS CIGARETTE SMOKING STATUS Figure 1 Relative risk for lung eancen according to number of cigarettes smoked per day and'whether or nor subject worked in a shipyard. Data from four case-control studies combined. Data from, Blot and Fraumeni (1981):

380 1'S. D. Stellman RR among nonshipyard workers. At the level ofl 0.5 to 1.5 packs per day, the actual RR among shipyard workers is higher than the additive model predicts, but less than the multiplicative modcl, whereas among the heaviest smokers, it is nearly the same as the mtdtiplicative ntodel:. Tahlo2shows tthe lu.ngcancer death ratess determined by Hammond a al. (1979) among asbestos insulation workers ac- cording to whether or not they smoked and according to the source of information on cause of death (death eertificate versus bestevidence). In either case„the actual rate among those exposed to both asbestos and cigarette smoking is well above the additive model, and, in the case of death certificate ascertainment, is nearly multi- plicative (601.6 observed versus 633.6 predicted). Stellman and Garfinkel (1984) recently reported on the mortality experience of 10,322 men employed in woodworking industries and followed up for 12 years in an American Cancer Society study. Figure 2 shows the standerdized mortality ratio (SMR) for lung cancer„according to smoking habit and usual employment as nonwoodworker, woodworker, or in the carpenter-joiner subgroup. The mortality rates for woodworkers in general and for the carpenters among them were higher than in the nonwoodworkers only among smokers of 20 or more cigarettes per day. Figure 3 shows the analogous SMR pattern for bladder cancer„with similar findings. In the case of lung caneers the risks of smoking and woodworking seem to be additive, but with bladder cancer they are more nearly multiplicative. Whether or not exposure to ionizing radiation in the form of radon daughters, particularly through underground mining, increases lung cancer risk multiplicatively has yet to be resolved. Relative risks for lung cancer among smoking Swedish under- ground miners of iron ore, computed by Damber and Larsson (1985), shown in Table 3„agree well with those predicted with a multiplicative model. On the other i Table 2 i Comparison of Observed Cancer Death Rates with Predictions of Additive and Multiplicative Models Exposure Lung cancer dlath rate Best evidence Death~ certificate Neither 11.3 11.3 Smoking 122.6 122.6 Asbestos 80.2 58.4 Both Actual 693.8 601.6 Predicted Additive model 191.5 1!69.7 Multiplicative model 870.1 633.6 Basu:d on data from llammond et al. (1979)

Smoking intxactions with Occupation and lDi.t / 381 500 250 C NON-WOOOWORKER. E» mOOOMORKER t= CKMCBTRY AMO AOINERY F/ V. P" NEvER SY0KE0 REGUUARIn •A C!•8 ::i_.__ 11_i1 O 1-i9 20. 21-39 40.CURRENT C"RETTE SYOK[RS,AMT vER DAY Er- wvf OR R SNOKERS CiGA SNOKERS Figure 2 Standardized mortality ratios (SMR) t'ur lung cancer by occupation lwcxxiwurker or nut) and srnoking habits. All categories arc relative to nunwocKlworkcrs whn were current smokers uf 20 ciylrettos per day (= 100). Reprinted, with pcrmiticion„frnmiStelhnanand Garfinkel (d984)1 hand„several other studies (Edling 1982; Radford and St Clair Renard 1984) show risks that are additive. Hirayama (1981) has reported age•stand'ardized death rates from all cancer, lung caneen, and stomach cancer (Table 4)!among material metal workers, in the context of a very large prospective study of the general population in Japan. According to his data, the actual rates for all cancers and fon stomach~cancer are considerably above those predicted by either modbl, whereas the lung cancer rate is consistent witit an additive model. In a case-control study in an industrialized area of Mbrthern Italy,,Pastorino et al. (1984) computed relative risks for lung cancer in relation to smoking and em- ployment in occupations in which exposures to known carcinogens are likely. Such exposures included asbestos, polycyclic aromatic hydrocarbons, chromium, nickel, and arsenic compounds, bis-chloromethyl ether, chloromethyl methyl'ether, and vinyl chloride. Subjects were classified as not exposed (-), or as definitely or potentially exposed (+,?). Figure 4 shows the relative risks. For the heaviest smokers, the RR was 20, which is higher than that predicted by an additive model (112.5) but Ics.c'thana multiplicative nuxlcl ('_7.5):

382 / S. D. Stellman Exposure Relative risk for lung cancen Neither 1.0 Smoking 7.4: Underground mining 5.5 Both Actual 40.6 Predicted Additive model 11.9 Multiplicative model 40.7 Table 3 Comparison of Observed Cancer Death Rates with Predictions of Adtlitive and Multiplicative Models Based on data from Damber and'larsson (1:985) NEYER SNOKED REOUIA/lT 1-1e 20 21-3f 10• CURRENT CIWRETTE SYOKER3,4YT.7ER.DA7 Ex- M0E OR SYOKERS CIDA R. SYOKERS Figure 3 Standardized mortality ratios (SMR) for bladder cancer by occupation (woodworker on not) and smoking habits. All categories are relative to nonwoodworkers who were currenn smokers of 20 cigarettes per day (- 100). Reprinted, with permission, from Steilman and Garfinkel (1984)1

Smoking Interactions with Occupation and Di.t /, 383 Table 4 Comparison of Observed Cancer Death Rates with Predictions of Additive and Multiplicative Models Age-standardized death rate per 100,000 from: Exposure All cancers Lung cancer Stomach cancer Neither' 304.3 20.7 136.5 Daily smoking 495.9 85.5 200.7. Material metal' workers 305.1 62.5 180.8 Both Actual 851.7 142:1 400.3 Predicted Additive model 496.8' 127.3 244.3 Multiplicative model 497.5 258:2 264.9 aApproximated'by total ttudy population nonsmokers Based on data l'rum llirsyanta (1981) . 0 1-9 10-19 20-29 30 + ugarettas Figure 4 Relative risk for lung eancer, by definite (+) or possible (?) occupational exposure to industrial' carcinogens, according,to number of cigarettes smoked'per day. Reprinted, with permission, from Pastorino et al. (1i984);.

384 / S. D. Sta/tman SMOKING AND DIET In addition to interactions between smoking and occupation„the interactions with alcohol consumption have been studied extensively. The recent surge,of interest in nutritional aspects of' carcinogenesis has also led to a realization that the samee principles may also appiy: In studies of diet and lung' cancer, Ilur example, it iss essential to take smoking habits into account, so as not to mistakenly report an apparent diet-lung cancer effect that might actually be due to confounding byy smoking. Alcohol plays an etiologic role in cancers of the mouth, larynx, oral cavity, and esophagus (Wynder and Stellman 1977, 1979). It has been widely stated that alcohol is not: a carcinogen by itself but that it promotes titie carcinogenic effects of tobacco smoke. However, it has always been difficult to settle this point„because of the cultural nature of heavy drinking: ltis very'rare to find a heavy drinker who is not also a smoker. Consequently,, the error bounds for estimates of' RR in heavy drinkers in the absence of smoking are usually too large to permit drawing firim conclusions about the carcinogenicity of alcohol alone. Further difficulties im estimating the relative risks of alcohol accurately arise from uncertainties in assigning dosages. Some heavy drinkers consume similar quantities of alcohol each day, as with cigarette smoking; but others drink in binges, spaced by short or long periods of time. The epidbrniologist musY conse• quently make drastic assumptions about exposure in order to aggregate sufficient numbers of cases into a small enough number of categories for useful analysis. Figure 5 shows a reasonable way in which such categorization has been done (Mashberg et al. 1981). Here, drinking has been classified according to the number of "whisky equivalents" (we) ~ consumed per day, as a way of normalizing beer, wine, and spirit consumption, on the same scale. An interesting feature of this case- control study of oral squamous carcinoma is the use as a reference gruup of "minimal" smokers and. "minimal" drinkers, rather than nonsmoking nondrinkers. This distinctioniwas necessary because of the high prcvalence ofbuth smoking andi drinking in: the entire study population. Among "minimal" smokers, the RR ruse with the number of we per day. At each higher level of smoking, the RR among heavier drinkers was higher than among "minimal"drinkers, but the relationships are not consistent. Table 5 shows the interaction models at the highest levels of smoking anddruiking; The RR (104.7) is somewhat between, the additive (30?), and multiplicative ('185.6) models. An alternative classification scheme for aloohol' consumption was proposed by Olsen et al. (1985) in a case-control study of cancer of the larynx. Figure 6 shows smoking- and alcohol-specific RRs, at levels of 0-100. 101-?00, 201-300, an& 301+ g of alcohol per week. As in the preceding study, the reference levels of both alcohol and tobacco usage were not restricted to total abstainers. A number of investigators have recently begun to examine the interaction be- tween smoking and specific food item consumption: Table 6 shows the SMR for

Smokinp Interactions with Occupation and Diet /'385 < O 160 I 2 J cc 1 40 I ' +' < ~ .' J M 120 • ' IDO Cl N ip1M ~ W N / t •e I % 0 e: • % 0 so Y 40 `%• • ~ • ' Q 20 0~.. • ~ l{ p / ` 0. _ YYi.a_ WNIMML CIGAR 10-19 20-3f 40+ J W %PE 0: DALYpGARETTE SMOKING NIkBIT Figure 5 Relative risk for ural afuantous earcinuma. aceording to number uf cigyrottcs vnuked per day and number of "whisky equivalents" consumed per day. Reference group cunsists of "minonal" drinkers and smokers„radier than nondrinkers and nonsmokers. (we) Whisky equivslents. Data from Mashberg et al. (1981): Table 5 Comparison ofl Relative Risks with Predictions of Additive and Multiplicative , Models Exposure Relative risk for oral' squamous.carcinoma "Minimal"smoking andidrinking 1'.0 Smoking 40 or moru vigJrettes per day 8.0 Drinking 10 or more whisky equivalents per day 23.2' Both Actual 104.7 Predicted Additive model 30.2 Multiplicative model 185.6 IMs.d on data from Mashberg,et al. (1981)

388 1 S. 0. Stellm6n .,«.., 1.~....,. xoi-~oa ~ ~oi-zoc o-10 0 0-10 I I+20 21+ TOBACCO (g/doy) Figure 6 Age- and sex-adjusted~rclative risk for larynx cancer according to numberof cigarettes smoked porday and number uf g uf alcohol consumed per week. Data from Ul.cun cti•rll (1985). Table 6 Comparison of Observedi Cancer Death Rates with Predictions of Additive and Multiplicative Models Age-standardized deathirate per 100,000 from lung cancer E xposure Males F ern ales Neither 14.5 1!1.7 Daily smoking 66.3 1i9:0 Avoidance of green-yellow vegetables 32.4 1i6:7 Both Actual 79.5 33.9 Predicted Additive model 84.1 24.0 Multiplicative model 147.8 27.1 Based on data from Hirayama (1979)

Smoking tntersetions with Occupation and Di.t / 387 lung cancer in both nten and women in Illrayallia's large-scale prospective study (Hirayama 1979) in relation to smoking an&avoidance of green-yellow vegetables, which was considered to be the risk factor of interest. The observed RR for men was close to the additive, whereas for women the observed RR exceeded both the additive and~ multiplicative models. The relationship between consumption of fruit and fruit juice was examined by Wang and Hammond (1985) using data from the American Cancer Society's 12-year follow-up study. Figure 7 shows the relative risk for lung cancer according to the number of days per week that fruit or fruit juice was consumed, with additional information on the regular use of vitamin supplements. The left half of the figure refers to all subjects while the analysis in the right half is restricted to current smokers of one pack or more per day: Lack of consumption of fruiti or fruit juice was seen to be a powerful risk factor for lung cancer in both groups. To examine the interaction between fruit consumption and smoking, however„it was necessary to recompute the RRs in relation to a common reference group. Figure 8 shows the result of this recalculation. Meavy smokers who consumed little or no I'ruit or juice 2•oi ALL SUBJECTS CURRENT SMOKERS OF 1.91 2 132 1.27 ONE PACK OR MORE DAILY 4r ::01~ I I' I II I I.o,~ I ~ ~, Io.l" r-__ID0VRUMNPLI.S: YN Y N Y N. Y' N YN Y N 0-2 3-4 5-7 0-2 3-4 5-7 FRUIT AND JUICE CONSUMPTION, DAYS/WEEK Figure 7 Relative risk for lung cancer among nearly 500;000 men survcycd~by American Cancer Society and'followed up 1960-1966„accurding to number of tunm pcr wcek subjcctscunsumed fruits or fruit juices, and whether or not subjects took regular vitamin supptements. Data from Wang, and Hammond (1985).

388 J S. 0. Stalmsm 0.0 FRUITS OR FRUIT JUICES DAYS PER WEEK ..~~. o-a ~. s. r H9 ~ 21-39 CIGARETTES CURRENTLY SMOKED PER DAY NON- SMOKER9 40+ Fiyure 8 Standardized mortality ratio for lung cancer among the same subjects as in Figure 7, according to number of cigarettes smoked per day (current smokers onty),and the number of times per week subjects oonsumed fruits or fruit juices. The reference group is men who smoked 20 cigarettes per day and who consumed fruits or juice 3-4 times per week. had RRs higher than those of frequent consumers. The joint risk was slightly greater than either additive or multiplicative, which in fact were equal~to each other because the relative risk in nonsmokers was indcpendent of the second factor (fruit and juice conswnption). DISCUSSION Table 7 compares the predictions front additive and multiplicative models in a number of the exaluplos eitcd'above, In four oll the studies cited, including two ufl asbestos-exposed workers, the multiplicative model seemed to predict the effect of jpin[ exposure. ln one study (avoidance of green-yellow vegctablcs; females oniy), the additive model appeared bests In two studies, the observed RR lay in bethween the two models, and in two others the prediction far exceeded evemthe muitiplica- tive model! It is important not to take su& findings too literally. This type of analysis is fairly new and is fraught with many methodological pitfalls. The most serious

Table 7 Additive Versus Multiplicative Models in Studies of Smoking, Other Exposures, and Cancer Exposures Outcome Observed Predicted References Smoking Other _ Ever smoked Asbestos insulator Lung cancer mortality Smoked 2+ packs/day Shipyard worker Lung cancer RR Ever smoked Underground miner Lung cancer RR Daily smoker Matcrial metal worker Cancer mortality Smoked 30+ cigarettes Occupational exposure to per day likely carcinogens Smoked 40+ cigarettes Drank 10+ whisky eyuiv per day alents/day Daily smoker Lack of grcen-ycllow vege- tables Smoked 40+ci6arettes 1=ruit or juice 0-2 tir»es/ wcek A.M., additive model; M.!st., multiplicative model Lung cancer mortality Stomach cancer mor- tality Lung cancer RR Oral cancer RR Lung cancer mortality, male Lunc cancer mortality, female Lung cancer RR A.M. M.M. 601.6 169.7 633.6 21.7 11.5 22.7 Hammond et al. t1979) Blot and Fraumeni (1981) 40.6 11.9 40.7 Damber and Larsson (1985) 851.7 496.8 497.5 Hirayama (1981) 142.1 127.3 258.2 Hirayama (1981) 400.3 244.3 264.9 Hirayama (1981) 20.0 12.5 27.5 Pastorino et al. (1984) 104.7 30.2 185.6 Mashberg et al. (1981) 79.5 84.1 147.8 Hirayanu (1979) 33.9 24.0 27.1 Hirayama (1979) 24.0 20.3 20.3 American Cancer Society (unpubl.)

390 /'S. D. Ste4lman problem~ is classification of exposure. Figure 9 shows the distribution of smoking and alcohol consumption in mouth cancer cases and controlfr reported by Wynder and Stellmam(1977). These exposures are highly correlated so that while any large study can be expected to observe statistically usefull numbcrs of nundrinking non- smokers and heavy-drinking heavy smokers the number of heavy-drinking non- smokers will invariably be very small (and subjects falling into that category by self: report possibly unreliable), making it difficult to estimate the parameter a in Table 1. Misclassification of just a few subjects can lead to a large bias in a, as well as in the predictions of the models. In such cases it is prudent to rely on qualitative examination of smoking- and second-factor specific rates or risks, as in the figures,, and not put too much weight on models. On the other hand, in such clear-cut cases as asbestos where there is no doubt as to the synergistic effect (Frank 1979), such information is of tremendous public health importance and is vital for establishing prevention strategies. The most challenging,task of all is to collect the needed data in the fusGplace. Table 8 shows the number of epidemiologic studies of tobacco-related diseases that ALCOHOL CONSUMPTION, OZ. Q! nMne„Occosanor ® 7.10c, '. a 10 0 12% . 2)~. 1t6~. 31 ~ 250 i2Q I t'il 22 cr.aM cn. tm..b cow. cor.wa. c.+.. Cax.rs cmn Camw Caxs MOrt-SM0KER5 I -© i- 20 21- 40 41, CIGARETTES SMOKED PER DAY Figure 9 Correlation of alcohol and tobacco consumptiun amung male mouth cancer caus and con- trols. Reprinted, with permission, from Wynder and Stellman (1i977).

Smoking Interactions w7th Occupation and Diet 1391 . Table 8 Smoking lnformation im32 Original Articles in Volume 26 of the Journal of Occu- pational Medicine (1984) on Epidemiology of Tobacco-related Diseases Subject matter Number of studies Number with smoking data Respiratory effects, including pulmonary function and radiographic testing 12 8 All cause SM R 5 0 Lung cancer 2 2 Coronary artery disease I 1 Reproductive outcomes 3 1 Sleep I I Reviews of studies which included tobacco- related outcomes 2' I Surveillance and'corporate medical programs 6 2 Total 32 16 were published in a recent volume (1984) of the Jouriwl of Occupational Medicine according to whether or not smoking information was also collected. Smoking data was available in only 16 of 32 such articles. These omissions were not necessarily the fault of the authors since many studies were based upon historicalisources in which no smoking data were present. Nevertheless, one obviously cannot study interactions between occupation and smoking data unless one has the smoking data first. Without such data, one cannot tell for certain to what extent smoking has in- fluenced the results. Similar remarks may be made concerning dietary studies. Figure 10 shows the distribution of smoking habits among men who scored high, mediuni, and low on scales which measured the frequency of consumption of foods high im vitamins A and C. Data were collected in 1982 by the American Cancer Society as parG of a follow-up study of over 1.2 million men and1women. Men who scored low on the vitamin scale were twice as likely to smoke cigarettes as men who scored! high. It is clear that any analysis of lung cancer mortality (or any other outcome which is thought to be related to vitamin A or C intake) must simultaneously control for cigarette smoking: REFERENCES Blot, W.J. and~ J.F. Fraumeni. 1!981. Cancer among shipyard workers. In Ban- bury report 9: Quantification of occuparional' cancer (ed. R. Peto and M. Schneiderman), p. 37. Cold~ Spring Harbor Laboratory,,Cold Spring Harbor,, New York.

392 / S. 0. Stollmen. VITAMIN-RICH FOODS L or+ M .dium Hiqh, NEVER SMOKED 1 n 76.8 1 27.6 I 31A REGULARLY CURRENT 19.4 14 7 CIGARETTE 27.5 . 45.8 45 5 FORMER 39.0 . CIGARETTE I PIPE/CIGAR 6 6 7.4 9.0 l . L Figure 10 Disttibution of smoking habits, adjusted for age among men ~surveyed in 1982 American Cancer Society follow-up study, according to whether they scored high, medium, or low on a scale which measures frequcncy of consumption of foods rich in vitamins A. and C:, Damber, L. and L.-G. Larsson. 11985. Underground mining, smoking, and lung can- cer: A case-control study in the iron ore municipalities in Northern Sweden. J. Natl. Cancer Inst. 74: 1207, Edling„ C. 1'982. Lung cancer and smoking ia a group of' iron ore miners. Am. J:, Ind: Med. 3: 191. Frank, A.L. 1979. Public health significance of smoking-asbestos interactions. Ann. N:Y: Acad. Sci. 330: 791. Hammond, E.C., l.J. Selikoff, and H. Seidman. 1979. Asbestos exposure, cigarette smoking and death rates. Ann. N. Y. Acad: Sci. 330: 473, Hirayama, T. 1979. Diet and cancer. Nutr. Cancer 1: 67. -. 1981. Proportion of cancer attributable to occupation obtained from a census, population-based, large cohort study in Japan. In Banbury report 9: Quantification of occupational cancer (ed. R. Peto and M. Schneiderman), p. 63l . Cold Spring Harbor Laboratory,, Coli1 Spring Ilarbor, New York.

Smoking Interactions with Occupation and Diet 1'393 Mashberg, A.,, I.. Garfinkrl, and S. Harris. l''tK1 Aic0linllas 3 priinary ri,k fnclur in oral squamous carcinoma. CA:A Cancer Journal jor Clinicians 31 : 146: Olsen, J., S. Sahreo, and U. Fasting. l'985, Interaction of alcohol and tobacco as risk factors in cancer of the laryngeal region. J. l:'piderninl: ('ontntunitv llealth 39: 165. Pastorino, U., F. Berrino, A. Gervasio, V. Pesenti, F. Riholi, and 1'. Crosigpani. 19841. Proportion of lung cancers due to occupational exposure. Int: J. Can- cer 33: 231. Radford, E.P. and K.G. St Clair Renard. 1984. Lung cancer in Swedish iron miners exposed to low doses of radon daughters. N. lingl. J. Med. 310: 1485.. Rothman, K.J. 1976. Causes. Am. J: Epidemiol. 104: 587. -. 1981. Occam's razor pares the choice among statistical models. A m. J. Epidemiol. 108: 347. Rothmani K.J.,, S. Greenland, and A.M. Walker. 1978. Concepts of interaction. Am: J. EpfdernioC 112: 467. Saracci, R. 1977'. Asbestos and lung cancer: An analysis of the epidemiological evidence on the asbestos-smoking interactiont Int. J: Cancer 20- 323. -. 1980: Interaction and synergism. Am. J. t'pidemiol, 112: 465_ Selikoff', 1J. 1981. Constraints in estimating occupational contributions to cur- rent cancer mortality in the United States. In Banbury report 9: Quantifica- tion of occupational cancer (ed. R. Peto and M. Schneiderman), p. 3. Cold Spring I larhor Laboratory, ('old Spring Ilarhori, New York. Stellman, S.D: and L. Garfinkel. 1984. Cancer mortality among woodworkers. Am. J. Ind. Med:, 5: 343. -. 1986. Smoking habits and' tar levels in a new American Cancer Society prospective study of 1,200,000 men and womcn~ J. Nutl: Cancer Inst. 76: 1057. U.S. Department of Health, F.ducation„and Welfare. 1979. Smoking and healtlt:, A report of'thc Surgeon General. U.S. Department uf Healt'h, Hducatiirn, and Welfare, Public Health Service„Office of the AssistanV Secretary for Health, Office on Smoking and Health, DHI?W Publication No. (I'HS)+79-50066. Walker, A.N. and K.1. Rothman. 1972. Models of varying parametric fnrm in ea,se- referunt studies. Ami J, L'prdcnziul. 115: I'9 Walter, S:D. and T.R. Holford. 1981. Additive, multiplicative, and other models for disease risks: Anr, J: lipidcminl: 108: 341. Wang, L.-D. and G.C: Ilammond. 1985. Lung cancer, truit, green salad, and vita- min pills. Chin. Med: J. 98: 206. Wynder, E.L. and S.D. Stelltnan. 1977. Comparative epidemiology' of tobacco- related cancers. Cancer Res. 37: 4608. -. 1979. ItnpacG of long-term filter cigarette usage on lung and! larynx cancer risk. A case-control study. J. Natl: Cancer Inst. 62: 471. COMMENTS TANNENBAUM: Jeff Harris has published a paper showing different! ways of analyzing the cigarette smoking data. I'm sure you're familiar with it. He

394 /, S. D. Stsllman showed that the cancers which appear at am earlier age as a result of risks, don't show up as dramatically when you use age standardization. When you just say what the relative risk is to cigarette smoking, you do not show that the deaths may also be occurring at an earlier age. STELLMAN: That's right. That's a problem, caused by our practice of age stan- dardization. When we do our initial runs we always analyze mortality by 5- year cohorts. But this generates stacks and stacks of paper, and we must have a way of summarizing the age-specific rates. Unless there are major obvious differences between the rates in different cohorts, we generally feel that combining them is a reasonable way of reducing the data; but obviously it can obscure data as well. WINKELSTEIN: I think,, generally speaking, that epidemiologists feel that en, vironmental associations get weaker as age goes up. This is a sort of general rule. Probably the reason is that people get sick in more ways as they get older; so there are more factors at work. Some specific factors are likely to show a lower association. TANNENBAUM: A paper was just published this year from Walter Willett, show- ing that even the consumption of green and yellow vegetables made a differ- ence, even above age 65. WINKELSTEQV: It may, but if you look at the Framingham data, for example, look at smoking in coronary disease, there's no association shown in people who entered the study after, the age of 45-in the cohort study. TANNENBAUM: You may be right. There may be some factors that may not make so much of a difference. For other factors it may make a difference. Actually, maybe that's part of the problem when you're dealing with two variables each of which may have an associated risk. ll guess one of the problems is that~ your model assumes that dose and time are contem• poraneous. STELLMAN: For cohort studies like the ones that we do, that is usually true. But in case-control studies, that is not often true; andl for mosti industrial cohort studies it is not true either. TANNENBAUM: No, but when you showed your analysis of whether it was an arithmetic or a multiplicative model, there were several different ex- posure models which you eoul& have used. I mean, where you have the two variables: A occurred' before B; A occurring at the same time as B; A oc- curring after B. STELLMAN: This is such an idealistic simplification of a terribly complex situa- tion that it's altnost embarrassing to present it that way.

Smoking Interactions with Occupation and Diet7 395 i TANNENBAUM: No, it's interesting. It's an interesting start. Let's take a specific case: wood workers and smoking. Is the number of years that they worked! as woodworkers overlapped with the number of years that they smoked? STELLMAN: Yes. CORREA: I think that the point that you are making is the sequence. If you have a promoter, it will show an effect only if it is present after an initiator and not if it is present before that factor. STELIMAN: For most of the examples that I presented, these exposures were contemporaneous for many years. So it's not like a two-stage skin painting experiment. In the United States, people in the generation we studied started smoking around the age of 20, and they smoked the resU of their lives. They died around 60 an& 70. So we have 40 years of smoking, and we have at least 30 years of the same occupation.. TANNENBAUM: So it is mostly that model, is what you're saying?' STELLMAN: That is correct. HOFFMANN': Do you have data, or are there too few cases on nasal' cancer in wood workers and their smoking haliits?. STELLMAN: Yes, there are data, but too few cases to analyze. The expected number of nasal cancers was 0:03, and the observed number was 2. So, the relative risk was very large, but the confidence limits were very wide. WINKEiSTEIN: I think that the point that Steve [Stellman] made is a very im- portant one. Epidemiologists musti begin to look at the joint action of factors, rather than spending so much of their, efforts on trying to separate out all of the actions.