Observations on the dose-response curve for arsenic exposure and lung cancer.

Observations on the dose-response curve for arsenic exposure and Occupationalstudies in threecountries have related quantitative estimates of arsenic exposure to lung cancer risks. Mine exposures in China appear to incur a higher relative risk than arsenic exposures elsewhere. All of the studies with quan titative data are consistent with a supralinear dose-response relationship. Two studies are also con sistent with a linear relationship over an elevated background risk of lung cancer among arsenic-ex posed workers. Neither toxicokinetic mechanisms nor confounding from age, smoking, or other work place carcinogens that differ by exposure level appear likely to explain this curvilinearity. Plausible explanations include (i) synergism (with smoking) which varies in magnitude according to the level of arsenic exposure, (ii) long-term survivorship in higher exposure jobs among the healthier, less sus ceptible individuals, (iii) exposure estimate errors that were more prominent at higher exposure levels as a result of past industrial hygiene sampling or worker protection practices.

A lar ge body of literature consistently link s arsenic inhalation exposure to the risk of lung cancer in humans (1)(2)(3). Data relating quantified estimates of arsenic exposure to lung can cer mortality have been reported for workers representing six occupational cohorts (4)(5)(6)(7)(8)(9)(10). In a brief report (11), we noted that data from three smelter studies (4-6) were con sistent with a curvilinear do se-re sponse in which slopes were steeper at low than at high cumulative exposures, and we di scu ssed the implications of this supralinearity for risk ext rapolations.
In this paper, we evaluate all published reports which relate lung cancer risk to quantified occupational arsenic expo sure. In particular, the goal s of this paper are to synthesize the findings; reconcile discrepancies among studies, where possible; generate " tes table" scientific hypotheses reg arding the magnitude and shape of the dose-response cur ves; assess the strength of evidence for each of these hypotheses; and pro vide guidance for future research on this Issue .
The pre senc e of a monot onic dose-response rel ation between e xpo sure and outcome is one of several criteria used in assess ing the strength of eviden ce for causal infe rence (12 ). In add ition, dos e-respon se data from epidemiologic studies increasingly playa role in standard setting and risk assessm ent for both occupational and en vironmental exposures ( 13). To extrapolate from high to low exposure scenarios, linear models are frequentl y used . That the data for arsenic-exposed workers suggest nonlinear relations underscores the need to understand what factors serve to influence or distort do se-respon se curves in order that the most scientifically sound risk estimates be obtained.
To begin, we (i) provide background information on these studies and pre sent the dose-response data in graphic form; we then (ii) discuss var iation in the magnitude of the risks, (iii ) evaluate the shape of the do se-respon se curves, and (iv) explore po ssible explanation s for the curvature ob served in several of the studies.

Background and dose-response relationships
Studies pro viding do se-respon se data for occupation al arsenic exposure and lung cancer risk have been conducted on large cohor ts of sme lter workers in Tacoma, Washington (4), Ana conda, Montana (5), and Ronnskar, Sweden (6,7); workers at six smaller smelters in the United States, one at Garfield, Utah , and five at other sites (10 ); a cohort of insecticide manufacturing workers in Midland, Michigan (8); and a cohort of mainly miners but also some smelter work ers employed by a tin corporation at numerous sites in China (9) . Table I summarizes these studies, including their des ign, location, industry, number of lung cancer cases, type of analysis, variables controlled for , and choice of referents. Except for the study of miners in China (9), all of the studies ex-Scand J Work Environ Health 199 3, vol 19, no 4 Table1. Six occupational studies with quantified dose-response data on lung cancer and arsen ic exposure. (SMR =standardized mort alit y rati o, PMR =proportionate mo rtality rati o, DSDR = direc tly standardized death rate, OR =odds ratio , RR =risk ratio ) • Cases were defi ned as death s from lung cancer for all stud ies except that of Taylor et al (9), who defined cases as memb ers of the cohort who were diagn osed with lung cancer and were living in 1985. b Standard ized for age and calendar year. c Nested wit hin the cohort of Jarup et al (6). • A proport ionat e mortality analysis was cond uct ed, that is, deceden ts from oth er causes wh o were at the same plant co nstitu ted the referents.
Workers exposed to asbestos were excl uded.
amined mortality. Within the two nonsmelter cohorts, a nested case-refere nt design was applied: one using living lung cancer cases and living refere nts (9) and the other using deaths from lung ca ncer as cases and deaths from other causes as referents (8). Both a cohort (6) and a nested case-referent (7) study analyzed quantified exposure data for employees at the Ronnskar smelter. The study of six sma ll smelters also used internal compariso ns. The number of observed deaths was compared with the number expected on the basis of the experience of all white smelter workers in the cohort (10) . Although other reports have also appeared for some of these cohorts, they included shorter follow-up (14,15), lacked quantified air measurements (15)(16)(17)(18)(19), or used an incorrect method of analysis (20,21). In figure I, lung cancer risk is plotted against arsenic exposures for the three large cohorts of smelter workers which were compared with external referents. Arsenic expo sure, when originally repor ted as cumulative micrograms per cubic meter-years or milligrams per cubic meter-months, was converted into units of cumulative milligrams per cubic meter-years. In these studies, risk ratios for lung cancer were estimated by the standardized mort ality ratio (SMR), for which expected deaths were derived from the application of death rates from a genera l population in the same state or region to the age, calendar year, and gender distribution in the cohort of interest. Note that the axes in figure I differ in the ranges for both the independent and dependent varia bles due to dif-218 ferences among the studies in the range of estimated exposures and in the range of observed SMR values. At the Anaconda smelter, workers were divided into cohorts by date of first employment. The cohort hired prior to 1925 and the one hired in 1925-1947 are shown. Figure 2 presents similar plots, with arsenic exposure in the same units, for two nested case-referent studies (8,9) and for the cohort study which relied on internal comparisons ( 10). In these studies, the risk ratios for lung cancer were estimated with the use of different measures. The report on miners in China provided multivariate odds ratios compar ing the upper three quartiles of exposure to the bottom quartile (9); the report on insecticide manufactur ing workers calc ulated proportionate mortality ratios for lung cancer standardize d to a cohort of unexposed decedents in the same industry (8); the study of six small-to-mo dera te sized smelters reported SMR values for which the expected numbers of deaths were based on the overall cohort, adjusted for age, calendar year, and latency (10). Thus the analysis for this group of smelters involved only internal comparisons among expos ure levels. For this last-mentioned study, we divi ded all of the risk ratios by the risk ratio in the lowest exposure group so that a value of 1.0 on the y-axis corresponded to the risk for those with < 100 ug . m-3-y ears of exposure. Thi s procedure transforms the effect measures to estimated risk ratios that are more comparable to those presented in the other studies.

Variation in the magnitude of risk
The data from the miners' study (9) and the study of insec ticide manufac turing wo rkers (8) appear to show a greater potency of arsenic (ie, a greater increase in risk for a given level of cumulative exposure) than the three large smelter studies . For insta nce, doses in the range of 4-15 cu mulat ive mg . m-3 -years induced a two-to fourfo ld increase in each of the three smelter studies (4---6), while similar exposures were associated with more than a 20fold increase in lung cancer risk amon g the miners in China (9) and a four-to seve nfold increase among insectic ide manufact uring workers (8).
Part of this discre pancy may be due to an underestimation of expos ures, at least in the mines. Table  2 summarizes data reported in the six studies on es-timated arsenic expos ure based on historical measurements. Rang es are shown where department-specific data were used . The arsenic measu rements in the mines in China (9) appear to be substantially lower than those in the three large smelters. For instance, afte r 1951 they are orders of magnitude lower than the high exposures for similar periods in the sme lter studies and appear to be similar to concentrations reported for the six smaller sme lters . These dra matic diffe rences suggest that the exposu re estimates for the mines in China may have been too low. In the insecticide manufacturing plant (8), concentrations were sim ilar to those in the large smelters; the authors in fact asserted that the exposures may have been overes timated.
Statistica l biases were probably not suffic ient to explain the higher risk ratios in nonsme lting indus- The large difference in apparent potency between the Chine se mine exposures and exposures in smelters and mines elsewhere remains enigmatic. Underestimation of exposure , interactions with other carcinogenic exposures (eg, radon, nickel, etc), and a higher proportion of susceptible individuals are the most likely factors to have contributed to the apparent discrepancy . The apparentl y higher poten cy at one insecticide manufacturing plant may also have been a result of an underestimation of exposure or interactions with other carcinogenic substances in the workplace.

Shape of the dose-response curve
The data from several of the studies shown in the figures suggest a possibly nonlinear relation ship between occupational arsenic exposures and the risk of lung cancer. Enter line et al (4) found that the data for the Tacoma smelter were best fit by a power function in which the SMR rises at roughly one-third the power of the cumul ative dose. To examine whether the data from the other two large smelter studies (5,6) were consistent with a curvat ure of this type, we did the following analyses: (i) power models were fit to the three smelter cohorts and compared with similar linear models, and (ii) P-value function s (26) were constructed around the SMR at each exposure level in the Ronnskar and Anaconda smelter cohorts and compared with the predicted response based on the model fit to the Washington smelter.
The power and linear models, respectively, took the forms: tries. For instance, upward bias from the use of odds ratios (where lung cancer may not have been rare and referents were not incidence-density sampled) or attenuation of SMR values in the cohort studies (as a result of the health y worker effect when external comparisons were used) would have had a small net impact on the measures of association. Lower background rates, however, may expla in the higher risk ratio s in the miners' study. A high er relative risk could represent a similar absolute risk from arsenic exposure across studies .
The potency of arsenic exposure in the mines may have appeared inflated by confoundin g from or interaction with radon , nickel, cadmium , and chromium. All of these carcinogens were present and had been measured in the mines (9). Ninety-three percent of the workers had some radon exposure (22). A smaller contribution to the discrepancy could have resulted from differences in particl e size, proportion in the trivalent versus pentavalent forms, or, alternative ly, personal work practices. Smelter exposures were mostly trivalent arsenic (23), but data were not reported on particle size or valence speciation in the study from China . Some of the smelter studies reported the use of respirators [at the Tacoma smelter respirators were introduced in the 1940s and tested as "99 percent effective" (16)], but protective devices were not mentioned for the miner s in China. Similar issues can be raised with respect to the insecticide manufacturing study, in which workers were exposed primarily to lead arsenate and calcium arsenate (8).
Taylor et al (9) could not successfully separate the effects of mining from smelting exposures because too few study subjects were employed solely in the smelter. Other studies conducted in Sweden (24) and Utah (25) have not demonstrated higher risks from arsenic exposures in mining as compared with smelting. These studies do not, however, preclude differences, for example, in silica content , when ores 220 where E[] indicates the expectation of a random variable (in this case from a Poisson distribution), Obs, represents the observed deaths at exposure level i, EXPj represents the expected deaths in the absence of exposure, at exposure level i, b is the slope parameter, and a is the power parameter. These models constrain the intercept. Models were also fit with an intercept parameter. Parameters were estimated by iteratively reweighted least squares regression, predicted deaths serving as weights. The best fitting power curve using all points from the three cohorts combined was a poor fit. The major contribution to the lack of fit was two points in the Anaconda smelter data . Linear models provided an even worse fit. When each data set was analyzed separately, better fitting models were obtained, and, in all cases , the power relations provided as good a fit as or a better fit than linear relations . The improvement ranged from minor to substantial. The comparisons of P-value functions with predictions from the model for the Tacoma smelter gave similar results. With the exception of one point (at 125 mg . m-3 in the Anaconda cohort hired 1925-1947), the curve fell in the central portion of these functions, suggesting general consistency with the fitted curvature of decreasing slope (4). On the other hand, from visual inspection and from the models with unconstrained intercepts, one sees that the data from these two smelter studies (5, 6) are also consistent with a linear relationship in which the intercept is above the background level. Finally, a casereferent study nested within the Swedish smelter cohort (6) showed little evidence for a supralinear curvature (7). In yet another study of smelter workers a decreasing slope in the dose-response relationship was noted (10), though the range of exposures was extremely low relative to the ranges in the other occupational cohorts, and the curvature was slight. In fact, this study suggests that the dose-response relationship is much closer to linear at low than at high cumulative exposures. Thus, with one exception, the sugge stion of a declining slope for the risk of lung cancer at higher cumulative exposures of arsenic is observed for all of the occupational studies with quantified dose-response data.
By plotting the relative risks rather than their logarithm s, we have evaluated dose-response on an additive, rather than on a multiplicative scale. On a multiplicative scale (ie, assuming that a given dose will multiply the risk by the same factor anywhere in the exposure range), the departure from linearit y in these data becomes more dramatic. Even those studies which, on an additive scale, appear to be compatible with a linear relationship and an elevated intercept show a clear curvilinear relationship with a decreasing slope on a multiplicative scale .
The dose-response curves for lung cancer risk among arsenic-exposed workers in several studies appear to increase more steeply at lower than at higher exposures. For two smelters (5,6), the data may Scand J Work Environ Health 1993, vol 19, no 4 be consistent with a linear dose-response, but do not contradict such a curvature, and for one of these (6) adjustment for age in a nested case-referent study appears to eliminate any curvature of decreasing slope (7) . One must ask "what mechanisms might account for the nonlinearity in the remaining studies?"

Explanations for the nonlinearity
In this section we postulate several possibilities for the nonlinearity and review the evidence and plausibility of each. Table 3 summarizes our findings , as discussed below.

Artifact of statistical method; confounding by age
Strictly speaking, SMR values are not comparable if the following two conditions are met: different exposure groups have different age distributions and age-specific mortality ratios with respect to the reference population are heterogeneous. Groups defined by cumulative exposure level are likely to have different age distributions for two reasons. First, those who are older will have had a longer opportunity for employment and hence for accumulating exposure. Since exposure groups refer to person-years rather than persons, a more precise way to say this is that person-years at older ages tend to dominate in the higher exposure groups. Secondly, in many industrial settings, exposures have been reduced over time so that even with the same duration of employment, older workers (more precisely, person-years at older ages) will be overrepresented in the higher exposure categories. Of course, noncomparable age distributions can be addressed through direct standardization, the reporting of age-and dose-specific mortality ratios, multivariate regression (eg, Poisson or logistic) control for age, or a nested case-referent study matching on age.
These numerous alternative analytic strategies were used in different studies, for example, "directly standardized death rates" (5), Poisson modeling (27), and nested case-referent studies that adjusted for age and calendar year (8,9,10). Such analyses did not alter the shapes of the observed dose-response curves , with one exception. Jarup & Pershagen (7) compared dose-specific odds ratios adjusted for age to the dose-specific SMR values from the original cohort study (6) after standardizing the SMR values to the lowest exposure level and found that the curvature seen in the SMR analysis was absent in the case-referent study after age adjustment.
In case-referent studies, the use of internal comparisons precludes the possibility that the apparent curvature is actually due to fluctuations around a linear relationship with an elevated intercept. That is, the intercept does not represent a background level of risk, since the value of 1.0 was based on a lowexposure group. Furthermore, all of the case-referent studies of arsenic-exposed workers adjusted for  • Each column represents a given occupatlonatly-exoosed cohort. For some cohorts, different publicatio ns assessed different aspects of the arsenic -lunq cancer assoc iation. Reference numbers in the tab le designate the publication that addressed the issue for that row, as described in the left column.
age and calendar year through multi variate modeling or standardization in the estimation of the risk ratios. Thu s co mparisons of the odds rati os among different exposure level s are valid. Two of these studies (8,9) showed unequivocal supralinearity.
Most of the studies did not addres s the second criteri on for noncomp arability of do se-specific measures of assoc iation, that is, homo geneity across age gro ups (7,8,9). In one cohort in whic h the age-arsenic interac tion was examined (the Tacoma cohort), the ev idence did not support heterogeneity (27). Overall , a variety of analyses indicated that any curvature in the dose-response relationship in five of the six study populations (4,5,8,9,10) was not a result of heterogeneous age distributions among different exposure level s (table 3).

Conf ounding by oth er carcinogens
Confound ing from other workplace exposures could raise the background rates and may be co nsistent with some of the studies. However, the higher slopes at lower exposures would suggest that these other exposures decline with increasing arse nic exposure. Generally, exposure to other carcinoge ns tends to be posit ively associ ated with arsenic (9, 10, 15) and therefore implies a slope that would increase rather than decrease with higher exposure. In one study (9), multiv ariate analysi s was used to cont rol for radon (which was noted by the authors to be highly correlated with arsenic exposure); the curvat ure shown in figure 2 is after such adjustment. A situation in which other carci nogenic exposures might be inversely related to arse nic exposure is when short-term workers find empl oyment elsew here in industries with 222 carcinogenic exposures . Unfortunately, occupatio nal studies usuall y fail to ascertain empl oyment histories after term ination at the plant under study.

Confounding by smoking
The possible contribution of confounding from smoking in the Swedish smelter study is evident fro m a comparison of the dose -res ponse curves with and without adj ustment for smoki ng (7). Thi s analysis revealed slight confounding as a result of smoking in the medium-exposure groups and "negative confounding" at the highe st level of arsenic exposure; that is, with an unadju sted analysis, the odd s ratios for arsenic were slightly inflated in the middle range of exposures and markedly depressed at the highest exposure. The percentage of smokers was similar across dose groups in the Anaconda cohort (2 1) and also amon g the insecticide manufacturin g workers . Thu s confound ing from smoki ng which is differential by dose group may at least partially explain the declining slopes in some of the studie s (eg, in refe rence 6) but co uld not be an explanat ion in others.

Synergism
A dose-dependent synergism with smoking that is larger at lower exposures to arsenic than at higher levels would also cause a reduction in slope. This explanation is supported by the findings that lung cancer risks are described by a multiplicati ve relation between smoking and residential (ie, low) exposure to arsenic (24) but a less than multipli cative (though still synergistic) relation between smo king and occupational (ie, high) arsenic exposure (3, 7, I I, 24, 25).

Healthy worker survivor eff ect
If employees with greater susceptibility to cancer are also those who leave employment earlier, then the resulting healthy worker survivorship could cause a progressively decreasing slope. Robins (28) found that a model incorporating nonrandom job-leaving (with the less susceptible remaining on the job) fit the Anaconda data better than a model that assumed random job-leaving. Jarup et al (6) noted that a stronger association of risk with intensity rather than duration of exposure could be evidence of healthy survivorship; indeed, in this study , short-term workers had mortality ratios as high or higher than those of long-term workers at the same average intensity, a pattern consistent with possible life-style differences. Arrighi & Hertz-Picciotto (27) used several methods to control for the healthy worker survivor effect in the Tacoma, Washington, smelter cohort data. Though some methods resulted in a stronger dose-response, none appeared to alter the curvilinear shape substantially.
If other cause s of death with shorter latency periods are also related to arsenic in a dose-response manner and occur more frequently among those more likely to develop lung cancer, then competing risks from other causes of death could result in a healthy worker survivor effect and distort an otherwise linear relationship. The first condition has been shown for cardiovascular mortality in the Anaconda cohort (21), but not among the Ronnskar workers (6). Neither study reported on nonmalignant respiratory disease, a cause of death that might show a stronger relationship with cumulative arsenic exposure than cardiovascular-related causes of death . However, the second condition, independence of censoring from the disea se of interest, is more difficult to establish. Further examination of competing risks and of a "healthy worker survivor effect" in other cohorts is in order.

Exposure misclassification
Misclassification of exposure that is more pronounced at higher absolute exposure levels could result in upwardly biased exposure estimates at the higher doses, or downwardly biased exposure estimates at lower levels . Either bias would distort a linear relationship to one that is supralinear. The finding by Enterline et al (4) that urinary measures of arsenic exposure appeared to predict lung cancer risk in a linear fashion suggests that the biological dose may not have been the measured air dose.
Although data on exposure misclassification are difficult to obtain, there are numerous ways in which such biases could occur. Nonrandom selection of locations for collecting industrial hygiene samples could have led to inflated exposure estimates in areas where arsenic levels were highest. In several stud-ies, exposures during certain periods were based on extrapolation from other periods. Rappaport (29) has shown, using empirical industrial hygiene data , that the contribution of variability, both within individuals over time and between individuals within a given job category, leads to significant misclassification.
At worksites where arsenic exposures were heavy, workers may have used respirators; failure to account for such reductions in target tissue doses would produce a spuriously low slope at the high end of the exposure range. Respirators were mentioned in relation to the Anaconda cohort (15) and the Tacoma cohort (16).
Finally , the use of cumulative exposure could be an inappropriate measure if short-term high concentration exposures induce a proportionally greater (or smaller) risk as compared with the same total exposure spread over a longer duration, although peak exposures are generally important only when damage operates via a threshold mechanism (29,30).
The previou sly described errors in exposure estimates were likely to have varied by the true exposure level: higher exposures would entail greater errors. The magnitude of errors may be proportional to the exposure level. For instance, at levels of 0.1 ug . m" , the error might be two-to fivefold too high or too low, and similarly at 100 ug . m" the error might again be two-to fivefold , but the magnitude is of course very different. Since larger measurement errors lead to a greater attenuation of relative risk estimates, error s that differ relative to the true exposure would distort a linear relationship to one that had a decreasing slope as exposure increased (31,32).

Toxicologic mechanisms
Inorganic trivalent arsenic is methylated to the less reactive monomethyl arsinic acid, and again to dimethyl arsinic acid (2 I, 33-35). If methylating enzymes were depleted at high doses, then inorganic arsenic could accumulate and, therefore, lead to a curve with an increasing, not decreasing, slope. Thus, if a metabolic pathway were being saturated, it would have to be a pathway leading to activation, not detoxification. Saturation of bronchial epithelial sites for cellular uptake seems unlikely, given the large surface area of the lung. Increased methylation capacity or other pathways of deactivation that operate only at high dose s is another candidate mechanism. The proportion of urinary arsenic excreted in the inorganic (unmethylated) form is about the same over a wide range of exposure levels (33)(34)(35)(36)(37)(38), but, since the data do not indicate what proportion of intake is excreted in the urine, the possibility of increased methylation rates at high chronic exposure levels remains . There might be a susceptible subgroup in the population who methylate arsenic less efficiently; this susceptible group might be depleted, with only slight increases in risk for those with a more efficient methylating capacity. At this point, evidence for other pathway s of deacti vation at high exposure levels , or for a bimodal distribution of methylating capacit y, is lacking. Both a saturation effect or an increa sed induction of methylating enzymes is contradicted by existing evidence. In addition, epidemiologic evidence provid es no basis for, and in fact contradicts, the hypothe sis of a threshold phenomenon (38), a hypothesis that has received attent ion in the literature (39).
A decrease in "potency" at higher exposures might also occur if increasing particul ate matter occurring with arsenic exposure induced more efficient tracheobronchial clearance due to greater mucociliary action and thereby resulted in lower expo sures to the target tissue. The induction of such clearance mechanisms is temporary and rever sible. Under this scenario, lower doses that do not induce clearance would be more potent; in other words, intensity of exposure would be less of a predictor of carcinogenic potency than duration of expo sure. Two studies observed a slightly stronger association of lung cancer risk with inten sity rather than with duration of exposure (4, 6), but another study found duration to be a much stronger predictor (9). Yet another study revealed no consistent pattern s (5).
Arsenic appears to be a co-mutagen. Arsenite has been shown to inhibit DNA repair (40)(41)(42), but a nonlinear relationship was noted between sodium arsenite or sodium arsenate and gene amplification activity (43,44). Enterline et al (4) observed a linear relationship between arsenic excreted in urine (inorganic and organic combined) and the SMR. These data suggest a possibility of a linear relationship with the internal dose and air measurements that are deficient as surrogates. Problems in using cumulative air exposure estimates were addressed in the previous subsection.
Several proposed toxicolo gic mechanisms appear to be refuted by the existing literature, and several others appear not to have been addressed by previous studies. At this point, evidence for a true biological curvilinear dose-response relation between occupational arsenic and lung cancer risk is far weaker than the evidence for seve ral epidemiologic explanations.

Concluding remarks
Studies in three countries evaluated the risk of lung cancer in relation to quantit ative information on occupational arsenic exposure. Mining exposures in China seem to confer a higher risk than smelting exposures in the United States or Sweden, but the reasons are not clear. Synergism with radon or with other metal carcinogens is plausible. An underestimation of exposure in the study of miners in China or a lower background rate of lung cancer may have contributed to this heterogeneity in the estimated carcinogenic potency of arsenic.

224
Studies with quantified exposure data are compatible with a supralinear (decreasing slope) dose-response relationship. Two of these studies may also be consistent with a linear relationship over a higher background risk of lung cancer among arsenic-exposed workers, but three other studies are clearly not. Yet another study found only a slight curvilinearity, but the exposures were far lower in this study than in the others.
Much of the known metaboli c and pharmacokinetic information gathered to date does not explain this curvature. For instance, the induction of methylating enzymes is contradicted by existing data. We have discussed several toxicologic mechanisms that may be operating, such as the induction of other pathways of deactivation at high chronic exposures or the depletion of a subgroup in the population who methylates arsenic less efficiently, but, at present, there appear to be no data addressing these questions. According to the available data, several epidemiologic explanations for the nonlinearity are unlikely. These include confounding by age, by workplace exposures, or by smoking. Prior studies and reanalyses of some existing data are consistent with a role for each of the following: dose-dependent synergism with smoking, dose-related competing risks of death and other contributors to a health y worker survivor effect, and dose-dependent exposure misclassification involving the overestimation of exposure in the highest exposure job categories. In several of the studies, more than one of these factors could have been distortin g the true biological dose-respon se, and it is also likely that the relative contributions of these explanation s differ among studies. Further toxicologic and epidemiologic studies of arseni c and lung cancer should be designed to evaluate these mechanisms.
The significance of the observed curvature in the dose-response curve extends beyond the specific case of arsenic and lung cancer. Several of the proposed mechanisms might apply equally to studies of any occupati onal carcinogen. For very few agents are there adequate data to allow an evaluation of doseresponse relationships over a wide range of quantified exposures in a variety of industrial settings. Thus the findings for arsenic are notable , particularly in light of the common assumption that a linear extrapolation to low doses is "health protective." From the public health standpoint, the much steeper slopes at low as opposed to high doses in these studies suggest that the use of linear models applied to occupational epidemiologi c data may, in some situations, result in an underestimation of the true risks at lower exposures.