Tobacco Documents Online

Risk Anal}-sis, Vol. 8, No. 2, 1988 How Do Cancer Risks Predicted From Animal Bioassays Compare with the Epidemiologic Evidence? The Case of Ethylene Dibromide Irva Hertz-I'icciotto,l Norman Gravitz,2 and Raymond Neutra2 Received June 20, 1985 Cancer risks for ethylene dibromide (EDB) were estimated by fitting several linear non- threshold additive models to data from a gavage bioassay. Risks predicted by these models were compared to the observed cancer mortality among a cohort of workers occupationally exposed to the same chemical. Models that accounted for the shortened latency period in the gavaged rats predicted upper bound risks that were within a factor of 3 of the observed cancer deaths. Data from an animal inhalation study of EDB also were compatible with the epidemiologic data. These findings contradict those of Ramsey et al. (1978), who reported that extrapolation from animal data produced highly exaggerated risk estimates for EDB- exposed workers. This paper explores the reasons for these discrepant findings. KEY WORDS: Ethylene dibromide; risk assessment; cancer; occupational exposure. 1. INTRODUCTION In the absence of adequate human data, quanti- tative cancer risk assessments have relied heavily on extrapolations from animal bioassays conducted at comparatively high doses.0) The validity of such extrapolations has, however, been a source of con- troversy.(2-s) A case in point is that of ethylene dibromid: (EDB), a fumigant that, until recently, was widely used on grain and citrus products. Re- sults of an animal bioassay(9) showed EDB to be an extremely potent carcinogen when administered by gavage. For regulatory purposes, the Carcinogen As- sessment Group (CAG) of the U.S. Environmental Protection Agency used these bioassay data to esti- mate human risks from consumption of EDB re- sidues in I'ood."°•it> I California Public Health Foundation, 2151 Berkeley Way, Room 515, Berkeley, California 94704. 2California 1Department of Health Services, Berkeley, California 94704. Ramsey and associates(2) applied the risk ex- trapolation model used in an early report of the regulatory agency(10) to a cohort of workers at two chemical manufacturing plants who were exposed to EDB by inhalation, and whose mortality was under study.(lZ) The results of Ramsey et al. suggested a wide discrepancy between the observed mortality and the risks predicted from the animal gavage data by a low-dose-linear extrapolation model. These results have been cited as evidence that extrapolations from animal bioassays to human real-world exposures are implausible and hence contraindicateVz-6•13) In this paper we investigate the reascns for the apparent discrepancy. Other nonthreshold models which are linear at low doses are fitted to the gavage data, including the one used in the final risk assess- ment of the CAWiI) Models are also fitted to data from a more recent inhalation bioassay.(14) The use of safety factors is not considered because of the limitations of this approach.(15) We then apply each fitted model to the cohort of EDB-exposed workers 205 0272-4332/88/0604o205SO6.0o/I CJ 1988 Society for Risk Analysis

tub and compare the predicted risks to the observed cancer mortality. 2. THE EIDB DATA AND THE REPORTED DISCREPANCY Cancer deaths among workers exposed to EDB were reported as not significantly elevated, unless a small group with additional exposure to arsenic was included.tlz> However, in a long-term gavage bioas- say,(4) and in two inhalation studies(1a.t6) published subsequent to the risk assessment of CAG, EDB proved highly carcinogenic. In the two assays (one gavage and one inhalation) conducted by the Na- tional Cancer Institute/National Toxicology Pro- gram (NCI/NTP), more than 50% of the high-dose animals exhibited contact-site tumors (squamous cell carcinomas of the stomach from gavage administra- tion, nasal cavity malignancies of several types from inhalation). Both low- and high-dose animals had s,atistically significant excesses of contact-site tumors aad a variety of tumors remote from the site of administration (i.e., systemic tumors). A one-hit, nonthreshold model was fitted by the CAG00> to the rat data from the NCI gavage bioas- say to assess risk from ingestion of EDB-contami- nated food. The CAG used squamous cell carcinomas of the stomach in male rats and an interspecies conversion based on surface area equivalence. In developing its risk assessment for public exposure via ingestion, the CAG specified that the parameter estimates were applicable only for intubation ex- posure. Different parameter estimates were recom- m+.r(ded for dietary exposure and for inhalation ex- pasuret10j (in a later risk assessment of EDB, the CAG scientists developed a more sophisticated model to deal with the irregularities in the gavage bio- assay(lt)). Etamsey et al.tz> applied the one-hit model fitted by the CAG to a cohort of 161 employees involved in the manufacture of EDB J12> Exposure was estimated by (i) assuming all workers were exposed to time- weighted average (TWA) concentrations based on measurements made during the 1970s at one of the two plants,0Z> and (ii) converting to a continuous lifetime equivalent dose using an average weight of 70 kg. Additionally, it was assumed that both potency and brologically effective dose were the same for inhalation as for intubation, i.e., no adjustments were made for route of exposure. I Hertz-Picciotto, Gravitz, 9nd [Yeft The risk of an EDB-induced cancer death wa calculated for each worker in the study by Ott et atO4 These were then summed to obtain the number d excess cancer deaths predicted by the model. In tht cohort of 161 workers, this model predicted over 80 excess cases of cancer from an exposure of 3.0 ppm, or about 50 cases from 0.9 ppm exposure.(Z> Thex predictions are for the partial lifetimes of the workett. Given that only eight cancer deaths were observed, with a 95% upper bound of 16, these predictions are clearly inconsistent with the observed mortality. Figure la displays a comparison of (a) observed cancer deaths, (b) expected cancer deaths based on U.S. white male age-specific rates, and (c) cancer deaths predicted by this model. Predictions are shown for each of the two assumed exposure levels. Since measurements were taken at the Michigan plant only, results for the two plants are presented separately. In light of this discrepancy between predicted and observed cancer mortality in EDB-exposed workers, some authors have suggested that extrapo- s0 55 0 30 to 0 a 1 KEY: } OBSERVED I 95'b CONFIDENCE INTERVAL ~ EXPECTED (AGE ADJUSTED U.S. WHITE MALE RATE) • PREDICTED AT 3.Oppm o PREDICTED AT .9ppm 0 0 + gl A TEXAS MICHIGAN B TEXAS MICHIGAN GAVAGE BIOASSAY INHALATION BIOASSAY Fig. 1. Observed, expected, and predicted cancer deaths using ~ one-hit model and two animal bioassays of ethylene dibromide. 0

R6k. Assessments of EDB and Epidemiology lating from animal bioassay data to predict human cancer risks is inappropriate.(3-6) Because data from such assays are often the strongest evidence of carcinogenicity and usually the only basis available for quantification of human risks, alternative ex- planations should be investigated before drawing such a co aclusion. These should include the possibilities that the assumed exposures in the occupational study were too high, that some aspect of the model or its application to the workers was inappropriate, or that the method for scaling doses between species was incorrect. lit seems unlikely that the workers' exposures were greatly overestimated. Most of the exposures to EDB occurred decades before the measurements were taken, suggesting underestimation of exposure. While indi.vi'idual work histories or job activities were not taken into account, and while it is possible that heavily exposed short-term workers were excluded and that long-term employees experienced lower ex- posures than the TWAs, biases due to these factors are likely to be less significant than the changes in exposure over time. With risk predictions over seven- fold too high (even using the lower of two estimated TWAs), inaccuracies in exposure assessment are un- likely to explain the inconsistency between observed and predicted cancer deaths. We have therefore ex- plored other potential explanations: deficiencies of the model, problems in its application to the workers including the different route of exposure, and the interspecies conversion factor. 3. MI:THODS AND RESULTS A crude extrapolation using direct proportional- ity from the lowest dosed animals in the gavage experiment indicated compatibility between the ani- mal aiad occupational data. The low-dose gavaged rats )received 5.37 mg/kg/day (in human equivalent) and developed 60% more tumors than the control rats. The average worktime dose in the occupational study was 4.6 mg/kg/day, which amounts to only 0.35 :mg/kg/day when averaged over their lives. This represe.nts about 0.065 of the rats' dose, implying an excess risk of 0.04 for each worker (0.065 X 0.6), or about six extra cancer deaths in the cohort of 161, where three excess cancer deaths were seen. 'We further compared the observed cancer deaths among workers in the study by Ott et al.(lZ) with predictions from several linear nonthreshold models, 207 using data from both the NCI inhalation bioassay(t`) and the NCI gavage bioassay.(9) The following mod- els were used: the one-hit model and the multistage` model were fitted to the inhalation bioassay data; the multistage model incorporating time-to-tumor data, the multistage model with variable dosing, and the proportional hazards model were fitted to the gavage data. In each case the fitted extrapolation model was applied to the workers' exposure to obtain predicted cancer risks, which were then compared with ob- served cancer deaths in the EDB-exposed cohort. For the inhalation bioassay, nasal cavity malignancies in male rats represented the most sensitive site, sex, and species. To simplify comparison, we made the same exposure assumptions as Ramsey et al.,(2) with the exception that the model incorporating variable dos- ing and the Cox model do not assume that average lifetime dose is the determinant of risk. Also, while CAG used data from only the low-dose animals in the gavage study, these analyses used data from all dosed animals. 3.1. Inhalation Data: Two Models The one-hit, nonthreshold model takes the form P(d)-P(0) =1-exp(-P•d) where P(d) represents average lifetime cancer risk for an individual exposed to dose d, P(0) is the background lifetime cancer risk, and P is the un- known parameter for carcinogenic potency (i.e., mortality per unit dose) of the substance. When fitted to the inhalation bioassay data using Global 82 software,(t7) this model predicted upper limits of 1.2 and 0.7 excess cancer deaths among . the exposed workers at the Texas and Michigan plants, respec- tively, assuming EDB concentrations averaged 3.0 ppm for all workers at both plants. Since the inhala- tion experiment ran for the full two years, and since most of the animals were sacrificed at term, the calculation of partial lifetime risks for the workers was based on the exponent for time (or age) depen- dence of cancer risk obtained from the gavage data. Other researchers have estimated similar values for the age dependence of human cancer(l8•19) ; using smaller values as reported for lung cancer by Doll and Peto(1O) did not substantially alter the results. As shown in Table I and Fig. lb, when the small excess risks predicted by this model were added to the expected deaths, the resulting total predicted cancer

208 Hertz-Picciotto, Gravitz, and Neutrlt Table L Numbers of Cancer Deaths Predicted by Linear Nonthreshold Models Fitted to Inhalation Data Excess predicted by one-hit model Observed Onutted With tumors found at terminal sacrifice: Included (95% Cl) Expected° 0.9 ppm 3.0 ppm 0.9 ppm 3.0 ppm Texas 3` (0.6-8.8) 3.6 0.13° (0.17)` 0.44 (0.57) 0.30 (0.38) 1.00 (1.25) Michigan 5 (1.6-11.7) 2.2 0.08 (0.10) 0.26 (0.33) 0.17 (0.22) 0.58 (0.72) °From t1.S. white male age- and calendar-year-specific mortality rates. hModels were fitted to inhalation data using Global 82 software published by Crump and Howe (1982). Proportion with tumors was based on life table adjustment. Animals who died prior to the first tumor were not included. `Included one arteriosclerotic heart disease death with lymph node malignancy (see Ref. 2). dMaximum likelihood estimates. `Numbers in parentheses are upper 95% confidence limits. mortality was close to the observed mortality among the EDB-exposed workers. 'The multistage model generalizes the one-hit model by allowing for nonlinear terms in the hazard rate of cancer death. This model is of the form P(d)-P(0)=1-exp[ -(iBt.d+RZ.d2+ .-..)] with E3; the unknown parameters. When fitted to the inhalation data, the linearized multistage model pre- dicted identical risks, that is, the strong linearity in the data dictated that the best fit was the one-hit model. 3.2. Gavage Data: Three Models The gavage bioassay was marked by severe early mortalhy from both toxic and carcinogenic effects of the high doses of EDB: one-third of the high-dose animals died by the 15th week:'The two-year bioas- SQY Wils, iht'.Bt'ft)rC tt'fn)inJlCtf hefi!(C /I):r rilj!# 11` 1})X' I/r t;, l,~ 1'1, !udlw.t joir 1Ir, 0u,rlr•ru'll (Ilhsprul~ sttr4 4Nii1 , 1111, 1 # t///i N, /t ri. 11 'l////Nt{oN 'if (hl' lf//1~~14t:)j!l tt/~n/i 1/ti~ ~~tlutttlflft~ tht,• Sltfv/vaI t/rr1cS of the animals (tlcnoted tttultistage with time-to-tumor model) was fitted to these data using Weibull 82 software. The form of this model is P(d)--P(0) =1-expl-'(Pt-d+P2•d' + ..-)(t-to)r~ where t represents time since first exposure and to represents the latency period. Thus, (3i and to are parameters to be estimated. Using this model, the upper-limit predictions were two to three times the observed cancer deaths assuming exposures of 3.0 ppm: 11.9 and 3.9 excess cancer deaths at the Texas and Michigan plants, respectively, In response to the early mortality, dosing was stopped for the high-dose animals, and subsequently a variable dosing pattern was instituted for this group. Zeise and Crouch,t1I Thorslund,t23t and Crump and Howetz4> developed a special case of the Armitage- Doll multistage model for carcinogenesis, which in- corporated such a variable dosing regimen. This model was used in the CAG's final risk assessment for EDB,tII> and took the form P(d) -P(0) =1-expt-(fl•d)[(t-s)'' -(t- f )Y11 where s is the age at start of exposure, f is the age at end of exposure, t is the age at end of observation period, and d is the daily dose from age s tn f. Y/Jtxtr t/tt~•f t/, tdo; pvage data, this model gaVe predictiorts that were similar to those of the ti.me-to- turrurr rru,del: 10.4 and 5.5 excess cancer deaths at the Texas and Michigan plants, respectively, assum- ing concentrations of EDB averaged 3.0 ppm (see Table II). Thus, a variable dosing schedule did not significantly influence the risk projections. The Cox proportional hazards model differs from the previously described models by treating the in- crease in risk as a multiplicative rather ~ than an additive effect. The model takes the form P(d, t) =1-expI- f tX(d,,. u) 3.1 0

Risk Assessments of EDB and Epidemiology 209 T'slae Il. Numbers of Cancer Deaths Predicted by Multistage race-, sex-, and calendar-year-specific rates Model Adapted for Variable Dosing Fitted to Gavage Data for U.S. males. Number of Cancer Deaths c. The predicted excess probability of a cancer Observed Excess predicted by model" death for worker j at dose = d was derived (95% CI) Expected° 0.9 ppm 3.0 ppm using Hj(0), the estimate for fl, and the pro- T:xas 3` (0.6-8.8) 3.6 2.84d (3.49)° 8.65 (10.42) portional hazards assumption: MSchigan 5 (1.6-11.7) 2.2 1.95 (2.33) 4.84 (5.53) "From U.S. white male age- and calendar-year-specific mortality rates. hMttltistage model adapted for variable dosing (Ref. 11). `Includes one arteriosclerotic heart disease death with lymph node malignancy (see Ref. 2). dMeximum likelihood estimates. 'Numbers in parentheses are upper 95% confidence fimits. wfaere P(d, [) represents the risk by time = t, for a dosing pattern d, and X(d, u) represents the in- st,antaneous hazard at time u due to dose = d„ (_= dose between 0 and u). Furthermore, this model assumes that the hazard for an exposed individual is proportional to the base-line hazard at all times: X(d, t) =f(d )'a(0,r) (Dose can be a function of time or not.) This model has been used recently as a basis for developing and' comparing potencies from animal carcinogenicity bioassays. t16-11' The function f(d) is taken to be linear: (1 +R• d). The following steps implemented this model: a. The parameter ~B was estimated from the animal gavage data. b. For each worker j, the integrated base-line hazard, Hj(0), was determined using age-, P(dj) + P(0) =1-exp{ -I,8dj] - [H,(0)I} d. The predicted number of excess cancer deaths for the cohort was obtained by summing the risks over all workers in the cohort. Unlike the previously discussed models, the Cox proportional hazards model assumes that the excess risk is a function of the background rates. Since EDB is not expected to affect all cancer sites uniformly, two potential sites were selected for extrapolation: lung and stomach. The excess lung cancer deaths predicted by the Cox model assuming that EDB concentrations aver- aged 3.0 ppm were 46.8 and 22.1 for the Texas and Michigan plants, respectively; predicted excess stom- ach cancer deaths were 17.8 and 9.5. These predict- ions are maximum likelihood estimates; because of the strong monotonicity of the gavage data, the vari- ance was too unstable to derive a reliable confidence interval. The high lung cancer predictions were simi- lar to those of the one-hit model fitted to the gavage data. Tables IIY and IV summarize the risk predic- tions from all of the models discussed above. For simplicity of presentation, Table III assumes 0.9 ppm exposure, and Table IV assumes 3.0 ppm exposure. Table IA. Total' Cancer Deaths Predictedh by Several Models for EDB-Exposed Workers'` Models fitted to gavage data Texas Michigan Overall Observed cancer deaths 3 5 Proportional hazards Multistage with Multistage with One-hit Stomach Lung time-to-tumor variable dosing 38.6 10.4 31.4 7.7 7.1 21.2 5.9 13.0 3.5 4.5 59.8 16.3 44.4 11.2 11.6 Models fitted to inhalation data One-hit: terminal sacrifice tumor: Omitted Included 3.8 2.3 6.1 4.0 2.4 6.4 'Total cancer deaths s Iexpected+predicted excessl. °Otlter assumptions are described in the text. Values in the table represent upper 95% confidence limits, except for the proportional hazards modcl. : whidh the variance estimates were too unstable to derive an upper confidence limit. `tissurnes 0.9 ppm exposure during time employed. 2025545876

210 Hertz-Picciotto, Gravitz, and Neutra Table IV. For Total° Cancer Deaths Predicted° by Several Models for EDB-Exposed Workers` Texes MicEug,an Overal l Observed cancer deaths 3 5 8 Models fitted to gavage data Proportional hazards Multistage with Multistage with One-hit Stomach Lung time-to-tumor variable dosing 56.6 21.4 50.4 15.5 14.0 34.2 11.7 24.3 6.1 7.7 90.8 34.5 74.7 21.6 21.7 Models fitted to inhalation diu One-hit: terminal sacrifice tumor Omitted Included 4.2 2.5 6.7 4.9 2.9 7.8 °Tot,al cancer deaths=(expected+predicted excessl. ^ Other assumptions are described in the text. Values in the table represent upper 95% confidence limits, except for the proportional hazards model, which the variance estimates were too unstable to derive an upper confidence limit. `Assumes 3.0 ppm exposure during time employed. Table V. Effect of Interspecies Dose Conversion Factor on Cancer Risk Predictions° a Dose equivalence by Predicted excess cancer deaths mg/kg/day mg/kg2/3/day Texas 0.9 pm 0.8 4.1 3.0 ppm 2.6 11.9 Michil att 0.9 ppim 0.3 1.3 3.0ppm 0.8 3.9 °Time-tatumor model fitted to gavage data. Values in the table represent upper 95% confidence limits. 3.3. Int:rspecie§ Scaling Factor As noted, the gavage analyses used surface area to scale the doses from animals to man. To investi- gate the role of the interspecies scaling factor we repeated the analysis that fitted the multistage with time-to-tumor model to the NCI gavage bioassay data, utsiing mg/(kg body weight)/day equivalence rather than surface area equivalence. A compari- son of these two analyses is shown in Table V. Sur- face area equivalence yielded risk estimates that were about five times larger than those based on mg/kg/day equivalence. 4. DISCUSSION The large uncertainty in risk assessment due to extrapolati,ng between high and low doses is of con- siderable concern. Unfortunately, the only feasible way to conduct a sensitive animal bioassay is to use high doses, since the risks at low doses generally cannot be detected unless many thousands of animals are treated; nevertheless, such risks may be of con- siderable public health concern if exposures are widespread. Thus, the uncertainty of high-to-low dose extrapolation is unavoidable. One result of this inves- tigation was to develop a means of narrowing the range of uncertainty by comparing model-based estimates derived from animal data to the observa- tions in epidemiologic studies. All the linear nonthreshold additive risk models considered here for extrapolating human cancer risks from animal bioassay data performed well when validated against the mortality of EDB-exposed workers. That is, the predictions are compatible with the reported cancer deaths in the occupational study of Ott and colleagues.t12> Even a crud`e extrapolation using direct proportionality, which is equivalent to drawing a straight line between zero excess risk at zero dose and the excess tumor rate in the low-dose rats, gave plausible risk estimates when compared to the epidemiologic data. This compatibility contrasts sharply with the findings of Ramsey and co- workers.S2> We discuss the reasons for this difference and the implications of our findings. 4.1. Choice of Model On theoretical grounds, the multistage model for variable dosing and time-dependent risk appears to be the most appropriate of the additive models for the analysis of the EDB gavage data. This is because it utilizes the full information on both survival times and dosing pattern and makes no assumptions re- garding dose rate. Considering the uncertainties in- volved in risk extrapolation it is apparent that the

Risk Assessments of EDB and Epidemiology predictions of this model are fully compatible with the observed mortality in the EDB-exposed cohort. The proportional hazards model, also sensitive to both the survival times and the variable dosing, gav: rather high risks, particularly if one assumes its effect to be on lung cancer, the site of contact for occupational exposure. This was due to the depen- dence of the model predictions on background rates in humans, when the model parameters were esti- mated from an animal experiment in which the back- ground rate was zero. It is also rare for an agent that inducs lung cancer to affect the background rates, which are primarily due to smoking, in a multiplica- tive way. Asbestos is a notable exception.('9) On the other hand, at the low exposure estimate (0.9 ppm), assuming EDB's effect is on stomach cancer, the predictions are compatible with the observed cancer deat#ls, at the one plant where measurements were taken (Michigan). Without knowledge of the site of EDE's carcinogenic activity in humans, it is difficult to say whether this model is compatible with the epidemiologic data. Tlae multistage model with time-to-tumor yields very similar risk estimates as the model that also incorporates the variable dosing pattern, suggesting that ~the variable dosing was not an important factor in th : potency of EDB in the gavage bioassay. With the 3.0 exposure estimate, overall predictions were aboun ithree times the observed mortality among the work .rs. Thus, because the workers' exposure begins comparatively late in life, the additive models which incorhorate information on time since exposure be- gins provide a far better fit to the worker data than those models that do not. NVlule numerous other models (probit, Weibull, logit, etc.) have been advocated for extrapolation, we have limited our analyses to those with the property of being linear at low doses. Such curves will yield an upper bound for the risk at low doses(3°-32) and thus provide a health-protective basis for regulatory deci- sion making. Itt should be emphasized that risk assessment makes no claim to providing precise predictions, but rather seeks to generate ball-park estimates. These estimates are intended as plausible upper bounds of risk. Restricting discussion to models used with the gavage data, the one-hit model predicts implausible risks, as does the proportional hazards model using lung cancer deaths, while both the variable dosing and the time-to-tumor forms of the multistage model are compatible with the epidemiologic data. We conclude '211 that the gavage data are not inherently incompatible with the workers' cancer mortality experience. The inhalation bioassay was not fraught with the complications -of high early mortality and a re- duced latency period. This may partially explain why the one-hit model (without an adjustment for latency) applied to the inhalation data predicted cancer risks that were fully compatible with the epidemiologic study. The importance of the latency period is under- scored by the fact that most of the tumors observed in the inhalation bioassay were discovered at termi- nal sacrifice. If we were to exclude such tumors, the predicted excess risks from this bioassay would be halved. 4.2. Route of Exposure When the doses in the inhalation study were converted to units of mg/kg/day, they were in fact larger than the doses of the gavage study (see Table VI). The time on the study was about double the duration of the gavage bioassay, while the probabil- ity of developing tumors was comparable to the probability in the gavage study if one includes the tumors found at terminal sacrifice (104 weeks in the inhalation bioassay). Thus, the gavage study is dis- tinguished from the inhalation study by lower doses, 'rable VI. A Comparison of Two Carcinogenicity Bioassays: Gavage and Inhalation Controls Low dose High dose Gavage Dosc° 0 5.37 6.66 Response (crude) 0/20 30/50 25/49 P (response)b 0.0 0.61 0.75 Weeks on study Mean 53 45 31 Median 49 47 36 Inhalation Dose° 0 9.56 38.24 Respotsse` (crude) 0/20 2/48 24/48 39/49 P(response)b 0.0 0.05 0.65 0.91 Weeks on study Mean 103 98 76 Median 104 104 80 °Doses are in mg/kg/day averaged over the lifetime, convened to human equivalent using surface area as the basis for interspecies scaling. "Kaplan-Meier probabilities. `Response rates for low-dose rats in the inhalation study were derived excluding (2/48) and including (24/48) tumors found at terminal sacrifice.

212 a higher carcinogenic potency manifested as much shorter latencies, and greater subchronic toxicity. The high carcinogenic potency in the gavage bioassay was driven largely by the shortened laten- cies. Adjusting for latency was especially crucial for accurately estimating risks to those whose exposure began late in life. Since even those models incorpo- ra'.ing latency period produced larger gavage-based risk estimates than models fitted to the inhalation data, it is possible that extrapolating from the inhala- tion bioassay underestimates the effects of EDB in humans. That is, humans are potentially as sensitive as the most sensitive site, sex, and species observed in a bioassay using any route. When data from more than one route are available, a risk assessment for human exposure can be based on data from the same route of exposure as that of the humans, if this same route leads to a positive dose-response and other factors are equal (e.g., statistical power of the study, species or strain sensitivity). If, however, bioassay da.ta using a route different from that of the human exposures show a higher potency, these data should not be rejected outright. In the interest of protecting the public health, the bioassay data showing the higher potency should be considered carefully, in conj unction with pharmacokinetic data that may shed light on species and route differences. 4.3. ]lnterspecies Conversion As demonstrated in Table VI, the mg/unit surface area basis for scaling doses between animals and man yields higher risks for humans than the use of mg/kg body weight. Thus, even the one-hit model would have predicted risks compatible with the workers' mortality, had body weight been used as the scaling factor. In a comprehensive review of the interspecies scaling issue, Davidson et a1.(33) conclude that surface area scaling is most likely to provide the correct scaling for carcinogenicity because toxico- logic,, metabolic, and pharmacokinetic data correlate best when body weight is raised to the power 2/3 or 3/4. With respect to EDB, two lines of argument lead to the conclusion that surface area is likely to be the appropriate basis. (i) EDB acts as a carcinogen at sites af art from point of contact; based on experi- mental data, at least one pathway involves activation by the cytochrome P-450 mixed function oxidase system.j34) (ii) Contact-site tumors were the most sensitive site for EDB, and the surface area of this target site is proportional to the body surface area. Hertz-Picciotto, Gravitz, and Neutra 5. GENERAL IMPLICATIONS While the relationship between the quantitative aspects of laboratory animal carcinogenesis and hu- man carcinogenesis remains to be delineated, there is evidence that the two may not be far apart for at least some agents. Rowe and Springer(3i) showed animal-based estimates of asbestos-induced carcino- genic potency to be within the range of human-based estimates from several studies. Similarly, animal- and human-based estimates derived for the carcinogenic potency of benzene(3b) and gasoline(37) were re- markably close. An analysis similar to the one pre- sented here indicated compatibility between a risk assessment for ethylene oxide based on rat mono- nuclear cell leukemias and leukemias observed in two cohorts of workers involved in the manufacture of ethylene oxide.08) Crouch and Wilson compare potency estimates based on human data to estimates based on rat and mouse data, and found that in about 2/3 of the comparisons, the estimates differed by less than one order of magnitude.(39) These find- ings are in direct contradiction to the claims of some scientists<3•4•',IIj that animal-to-human extrapolations have no scientific basis. In response to reports of a high correlatioL, between rat and mouse carcinogenic potencies,(39-41) Bernstein et a1.(42) have shown that rat-mice potency correlations are an artifact of the way doses are determined for the bioassays. Since human doses are not established by experimental protocol, similar potencies for animals and humans are unlikely to be an artifact. The present paper adds further empirical evi- dence that quantitative data from animal carcinogen- icity studies are a reasonable basis for estimating human cancer risks, and that linear nonthreshold additive models provide a practical means for such risk estimation. This should not be construed to mean that human and animal data will necessarily be consistent. Species differences for some compounds are supported on both theoretical and empirical grounds. When comparing bioassays of the same chemical performed in different species, potencies may differ by more than an order of magni- tude.(39-a1,a3•`w) On the other hand, even if the true carcinogenic potencies for two species are close, the estimated potencies may not be. This is because assumptions are required wherever the data are lack- ing. Well-conducted epidemiologic studies of those occupationally exposed to compounds present at X

Rislk Assessments of EDB and Epidemiology muc:h lower levels in the environment provide crucial information to environmental health professionals and risk assessors. Even when such studies yield negative results in a hypothesis test, they can serve as a check on the plausibility of animal-based risk estimates. Clearly inappropriate models or assump- tions can be discarded, and greater confidence can ther be placed in the final risk assessment. Towards thi"is end, occupational studies require more attention to exposure estimation than has generally been the case in the past, and continuing foflow-up of exposed cohorts. Twelve years have passed since the closing date of follow-up in the study by Ott et a02) While we urge the collection and analysis of such data, we would also emphasize that even in the absence of human data, the continued use of animal data is appropriate. The field of carcinogenic risk assessment is in its infancy. The primitiveness of methodology echoes the lack of a clear theory of carcinogenesis. However, the gaps in knowledge and the uncertainties in meth- ods do not constitute sufficient justification for abandoning efforts to provide the public with plausi- ble upper bounds for cancer risks due to environ- mental chemical exposures. For a large number of such exposures, these estimates will necessarily be based on animal data. When quantified human ex- posure data are available and are related to cancer risk, these data can be useful either as a basis for exarapolation or as a standard for assessing the plausibility of risk estimates based on animal data alone. 6. SUMMARY Critics of cancer risk projections based on animal bioassays frequently make reference to negative epi- de niologic findings, and to reports such as that by Ram sey et aL(2) The analyses presented here demon- strate that low-dose extrapolations using linear non- thre;,hold additive models are not intrinsically dis- crepant with epidemiologic observations of cancer mortality. Additive risk models fitted to data from both gavage and inhalation bioassays predicted risks that were plausible when compared to published data from. an epidemiologic study of EDB-exposed work- ers. However, in rats, EDB is a more potent carcino- gen by gavage than by inhalation, with the higher potency manifested in shortened latency periods. Be- caus: of the shortened latency period, only models incorporating age at start of exposure were ap- propriate for the purpose of applying a risk assess- ment based on the gavage data to workers whose exposure began late in life. Application of a multi- plicative model gave implausibly high risk estimates when using lung cancers, though this may have been due to the choice of the wrong target site in humans. Thus, the previously reported overestimate of risk to workers occupationally exposed to EDB was due to a failure to consider their age at start of exposure when extrapolating from an animal bioassay with an ex- ceedingly short latency period. In the absence of viable alternatives, the results of this investigation support continued use of animal extrapolations to predict human cancer risks from environmental chemicals. Epidemiologic data with quantified exposure estimates can serve as an em- pirical standard for assessing the plausibility of ex- trapolation models. Linear nonthreshold additive models have been shown to provide plausible upper bounds when applied with due consideration to the quality of the data from the animal bioassays. ACKNOWLEDGMENTS We would like to thank Dow Chemical for pro- viding the mortality data for the employees exposed to ethylene dibromide, and Dr. Todd Thorslund for reviewing the history of the CAG's development of a risk assessment for EDB. REFERENCES 213 1. E. L. Anderson and the Carcinogen Assessment Group of the U.S. Environmental Protection Agency. "Quantitative Ap- proaches in Use to Assess Cancer Risk" Risk Analvsis 3, 277-295 (1983). 2. J. C. Ramsey, C. N. Park, M. G. Ott, and P. J. Gehring. "Carcinogenic Risk Assessment: Ethylene Dibromide," Toxi- cology and Applied Pharmacology 47, 411-414 (1978). 3. B. N. Ames, "Cancer and Diet-Reply," Science 21.4, 668-670, 757-760 (1984). 4. F. J. Stare, "Controversy About the Risks of EDB (Letter)," New England Journal of Medicine 310, 1387 (1984). 5. W. R. Havendar, Editorial, "EDB and the Marigold Option," Regulation, AEI Journal on Government and Society 16, ~ 13 -17 (1984). O 6. P. J. Gehring, "The Chemical Industry's Record in Environ- .~~ mental Health," Journal of Enuironmental Health 47, 58-61 (1984). ~ 7. D. A. Freedman and H. Zeisel, "From Mouse to Man: The ~ Quantitative Assessment of Cancer Risks," Technical Report ~ No. 79, Department of Statistics, University of California. Berkeley, California (1986). ~ 8. B. N. Ames, R. Magaw, and L. S. Gold, "Ranking Possiblc ~ Carcinogenic Hazards," Science 236, 271-280 (1987). ~..~'

214 9. National Cancer Institute (NCI), "Bioassay of 1,2- Dibromoethane for Possible Carcinogenicity," NCI Technical Report Series No. 86, Bethesda, Maryland (1978). 10. Environmental Protection Agency, "The Carcinogen Assess- ment Group's Risk Assessment for Ethylene Dibromide (EDB)" (1978). 11. Environmental Protection Agency, Office of Pesticide Pro- grams "Ethylene Dibromide" Position Document 4, Washing- ton, D.C. (1983). 12. M. G. Ott, H. C. Scharnweber, and R. R. Langner, "Mortality ,.xperience of 161 employees exposed to ethylene dibromide in ;wo production units." British Journal of Industrial Medicine :37, 163-168 (1980). 13. G. W. Gribble, "Ethylene Dibromide Uproar" (Letter), Chem- ical and Engineering News 62, 63 (1984). 14. National Cancer Institute (NCI), "Carcinogenicity Bioassay of 1,2-Dibromoethane in F344 Rats and B6C3F1 Mice (Inha- lation Study)," NCI Technical Report Series No. 210, Re- search Triangle Park, North Carolina and Bethesda, Maryland (1982). 15. I). Krewski, C. Brown, and D. Murdoch, "Determining 'Safe' Levels of Exposure: Safety Factors or Mathematical Models?" 1'undamentals of Applied Toxicology 4, S383-S394 (1984). 16. C. K. Wong, J. M. Winston, C. B. Hong, and H. Plotnick, "Carcinogenicity and Toxicity of 1,2-Dibromoethane in the P'.at;" Toxicology and Applied Pharmacology 63, 155-162 (1982). 17. K. S. Crump, "An Improved Procedure for Low-Dose Carcinogenic Risk Assessment from Animal Data," Journal of Environmental Pathology, Toxicology, and Oncology 5, 339-345 (1984). 18. P. Armitage and R. Doll, "The Age Distribution of Cancer and a Multistage Theory of Carcinogenesis," British Journal of Cancer 8, 1-12 (1954). 19. R. Doll, "The Age Distribution of Cancer: Implications for Models of Carcinogenesis," Journal of the Royal Statistical S xiety 134, 133-166 (1971). 20. R. Doll and R. Peto, "Cigarette Smoking and Bronchial Carcinoma: Dose and Time Relationships Among Regular Smokers and Lifelong Nonsmokers," Journal of Epidemiology and Community Health 32, 303-313 (1978). 21. IC, S. Crump, "Dose Response Problems in Carcinogenesis," Biometrics 35, 157-167 (1979). 22. L. Zeise and E. A. C. Crouch, "Reconciling the Results of Carcinogenesis Bioassays with Benzo(a)pyrene for the Purpose of Risk Assessment," Society for Environmental Toxicology and Chemistry Annual Meeting, Arlington, Virginia (1983). 23. T. W. Thorslund, "Estirnation of the Effects of Exposure to a Carcinogen that Fluctuates over Time on the Lifetime Risk of Cancer Death," AAAS Pacific Division 63rd Annual Meeting (1982). 24. K. S. Crump and R. B. Howe, "The Multistage Model with a 'lirne-Dependent Dose Pattern: Applications to Carcinogenic Risk Assessment," Risk Analysis 4, 163-176 (1984). 25. 1. I). Kalbfleisch and R. L. Prentice, The Statistical Analysis of Failure Time Data (Wiley, New York, 1984). 26. C. Sawyer, R. Peto, L. Bernstein, and M. C. Pike, "Calcula- tion of Carcinogenic Potency from Long-Term Animal Carcinogensis Experiments," Biometrics 40, 27-40 (1984). 27. R. Peto, M. C. Pike, L. Bernstein, L S. Gold, and B. N. Ames, "'i'he TD,,: A Proposed General Convention for the Numeri- Hertz-Picciotto, Gravitz, and Neutra cal Description of the Carcinogenic Potency of Chemicals in Chronic-Exposure Animal Experiments;' Environmental Health Perspectives 58, 1-12 (1984). 28. L. S. Gold, L. Bernstein, J. Kaldor, G. Backman, and D. Hoel, "An Empirical Comparison of Methods Used to Estimate Carcinogenic Potency in Long-Term Animal Bioassays: Life- table vs. Summary Incidence Data," Fundamentals of Applied Toxicology 6, 263-269 (1986). 29. E. C. Hammond, I. J. Selikoff, and H. Seidman, "Asbestos Exposure, Cigarette Smoking and Death Rates," Annals of the New York Academy of Science 330, 473-490 (1979). 30. R. Peto, "Carcinogenic Effects of Chronic Exposure to Very Low Levels of Toxic Substances," Environmental Health Per- spectives 22, 155-159 (1978). -31. K. S. Crump, D. G. Hoel, C. H. Langley, and R. Peto, "Fundamental Carcinogenic Processes and their Implications for Low Dose Risk Assessment," Cancer Research 36, 2973-2979 (1976). 32. P. Armitage, "The Assessment of Low-Dose Carcinogenicity:" Biometrics Suppl. 38, 119 -129 (1982). 33. I. W. F. Davidson, J. C. Parker, and R. P. Beliles, "Biological Basis for Extrapolation Across Mammalian Species;" Regul Toxicol Pharmacology 6, 211-237 (1986). 34. P. J. van Bladeren, J. J. Hoogeterp, D. D. Breimer, and A. van der Glen, "The Influence of Disulfiram and Other Inhibi- tors of Oxidative Metabolism on the Formation of 2-Hydroxy- ethyl-mercapturic Acid from 1,2-Dibromoethane by the Rat " Biochemical Pharmacology 30, 2983-2987 (1981). 35. J. N. Rowe and J. A. Springer, "Asbestos Lung Cancer Risks: Comparison of Animal and Human Extrapolations," Risk Analysis 6, 171-180 (1988). 36. California Department of Health Services, "Health Effects of Benzene, Report to the Scientific Review Panel, Part ' (1984). 37. P. E. Enterline, "A Method for Estimating Lifetime Cancar Risks from Limited Epidemiologic Data," Risk Analvsis 7, 91-96 (1987). 38. 1. Hertz-Picciotto, R. R. Neutra, and J. F. Collins, "Carcino- genicity of Ethylene Oxide: A Comparison of Dose-Response Data in Animals and Humans" (letter), Journal of the Ameri- can Medical Association 257, 2290 (1987). 39. E. Crouch and R. Wilson, "Interspecies Comparison of Carcinogenic Potency; " Journal of Toxicology and Environmen- tal Health 5, 1095-1118 (1979). 40. E. A. C. Crouch, "Uncertainties in Interspecies Extrapolations of Carcinogenicity," Environmental Health Perspecives '50, 321-327 (1983). 41. D. W. Gaylor and J. J. Chen, "Relative Potency of Chemical Carcinogens in Rodents," Risk Analysis 6, 283-290 (1986). 42. L Bernstein, L. S. Gold, B. N. Ames, M. C. Pike, and D. G. Hoel, "Some Tautologous Aspects of the Comparison of Carcinogenic Potency in Rats and Mice," Fundamentals of Applied Toxicology 5, 79-86 (1985). 43. D. M. Siegel, I. Hertz-Picciotto, M. Lipsett, and R. Neutra, "Health Effects of Cadmium. Part B. Report to the Air Resources Board." California Department of Health Services (1986). 44. I. Hertz-Picciotto, "The Role of Assumptions in Quantitative Cancer Risk Estimation," American Public Health Association 114th Annual Meeting (1986).