[CANCER RESEARCH 38, 1835-1838, 0008-5472/78/0038-OOOOS02.00 June 1978] Letter to the Editor Correspondence re: K. S. Crump, D. G. Hoel, C. H. Langley, and R. Peto. Fundamental Carcinogenic Processes and Their Implications for Low-Dose Risk Assessment. Cancer Res., 36: 2973-2979, 1976, and H. Guess, K. Crump, and R. Peto. Uncertainty Estimates for Low-Dose-Rate Extrapolations of Animal Carcinogenicity Data. Cancer Res., 37: 3475-3483, 1977. In recent years 2 papers (4, 7) from the National Institute of Environmental Health Sciences (including coauthors from other institutions) have appeared in Cancer Research on the subject of low-dose risk assessment for carcinogens. The authors of these papers have made related points in other publications (3, 5, 6). Related papers with myself as coauthor (13, 14) have appeared in Cancer Research and the Journal of the National Cancer Institute. A main thesis of the first of the papers cited (4) was the justification of a linear model (at least at very low doses) relating added cancer risk to dose of carcinogen; cancer risk above background is proportional to any low level for the dosage of the agent used. This is shown to apply for multihit, multistage, and multievent models. For arbitrary models a linear relationship for small added doses is justi fied on the basis that the high background rate of cancer in humans reflects the presence of high levels of carcinogens that are already in the environment, so that the effect of any small increment in the dosage should be approximately proportional to the size of the increment. This argument would hold on the assumption that the carcinogen being considered is interchangeable with the carcinogens already in the environment. In the more recent article in Cancer Research (7), the point is pursued still further, and the much lower "safe" levels given by this approach than those by the M-B1 approach are demonstrated. This actually had already been anticipated by us (14), and we had indicated that, for the linear "1-hit" model, the 1/100,000,000 risk level dosage would be approximately 1/1,000,000 dosage at a risk of 1%. However, by the unit probit slope M-B model, the corre sponding dosage fraction would be about 1/2,000 (more closely, 1/1,830). We suggested that the 1-hit model pro vided a most conservative rule for extrapolation. [To indi cate that the M-B model would be conservative for hitness models higher than 1, we stated that, for a 2-hit or 2-stage process, the dosage fraction, 1/100,000,000 versus 1%, would be 1/1,000 (see also Ref. 9)]. While the paper by Guess et al. (7) was under considera tion for publication, I had some discussions about it with one of the coauthors, Dr. Richard Peto. Peto indicated that one of the functions of the paper was to bring out that under the circumstances the M-B method might fail to be conservative. Some general impression had arisen that the M-B method was always conservative and, in Peto's view, slope under the linear model), it still provided high conserv atism relative to the threshold model that had earlier pre dominated. Under a threshold model the zero slope would apply over a finite dose interval in contrast to the infinitesi mal interval under the M-B model. Peto did not consider his work with Guess and Crump (7) a scoring of the M-B procedure, but only a clarification of one aspect. However, one aspect of the work of this group leaves me concerned. They have created the impression that they have scientifically established the validity of their linear model, particularly for "direct-acting carcinogens." If hu mans or mammals were molecules or, possibly, even singlecelled organisms, it could be that they would show a linear dose response to direct carcinogens. However, humans and mammals are complex organisms and do not behave according to molecular theory. It may be that a particular animal can cope with some moderate amount of an agent, whether by elimination, repair, or any other process. In a sense the individual animal has a threshold, but there is no general threshold since some individual thresholds could be rather small. It would not be unreasonable for the response versus, dose slope to be zero in the neighborhood of zero dose for complex organisms exposed to a direct carcinogen. Cornfield (1) describes how such response curves with zero slope at the origin can result even if the underlying process is a linear one. If such a scientific basis for a linear model does not apply, these authors must rely on the justification that the dosage applied is a small increment to high levels of interchangea ble carcinogens that are already in the environment. How ever, this is not a scientifically established fact. Any inves tigator attempting to establish a threshold for a carcinogen would be denying this possibility implicitly, for such thresh olds could not exist for agents that were already at effec tively high levels (and it was in a milieu of threshold-seeking attempts that Bryan and I had been working). Indicative of the noninterchangeability of carcinogens is the fact that some, like vinyl chloride, induce unique neoplasms that occur only rarely, if at all, in the general population. It would seem that in Ref. 7 the authors were only expressing an opinion and were not establishing a scientific result, and their demonstration of low safe levels under a linear model had long before been conceded. Peto's acceptance of the contributions by Bryan and this required correction. Peto recognized that, although the M-B model called for a response versus dose slope of zero in the neighborhood of zero dose (in contrast to a positive myself would seem not to be shared by all his coauthors. Thus, when Salsburg (16) published an open query about theoretical problems in the modified M-B procedure, Crump (2) responded by raising his own questions about the M-B procedure while answering none of Salsburg's Supported by USPHS Grant CA-15686 from the National Cancer Institute. 1 The abbreviation used is: M-B. Mantel-Bryan. questions. (Although Salsburg and Crump may be in con cert in faulting M-B, they are in serious disagreement with each other. Salsburg was motivated by a concern that the Received December 22, 1977; accepted March 17, 1978. JUNE 1978 Downloaded from cancerres.aacrjournals.org on June 15, 2017. © 1978 American Association for Cancer Research. 1835 N. Mantel M-B procedure is grossly overconservative, whereas Crump is concerned that it is insufficiently conservative.) Were all issues raised by Crump resolved, the causes for Salsburg's concern would still remain, and to these I have provided my own reply (12). Crump's questions relate primarily to the poor fit to data provided by the M-B model and the arbitrariness of the model. Crump contends that, relative to a model incorpo rating reasonable biological assumptions, use of the M-B would be decidedly nonconservative. He further questions the M-B device of removing the data successively, starting from the highest dose down and then selecting the highest resulting safe dose (except in special situations as indicated in Ref. 10). However, corresponding issues arise with meth ods that Crump and Guess advocate. The arbitrariness and poor fit to data for the M-B probit model are conceded properties of the M-B procedure. It was never claimed to provide a fit to the data; it was only claimed to provide an upper limit, under stated assump tions, to the true dose response at low dosage levels, and the plausible exception for 1-hit models was indicated. Crump cannot verify any better fit for the polynomial model he uses; any goodness of fit that he may observe reflects only that he has fitted so many parameters as to force the fit. If it is a model advocated by others, such advocacy has been based largely on mathematical rather than biological considerations. One justification of the polynomial model could even be that, whatever the true response curve might be, it can be mimicked as closely as desired by a polynomial of sufficiently high degree; however, use of a high-degree polynomial on such a basis would not consequently make valid the safe dose estimates resulting. For that matter, if the fitted polynomial were one in the square root of dose (a possibility raised in the following discussion) (6), the result ing safe levels would make those obtained with the polyno mial seem nonconservative by comparison. I would not so term them, however, since some way can be found for characterizing any safe level except zero as nonconservative. If the M-B procedure involves successive removal of data at higher doses, the polynomial approach of Crump and Guess does the same effectively but subtly. The M-B ap proach can be applied by substituting the 1-hit model for the probit model. If we did so apply it and found that, at higher doses, responses observed rose more rapidly than called for by the 1-hit model, the procedure of removing higher dose data would discount such observations. Since our interest is in what might happen at an extremely lowdose level, sharp increases in what happens on the right would be irrelevant, and only the lower dose observations would matter. However, if one specifically used a polyno mial model, sharp increases on the right would lead to fitted positive coefficients for the higher powers of dose. Yet, in going down to extremely low-dose levels, only the linear term in the polynomial would really matter, and the higher-degree positive coefficients would play virtually no role. Much the same safe levels would result if we were to drop deliberately the high response data at the higher doses or if we explain such data by positive high-degree coeffi cients that do not influence the safe dose estimate. (I note that, when Crump and Guess use polynomial fitting, they 1836 put certain limitations on the fitted coefficients to make for convexity. Thus, the fitted linear coefficient may not be negative, although with unrestricted fitting it might be. The statistical evaluation of such restricted fitting is not com pletely straightforward.) Whereas Crump has questioned the conservatism of the M-B probit model, Jerome Cornfield in private discussions with me has expressed concern that it is too conservative. Cornfield and Crump come to diametrically opposite posi tions, starting from similar multihit models. For a true high multihit model, the safe dose could be much higher than called for by the M-B probit model (except at fantastically low levels of risk). Even for a 2-hit model, the safe dose would be moderately higher. Cornfield would be willing to use whatever degree of hitness the data suggest in extrap olating to safe levels, foregoing any low or linear coeffi cients unless clearly indicated by the data. However, by Crump's approach, even if the linear coefficient is not clearly indicated in the data, that linear coefficient and other low-order coefficients might still be positive, and lim its on those coefficients, more precisely on the fitted curve, must be considered. The setting of limits in a way that allows for positive linear coefficients makes for the much lower safe levels by Crump's interpretation than by Corn field's interpretation of multihit models. (Crump's interpre tation is also questionable by the concept of interchangeability of carcinogens, to which I alluded previously.) To understand what is involved in the setting of limits, consider that we have performed an experiment involving several dosage levels, in which the response measured has the same known or unknown variability at all levels. The variability of the fitted curve and the fitted coefficients would depend only on the experimental design and could be relatively high for some designs. If the true linear coefficient were indeed zero or, if not zero, comparatively small, any statistical upper limit on that coefficient would then depend primarily on its statistical variability and thus primarily on the experimental design. Since, at doses close to zero, it is virtually only the linear coefficient that matters, the calculated safe dose would depend on the calculated upper limit on the linear coefficient and so also would depend only on the experimental design. It could even be predictable without experimentation if the response vari ance is known in advance. All of this, however, must be altered for the more real case in which there is a yes/no response to whether cancer is present or absent in a treated animal. For such binomial responses, the variability depends on the true response rate. However, it will be the outcome, among control animals and among animals receiving the smallest nonzero dosage, that will effectively determine the limiting value for the linear coefficient in a polynomial model. Conceptually, if we had several different response curves that agreed at the lowest nonzero dose level used and also agreed by definition at the zero dosage, all of these would tend to yield the same safe level by Crump's approach. Whether the fraction of the 1% dosage needed to get a 1/100,000,000 risk should be 1/3.16 for a pure although true 12-hit model, 1/10 for a true 6-hit model, 1/100 for a true 3-hit model, or 1/1,000 for a true 2-hit model, the Crump approach would yield the fraction 1/1,000,000 corresponding to the 1-hit CANCER RESEARCH VOL. 38 Downloaded from cancerres.aacrjournals.org on June 15, 2017. © 1978 American Association for Cancer Research. Letter fo the Editor model, which is the only one of these models relative to which the M-B probit model is not conservative. That Crump's safe levels invariably correspond to the 1-hit model can be verified in Fig. 2 of his response to Salsburg (2) and in various figures by Crump and Guess in other publica tions. Invariably, there is about a 45°line shown for the log-log relationship between added risk and dose for the safe dose curve, signifying simple proportionality between the 2. In one of his papers with Guess as senior author (6), Crump takes a position that would suggest incomplete understanding of the concepts. Since the 1-hit model would yield impractically low safe levels, as stated by Mantel and Schneiderman (15), Guess and Crump suggested that their use of a polynomial form model would be remedial. At very low doses the increased risk over background would be dependent essentially on the lowest power of dosage with a positive coefficient in the polynomial expression and, if that power is sufficiently large, higher safe levels would result. Such safe levels, they state, might be as low as the ones for the 1-hit model or as high as the ones for the probit model. (Actually, if the power is 2 or greater, the safe levels would be higher than the ones yielded by the M-B probit model.) It would seem at first that Guess and Crump are taking a position similar to that of Cornfield, which was indicated previously. Instead, in the illustrations given by Guess and Crump, all of the safe level curves end up of necessity as curves corresponding to the 1-hit model. Perhaps these authors did intend to convey that higher than 1-hit safe levels could result only if the lowest power of dosage with positive coefficient was stipulated in advance to take some particular value in excess of unity, but I cannot read this into their text. They do seem to make a promise of possibly coming up with such higher safe levels, yet they cannot do so once they admit that the linear coefficient may be positive and, if this is denied, they must concede practical conservatism for the M-B method. Also, if the concept of interchangeability of environmental carcinogens with the test carcinogen is held, a concept with which Guess and Crump, among others, are identified, these authors were not in a position to promise higher safe levels. Under the interchangeability concept, any safe level would be virtually independent of the mathematical model for dose response, as Richard Peto understands. A word on the approach of Hartley and Sielken (8) is in order here. Hartley and Sielken suggest the use of a polynomial model that is not unlike that of Guess and Crump, but they use a different conditioning when fitting to impose convexity. If statistical variation is taken into ac count, that too should reduce to getting safe levels corre sponding to the 1-hit model, so that the results should be essentially the same as those obtained by Guess and Crump Professor Mantel has raised a number of issues with regard to the estimation of risks from low doses of carcino gens. My point of view and that of various coauthors with regard to some of these issues has been stated elsewhere (2-7). Nevertheless, there are a few points made here by Mantel for which I feel further comments could possibly be useful. or those that would be obtained if the M-B approach were coupled with the 1-hit model. If Hartley and Sielken have derived different safe levels, it is because they have used a different and, to my mind, questionable way of taking statistical variation into account (11). As a matter of interest, Hartley discussed in advance, with myself and Cornfield, his plans for using a polynomial regression form. We predicted that the results would be essentially equivalent to those for a 1-hit model. Crump and Guess have effectively verified this prediction. REFERENCES 1. Cornfield. J. Carcinogenic Risk Assessment. Science, 798. 693-699, 1977. 2. Crump, K. S. Response to Open Query: Theoretical Problems in the Modified Mantel-Bryan Procedure. Biometrics, 33. 752-755, 1977. 3. Crump, K. S., Guess, H. A., and Deal, K. L. Confidence Intervals and Tests of Hypotheses concerning Dose Response Relations Inferred from Animal Carcinogenicity Data. Biometrics, 33. 437-451, 1977. 4. Crump, K. S., Hoel. D. G., Langley, C. H., and Peto, R. Fundamental Carcinogenic Processes and Their Implications for Low-Dose Risk As sessment. Cancer Res., 36. 2973-2979. 1976. 5. Guess, H. A., and Crump, K. S. Low Dose-Rate Extrapolation of Data from Animal Carcinogenicity Experiments —Analysisof a New Statistical Technique. Math. Biosci., 30: 15-36, 1976. 6. Guess, H. A., and Crump, K. S. Can We Use Animal Data to Estimate "Safe" Doses for Chemical Carcinogens? In: A. S. Whittemore (ed.), Environmental Health. Quantitative Methods, pp. 13-30. Philadelphia: Society for Industrial and Applied Mathematics, 1977. 7. Guess. H., Crump, K.. and Peto, R. Uncertainty Estimates for Low-DoseRate Extrapolations of Animal Carcinogenicity Data. Cancer Res.. 37. 3475-3483, 1977. 8. Hartley. H. O., and Sielken, R. L., Jr. Estimation of "Safe Doses" in Carcinogenic Experiments. Biometrics. 33. 1-30, 1977. 9. Mantel. N. The Concept of Threshold in Carcinogenesis. Clin. Pharmacol. Therap.,4. 104-109. 1963. 10. Mantel, N. Estimating Limiting Risk Levels from Orally Ingested DDT and Dieldrin Using an Updated Version of the Mantel-Bryan Procedure. Report Prepared for the Office of Toxic Substances, Environmental Protection Agency. April 9, 1974. Supplementary Report, December 27, 1974. Washington, D. C.: Environmental Protection Agency. 11. Mantel, N. Aspects of the Hartley-Sielker Approach for Setting "Safe Doses" of Carcinogens. In: H. H. Hiatt, J. D. Watson, and J. A. Winsten 12. 13. 14. 15. 16. (eds.). Origins of Human Cancer. Book C, Human Risk Assessment, pp 1397-1401. Cold Spring Harbor, N. Y.: Cold Spring Harbor Laboratory. 1977. Mantel, N. Response to Open Query: Theoretical Problems in the Modified Mantel-Bryan Procedure. Biometrics, 33: 755-757, 1977. Mantel, N.. Bohidar, N. R., Brown, C. C., Ciminera, J. L., and Tukey, J. W. An Improved Mantel-Bryan Procedure for "Safety" Testing of Carcin ogens. Cancer Res., 35: 865-872, 1975. Mantel, N., and Bryan, W. R. "Safety" Testing of Carcinogens. J. Nati. Cancer Inst.. 27: 455-470, 1961. Mantel, N., and Schneiderman, M. A. Estimating "Safe" Levels, a Hazardous Undertaking. Cancer Res., 35: 1379-1386, 1975. Salsburg, D. S. Open Query: Theoretical Problems in the Modified Mantel-Bryan Procedure. Biometrics, 33: 419-421, 1977. Nathan Mantel Biostatistics Center George Washington University 7979 Old Georgetown Road Bethesda, Maryland 20014 First, I would like to acknowledge the real pioneering contribution of Mantel through the introduction of the Mantel-Bryan procedure (8) in 1961. This procedure, pro posed in an age when threshold seeking was very much in vogue, broke new and significant ground by utilization of a model in which no dose could be considered completely without risk. JUNE 1978 Downloaded from cancerres.aacrjournals.org on June 15, 2017. © 1978 American Association for Cancer Research. 1837 Correspondence re: K. S. Crump, D. G. Hoel, C. H. Langley, and R. Peto. Fundamental Carcinogenic Processes and Their Implications for Low-Dose Risk Assessment. Cancer Res., 36: 2973−2979, 1976, and H. Guess, K. Crump, and R. Peto. Uncertainty Estimates for Low-Dose-Rate Extrapolations of Animal Carcinogenicity Data. Cancer Res., 37: 3475−3483, 1977. Nathan Mantel Cancer Res 1978;38:1835-1837. Updated version E-mail alerts Reprints and Subscriptions Permissions Access the most recent version of this article at: http://cancerres.aacrjournals.org/content/38/6/1835.citation Sign up to receive free email-alerts related to this article or journal. To order reprints of this article or to subscribe to the journal, contact the AACR Publications Department at [email protected]. To request permission to re-use all or part of this article, contact the AACR Publications Department at [email protected]. Downloaded from cancerres.aacrjournals.org on June 15, 2017. © 1978 American Association for Cancer Research.
© Copyright 2026 Paperzz