Is Risk Assessment Really Too Conservative?: Revising the Revisionists Adam M. Finkel* INTRODUCTION A growing chorus of voices is convinced that quantitative risk assessment (QRA) has evolved into a caricature of itself. Accord- ing to this view, quantitative estimates of human health risk from environmental pollution (particularly from carcinogenic sub- stances) have become so dependent on unreasonable worst-case assumptions as to be meaningless, alarmist, and counterproduc- tive. The articles in this symposium issue by Dr. Elizabeth Ander- son' and Dr. Dennis Paustenbach2 discuss many of the main arguments made by a number of health scientists,3 policy ana- lysts,4 and regulators.5 Dr. Finkel is a research fellow at the Center for Risk Management at Resources for the Future in Washington, D.C. He has an Sc.D. in environmental health sciences from the Harvard School of Public Health and a master's degree in public policy from the Kennedy School of Government. He has served as a technical advisor to the National Academy of Sciences, the House Science and Technology Committee, and an environmental research organization in Mexico City, and was editor-in-chief of a weekly newsletter on hazardous waste issues. The contributions of his colleagues Michael Gough, Paul Portney, and John Mankin are gratefully acknowledged. The opinions expressed herein are those of the au- thor and do not necessarily reflect the views of Resources for the Future or the Center for Risk Management. 1. Anderson, Scientific Developments in Risk Assessment: Legal Implications, this volume. 2. Paustenbach, Health Risk Assessments in Toxic Tort Litigation: Opportunities and Pitfalls, this volume. 3. Maxim, Problems Associated with the Use of Conservative Assumptions in Exposure and Risk Analysis, in THE RISK AsSESSMENT OF ENVIRONMENTAL HAZARDS: A TEXTBOOK OF CASE STUDIES 526-60 (D.J. Paustenbach ed. 1989); Sielken, Pitfalls of Needlessly Conservative Aspects of Cancer Risk Quantification and How to Avoid Them, (Jan. 15, 1989) (presented to the American Association for the Advancement of Science, San Francisco, Calif.). 4. Nichols & Zeckhauser, The Dangers of Caution: Conservatism in Assessment and the Misman- agement of Risk, in 4 ADVANCES IN APPLIED MICRo-EcoNoMIcs 55-82 (1986); Huber, Exorcists Versus Gatekeepers in Risk Regulation, REGULATION, November/December, 1983, at 23-32. 5. T. Yosie, Speech to the Society for Risk Analysis, Houston, Texas (Nov. 2 1987) (entitled Science and Sociology: The Persistence of Conservative Assumptions in Risk Assessment)(as reported in Society for Risk Analysis Risk Newsletter). 427 COLUMBIA JOURNAL OF ENVIRONMENTAL LAw [Vol. 14:427 The denunciation of QRA as an ideologically-motivated exer- cise in exaggeration has cropped up from time to time,6 begin- ning even before United States federal agencies started to codify their QRA procedures.7 Only in the past year or two, however, has a critical mass of scientific experts converged to the view that the time has come to reevaluate these procedures, and the first tangible results of this "revisionist" stance are even more recent.8 Linked temporally and perhaps causally to this growing momen- tum against "conservatism" in QRA is the view that science has at last begun to offer practical and defensible alternatives to what Anderson calls "overestimated theoretical risk," that is, the pro- duction of risk estimates via a process critics claim is rife with sci- entific shortcomings.9 The critics of conservatism are a diverse group. They share the belief that some or all of the inferences central to QRA are overly timid and at variance with new theories or data to the contrary. They differ widely, however, in the scope and intensity of their enthusiasm for specific alternatives, and in their degree of dissat- isfaction with the status quo. In order to respond both to Ander- son and Paustenbach and to the growing number of published and unpublished critiques of conservatism, I have coined the catchall term "revisionist." I emphasize that this is a stereotype which describes none of these critics precisely, but which is a composite of documented views rather than a "straw man." My hypothetical revisionist believes that at virtually all phases of QRA, the preoccupation with the "worst case" has driven out scientific rules of reason, and tends to believe that all new theo- ries or data sets, if valid, will likely reduce existing estimates of biologic potency or human exposure. Not everyone who advo- cates the rollback of particular conservative assumptions is a revi- sionist, and probably no individual critic of QRA could fall prey to all of the misperceptions I mention in this article. Neverthe- less, this stereotype is not so limited that it refers only to some- 6. E. EFRON, THE APOCALYPTICS: CANCER AND THE BiG LIE (1984). 7. E.P.A. Guidelines for Carcinogen Risk Assessment, 51 Fed. Reg. 33,992-34,003 (1986). 8. Shabecoff, E.P.A. Reassesses the Cancer Risks of Many Chemicals: Hazards Seen as Lower, N.Y. Times, Jan. 4, 1988, at Al; EPA, Office of Health and Environmental Assessment, Update to the Health Assessment Document and Addendum for Dichloromethane (Methy- lene Chloride): Pharmacokinetics, Mechanism of Action, and Epidemiology (1987) (Exter- nal Review Draft EPA/600/8-87/030A). 9. Paustenbach, supra note 2, at Section I. 428 Revising the Revisionists one who would prefer never to assess or control risks in the absence of unequivocal evidence of harm. Finally, from the per- spective of an EPA official charged with responding to calls for QRA revision, all of the revisionist arguments must be equally dealt with, even though they come from individuals with diverse perspectives, backgrounds, and motives. While I advocate certain changes in QRA, I do not share the revisionist disdain for risk assessment as currently practiced. I am not so confident there exists systematic conservatism in QRA, that if so it would be a scourge we must repudiate, or that the new science necessarily provides a more appropriate or even a more credible alternative.'0 I reject as disingenuous the characteriza- tion that "[conservatism] allows for policy choices to masquerade as if they were scientific facts."" Revisionists also tend to selec- tively marshall and interpret facts to support or supplant policy choices. Moreover, I question the pejorative description of con- servatism as persisting solely because bureaucrats appreciate its rigidity or because they live in fear of the "accusation that their [risk] assessments are not sufficiently health protective."' 2 While I agree with many of the revisionists that ideally, QRA would not subsume covert value judgments that ought to be part of the overt balancing process that is risk management, I believe there is no workable alternative to the current mingling of these related activities. I welcome efforts to make the assumptions and uncer- tainties in QRA more explicit, so that as a society we can discern how conservative regulatory proposals actually are. I do not be- lieve, however, that merely scaling back some or all of these as- sumptions will achieve a value-free process with the desired separation of risk assessment and risk management. In short, I believe that some of the critics of conservatism have fallen prey to some or all of a series of misperceptions about the science and craft of QRA. These most often result from insuffi- cient attention to the twin influences of scientific uncertainty and human interindividual variability, so the lion's share of the argu- ments raised in this paper will concern these two under-appreci- ated phenomena. Indeed, I will conclude in part that when risk assessment and management are. reevaluated in light of uncer- tainty and variability, the narrowness of the entire debate over 10. Id. 11. Sielken, supra note 3. 12. Paustenbach, supra note 2, at Section III. 1989] 429 COLUMBIA JOURNAL OF ENVIRONMENTAL LAW [Vol. 14:427 whether the risk numbers we generate are too high or too low will be revealed. At the outset, I emphasize that I do not share the belief, which Paustenbach attributes to Commoner,'3 that QRA is a method for legitimizing the discharge of pollutants in quantities preferred by industry. Like Paustenbach, I look forward to a time when QRA will be even more influential in helping regulators and the public discriminate between significant and de minimis risks. Moreover, I share his apparent view that blanket prejudices against the use of QRA threaten to make us all less safe, and may lead us to squan- der the finite resources available for risk reduction. Nor do I dis- pute that we need to guard against the reflexive and exclusive use of worst-case assumptions. In fact, neither Anderson nor Paus- tenbach mentions two very potent arguments against conserva- tism that I think the scientific and regulatory communities must address. First, we need to recognize an inescapable paradox about con- servatism-in its attempt to impose uniformity on the mechanics of risk assessment, it virtually guarantees marked non-uniformity in the outputs of risk assessment. Some worst cases are simply "worse" than others, in the sense of being less plausible to occur (or less frequent in occurrence), but conservatism tends to ob- scure these differences. For example, even a seemingly innocu- ous contrivance such as "assume that at all hazardous waste sites, the maximally exposed individual (MEI) is exposed to the concen- tration measured directly downgradient at the property boundary of the facility" introduces what Nichols and Zeckhauser term "asymmetric conservatism." 4 In some cases, the resulting risk estimate will be quite conserva- tive if the geography is such that the MEI lives hundreds of yards away from the boundary and/or is not directly downgradient of the pollutant source. In other cases, however, the assumption may be nearly correct and thus barely conservative. In statistical terms, regulators might then be forced to compare a pair of risk estimates, perhaps one that had only a 1 in 1000 chance of being too low and another that had a 1 in 10 chance of being too low, yet this information about the different character of the estimates would be unavailable. Such a comparison might well result in the 13. Id. 14. Nichols & Zeckhauser, supra note 4, at 69. 430 Revising the Revisionists allocation of greater resources to address the relatively less im- portant risk, at the expense of the more significant one.'5 Second, certainaspects of conservatism can provide a powerful disincentive for needed research. For instance, if one or more animal carcinogenicity bioassays already exists for a substance, and if the relevant regulatory agency is on record stating it will use only the assay yielding the highest estimate of the cancer po- tency factor, essentially all incentive for an industrial concern to conduct or fund additional animal tests is removed by virtue of this "stacked deck." I have no doubt that as QRA matures, we will discover that con- servatism has caused some of our existing assessments to enor- mously exaggerate particular human risks. Some fragmentary evidence has already accrued in this regard. For example, the dis- covery of a protein (alpha2u-globulin), arguably unique to male rats and essential for carcinogenesis to occur in the male rat kid- ney following certain stimuli,'6 suggests that our estimates of the human carcinogenic potency of certain substances (perhaps un- leaded gasoline) were qualitatively in error (the true human can- cer risks may in fact be zero). In these cases, the discarding of all other test results and the preoccupation with the single positive response may have been the undesirable outcome of a conserva- tive stance. TABLE I NINE PERVASIVE MISPERCEPTIONS ABoUT RISK ASSESSMENT "CONSERVATISM" 1) There is no such thing as an "actual risk." 2) Conservatism. is inherently no more or less biased a method than alternative approaches. 3) Only some conservative assumptions are gratuitous. 4) Not all of the inferences we make are in fact conservative. 5) A cascade of truly conservative steps may still yield a rea- sonable estimate of risk. 15. In statistical terms, the first estimate would lie at the 99.99th percentile of an uncer- tainty distribution about the (unknown) true population risk, and the other would lie at the 90th percentile of its distribution. See Section X infra, however, for a discussion of how eliminating conservatism would not necessarily ameliorate this ranking problem. 16. Charbonneau, Short, Lock & Swenberg, Mechanism of Petroleum-Induced Sex-Specific Protein Droplet Nephropathy and Renal Cell Prohferation in Fischer-344 Rats: Relevance to Humans, in 21 TRACE SUBSTANCES IN ENVIRONMENTAL HEALTH 263-73 (1987). 1989) 431 COLUMBIA JOURNAL OF ENVIRONMENTAL LAw [Vol. 14:427 6) Conservative assumptions at some stages of risk quantifica- tion may correct for the omission of other stages altogether. 7) More science does not always mean less risk, nor does it always "reduce uncertainty." 8) Costs, like risks, can be biased or de minimis. 9) There is a major difference between the pooling of data and the averaging of irreconcilable theories or results; the latter is a perilous process laden with hidden value judgments. Nevertheless, these objections to current QRA procedures or results pale by comparison to my skepticism that all the revisionist positions are fully thought through. The remainder of this article constitutes a cautionary note warning against hasty or piecemeal revision of existing risk assessment procedures. A closer look at the following nine common misperceptions about QRA and con- servatism (see Table I) may cast doubt on Paustenbach's conten- tion that "something went wrong"' 7 with risk assessment, and suggest that much of the revisionist "wish list" for the future is objectively incomplete, potentially imprudent, and at best no less problematic than the status quo. I. THERE Is No SUCH THING AS AN "ACTUAL RISK" Some revisionists suggest that all of the problems with contem- porary risk assessment stem from the "unreality" of its outputs. They claim that instead, scientists ought to provide risk managers with estimates of "actual risk," or what Anderson calls "real risk."' 8 In this view, now that science has begun to develop tech- niques for making risk estimates more accurate, these should re- place conservative estimates, as surely as the Copernican model of the solar system replaced the Ptolemaic model. As Section II of this paper will suggest, the desire to promote decisions based on actual estimates of risk may not be free of subtle and poten- tially troubling value judgments. But more fundamentally, there exists widespread confusion about the concept of real risk itself, which is, in many ways, nonsensical. All risks are probabilistic summaries of unknowable future events-they can describe the long-run or average behavior of simple systems, but the behavior of subsystems or of individual 17. Paustenbach, supra note 2, at Section II. 18. Anderson, supra note 1, at Section 1. 432 Revising the Revisionists members of the system (as well as the short-run behavior of the whole) may diverge markedly from any prediction based on a risk estimate. Consider a simple homogeneous system obeying well- understood but stochastic physical laws-a chunk of the radioac- tive isotope uranium-238. Science may determine virtually ex- actly that the half-life of the isotope is 4.51 billion years, so that in that time period, almost exactly half of the atoms in an initially pure sample would have decayed into lighter isotopes. It makes no sense, however, to try to predict when a particular atom in the sample will decay, or whether it will still be a U238 atom at a partic- ular time in the future, or even to bank on a precise estimate of how many decay events will occur in a short time period and/or among a small subpopulation of atoms. Yet these are some of the kinds of inferences an actual risk would have to allow us to make in order to be qualitatively better than the "unreal" estimates cur- rent QRA provides. Of course, in moving from this idealized example to the practi- cal area of human health risk assessment, the system becomes even more complicated along several dimensions, further defeat- ing the goal of reaching actual risk. First, there are considerable difficulties even in estimating the true long-run average probability in real situations. The uncertainty in the half-life is primarily due to simple measurement error, which we can reduce further by improving our analytical devices.'9 Uncertainties in exposure estimation and carcinogenic potency assessment take many different forms and can be quite recalcitrant to brute force methods of data acquisition. Often, obtaining enough exposure or potency observations to make sampling error manageable is impossible or quite expensive, and then a host of non-random parameter uncertainties (e.g., the "healthy worker effect" in occu- pational studies) and fundamental modeling uncertainties (e.g., are Gaussian dispersion or linear dose-response models appropri- ate?) further confound efforts to converge on a single probability or risk number.20 Second, the physicist can rely on every atom of U2 ` being alike-they will not all decay at the same time, but they all face the same underlying chance of decaying at any moment in time. 19. Presumably, there is virtually no binomial or sampling error in the calculation be- cause the number of observations (decaying atoms) is so large. 20. Finkel, Perspectives on Uncertainty in Risk Assessment: A Guide for Decision-Mak- ers (1989) (Center for Risk Management Report), at 9-39. 1989] 433 COLUMBIA JOURNAL OF ENVIRONMENTAL LAw [Vol. 14:427 Human beings, however, exhibit enormous interindividual varia- bility, both in the exposures they receive (due to geographic, lifestyle, and genetic factors)2 1 and in their susceptibilities to each exposure (due to other genetic, temporal, and lifestyle factors).22 So, even if an individual could know his exposure pattern pre- cisely, a potency estimate that may by chance accurately describe his probability of death or disease at a given moment will cer- tainly fail to yield a real risk for some other person, or even for the original person as his age or environment changes. Finally, for a number to be an actual risk, one must believe that, both for the regulator and the exposed public, the concept of risk can be reduced to an unalloyed numerical value. This conclusion is tempting in light of the classical theory of individual utility, which holds that while people can be risk-averse with respect to an outcome (e.g., decline a chance to flip a coin to win $100 or lose $50, even though the expected value of the wager is posi- tive), rational thought precludes any special weighting of the probabilities of an outcome.23 In other words, risk numbers only modify the ultimate outcome (disease or death), and therefore one ought to view a risk of 10-3 as exactly ten times as bad as one of 10-'. Furthermore, one should be indifferent between the fol- lowing two descriptions of risk, whose expected values are equal: 1) one's risk is determined to be 10-, and this value is known with certainty; and 2) there is a fifty-fifty chance one's true risk is either 2 X 10-5 or zero. This view has recently been challenged by researchers con- cerned with human cognition and perception,24 who reject on both empirical and theoretical grounds the "linearity assump- tion" 25 that allows risks to be averaged and otherwise manipu- lated as if they had no special connotative meanings. Therefore, in the general case when the value of a probability is uncertain, 21. Ott, Total Human Exposure-An Emerging Science Focuses on Humans as Receptors of Envi- ronmental Pollution, 19 ENVTL. Sci. & TECH. 880-86 (1985); Wallace, The Influence of Personal Activities on Exposure to Volatile Organic Compounds, ENVTL. RESEARCH (submitted 1989). 22. Finkel, Estimating the Extent of Human Variability in Susceptibility to Carcinogenesis, RISK ANALYSIS (1988) (in press). 23. H. RAIFFA, DECISION ANALYSIS: INTRODUCTORY LECTURES ON CHOICES UNDER UN- CERTAINTY 57-61 (1968). 24. Tversky & Kahnernan, The Framing of Decisions and the Psychology of Choice, 211 ScI- ENCE 453 (1981); Lopes, Some Thoughts on the Psychological Concept of Risk, in J. EXPERIMEN- TAL PSYCHOLOGY: HUMAN PERCEFTION AND PERFORMANCE 137-44 (1983). 25. K. SHRADER-FRECHETTE. RISK ANALYSIS AND SCIENTIFIC METHOD 157-95 (1985). 434 1989] Revising the Revisionists 435 calling any single numerical summary of that uncertainty an actual risk may ignore human values and perceptions; even if that probability were somehow known with certainty, it may be simi- larly unresponsive to compare it with another actual risk, or to convert the risk to a measure of social cost, without considering factors other than the probability itself. None of this discussion is intended to deny that certain incre- mental changes in QRA procedures may be motivated by dispari- ties between existing predictions and observational experience, or to deny that they may move particular risk estimates closer to values we deem more likely to be borne out in the future. Propo- nents of change "oversell" these adjustments, however, if they equate "more likely to be borne out" with "real." First, unlike Copernican celestial mechanics, we will never actually be able to test the accuracy of alternative risk estimates against real human experience, although this alone does not militate against the quest for better numbers. Second, the risk assessment process is sufficiently complex that we cannot be confident that adding a dose of reality at one stage will necessarily improve the reality content of the overall output. Third, uncertainty is sufficiently pervasive that no new theory is likely to yield an estimate that automatically supplants the standard one; such improvements provide more information, not definitive information. Again, the claim that new approaches "provide the regulatory or legal arena with an indication of the extent to which the plausible upper bound may be overestimating risk for particular chemicals,"2 6 which may be tantamount to saying they yield real estimates, overstates the case for revision. Ideally, conservative estimates have the virtue of not being advertised and utilized with reference to their reality content. In practice, although EPA often fails to communicate that its risk estimates. are construed to be upper bounds, on occasion it overstates the case and implies there is virtually zero possibility that risks could be higher than stated. II. CONSERVATISM IS INHERENTLY NO MORE OR LESS BIASED A METHOD THAN ALTERNATIVE APPROACHES Critics of current QRA procedures often equate the various conservative assumptions with the statistical term "bias," a word which in everyday parlance connotes intentional disregard for 26. Anderson, supra note 1, at Section II(B). COLUMBIA JOURNAL OF ENVIRONMENTAL LAW [Vol. 14:427 facts or conventions of fairness. The implication of descriptions such as "assumptions that bias the estimates upward"2 7 is that lowering the risk estimates would reduce or remove the bias. Again, just as with real and unreal, some revisionists prefer to frame the debate in semantic terms that contrast laudable versus suspicious motives. The statistical concept of bias, however, refers to an estimator (i.e., a procedure for producing an estimate) which on average does not equal the parameter it is supposed to estimate.28 This descriptive notion has nothing to do with evaluating the choice of the estimator itself on prescriptive grounds. Using a specific sta- tistical upper confidence limit (UCL) to describe the results of an animal carcinogenicity bioassay, or basing population exposure estimates on the parameter values applicable to a highly exposed individual, can be criticized as a value-laden estimator. But it is not necessarily a biased procedure, unless the critic can show that the analyst's number is systematically different (higher) than the quantity he believes he is estimating (e.g., the desired UCL on po- tency or the exposure the individual faces). I have no doubt such true bias occasionally does occur, when analysts or regulators treat numbers meant to be particular UCLs as if they were central or real estimates, and then explicitly or tacitly hedge further to- wards overcautiousness by inflating these numbers further. Such behavior is a problem of education, not an acknowledgement that the estimators themselves are biased. Besides, true statistical bi- ases are probably more likely to occur with central estimators of risk. It is often easier to estimate, to a desired amount of preci- sion, more extreme values than central ones; the average speed of a sample of 10,000 cars on an interstate highway may be difficult to predict (and highly variable depending on conditions), but one might predict rather confidently that the 500th-fastest car travels at between 72 and 76 mph. But what of the description of conservatism as "value-laden," a term that can turn pejorative when critics allege these values are covert and are "masquerading as facts"? There is no sense deny- ing the judgmental content of conservatism, but many critics seem not to accept that all summary estimators of an uncertain quantity are value-laden. Summary measures are little more than 27. Nichols & Zeckhauser, supra note 4. at 57. 28. J. FREUND, MATHEMATICAL STATISTIcs 256 (1971). 436 Revising the Revisionists ways to interpret facts in light of a subjective calculus of the costs of error--differences among values yield differences in the inter- pretive calculus, but all values can masquerade as (or preferably, enrich) the facts to the same degree. For example, choosing to describe an uncertain quantity by the 95th percentile of its probability distribution merely reflects the conscious or tacit eval- uation that an error of underestimation (the five percent chance the "truth" exceeds the summary value) is nineteen times as bad as an error of overestimation. Most other summary measures simply strike this balance be- tween probabilities and social costs in a different way. For exam- ple, the lognormal distribution is one of the most common mathematical formulations of uncertainty used in risk analysis;29 it can be summarized as a best estimate with a "factor of X" un- certainty surrounding it in both directions.30 One could use this best estimate (which is the statistical median) to describe the risk, and this would embody the value judgment that the costs of the two types of errors are exactly equivalent (as the probability of each error is fifty percent when the median is chosen). Another common estimator is the mode of the uncertainty distribution, the single value deemed more likely to occur than any other. The maximum likelihood estimator (MLE) advocated by several revi- sionists3' is, in certain contexts such as sampling error in animal bioassay data, the mode of the relevant uncertainty distribution. The mode reflects a different value judgment-that one should minimize the probability of an error, without regard to its type (over- or underestimation) or its magnitude. Indeed, none of these probability-based estimators consider how large the errors might be, except indirectly. As Section V will demonstrate, the mean of an uncertainty distribution is fre- quently much larger than the median, which in turn is generally larger than the mode. The mean may even be comparable in magnitude to the 95th percentile (or more extreme) "upper 29. Crouch & Wilson, Regulation of Carcinogens, I RISK ANALYSIS 48 (1981); Finkel, supra note 22. 30. If, as is commonly assumed, the "factor of X" is intended to demarcate a 97 percent confidence region (such that there is a 97 percent chance the true value will lie between the best estimate (B) divided by X and B times X), then there is a 68 percent chance the true value will lie between B/X" 2 and BX'n. 31. Sielken, supra note 3; Nichols & Zeckhauser, supra note 4. 1989] 437 COLUMBIA JOURNAL OF ENVIRONMENTAL L.Aw [Vol. 14:427 bounds."3 2 The mean is the sensible estimator if one believes that the social cost of an error increases with the magnitude of that error (e.g., in the classic "newsboy's problem,"33 if the true demand is for 100 papers, ordering 10 or 190 papers from the supplier is a more costly mistake than ordering 99 or 101). Es- timators even larger than the mean may be more appropriate if in addition to caring about the size of errors, there is also a per- ceived asymmetry in social cost (e.g., an excess newspaper is not as bad as an unsatisfied customer, or an error of overestimating risk is not as bad as an equally large error of underestimation). Therefore, to the extent that some of the central estimates put forth as alternatives to conservatism are not as large as the true mean, and to the extent that even the mean may not capture asymmetries in social cost, these estimators are not only as value- laden as conservative ones-they may be laden with value judg- ments at odds with those of society as a whole. III. ONLY SOME CONSERVATIVE AssumprIoNs ARE GRATUITOUS Certain results of a conservative stance are easy targets for revi- sionist observers. In cases such as the one Paustenbach de- scribes34 about the dioxin hazard from municipal waste incinerators, a strong case can be made that some risk assess- ments strain the bounds of credulity. As a practitioner and pro- ponent of QRA, I too am concerned that Congress and the public not lose faith in the process because they construe QRA as an exercise in constructing bizarre hypotheticals, like the child who seems to be getting most of his nourishment from incinerator fall- out. Nearly as damaging to the reputation of QRA are instances such as the one Anderson points out,3 5 where seemingly obvious questions (how much dioxin is irreversibly bound to soil and thus not bioavailable upon ingestion?) are not asked, and the assessor instead chooses to assume the worst. 32. For example, if the median or "best estimate" is 10.0 in a lognormal distribution with a rather modest "factor of 25" uncertainty, the mean is approximately 36.5, the mode approximately 0.75, and the 95th percentile value approximately 141.0. 33. M. MORGAN & M. HENRION, UNCERTAINTY: A GUIDE FOR DEALING WrH UNCERTAINTY IN QUANTITATIVE RisK AND POLICY ANALYSIS 384-86 (1988)(preprint). 34. Paustenbach, supra note 2, at Section IV(D). 35. Anderson, supra note 1, at Section II(C). 438 1989] Revising the Revisionists 439 Such reflexive use of pessimistic assumptions can be termed gratuitous, in that it suggests at least the appearance of laziness on the part of the assessor. To a varying degree, a bit more intro- spection or research would presumably reveal that such assump- tions are simply short-cuts. Moreover, in many cases the assessor would see that if he continued to gather more data, the original conservative estimate would become less and less plausible, in an orderly and predictable fashion. It would be a serious mistake, however, to malign all of the al- legedly conservative aspects of QRA with the same broad brush. In fact, many of the basic assumptions of QRA are conservative both out of respect for how little we know and out of a recogni- tion that as we learn more, the outputs of these procedures may not converge towards lower results. Perhaps the best example is the use of the statistical 95th percentile UCL, rather than the. MLE, in the analysis of rodent bioassay data. Both the MLE and the UCL are converted into estimators of the carcinogenic po- tency of the substance; to a very rough approximation, the MLE estimator of potency is the slope of the straight line that gives the "best fit" to the rodent data.36 Again very roughly, the UCL esti- mator of potency is the slope of a steeper straight line fitting an alternative set of data points, wherein we hypothesize that if we had repeated the bioassay, the same underlying risks to the ani- mals might have yielded a more pronounced tumor response. If we observe two tumors in fifty rodents, the best estimate is that each rodent had a cancer risk of 0.04 (2 + 50) at that dose. How- ever, we know that even if the true risk was 0.04, if we could re- peat the bioassay 100 times, on about five occasions we would observe five or more tumors in this group. If we only had one of those bioassays, we would have concluded that the best estimate of risk was at least 0.1 (5 - 50). In other words, the true risk to each animal could be 0.1 or higher; the chance observation of only two tumors out of fifty is not at all inconsistent with this estimate. It is well known that depending on the arrangement of the doses and the observed responses, the UCL slope may be many 36. The bioassay data are as follows: at each dose, located on the horizontal axis, the response (on the vertical axis) is the fraction of animals tested which developed tumors in a particular tissue. COLUMBIA JOURNAL OF ENVIRONMENTAL LAw [Vol. 14:427 times greater than the MLE slope.3 7 But this does not mean that the UCL is an outrageous value waiting to be refuted by more data. It is nothing more or less than the lower bound on the exact value we believe we would call the best estimate on five occasions if we repeated the animal experiment 100 more times. Consider this analogy to highlight the distinction between gratuitousness and prudence. If a baseball player approached his team's owner after the first two games of the season and asked for a million- dollar raise because he was batting .800 at the time, the owner would probably think it prudent to wait for the player to amass 100 or more at-bats before caving in. This is because at the time, the owner might believe there was a reasonable chance the player's true average might end up being as low as .200. We are in the same situation with potential carcinogens, except that bioassays are so expensive and time-consuming that we will never get a "full season's" worth of data on a chemical. While I agree with Paustenbach that the "degree of potential conserva- tism of the bounding procedure" (that is, the ratio of the UCL to MLE estimators) should be reported in risk characterizations, it is not generally true that "zero risk is as likely as the upper bound value of risk." 38 In many cases, the MLE slope is greater than zero, indicating that we would have to assume the observed data were by chance an unusually strong result of a weak underlying risk (akin to the baseball owner assuming the .800 hitter was actu- ally on a "cold streak" at the time!) cto believe that zero potency was a plausible conclusion. Besides, Paustenbach fails to make the distinction between the slope of a linear dose-response func- tion and the potency. Even in cases where an estimator of the linear slope is zero (and the UCL on the linear slope is always positive),39 this does not mean that the chemical is not a carcino- gen, only that its risk decreases more rapidly than linearly at lower doses.40 37. T. Thorslund, Estimation of Lifetime Risks Using a Multistage Theory of Carcino- genesis (Feb. 20, 1985)(EPA Carcinogen Assessment Group typescript). 38. Paustenbach, supra note 2, at Section IV(B). 39. Guess, Crump & Peto, Uncertainty Estimates for Low-Dose-Rate Extrapolation of Animal Carcinogenicity Data, 37 CANCER RESEARCH 3475-83 (1977). 40. This oversight is encouraged by EPA's current reporting procedures, which some- times treat "potency" as if it refers only to the linear part of the function. If the MLE of the linear part is zero, the MLE for "potency" (really the risk at an arbitrary low dose) is not zero, but comes from the dose squared times the MLE for the "dose-squared term" of the function. 440 Revising the Revisionists In addition to placing too much confidence in a limited set of data which may not be borne out by further experience, the MLE has the added disadvantage of being extremely fragile to small fluctuations in those data. Paustenbach's example of a one-tumor difference between otherwise identical bioassays conducted in a "New York" and a "San Francisco" laboratory"l actually demon- strates why EPA and other agencies shun the MLE in favor of the UCL. The exact values for the added risk at a dose of 0.01 mg/kg-day, using the MLEs,42 are 5.704 X 10-7 in "New York" (or about one chance in 1.7 million) and 9.88 X 10' in "San Francisco" (almost exactly one chance in 10,000). These risks do indeed differ by a factor of 174. However, the UCL on added risk is 2.42 X 10' in "New York" and 3.02 X 10- in "San Francisco," a trivial difference of only twenty-five percent. If anything, this example underscores the folly of using a best estimate that might so dramatically underestimate true risk due to a capricious event (perhaps the rod'ents in the middle dose group in "New York" were just lucky, or perhaps the pathologist who examined these animals failed to correctly diagnose one (or more) that had tu- mors). While the UCL may sometimes obscure real differences among chemicals,4 3 the MLE is really the more gratuitous estima- tor, for it ignores how much the risk estimate might differ if more data (or more thorough analysis. of existing data) were available.4 4 IV. NOT ALL OF THE INFERENCES WE MAKE ARE IN FACT CONSERVATIVE In addition to criticizing all of the truly conservative assump- tions regardless of their underlying rationales, the revisionist tends to portray all inferences used in QRA as conservative, re- gardless of whether this description is numerically apt. State- ments such as "wherever there was scientific uncertainty, the 41. Paustenbach, supra note 2, at Section IV(B). 42. MLE and UCL slopes were calculated via the computer code "MSTAGE87," pro- vided courtesy of E. Crouch. 43. H. Ozkaynak & A. Finkel, Potencies and Unit Risk Values for Suspected Human Carcinogens as Input to Health Risk Assessment (Dec. 18, 1986)(report to Energy and Environmental Systems Division, Argonne National Laboratory). 44. A. Finkel, Computing Uncertainty in Carcinogenic Potency: A "Bootstrap" Ap- proach Incorporating Bayesian Prior Information (Aug. 8, 1988)(report to EPA Office of Policy, Planning and Evaluation). This report suggests alternative ways of characterizing "potency" that do not rely on point estimators such as the MLE or UCL. 1989] 441 COLUMBIA JOURNAL OF ENVIRONMENTAL LAw [Vol. 14:427 most ... conservative assumption was always chosen"4 5 contrib- ute to an exaggerated picture of the risk assessment process as a rote series of timid choices. In some cases, neither the regulator nor the critic knows whether a particular assumption is conserva- tive, yet both now accept the premise that the choice is in fact a biased one. For example, both EPA and its critics now describe EPA's choice of the animal data set from which it calculates its cancer potency factors as a routine of "selecting the most sensi- tive sex/species combination" available. This statement is objec- tively true (with the caveat that this sex/strain combination must show a statistically significant tumor increase), yet it conveys too high a degree of confidence that this choice leads to conservative human risk estimates. If EPA generally had access to data on doz- ens of animal species, and then chose the single sex/species com- bination yielding the largest potency factor, one might justifiably compare this to the "multiple comparisons" fallacy in epidemiol- ogy (wherein data are gathered on so many health effects, without regard to theories of causal association, that by chance at least one effect is likely to show a "significant," but spurious, elevation).46 But the connotation of picking the "one bad apple from the barrel" is certainly less justified when one recalls that except in rare instances, EPA has but four sex/species combinations to pore over (males and females, rats and mice). Just because the female mouse may be the most sensitive test animal available does not necessarily imply that it is more sensitive than the human. Paustenbach's accurate statement that current potency estimates are often "more applicable to a rat than a human"47 is in fact an indictment of the process as nonconservative if in general or in particular cases rats are actually less sensitive than humans. In several other instances, actual data exist to bolster the argu- ment that particular assumptions are not always conservative. Very few scientists or policy analysts are calling for revision of these procedures to increase their degree of conservatism, yet in at least three major portions of the process, such revision might be justified, if indeed we always wanted to guarantee the most conservative plausible stance: 45. Anderson, supra note 1. at Section 1. 46. Feinstein, Scientific Standards in Epidemiologic Studies of the Menace of Daily Life, 242 Sci- IF ENCE 1253 (1988). 47. Paustenbach, supra note 2, at Section IV(D). 442 Revising the Revisionists * Use of Linear Dose-Response Models. This inference has been characterized almost universally as the "most conservative plausi- ble model,"48 yet- until recently no attempts had been made to validate this view. Bailar and his colleagues,49 risk analysts from the Harvard School of Public Health, recently showed that in a significant number of cases, superlinear functions that are steeper at low doses than at higher ones fit observed animal data better than linear functions do. One familiar example where super- linearity has been verified via extensive testing in the low-dose region concerns the human liver carcinogen vinyl chloride (VC). Bailar noted that linear extrapolation using only the control group and the two "high-dose" groups tested in the original bioassay (inhalation of 2500 and 6000 ppm VC) produced a po- tency estimate nine times lower than that obtained by linear ex- trapolation using eight "low-dose" groups (between 0 and 200 ppm inhalation exposure). In this case, the "plateau" of response at higher doses is known to result from the saturation of an en- zyme system that converts VC to a potent carcinogenic intermedi- ate. Bailar discusses five other biological reasons why the true dose-response relation for a given chemical may "plateau," mak- ing it more likely that the EPA procedure yields nonconservative estimates of low-dose potency. * Use of Standard Air or Water Dispersion Models. Although the exposure scenarios risk analysts employ are often conservative (e.g., they assume that all persons drink two liters of tap water daily50), the dispersion models used to predict the pollutant con- centrations in the water or air are often nonconservative. For ex- ample, the various short- and long-term air dispersion models were developed for use in simple terrain, without hills, valleys, or buildings in between the pollutant source and the human "recep- tor." Such complexities of terrain will cause the models to over- predict exposures in some situations, but underpredict them in others, particularly near the source where concentrations are highest. Also, meteorological conditions not accounted for in ex- isting models, notably the common phenomenon of "fumigation" (the rapid shift from "stable" to "unstable" stratification of the 48. Lave, Estimating the Risk of Carcinogens, I RISK ANALYSis 60 (1981). 49. Bailar, Crouch, Shaikh & Spiegelman, One-Hit Models of Carcinogenesis: Conservative or Not?, 8 RISK ANALYSIs 485-97 (1988). 50. Maxim, supra note 3, at 542. 1989] 443 COLUMBIA JOURNAL OF ENVIRONMENTAL LAW [Vol. 14:427 atmosphere)5 ' can cause transient but substantial elevations in pollutant concentration. Moreover, as Anderson acknowledges, the "source term" (the amount of pollutant emitted or created) in these models may be underestimated. She cites the case of chemical conversion of one pollutant to a more toxic one,52 a process which can also involve biochemical transformation, as in the generation by microbes of highly toxic methylmercury from emissions of inorganic mer- cury.53 Both Anderson and Paustenbach cite examples where the source term is decreasing with time, to show that simply multiply- ing the current emission rate by the time horizon is too conserva- tive. Counterexamples exist, however, where by virtue of chemical transformation or increased contributions to the source term (as in groundwater contamination from a hazardous waste site where disposal is ongoing) such simple extrapolation is equally nonconservative. Finally, contrary to Anderson's refer- ence to the Tacoma smelter (ASARCO) case,54 more intensive data collection on emission rates does not necessarily reveal sys- tematic conservatism. In that case, EPA lowered (by roughly four-fold) only its source term estimate for the "process emis- sions" of arsenic from the main stacks; simultaneously, it discov- ered that the source term for the "fugitive emissions" from valves, leaks, etc., was underestimated by about a factor of two (since these latter emissions occur much closer to ground level, the net effect on human risk of these adjustments tended towards a balance).55 The point is not to quibble with the specific exam- ple, but merely to show that it is unfair to automatically equate more data with lower exposure estimates. * Assumption that Multiple Risks are "Additive" in Nature. We know that in some cases, exposures to two different substances can have a synergistic effect in elevating excess cancer risk. For example, the cancer rate among asbestos workers with a history of cigarette smoking was found to be more than three times higher 51. F. PASQUILL & F. SMITH, ATMOSPHERIc DIFFusIoN 282 (3d ed. 1983). 52. Anderson, supra note 1, at Section II(E). 53. The Impact of Mercury Releases at the Oak Ridge Complex Hearing Before the Subcomm. on Investigations and Oversight of the House Comm. on Science and Technology, 98th Cong., Ist Sess. 211-18 (1983). 54. Anderson, supra note 1, at Section I(E). 55. Finkel & Evans, Towards Cost-Effective Methods for Reducing Uncertainty in Envi- ronmental Health Decision Processes (1985)(proceedings of the Annual Meeting of the Society for Risk Analysis at 543-553.). 444 Revising the Revisionists i i . 1 C; C~C)6 C~J k~ 0 0 -La) AKJpquqoid ;)AIIEI;)-J o 0 0 0 0 o o - : 0 on -0 0