Below are the first 10 and last 10 pages of uncorrected machine-read text (when available) of this chapter, followed by the top 30 algorithmically extracted key phrases from the chapter as a whole.
Intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text on the opening pages of each chapter. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
Do not use for reproduction, copying, pasting, or reading; exclusively for search engines.
OCR for page 188
Page 188 10 Variability Introduction And Background It is always difficult to identify the true level of risk in an endeavor like health risk assessment, which combines measurement, modeling, and inference or educated guesswork. Uncertainty analysis, the subject of Chapter 9, enables one to come to grips with how far away from the desired answer one's best estimate of an unknown quantity might be. Before we can complete an assessment of the uncertainty in an answer, however, we must recognize that many of our questions in risk assessment have more than one useful answer. Variabilitytypically, either across space, in time, or among individualscomplicates the search for the desired value of many important risk-assessment quantities. Chapter 11 and Appendix I-3 discuss the issue of how to aggregate uncertainties and interindividual differences in each of the components of risk assessment. This chapter describes the sources of variability1and appropriate ways to characterize these interindividual differences in quantities related to predicted risk. Variability is a very well-known "fact of life" in many fields of science, but its sources, effects, and ramifications are not yet routinely appreciated in environmental health risk assessment and management. Accordingly, the first section of this chapter will step back and deal with the general phenomenon (using some examples relevant to risk assessment, but not exclusively), and then for the remainder of the chapter focus only on variability in quantities that directly influence calculations of individual and population risk. When an important quantity is both uncertain and variable, opportunities
OCR for page 189
Page 189 are created to fundamentally misunderstand or misestimate the behavior of the quantity. To draw an analogy, the exact distance between the earth and the moon is both difficult to measure precisely (at least it was until the very recent past) and changeable, because the moon's orbit is elliptical, rather than circular. Thus, as seen in Figure 10-1, uncertainty and variability can complement or confound each other. When only scattered measurements of the earth-moon distance were available, the variation among them might have led astronomers to conclude that their measurements were faulty (i.e., ascribing to uncertainty what was actually caused by variability) or that the moon's orbit was random (i.e., not allowing for uncertainty to shed light on seemingly unexplainable differences that are in fact variable and predictable). The most basic flaw of all would be to simply misestimate the true distance (the third diagram in Figure 10-1) by assuming that a few observations were sufficient (after correcting for measurement error, if applicable). This is probably the pitfall that is most relevant for health risk assessment: treating a highly variable quantity as if it was invariant or only uncertain, thereby yielding an estimate that is incorrect for some of the population (or some of the time, or over some locations), or even one that is also an inaccurate estimate of the average over the entire population. In the risk-assessment paradigm, there are many sources of variability. Certainly, the regulation of air pollutants has long recognized that chemicals differ from each other in their physical and toxic properties and that sources differ from each other in their emission rates and characteristics; such variability is built into virtually any sensible question of risk assessment or control. However, even if we focus on a single substance emanating from a single stationary source, variability pervades each stage from emission to health or ecologic end point: • Emissions vary temporally, both in flux and in release characteristics, such as temperature and pressure. • The transport and fate of the pollutant vary with such well-understood factors as wind speed, wind direction, and exposure to sunlight (and such less-acknowledged factors as humidity and terrain), so its concentrations around its source vary spatially and temporally. • Individual human exposures vary according to individual differences in breathing rates, food consumption, and activity (e.g., time spent in each micro-environment). • The dose-response relationship (the "potency") varies for a single pollutant, because each human is uniquely susceptible to carcinogenic or other stimuli (and this inherent susceptibility might well vary during the lifetime of each person, or vary with such things as other illness or exposures to other agents). Each of these variabilities is in turn often composed of several underlying variable phenomena. For example, the natural variability in human weight is due to the interaction of genetic, nutritional, and other environmental factors.
OCR for page 190
Page 190 FIGURE 10-1 Effects of ignoring uncertainty versus ignoring variability in measuring the distance between the earth and the moon.
OCR for page 191
Page 191 According to the central limit theorem, variability that arises from independent factors that act multiplicatively will generally lead to an approximately lognormal distribution across the population or spatial/temporal dimension (as is commonly observed when concentrations of air pollutants are plotted). When there is more than one desired answer to a scientific question where the search for truth is the end in itself, only two responses are ultimately satisfactory: gather more data or rephrase the question. For example, the question "How far away is the moon from the earth?" cannot be answered both simply and correctly. Either enough data must be obtained to give an answer of the form "The distance ranges between 221,460 and 252,710 miles" or "The moon's orbit is approximately elliptical, with a minor axis of 442,920 miles, a major axis of 505,420 miles, and an eccentricity of 0.482," or the question must be reduced to one with a single right answer (e.g., "How far away is the moon from the earth at its perigee?"). When the question is not purely scientific, but is intended to support a social decision, the decision-maker has a few more options, although each course of action will have repercussions that might foreclose other courses. Briefly, variability in the substance of a regulatory or science-policy question can be dealt with in four basic ways: 1. Ignore the variability and hope for the best. This strategy tends to be most successful when the variability is small and any estimate that ignores it will not be far from the truth. For example, the Environmental Protection Agency's (EPA's) practice of assuming that all adults weight 70 kg is likely to be correct to within ±25% for most adults and probably valid to within a factor of 3 for virtually all adults. However, this approach may not be appropriate for children, where variability may be large (NRC, 1993e). 2. Explicitly disaggregate the variability. Where the quantity seems to change smoothly and predictably over some range, continuous mathematical models may be fitted to the data in place of a discrete step function. An example might be the fitting of sine waves to annual concentration cycles for a particular pollutant. In other cases, it is easier to disaggregate the data by considering all or the relevant subgroups or subpopulations. For interindividual variability, this involves dividing the population into as many subpopulations as deemed necessary. For example, one might perform a separate risk assessment for short-term exposure to high levels of ionizing radiation for each 10-year age interval in the population, to take account of age-related differences in susceptibility. For temporal variability, it involves modeling or measuring in a discrete, rather than a continuous, fashion, on an appropriate time scale. For example, a specific type of air-pollution monitor might collect air for 15 min of each hour and report the 15-min average concentration of some pollutant. Such values might then be further aggregated to produce summary values at an even coarser time scale. For spatial variability, it involves choosing an appropriate subregion, e.g., modeling
OCR for page 192
Page 192 the extent of global warming or cooling for each 10-deg swath of latitude around the globe, rather than predicting a single value for the entire planet, which might mask substantial and important regional differences. In each case, the common thread appears: when variability is "large" over the entire data set, the variability within each subset can become sufficiently "small" ("small" in the sense of the body-weight example in the paragraph above), if the data are disaggregated into an appropriate number of qualitatively distinct subsets. The strategy tends to be most successful when the stakes are so high (or the data or estimates so easy to obtain) that the proliferation of separate assessments does not consume inordinate amounts of resources. In contrast, in studies of a phenomenon such as global climate change, where the stakes are quite high, the estimates may also be quite hard to obtain on a highly disaggregated basis. In health risk assessment, the choice of the averaging time used to transform the variable quantity into a more manageable form is crucially important. In general, for the assessment of acute toxicity, estimates of the variability in exposure and/or uptake over relatively short periods (minutes, hours, days) are needed. For chronic effects such as cancer, one might model exposure and/or update over months or years without losing needed information, since short-term "peaks and valleys" would matter for cancer risk assessment only insofar as they affected the long-term or lifetime average exposure.2The longer-term variability will generally, though not always, be significantly less marked than the variation over the short-term (but see Note 3). Moreover, the shorter the averaging time, the more such periods will be contained in an individual's lifetime, and the more opportunity there will be for rare fluctuations in exposure or uptake to produce significant risks. This, for example, explains why regulators concerned with the health effects of tropospheric ozone consider the combination of peak short-term concentration and peak activity (e.g., the "exercising asthmatic"). In all cases, the exposure assessor needs to determine which time periods are relevant for which toxic effects, and then see whether available data measuring exposure, uptake, internal dose rates, etc., can provide estimates of both the average and the variability over the necessary averaging time. 3. Use the average value of a quantity that varies. This strategy is not the same as ignoring the variability; ideally, it follows from a decision that the average value can be estimated reliably in light of the variability, and that it is a good surrogate for the variable quantity. For example, EPA often uses 70 kg as the average body weight of an adult, presumably because although many adults weigh as little as 40 kg and as much as 100 kg, the average weight is almost as useful as (and less complicated than) three different "scenario" values or an entire distribution of weights. In the same vein, a layperson might be content in knowing the average value of the moon's distance from the earth, rather than the minimum, average, and maximum (let alone a complete mathematical description of its orbit)whereas the average alone would be useless, or even dangerous, to the National Aeronautics and Space Administration in
OCR for page 193
Page 193 planning an Apollo mission. Thus, this strategy tends to be most successful (and indeed might be the only sensible strategy) when the variability is small3or when the quantity is itself an input for a model or decision in which the average value of the end result (the combination of several quantities) is all that matters, either for scientific or policy reasons. An example of a scientific rationale for using the average value is the long-term average concentration of a carcinogen in air. If the dose-response function is linear (i.e., ''potency" is a single number), the end result (risk) is proportional to the average concentration. If the concentration is, say, 10 ppm higher than the average in one week and 10 ppm lower than the average in another week, this variability will have no effect on an exposed person's lifetime risk, so it is biologically unimportant. An example of a policy rationale is the use of the expected number of cancer cases in a population exposed to varying concentrations of an airborne carcinogen. If it is determined for a particular policy rationale that the distribution of individual risks across the population does not matter, then the product of average concentration, potency and population size equals the expected incidence, and the spread of concentrations about the average concentration is similarly unimportant. The average value is also the summary statistic of choice for social decisions when there is an opportunity for errors of underestimation and overestimation (which lead to underregulation and overregulation) to even out over a large set of similar choices over the long run. There are at least two reasons why large variabilities can lead to precarious decisions if the average value is used. The obvious problem is that individual characteristics of persons or situations far from the average are "averaged away" and can no longer be identified or reported. A less obvious pitfall occurs when the variability is dichotomous (or has several discrete values) and the average corresponds to a value that does not exist in nature. If men and women respond markedly differently to some exposure situation, for example, the decision that would be appropriate if there existed an "average person" (midway between man and woman) might be inappropriate for either category of real person (see Finkel, 1991). 4. Use a maximum or minimum of a quantity that varies. This is perhaps the most common way of dealing with variability in risk assessmentto focus attention on one period (e.g., the period of peak exposure), one spatial subregion (e.g., the location where the "maximally exposed individual" resides), or one subpopulation (e.g., exercising asthmatics or children who ingest pathologically large amounts of soil) and ignore the rest. This strategy tends to be most successful when the measures needed to protect or account for the person (or situation) with the extreme value will also suffice for the remainder of the distribution. It is also important to ensure that this strategy will not impose inordinate costs, compared with other approaches (such as using different controls for each subregion or population or simply controlling less stringently by using the average value instead of the extreme "tail").
OCR for page 194
Page 194 The crucial point to bear in mind about all four of those strategies for dealing with variability is that unless someone measures, estimates, or at least roughly models the extent and nature of the variability, any strategy will be precarious. It stands to reason that strategy 1 ("hope for the best") hinges on the assumption that the variability is smallan assumption whose verification requires at least some attention to variability. Similarly, strategy 2 requires the definition of subregions or subpopulations in each of which the variability is small, so care must be taken to avoid the same conundrum that applies to strategy 1. (It is difficult to be sure that you can ignore variability until you think about the possible consequences of ignoring it.) Less obviously, one still needs to be somewhat confident that one has a handle on the variability in order to reduce the distribution to either an average (strategy 3) or a "tail" value (strategy 4). We know that 70 kg is an average adult body weight (and that virtually no adults are above or below 70 kg by more than a factor of 3), because weight is directly observable and because we know the mechanism by which people grow and the biologic limits of either extreme. Armed with our senses and this knowledge, we might need only a few observations to pin down roughly the minimum, the average, and the maximum. But what about a variable like "the rate at which human liver cells metabolize ethylene dibromide into its glutathione conjugate"? Here a few direct measurements or a few extrapolations from animals may not be adequate, because in the absence of any firm notion of the spread of this distribution within the human population (or the mechanisms by which the spread occurs), we cannot know how reliably our estimate of the average value reflects the true average, nor how well the observed minimum and maximum mirror the true extremes. The distribution for an important variable such as metabolic rate should thus explicitly be considered in the risk assessment, and the reliability of the overall risk estimate should reflect knowledge about both the uncertainty and the variability in this characteristic. The importance of a more accurate risk estimate may motivate additional measurements of this variable, so that its distributions may be better defined with these additional data. This chapter concentrates on how EPA treats variability in emissions, exposures, and dose-response relationships, to identify which of the four strategies it typically uses and to assess how adequately it has considered each choice and its consequences. The goals of this chapter are three: (1) to indicate how EPA can increase its sophistication in defining variability and handling its effects; (2) to provide information as to how to improve risk communication, so that Congress and the public understand at least which variabilities are and which are not accounted for, and how EPA's handling of variability affects the "conservatism" (or lack thereof) inherent in its risk numbers; and (3) to recommend specific research whose results could lead to useful changes in risk-assessment procedures. In recent years, EPA has begun to increase its attention to variability. More-
OCR for page 195
Page 195 over, the lack of attention in the past was due in part to a set of choices to erect a set of conservative default options (strategy 4 above) instead of dealing with variability explicitly. In theory at least, the question "How do you determine the extreme of a distribution without knowing the whole distribution?" can be answered by setting a highly conservative default and placing the burden of proof on those who wish to relax the default by showing that the extreme is unrealistic even as a "worst case." For example, the concept of the MEI (someone who breathes pollutants from the source for 70 years, 24 hours per day, at a specified location near a plant boundary) has been criticized as unrealistic, but most agree that as a summary of the population distribution of "number of hours spent at a given location during a lifetime" it might be a reasonable place to start from as a conservative short-cut for the entire distribution. EPA has also tackled interindividual variability squarely in Exposure Factors Handbook (EPA, 1989c), which provides various percentiles (e.g., 5th, 25th, 50th, 75th, 95th) of the observed variability distributions for some components of exposure assessment, such as breathing rates, water ingestion, and consumption of particular foodstuffs. This document has not yet become a standard reference for many of EPA's offices, however. In addition, as we will discuss below, EPA has not dealt adequately with several other major sources of variability. As a result, EPA's methods to manage variability in risk assessment rely on an ill-characterized mix of some questionable distributions, some verified and unverified point values intended to be "averages," some verified and unverified point values intended to be "worst cases," and some "missing defaults," that is, hidden assumptions that ignore important sources of variability. Moreover, several trends in risk assessment and risk management are now increasing the urgency of a broad and well-considered strategy to deal with variability. The three most important of these trends are the following: • The emergence of more sophisticated biological models for risk assessment. As pharmacokinetic models replace the administered assumption and as cell-kinetics models (such as the Moolgavkar-Venzon-Knudson model) replace the linearized-multistage model, default models that ignored human variability or took conservative measures to sidestep it will be supplanted by models that explicitly contain values of biologic measures intended to represent the human population. If the latter models ignore variability or use unverified surrogates for presumed average or worst-case properties, risk assessment might take a step backwards, becoming either less or more conservative without anyone's knowledge. • The growing interest in detailed assessments of the actual exposures that people face, rather than hypothetical worst-case exposures. To be trustworthy, both average and worst-case surrogates for variability require some knowledge of the rest of the distribution, as mentioned above. However, it is not well recognized that the average might be more sensitive to the extreme portions of
OCR for page 196
Page 196 the whole distribution than an upper percentile might be, such as the 95th. In addition, the use of such terms as actual and best estimates carries an expectation of precision that might apply to only part of the exposure assessment, dose-response relationship, or risk assessment. If, for example, we could precisely measure the airborne concentration of a pollutant in a community around a stationary source (i.e., understand the spatial variability), but did not know the population distribution of breathing rates, we could not predict anyone's "actual exposure." In fact, even if we knew both distributions but could not superimpose them (i.e., know which breathing rates went with which concentrations), the predictions would be as variable as either of the underlying distributions. These circumstances speak to the need for progress in many kinds of research and data collection at once, if we wish to improve the power and the realism of risk assessment. • The growing interest in risk-reduction measures that target people, rather than sources. It should go without saying that if government or industry wishes to eliminate unacceptably high risks to particular persons by purchasing their homes, providing them with bottled water, restricting access to "hot spots" of risk, etc., it needs to know precisely who those persons are and where or when those hot spots are occurring. Even if such policies were not highly controversial and difficult to implement in an equitable and socially responsive way, merely identifying the prospective targets of such policies may well presuppose a command of variability beyond our current capabilities. Exposure Variability Variability in human response to pollutants emitted from a particular source or set of sources can arise from differences in characteristics of exposure, uptake, and personal dose-response relationships (susceptibility). Exposure variability in turn depends on variability in all the factors that affect exposure, including emissions, atmospheric processes (transport and transformation), personal activity, and the pollutant concentration in the microenvironments where the exposures occur. Information on those variabilities is not routinely included in EPA's exposure assessments, probably because it has been difficult to specify the distributions that describe the variations. Human exposure results from the contact of a person with a substance at some nonzero concentration. Thus, it is tied to personal activities that determine a person's location (e.g., outdoors vs. indoors, standing downwind of an industrial facility vs. riding in a car, in the kitchen vs. on a porch); the person's level of activity and breathing rate influences the uptake of airborne pollutants. Exposure is also tied to emission rates and atmospheric processes that affect pollutant concentrations in the microenvironment where the person is exposed. Such processes include infiltration of outside air indoors, atmospheric advection (i.e., transport by the prevailing wind), diffusion (i.e., transport by atmospheric turbu-
OCR for page 197
Page 197 lence, chemical and physical transformation, deposition, and re-entrainmentvariability in each process tends to increase the overall variability in exposure. The variabilities in emissions atmospheric processes, characteristics of the microenvironment, and personal activity are not necessarily independent of each other; for example, personal activities and pollutant concentrations at a specific location might change in response to outdoor temperature; they might also differ between weekends and weekdays because the level of industrial activity changes. Emissions Variability There are basically four categories of emission variability that may need separate assessment methods, depending on the circumstances: • Routinethis is the type most frequently covered by current approaches. • Ordinary maintenancespecial emissions may occur, for example, when the bag house is cleaned. In other cases certain emissions may only occur during maintenance, as when a specific volatile cleaner is routinely used to scour or wash out a reaction tank. These can be deliberately observed and monitored to obtain needed emissions information, if this mode is deemed likely to be significant. • Upsets and breakdownsunusual operating conditions that may recur within average periods of days, weeks, or months, depending on the facility/process. A combination of observations and modeling approaches may be needed here. • Catastrophic failureslarge explosions, ruptures of storage tanks, etc. The last category is addressed in a separate section of the Clean Air Act and is not discussed in this report. At least two major factors influence variability in emissions as it affects exposure assessment. First, a given source typically does not emit at a constant rate. It is subject to such things as load changes, upsets, fuel changes, process modifications, and environmental influences. Some sources are, by their nature, intermittent or cyclical. A second factor is that two similar sources (e.g., facilities in the same source category) can emit at different rates because of differences in such things as age, maintenance, or production details. The automobile is an excellent example of both causes. Consider a single, well-characterized car with an effective control system. When it is started, the catalyst has not warmed up, and emissions can be high. Almost half the total automobile emissions in, say, Los Angeles can occur during the cold-start period. After the catalyst reaches its appropriate temperature range, it is extremely effective (›90%) at removing organic substances, such as benzene and formaldehyde, during most of the driving period. However, hard accelerations can overwhelm the system's capabilities and lead to high emissions. Those variations can lead to spatial and temporal distributions of emissions in a city (e.g.,
OCR for page 198
Page 198 high emissions in areas with a large number of cold starts, particularly in the morning). The composition of the emissions, including the toxic content, differs between cold-start and driving periods. Emissions also differ between carsoften dramatically. Because of differences in control equipment, total emissions can vary, and emissions between cycles can vary between cars (e.g., cold-start vs. evaporative emissions). A final notable contribution to emission variability in automobiles is the presence of super-emitters, whose control systems have failed and may emit organic substances at a rate 10 times that of a comparable vehicle that is operating properly. Thus, an exposure analysis based on source-category average emissions will miss the variability in sources within that category. And, exposure analyses that do not account for temporal changes in emissions from a particular source will miss an important factor, especially to the extent that emissions are linked to meteorologic conditions. In many cases, it is difficult or impossible to know a priori how emissions will vary, particularly because of upsets in processes that could lead to high exposures over short periods. Atmospheric Process Variability Meteorologic conditions greatly influence the dispersion, transformation, and deposition of pollutants. For example, ozone concentrations are highest during summer afternoons, whereas carbon monoxide and benzene concentrations peak in the morning (because of the combination of large emissions and little dilution) and during the winter. Formaldehyde can peak in the afternoon during the summer (because of photochemical production) and in the morning in the winter (because of rush-hour emissions and little dilution). Concentrations of primary (i.e., emitted) pollutants, such as benzene and carbon monoxide, are higher in the winter in urban areas, whereas those of many secondary pollutants (i.e., those resulting from atmospheric transformations of primary pollutants), such as ozone, are higher in the summer. Meteorologic conditions may also play a role in regional variations. Some areas experience long periods of stagnant air, which lead to very high concentrations of both primary and secondary pollutants. An extreme example is the London smog that led to high death rates before the mid-1950s. Wind velocity and mixing height also influence pollutant concentrations. (Mixing height is the height to which pollutants are rapidly mixed due to atmospheric turbulence; in effect, it is one dimension of the atmospheric volume in which pollutants are diluted.) They are usually correlated; the prevailing winds and velocities in the winter, when the mixing height is low, can be very different from those in the summer. Some quantitative information is available about the impact of meteorologic variability on pollutant concentrations. Concentrations measured at one location over some period tend to follow a lognormal distribution. There are significant fluctuations in the concentrations about the medians (e.g., Seinfeld, 1986), which
OCR for page 213
Page 213 when, as is usually the case, there are both uncertainty and variability in key risk-assessment inputs. It is important, whenever possible, to separate the two phenomena conceptually, perhaps by presenting multiple analyses. For its full (as opposed to screening-level) risk assessments, EPA should acknowledge that all its risk numbers are made up of three components: the estimated risk itself (X), the level of confidence (Y) that the risk is no higher than X, and the percent of the population (Z) that X is intended to apply to in a variable population. EPA should use its present practice of saying that "the plausible upper-bound risk is X" only when it believes that Y and Z are both close to 100%. Otherwise, it should use statements like, "We are Y% certain that the risk is no more than X to Z% of the population," or use an equivalent pictorial representation (see Figure 10-2). As an alternative or supplement to estimating the value of Z, EPA can and should try to present multiple scenarios to explain variability. For example, EPA could present one risk number (or preferably, an uncertainty distributionsee Chapter 9) that explicitly applies to a "person selected at random from the population," one that applies to a person of reasonably high susceptibility but "average" behavior (mobility, breathing rate, food consumption, etc.), and one that applies to a person whose susceptibility and behavioral variables are both in the "reasonably high" portion of their distributions. Identifiability and Risk Assessment Not all the suggestions presented here, especially those regarding variation in susceptibility, might apply in every regulatory situation. The committee notes that in the past, whenever persons of high risk or susceptibility have been identified, society has tended to feel a far greater responsibility to inform and protect them. For such identifiable variability, the recommendations in this section are particularly salient. However, interindividual variability might be important even when the specific people with high and low values of the relevant characteristic cannot currently be identified 8 Regardless of whether the variability is now identifiable (e.g., consumption rates of a given foodstuff), difficult to identify (e.g., presence of a mutant allele of a tumor-suppressor gene), or unidentifiable (e.g., a person's net susceptibility to carcinogenesis), the committee agrees that it is important to think about its potential magnitude and extent, to make it possible to assess whether existing procedures to estimate average risks and population incidence are biased or needlessly imprecise. In contrast with issues involving average risk and incidence, however, some members of the committee consider the distribution of individual susceptibilities and the uncertainty as to where each person falls in that distribution to be irrelevant if the variation is and will remain unidentifiable. For example, some argue that people should be indifferent between a situation wherein their risk is determined to be precisely 10-5 or one wherein they have a 1% chance of being highly
OCR for page 214
Page 214 FIGURE 10-2 Communicating risk, uncertainty, and variability graphically.
OCR for page 215
Page 215 FIGURE 10-2 Continued
OCR for page 216
Page 216 susceptible (with risk = 10-3) and a 99% chance of being immune, with no way to know which applies to whom. In both cases, the expected value of individual risk is 10-5, and it can be argued that the distribution of risks is the same, in that without the prospect of identifiability no one actually faces a risk of 10-3, but just an equal chance of facing such a risk (Nichols and Zeckhauser, 1986). Some of the members also argue that as we learn more about individual susceptibility, we will eventually reach a point where we will know that some individuals are at extremely high risk (i.e., carried to its extreme, an average individual risk of 10-6 may really represent cases where one person in each million is guaranteed to develop cancer while everyone else is immune). As we approach this point, they contend, society will have to face up to the fact that in order to guarantee that everyone in the population faces ''acceptable" low levels of risk, we would have to reduce emissions to an impossibly low extent. Other committee members reject or deem irrelevant the notion that risk is ultimately either zero or 1; they believe that, both for an individual's assessment of how foreboding or tolerable a risky situation is and for society's assessment of how just or unjust the distribution of risks is, the information about the unidentifiable variability must be reportedthat it affects both judgments. To bolster their contentions, these members cite literature about the limitations of expected utility theory, which takes the view, contradicted by actual survey data, that the distribution of risky outcomes about their mean values should not affect the individual's evaluation of the situation (Schrader-Frechette, 1985; Machina, 1990), and empirical findings that the skewness of lotteries over risky outcomes matters to people even when the mean and variance are kept constant (Lopes, 1984). They also argue that EPA should maintain consistency in how it handles exposure variability, which it reports even when the precise persons at each exposure level cannot be identified; i.e., EPA reports the variation in air concentration and the maximal concentration from a source even when (as is usually the case) it cannot predict exactly where the maximum will occur. If susceptibility is in large part related to person-to-person differences in the amount of carcinogenic material that a person's cells are exposed to via metabolism, then it is essentially another form of exposure variability, and the parallel with ambient (outside-the-body) exposure is close. Finally, they claim that having agreed that issues of pure uncertainty are important, EPA (and the committee) must be consistent and regard unidentifiable variability as relevant (see Appendix I-3). Our recommendations in Chapter 9 reflect our view that uncertainty is important because individuals and decision-makers do regard values other than the mean as highly relevant. If susceptibility is unidentifiable, then to the individual it represents a source of uncertainty about his or her individual risk, and many members of the committee believe it must be communicated just as uncertainty should be. Social-science research aimed at clarifying the extent to which people care about unidentifiable variability in risk, the costs of accounting for it in risk management, and the extent to which people want government to take such
OCR for page 217
Page 217 variation and costs into account in making regulatory decisions and in setting priorities might be helpful in resolving these issues. Findings And Recommendations The committees findings and recommendations are briefly summarized below. Exposure Historically, EPA has defined the maximally exposed individual (MEI) as the worst-case scenarioa continuous 70-year exposure to the maximal estimated long-term average concentration of a hazardous air pollutant. Departing from this practice, EPA has recently published methods for calculating bounding and "reasonably high-end" estimates of the highest actual or possible exposures using a real or default distribution of exposure within a population. The new exposure guidelines do not explicitly define a point on this distribution corresponding to the highest expected exposure level of an individual. • The committee endorses the EPA's use of bounding estimates, but only in screening assessments to determine whether further levels of analysis are necessary. For further levels of analysis, the committee supports EPA's development of distributions of exposure values based on available measurements, modeling results, or both. These distributions can also be used to estimate the exposure of the maximally exposed person. For example, the most likely value of the exposure to the most exposed person is generally the 100[(N - 1)/N]th percentile of the cumulative probability distribution characterizing interindividual variability in exposure, where N is the number of persons used to construct the exposure distribution. This is a particularly convenient estimator to use because it is independent of the shape of the exposure distribution. The committee recommends that EPA explicitly and consistently use an estimator such as 100[(N - 1)/N], because it, and not a vague estimate "somewhere above the 90th percentile," is responsive to the language in CAAA-90 calling for the calculation of risk to "the individual most exposed to emissions. …" In recent times, EPA has begun incorporating into distributions of exposure assumptions that are based on a national average of years of residence in a home, as a replacement for its 70-year exposure assumption (e.g., an average lifetime). Proposals have been made for a similar "departure from defaults" for the time an individual spends at a residence each day, as a replacement for the 24 hours assumption. However, such analyses make the assumption that individuals move to a location of zero exposure when they change residences during their lifetime or leave the home each day. But, people moving from one place to another, whether it be changing the location of their residence or moving from the home to office, may vary greatly in their exposure to any one pollutant, from relatively
OCR for page 218
Page 218 high exposures to none. Further, some exposures to different pollutants may be considered as interchangeable: moving from one place to another may yield exposures to different pollutants which, being interchangeable in their effects, can be taken as an aggregate, single "exposure." This assumption of interchangeability may or may not be realistic; however, because people moving from place to place can be seen as being exposed, over time to a mixture of pollutants, some of them simultaneously and others at separate times, a simplistic analysis of residence times is not appropriate. The real problem is, in effect, a more complex problem of how to aggregate exposure to mixtures as well as one of multiple exposures of varying level of intensities to a single pollutant. Thus, a simplistic analysis based on a simple distribution of residence times is not appropriate. • EPA should use the mean of current life expectancy as the assumption for the duration of individual residence time in a high-exposure area, or a distribution of residence times which accounts for the likelihood that changing residences might not result in significantly lower exposure. Similarly, EPA should use a conservative estimate for the number of hours a day an individual is exposed, or develop a distribution of the number of hours per day an individual spends in different exposure situations. Such information can be gathered through neighborhood surveys, etc. in these high-exposure areas. Note that the distribution would correctly be used only for individual risk calculations, as total population risk is unaffected by the number of persons whose exposures sum to a given total value (if risk is linearly related to exposure rate). EPA has not provided sufficient documentation in its exposure-assessment guidelines to ensure that its point-estimation techniques used to determine the "high-end exposure estimate" (HEEE) when data are sparse reliably yield an estimate at the desired location within the overall distribution of exposure (which, according to these guidelines, lies above the 90th percentile but not beyond the confines of the entire distribution). • EPA should provide a clear method and rationale for determining when point estimators for the HEEE can or should be used instead of a full Monte Carlo (or similar) approach to choosing the desired percentile explicitly. The rationale should more clearly indicate how such estimators are to be generated, should offer more documentation that such point-estimation methods do yield reasonably consistent representations of the desired percentile, and should justify the choice of such a percentile if it differs from that which corresponds to the expected value of exposure to the "person most exposed to emissions". Potency EPA has dealt little with the issue of human variability in susceptibility; the limited efforts to date have focused exclusively on variability relative to noncar-
OCR for page 219
Page 219 cinogenic effects (e.g., normal versus asthmatic response to SO2). The appropriate response to variability for noncancer end points (i.e., identify the characteristics of "normal" and "hypersusceptible" individuals, and then decide whether or not to protect both groups) might not be appropriate for carcinogenesis, in which variability might well be continuous and unimodal, rather than either-or. • EPA, NIH, and other federal agencies should sponsor molecular epidemiologic and other research on the extent of interindividual variability in various factors that affect susceptibility and cancer, on the relationships between variability in each factor and in the health end point, and on the possible correlations between susceptibility and such covariates as age, race, ethnicity, and sex. Results of the research should be used to adjust and refine estimates of risks to individuals (identified, identifiable, or unidentifiable) and estimates of expected incidence in the general population. As this research progresses, the natural science and social science community should collaborate to explore the implications of any susceptibility factors that can be tested for or that strongly correlate with other genetic traits, so as to ensure that any findings are not misinterpreted or used outside of the environmental risk assessment arena without proper care. Susceptibility EPA does not account for person-to-person variations in susceptibility to cancer; it thereby treats all humans as identical in this respect in its risk calculations. • EPA should adopt a default assumption for susceptibility before it begins to implement those decisions called for in the Clean Air Act that require the calculation of risks to individuals. EPA could choose to incorporate into its cancer risk estimates for individual risk a "default susceptibility factor" greater than the implicit factor of 1 that results from treating all humans as identical. EPA should explicitly choose a default factor greater than 1 if it interprets the statutory language to apply to an individual with high exposure and above-average susceptibility. EPA could explicitly choose a default factor of 1 for this purpose, if it interprets the statutory language to apply to an individual with high exposure but average susceptibility. Preferably, EPA could develop a "default distribution" of susceptibility, and then generate the joint distribution of exposure and cancer potency (in light of susceptibility) to find the upper 95th percentile (or 99th percentile) of risk for each risk assessment. EPA makes its potency calculations on the assumption that, on average, humans have susceptibility similar to that of the particular sex-strain combination of rodent that responds most sensitively of those tested in bioassays or susceptibility identical with that of the particular groups of persons observed in epidemiologic studies.
OCR for page 220
Page 220 • EPA should continue and increase its efforts to validate or improve the default assumption that, on average, humans to be protected at the risk-management stage have susceptibility similar to that of humans included in relevant epidemiological studies, the most-sensitive rodents tested, or both. It is possible that ignoring variations in human susceptibility may cause significant underestimation of population risk, if both of two conditions hold: (1) current procedures to extrapolate results of laboratory bioassays or epidemiologic studies to the general population correctly map the observed risk in the test population to the human with median susceptibility, not to the expected value averaged over the entire general population; and (2) there is sufficient skewed variability in susceptibility in the general population to cause the expected value to exceed the median to a significant extent. • In addition to continuing to explore the assumption that interspecies scaling (or epidemiologic extrapolation) correctly predicts average human susceptibility, EPA should investigate whether the average that is predicted corresponds to the median or the expected value. If there is reason to suspect the former is true, EPA should consider whether it needs to adjust its estimates of population risk to account for this discrepancy. Children are a readily identifiable subpopulation with its own physiologic characteristics (e.g., body weight), uptake characteristics (e.g., food consumption patterns), and inherent susceptibilities. • If there is reason to believe that risk of adverse biological effects per unit dose depends on age, EPA should present separate risk estimates for adults and children. When excess lifetime risk is the desired measure, EPA should compute an integrated lifetime risk, taking into account all relevant age-dependent variables. EPA does not usually explore or consider interindividual variability in key biologic parameters when it uses or evaluates various physiologic or biologically based risk-assessment models (or else evaluates some data but does not report on this in its final public documents). In some other cases, EPA does gather or review data that bear on human variability, but tends to accept them at face value without ensuring that they are representative of the entire population. As a general rule, the larger the number of characteristics with an important effect on risk or the more variable those characteristics are, the larger the sample of the human population needed to establish confidently the mean and range of each of those characteristics. • When EPA proposes to adopt an alternative risk-assessment assumption (such as use of a PBPK model, use of a cell-kinetics model, or the determination that a given animal response is "not relevant to humans"), it should consider human interindividual variability in estimating the model parameters or verify-
OCR for page 221
Page 221 ing the assumption of "irrelevance." If the data are not available to take account of human variability, EPA should be free to make any reasonable inferences about its extent and impact (rather than having to collect or await such data), but should encourage other interested parties to collect and provide the necessary data. In general, in parallel to recommendation UAR4, EPA should ensure that a similar level of variability analysis is applied to both the default and the alternative risk assessment, so that it can compare equivalently conservative estimates from each procedure. Risk Communication EPA does not adequately communicate to its own decision-makers, to Congress, or to the public the variabilities that are and are not accounted for in any risk assessment and the implications for the conservatism and representativeness of the resulting risk numbers. • EPA should carefully state in each risk assessment what its particular assumptions about human behavior and biology do and do not account for. For its full (as opposed to screening-level) risk assessments, EPA makes risk-communication and risk-management decisions more difficult when, as is usually the case, both uncertainty and variability are important. • Whenever possible, EPA should separate uncertainty and variability conceptually, perhaps by presenting multiple analyses. EPA should acknowledge that all its risk numbers are made up of three components: the estimated risk itself (X), the level of confidence (Y) that the risk is no higher than X, and the percent of the population (Z) that X is intended to apply to in a variable population. In addition, rather than reporting both Y and Z, EPA can and should try to present multiple scenarios to explore and explain the variability dimension. Notes 1. Some specialists in different fields often use the term "variability" to refer to a dispersion of possible or actual values associated with a particular quantity, often with reference to random variability associated with any estimate of an unknown (i.e., uncertain) quantity. This report, unless stated otherwise, will use the terms interindividual variability, variability, and interindividual heterogeneity all to refer to individual-to-individual differences in quantities associated with predicted risk, such as in measures of or parameters used to model ambient concentration, uptake or exposure per unit ambient concentration, biologically effective dose per unit exposure, and increased risk per unit effective dose. 2. This assumes that risk is linear in long-term average dose, which is one of the bases of the classical models of carcinogenesis (e.g., the LMS dose-response model using administered dose). However, when one moves to more sophisticated models of the dose-exposure (i.e., PBPK) and exposure-response (i.e., biologically motivated or cell-kinetics models) relationships, shorter averaging times become important even though the health endpoint may manifest itself over the long-term. For example, the cancer risk from a chemical that is both metabolically activated and detoxified in
OCR for page 222
Page 222 vivo may not be a function of total exposure, but only of those periods of exposure during which detoxification pathways cannot keep pace with activating ones. In such cases, data on average long-term concentrations (and interindividual variability therein) may completely miss the only toxicologically relevant exposure periods. 3. As discussed above, in many cases variability that exists over a short averaging time may grow less and less important as the averaging time increases. For example, if on average, adults breathe 20m3 of air per day, then over any random 1-minute period, in a group of 1,000 adults there would probably be some (those involved in heavy exertion) breathing much more than the average value of 0.014 (m3/min), and other (those asleep) breathing much less. Over the course of a year, however, the variation around the average value of 7300 m3/yr would be much smaller, as periods of heavy exercise, sleep, and average activity "average out." On the other hand, some varying human characteristics do not substantially converge over longer averaging periods. For example, the daily variation in the amount of apple juice people drink probably mirrors the monthly and yearly variation as wellthose individuals who drink no apple juice on a random day are probably those who rarely or never drink it, while those at the other "tail" of the distribution (drinking perhaps three glasses per day) probably tend to repeat this pattern day after day (in other words, the distribution of "glasses drunk per year'' probably extends all the way from zero to 365 × 3, rather than varying narrowly around the midpoint of this range). 4. Similarly, the two persons might face equal cancer risks at exposures that were 10,000-fold different. However, an alternative definition, which would be more applicable for threshold effects, would be to call the difference in susceptibility the ratio of doses needed to produce the same effect in two different individuals. 5. The logarithmic standard deviation is equivalent to the standard deviation of the normal distribution corresponding to the particular lognormal distribution. If one takes the antilog of the logarithmic standard deviation, one obtains the "geometric standard deviation", or GSD, which has a more intuitively appealing definition: N standard deviations away from the median corresponds to multiplying or dividing the median by the GSD raised to the power N. 6. Moreover, existing studies of overall variations in susceptibility suggest that a factor of 10 probably subsumes one or perhaps 1.5 standard deviations above the median for the normal human population. That is, assuming (as EPA does via its explicit default) that the median human and the rodent strain used to estimate potency are of similar susceptibility, an additional factor of 10 would equate the rodent response to approximately the 85th or 90th percentiles of human response. That would be a protective, but not a highly conservative, safety factor, inasmuch as perhaps 10 percent or more of the population would be (much) more susceptible than this new reference point. Inclusion of a default factor of 10 could bring cancer risk assessment partway into line with the prevailing practice in noncancer risk assessment, wherein one of the factors of 10 that are often added is meant to account for person-to-person variations in sensitivity. However, if EPA decides to use a factor of 10, it should emphasize that this is a default procedure that tries to account for some of the interindividual variation in dose-response relationships, but that in specific cases may be too high or too low to provide the optimum degree of "protection" (or to reduce risks to "acceptable" levels) for persons of truly unusual susceptibility. Nor does it ensure that (in combination with exposure estimates that might actually correspond to a maximally exposed or reasonably high-end person) risk estimates are predictive or conservative for the actual "maximally-at-risk" person. In contrast, some persons of extremely high susceptibility might, as a consequence of their susceptibility, not face high exposures. It might also be the case that some risk factors for carcinogenesis also predispose those affected to other diseases from which it might be impossible to protect them. 7. For example, suppose the median income in a country was $10,000, but 5 percent of the population earned 25 times less or more than the median and an additional 1 percent earned 100 times less or more. Then the average income would be [(0.05)(400) + (0.05)(250,000) + (0.01)(100) + (0.01)(1,000,000) + (0.88)(10,000)] = $31,321, or more than three times the median income.
OCR for page 223
Page 223 8. "Currently" is an important qualifier given the rapid increases in our understanding of the molecular mechanisms of carcinogenesis. During the next several decades, science will doubtless become more adept at identifying individuals with greater susceptibility than average, and perhaps even pinpoint specific substances to which such individuals are particularly susceptible.
Representative terms from entire chapter: