Address correspondence to T.J. Woodruff, Public Health and Environmental Policy Team, National Center for Environmental Economics, U.S. EPA, 75 Hawthorne St., MC SPE-1, San Francisco, CA 94105 USA. Telephone: (415) 947-4277. Fax: (415) 947-3519. E-mail: Woodruff.Tracey@epamail.epa.gov

We thank D. Axelrad, V.J. Cogliano, J. Gift, J. Jinot, and T. McKone for their expert technical assistance and advice.

This work was supported in part by grants from the U.S. EPA and the National Institute of Environmental Health Sciences (R826709 and ES09605). The views expressed in this article are those of the authors and do not necessarily reflect those of the U.S. EPA.

The authors declare they have no conflict of interest.

Received 2 January 2003; accepted 12 May 2003.

Introduction

Methods for evaluating risks from exposure to toxic substances for noncancer health end points (such as birth defects, respiratory effects, and hepatotoxicity) are based on the theory that there is a threshold below which there is negligible risk of adverse health effects from environmental exposures. At the U.S. Environmental Protection Agency (U.S. EPA), this negligible risk is quantified through the use of reference doses (RfDs) and reference concentrations (RfCs). The RfD or RfC is defined as an estimate of daily or continuous exposure to the human population (including sensitive subgroups) that is likely to be without an appreciable risk of deleterious effects during a lifetime (U.S. EPA 1999a). The value of the RfD or RfC is derived by determining a point of departure divided by uncertainty factors (UFs), which are used to account for uncertainties in the available studies, such as limitations in the database, variability within humans, and differences in species response (i.e., animal-to-human extrapolation).

The point of departure in environmental health risk assessment is meant to represent the lowest dose within the range of experimental data. In past practices, the point of departure was exclusively based on a no-observed-adverse-effect level (NOAEL) or a lowest-observed-adverse-affect level (LOAEL), derived from animal or epidemiologic studies. The NOAEL is the highest dose for which there are no observed statistically or biologically significant increases in the frequency or severity of adverse effects between the exposed population and its control. Similarly, the LOAEL is the lowest dose at which there are statistically or biologically significant increases in the frequency or severity of adverse effects between the exposed population and its appropriate control group. The NOAEL/LOAEL structure, however, does not provide sufficient information to quantify the equivalent risk levels from exposure at the RfD/RfC because there is no estimated risk at the NOAEL or LOAEL itself. Several authors have criticized the use of the NOAEL because of its sensitivity to sample size, the high sampling variability from experiment to experiment, and the inability to use all of the available dose-response data (Barnes et al. 1995; Crump 1984; Gaylor et al. 1998; Leisenring and Ryan 1992; U.S. EPA 2000a). Leisenring and Ryan (1992) have shown that average risk levels associated with the NOAEL may be substantial. The true risk of exposure at the NOAEL can vary from zero to > 20%, depending on the end point, spacing of doses, and numbers of animals used (Leisenring and Ryan 1992). In many cases, an adverse effect may not be detected in a critical effect study because of insufficient statistical power.

Although the NOAEL/LOAEL structure does not provide sufficient information to quantify risk levels from exposure (Gaylor and Kodell 2000), the resulting RfDs and RfCs are assumed to be equivalent to negligible or de minimis risks. As a point of comparison, the U.S. EPA has defined 1 in 1,000,000 excess cancer risk as a de minimis risk level for cancer (Caldwell et al. 1998; Clean Air Act Amendments 1990; Fiori and Meyerhoff 2002; U.S. EPA 1991), although regulatory actions are sometimes limited to instances where risk exceeds 1 in 100,000.

Over the past several years, the U.S. EPA has been in the process of developing the benchmark dose (BMD), which is derived from modeling the exposure-response data, as an alternative to the NOAEL/LOAEL as the point of departure for noncancer risk assessments. The BMD is the dose that corresponds to a specified level of increased response [the benchmark response (BMR)] compared with background. This dose is calculated by fitting a mathematical model to the dose-response data, which can be continuous or quantal. The BMD method has several advantages over the NOAEL/LOAEL method, including making better use of dose-response information and reflecting sample size more appropriately (Barnes et al. 1995; Crump 1995). In single-chemical hazard assessments, the BMD allows for consideration of the dose-response over the entire exposure range, and furthermore, when a dose derived from benchmark modeling is used as the point of departure, actual risk levels can be calculated as an alternative to the hazard index (which is typically based on comparisons of human exposures with an RfD or RfC).

In 2000, the U.S. EPA published a draft guidance document on the application of the BMD approach in determining the point of departure for all types of health effects data (U.S. EPA 2000a). Although a BMR of 10% has most often been used by the U.S. EPA in its assessments, it is anticipated that 5% or 1% would be a more appropriate response level for some end points and designs. Furthermore, in these draft guidelines, a lower statistical confidence limit on the BMD (BMDL) is specifically proposed as a replacement for the NOAEL/LOAEL in setting the point of departure, which is used to determine acceptable levels of human exposure to environmental toxicants. A lower confidence limit is placed on the BMD to obtain a dose (BMDL) that assures with high confidence (e.g., 95%) that the BMR is not exceeded (U.S. EPA 2000a). In addition to ensuring an added measure of protection, this process rewards better experimental design and procedures that provide more precise estimates of the BMD.

Most new and revised RfDs and RfCs in the U.S. EPA Integrated Risk Information System (IRIS) assessments are based on BMD modeling. Certain health end points, however, are not amenable to modeling, and the NOAEL/LOAEL approach will continue to be used in some cases (U.S. EPA 2000a). For this article, we have reviewed and synthesized currently available risk assessment information on the chemicals for which U.S. EPA reference values are based on BMD modeling. We estimate the equivalent risk levels expected from hypothetical human exposures at established RfD and RfC values using the BMD dose-response information to investigate whether these levels represent negligible risks, and to underscore some of the potential strengths of using benchmark modeling in environmental health risk assessment.

Methods

Table 1.

Table 2.

Table 3.

Table 4.

Figure 1. Linear extrapolation of risk from exposure at the RfD when the RfD is derived from benchmark modeling of quantal dose-response data. Assuming the relationship between risk of adverse outcome and dose is linear at low doses, we estimated risk from exposure at concentrations equal to established RfD and RfC values by extrapolating from the point represented by the BMR at the BMDL or BMCL to established RfD and RfC values. POD, point of departure. aFor quantal data, the BMR represents a percentage increase in risk compared with controls.

Table 5.

Table 6.

Table 7.

Figure 2. Estimated risk levels from exposure at the RfD based on the BMD and the BMDL derived from linear or supralinear dose-response curves. EGBE, ethylene glycol monobutyl ether.

Figure 3. Estimated risk levels from exposure at the RfC based on the BMC and the BMCL derived from linear or supralinear dose-response curves. EGBE, ethylene glycol monobutyl ether. aBMC central estimate not provided by the U.S. EPA.

Figure 4. Distribution of estimated risk levels from human exposures at established RfD and RfC values for reference doses derived from linear or supralinear dose-response curves. aRisk estimates for assessments based on sublinear dose-response curves are not included.

Figure 5. Ratio of the points of departure (PODs) to empirically derived NOAELs. PODs based on benchmark modeling.

We searched the U.S. EPA IRIS database (U.S. EPA 2000b) to identify the chemicals for which current regulatory reference values relied on BMD modeling. We identified 11 chemicals with RfDs based on oral BMD values, and 12 chemicals with RfCs based on benchmark concentration (BMC) values. The BMD and BMC values and other risk assessment information for these chemicals are presented in Tables 1-4 (U.S. EPA 1995a, 1995b, 1995c, 1995d, 1997a, 1998a, 1998c, 1998e, 1998g, 1999b, 2000c, 2001a, 2001b, 2001d, 2002a, 2002c, 2002e, 2003). We have included risk assessment information for naphthalene, a chemical with an IRIS assessment containing an established RfD based on a NOAEL (RfD = 0.02 mg/kg/day) as well as a prospective RfD based on BMD modeling (RfD = 0.03 mg/kg/day) (U.S. EPA 1998i).

We determined whether the principal study identified for each chemical's IRIS assessment was derived from quantal (dichotomous) or continuous critical effect data (Tables 2 and 4). For compounds with BMD or BMC values based on quantal data, the BMR is expressed in terms of a percent increase in risk of adverse outcome compared with background. For compounds with BMD or BMC values based on continuous data, the BMR may be expressed as a percent change in mean response compared with control (e.g., immunosuppression with tributyltin oxide) or in terms of a 1 standard deviation change from the control mean response (e.g., decreased lymphocyte count with benzene). The BMR is a response level used to define a BMD, which is used as the point of departure, from which an RfD or RfC can be developed. The BMR is typically set at the lower end of the range of responses (e.g., 10% or 5% change) that can be detected experimentally. This can help to avoid uncertainties associated with low-dose extrapolation using models that may not reflect biologic realities (Crump 1995).

Using the benchmark modeling information described above from the IRIS assessments of 19 chemicals, we estimated the equivalent risk levels expected from hypothetical human exposures at the established RfDs and RfCs. We also evaluated whether each of the models used for BMD calculations was linear, sublinear, or supralinear in the observed dose range (Table 5) [National Research Council (NRC) 2000; U.S. EPA 1995a, 1995b, 1995c, 1995d, 1997b, 1998b, 1998c, 1998d, 1998f, 1998h, 1998j, 1999c, 2000d, 2001c, 2001e, 2002b, 2002d]). To estimate the risk level of the derived RfDs and RfCs, which have a linear or supralinear dose-response curve at the BMD/BMC, we assumed that the dose-response curves for these compounds are linear at doses below the BMD or BMC (Figure 1). For two assessments, sufficient information was not available to determine the shape of the dose-response curve (carbon disulfide and 1,1,1,2-tetrafluoroethane), and we assumed linearity [this assumption did not overly bias our results because linearity or supralinearity was the shape of the dose-response curve in approximately three-quarters (16 of 21) of cases in which the shape of the dose-response curve was discernable]. This assumption of linearity at the relevant part of the dose-response curve is necessary to extrapolate equivalent risk levels from U.S. EPA reference values derived from BMD modeling, and it is consistent with methods proposed in the U.S. EPA draft cancer risk assessment guidelines (U.S. EPA 1999d). Such risk-level extrapolation is not possible using the NOAEL/LOAEL approach.

Methods for risk-level estimation varied between reference values based on quantal end points and those based on continuous end points. For BMD/BMC values derived from quantal critical effect data, we estimated risk from exposure at concentrations equal to established RfD and RfC values by extrapolating linearly from the point represented by the BMR at the BMDL/BMCL to the established RfD and RfC values (Figure 1). We divided the risk at the BMR by the composite UF for those BMD models that were linear or supralinear. For example, to estimate risk from exposure to chloroform's RfC, we divided the estimated risk level at the point of departure (1 in 10 for BMR = 10%) by the composite UF of 100, to arrive at a risk of 1 in 1,000. For BMD/BMC values derived from continuous critical effect data (normally distributed), a change in response of 1 standard deviation from control is considered roughly equivalent to a 10% increase in risk of adverse response from exposure (e.g., benzene's BMR = change of 1 standard deviation in lymphocyte count compared with control mean) (U.S. EPA 2000a, 2003). Therefore, when data quality and distribution allowed, we treated the dose that resulted in a 1-standard-deviation change from control as equivalent to BMD₁₀/BMC₁₀ (BMD that equals a BMR of 10%/BMC that equals a BMR of 10%) values derived from quantal data. For assessments based on sublinear dose-response curves, we estimated risk of exposure at the RfD/RfC dose levels by extrapolating the BMD model response function to the RfD/RfC (i.e., using the BMD model, we estimated risk by putting the exposure equal to the RfD/RfC in the model).

For all chemicals in our assessment group with adequate data, we calculated the ratio of the central estimate (BMD or BMC) to the lower statistical confidence limit on the benchmark dose (BMDL) or concentration (BMCL) (Tables 1 and 3). This ratio (e.g., BMD/BMDL) provides a metric to compare the relative impact on estimated risk levels resulting from the selection of the BMD/BMC versus the BMDL/BMCL as the point of departure. Finally, among the studies that identified NOAELs, we compared the modeled BMDL and BMCL values with the empirical NOAELs as a means to investigate how using BMDL/BMCL values versus NOAELs compares with previous RfD/RfC methods based on the NOAEL approach.

Results

We found that 13 out of 23 (57%) of the BMD and BMC values were derived from quantal versus continuous data. A 10% additional risk or 10% change from control mean response was selected as the BMR for 17 of the 23 assessments, and a 5% BMR was selected for 3 of the remaining 6 assessments (Tables 6 and 7) [National Toxicology Program (NTP) 1993; U.S. EPA 1995a, 1995b, 1995c, 1995d, 1997a, 1998a, 1998c, 1998e, 1998g, 1998i, 1999b, 2000c, 2001a, 2001b, 2001d, 2002a, 2002e]. The BMR values for benzene's BMC and BMD and for phenol's BMD were based on a 1-standard-deviation change in acute lymphocyte count and maternal body weight, respectively, compared with the control mean (U.S. EPA 2002c, 2003).

Of the 21 BMD and BMC values for which sufficient dose-response information was available, we found that 16 (76%) were derived from dose-response data fitted to linear or supralinear models in the observed dose range (Table 5) (NRC 2000; NTP 1993, 1998; U.S. EPA 1995b, 1995d, 1997b, 1998b, 1998c, 1998d, 1998f, 1998h, 1998j, 1999c, 2000c, 2001c, 2001e, 2002b, 2002d, 2002e, 2003). Seven assessments were based on linear models (two linear models; three polynomial models with ß_i 0 for i > 1; two power models with k 1), nine were based on supralinear models, and five were based on sublinear models. Sufficient information was not available to determine the shape of the fitted model for two assessments, carbon disulfide and 1,1,1,2-tetrafluoroethane (i.e., the response function and model parameters were not provided) (U.S. EPA 1995a, 1995c).

We calculated 17 RfD and RfC equivalent risk levels (for 14 compounds) assuming linear dose-response curves. These risk estimates ranged from 1 in 10,000 to 5 in 1,000 for the oral route of exposure for compounds with RfDs based on BMD values, and from 1 in 10,000 to 3 in 1,000 for inhalation for compounds with RfCs based on BMC values (Tables 6 and 7). Figures 2 and 3 present the RfD and RfC equivalent risk estimates on a logarithmic scale for these chemicals with linear or supralinear dose-response curves at the BMD/BMC. For four RfD and RfC equivalent risk levels (for 3 compounds), we used the sublinear dose-response model to calculate risk at the RfD or RfC. These risk estimates ranged from 3 in 1,000,000,000 to 1 in 100,000 from the oral route of exposure for compounds with RfDs based on BMD values, and from 3 in 1,000,000 to 8 in 10,000 from inhalation for compounds with RfCs based on BMC values (Tables 6 and 7).

Five of 21 reference values (24%) reviewed for this assessment corresponded to estimated risk levels greater than 1 in 1,000. Insufficient information was available to estimate risk from exposure at two reference values that were based on continuous response data, chromium VI particulates and naphthalene (i.e., the response function, underlying distribution of the end point, or mean response and standard deviation of the treatment group and controls were not provided). Figure 4 presents the distribution of estimated risk levels from human exposures at established RfD and RfC values for compounds with a linear or supralinear dose-response curve at the BMD. Risk estimates for four assessments derived from sublinear dose-response curves are presented in Tables 6 and 7.

Among the chemicals for which the RfD was based on a BMD, the BMD/BMDL ratio ranged from 1.3 to 3.0. Among the chemicals for which the RfC was based on a BMC, the BMC/BMCL ratio ranged from 1.1 to 2.3 (Tables 1 and 3). Thus, using the central estimate of the BMD or BMC (maximum likelihood estimate) instead of the lower statistical confidence limit (BMDL or BMCL) as the point of departure would result in a 1- to 3-fold difference in the estimated risk levels (Figures 2 and 3).

The effect level extrapolation factor (ELE) is an UF analogous to the LOAEL-to-NOAEL extrapolation factor. ELEs were applied in the assessments of three compounds, 1,3-butadiene (RfC), benzene (RfC and RfD), and ethylene glycol monobutyl ether (RfC) (Tables 2 and 4) (U.S. EPA 1999b, 2002e, 2003). Thus, no ELE factor was assigned for 16 of 17 assessments that were based on a BMR of 10%.

When we compared the points of departure (i.e., BMDL/BMCL values) with the NOAELs, we found that the points of departure were higher than the NOAELs in 10 of the 14 studies with identified NOAELs (Figure 5). The BMDL values were up to 4.6 times higher than the empirical NOAELs (range, 0.5-4.6), and the BMCL values were up to 3.9 times higher than the empirical NOAELs (range, 0.2-3.9) (Tables 1 and 3).

Discussion

To determine whether U.S. EPA reference values represent negligibly small risk levels, we reviewed and synthesized currently available risk assessment information on chemicals for which established RfD and RfC values are based on BMD modeling. For RfDs and RfCs derived from linear or supralinear dose-response curves, our risk estimates ranged from 1 in 10,000 to 5 in 1,000 for the oral route of exposure, and from 1 in 10,000 to 3 in 1,000 for inhalation. Risk estimates for RfDs and RfCs derived from sublinear dose-response curves ranged from 3 in 1,000,000,000 to 8 in 10,000. Twenty-four percent of reference values reviewed for this assessment corresponded to estimated risk levels greater than 1 in 1,000. The estimated risk of exposure to 1,1-dichloroethylene at its RfC, for example, corresponded to a 3-in-1,000 risk of adverse effect [liver toxicity (fatty change)].

The BMD methodology is the first step in the development of continuous risk functions that can be used to estimate risks at different exposures rather than using an RfD/RfC approach, which has limited use in the decision-making process. For example, BMD and BMC values that are based on the same level of adverse response (e.g., BMR = 10%) can be used to rank the potential hazard of exposure to multiple toxicants. Further application of BMD models, such as has been done here, can be used for estimating adverse noncancer health outcomes from different exposures for other risk-ranking exercises, regulatory policy development, and cost/benefit analyses.

The U.S. EPA used a variety of fitted models to calculate the BMD/BMC values found in IRIS (e.g., K power, linear, quantal-linear, exponential polynomial, and Weibull). To compare RfD and RfC equivalent risk levels, we assumed that the dose-response curves for the chemicals in our assessment group are linear at doses below the point of departure. This assumption could have resulted in both underestimates and overestimates of risk. In the case of a supralinear dose-response curve at low doses, for example, this assumption may have resulted in an underestimate of risk. Among the chemicals we reviewed, 9 of the 21 assessments with sufficient information to determine the shape of the dose-response curve were based on supralinear functions. In the case of a sublinear dose-response, this assumption may have resulted in a marked overestimate of risk. For the 5 assessments based on sublinear dose-response curves, therefore, we did not assume linearity for low dose extrapolation and risk estimation.

We believe that the assumption of linearity in the relevant part of the dose-response curve is justified and useful to compare risk levels among this group of compounds. Seventy-six percent of BMD and BMC values considered in this assessment were derived from dose-response data fitted to linear or supralinear models. Furthermore, the range of extrapolation for the RfD/RfC calculations was not large among this group, with most based on points of departure extrapolated to 2 orders of magnitude or less (7 were extrapolated to 1 order of magnitude, 13 were extrapolated to 2 orders of magnitude, and 3 were extrapolated to 3 orders of magnitude). The average and median composite UFs among the compounds in our assessment are 340 and 100, respectively. This implies that even if the dose-response curve for a particular compound is not strictly linear at much lower doses, we could expect the potential impact on the risk estimate to be relatively small.

Current U.S. EPA methodology for reference value derivation assumes that the established RfDs/RfCs represent negligibly small risk levels. For assessments that have linear dose-response curves, the extrapolation from the point of departure is typically 2 orders of magnitude or less. Therefore, for the RfD/RfC values to represent risk levels that are below regulatory concern, the dose-response curve would have to drop off sharply after the point of departure. This assumption seems unlikely, especially given our finding that a large number of the assessments we reviewed (9 of 21) were based on supralinear dose-response functions. Although this supralinearity may carry significant implications for risk assessment, more research is needed to determine whether these dose-response relationships remain supralinear at very low doses. On the other hand, assessments based on dose-response curves that are not monotonic may have sublinear or stepwise relationships below the observed data.

Using the BMDL or BMCL as the point of departure in the risk assessment of noncarcinogenic compounds rather than the BMD or BMC central estimate is generally characterized as a conservative assumption (in the health-protective sense). We found that using the central estimate of the BMD (maximum likelihood estimate) instead of the lower bound estimate as the point of departure results in a 1- to 3-fold difference in the risk estimates. According to the U.S. EPA draft BMD guidelines (U.S. EPA 2000a), a lower confidence limit is placed on the BMD to obtain a dose (BMDL) that assures with high confidence (e.g., 95%) that the BMR is not exceeded. This process of using the BMDL rewards better experimental design and procedures that provide more precise estimates of the BMD, resulting in tighter confidence intervals and thus BMDLs that are closer to the central estimate. Our results suggest that the current practice of using the statistical lower bound estimate versus the maximum likelihood estimate as the point of departure is reasonable and does not substantially bias the risk estimate.

For carcinogens, the U.S. EPA typically develops a linear estimate of the slope of the dose-response curve, under the assumption that the curve is linear at low doses. This allows for quantification of risk at any given level of exposure. The U.S. EPA has defined 1 in 1,000,000 excess cancer risk as a de minimis risk level for cancer (Caldwell et al. 1998; Clean Air Act Amendments 1990; Fiori and Meyerhoff 2002; U.S. EPA 1991), although regulatory actions are sometimes limited to instances where risk exceeds 1 in 100,000. Among compounds in IRIS with RfDs and RfCs based on BMD modeling, however, we found risk estimates as great as 5 in 1,000. Although noncancer outcomes may in some instances be reversible and considered less severe, this finding calls into question the assumption that noncancer RfD and RfC values represent "acceptable levels" of exposure. In addition, some of the noncancer health end points considered in this assessment are severe and irreversible events, for example, ovarian atrophy (1,3-butadiene) and developmental neuropsychologic impairment (methylmercury), highlighting the need for a renewed discussion within the public health community about what should be considered an "acceptable level" of risk from exposure to toxicants with noncancer health end points.

Most of the BMDLs and BMCLs used as points of departure in IRIS are based on 10% BMRs, many with values higher than the empirically derived NOAELs. This research should help inform discussions about whether this level of BMR is adequately protective of the public health, and whether human exposures at concentrations equal to the resulting reference values do in fact represent negligibly small risk levels.