[Federal Register Volume 89, Number 79 (Tuesday, April 23, 2024)]
[Notices]
[Pages 30326-30336]
From the Federal Register Online via the Government Publishing Office [www.gpo.gov]
[FR Doc No: 2024-08604]


=======================================================================
-----------------------------------------------------------------------

CONSUMER PRODUCT SAFETY COMMISSION

[Docket No. CPSC-2023-0032]


Notice of Availability: Supplemental Guidance for CPSC Chronic 
Hazard Guidelines

AGENCY: U.S. Consumer Product Safety Commission.

ACTION: Notice of availability.

-----------------------------------------------------------------------

SUMMARY: The Consumer Product Safety Commission (Commission or CPSC) is 
announcing the availability of final supplemental guidance for its 
Chronic Hazard Guidelines. This supplemental guidance contains two 
guidance documents, one for the use of benchmark dose methodology in 
risk assessment and the other for the analysis of uncertainty and 
variability in risk assessment.

ADDRESSES: Docket: For access to the docket to read background 
documents or comments received, go to www.regulations.gov and insert 
the docket number, CPSC-2023-0032, in the ``Search'' box, and follow 
the prompts.

FOR FURTHER INFORMATION CONTACT: Eric Hooker, Directorate for Health 
Sciences, U.S. Consumer Product Safety Commission, 5 Research Place, 
Rockville, MD 20850; telephone: (301) 987-2516; email: 
[email protected].

SUPPLEMENTARY INFORMATION:

I. Background

    In 1992, the Commission issued guidelines for assessing chronic 
hazards (Chronic Hazard Guidelines or Guidelines) under the Federal 
Hazardous Substances Act (FHSA), 15 U.S.C. 1261-78, including 
carcinogenicity, neurotoxicity, reproductive/developmental toxicity, 
exposure, bioavailability, risk assessment, and acceptable risk. 57 FR 
46626. In August 2023, the Commission issued a Notice of Availability 
containing Proposed Supplemental Guidance for CPSC Chronic Hazard 
Guidelines and asked for comments on the proposed guidance. 88 FR 
57947. After reviewing those comments, the Commission is now issuing 
the final supplemental guidance contained below in sections III and 
IV.\1\
---------------------------------------------------------------------------

    \1\ On April 12, 2024, the Commission voted 5-0 to approve 
publication of this notice. Commissioners Feldman and Dziak 
submitted a joint statement, available at https://www.cpsc.gov/About-CPSC/Commissioner/Douglas-Dziak-Peter-A-Feldman/Statement/Statement-of-Commissioners-Peter-A-Feldman-and-Douglas-Dziak-on-CPSC-Chronic-Hazard-Guidelines. Commissioner Trumka submitted a 
statement, available at https://www.cpsc.gov/About-CPSC/Commissioner/Richard-Trumka/Statement/CPSC-Revamps-Chronic-Hazards-Guidelines-Making-It-Easier-to-Protect-You-From-Toxic-Chemicals-in-Your-Home.
---------------------------------------------------------------------------

    Determining whether a product is or contains a hazardous substance 
involves scientific analysis, legal interpretation, and the application 
of policy judgment. The Guidelines are intended to assist firms in 
identifying products that present chronic hazards, to meet their 
labeling obligations under the FHSA and the Labeling of Hazardous Art 
Materials Act (LHAMA). 15 U.S.C. 1277. They are not binding on industry 
or the Commission. Indeed, chronic toxicity may be established in 
various ways. The Commission may determine that a product is a 
hazardous substance due to a chronic hazard based on any evidence that 
is relevant and material to such a determination.

[[Page 30327]]

    For example, peer-reviewed scientific studies by third parties and 
toxicity assessments from CPSC's peer agencies may be relevant and 
material evidence to establish chronic toxicity and that a substance is 
a ``hazardous substance'' under the FHSA. Likewise, evidence from third 
parties may be useful to determine chronic toxicity. For instance, 
third party studies may indicate that chronic adverse health effects 
are associated with foreseeable levels of consumer exposure, allowing 
the Commission to conclude that the FHSA's criteria for a ``hazardous 
substance'' are satisfied. Other cases, however, may require original 
research to fill gaps in knowledge.
    In addition, while the Guidelines describe certain toxic endpoints, 
they do not limit the toxic endpoints the Commission may consider. The 
Commission may consider all forms of personal injury or illness as 
potential toxic endpoints.
    The Chronic Hazard Guidelines, which should be understood as a set 
of best practices, are not mandatory for the Commission or for 
stakeholders. The guidelines describe methods that CPSC staff may use 
to assess chronic hazards under the FHSA. Furthermore, the guidelines 
are intended to be sufficiently flexible to incorporate the latest 
scientific information, such as advances in risk assessment 
methodology. Risk assessors may deviate from the default assumptions 
described in the guidelines, provided that their methods and 
assumptions are documented, scientifically defensible, and supported by 
appropriate data as indicated in section VI.A.2 of the preamble of the 
guidelines. 57 FR 46633. However, given that the guidelines represent 
an available set of best practices, risk assessors are encouraged to 
use the information and approaches outlined therein where appropriate.
    In the years since the guidelines were issued, there have been 
numerous advances in the basic science underlying the guidelines, such 
as the use of transgenic animals to elucidate mechanisms of 
carcinogenicity and toxicity. There also have been several changes in 
the practice of risk assessment, including wider acceptance and use of 
risk assessment methods such as the benchmark dose approach and 
probabilistic exposure assessment. Therefore, CPSC is finalizing two 
guidance documents to supplement the 1992 guidelines.
    The first supplement provides guidance for the application of 
benchmark dose methodology (BMD) to risk assessment. This supplement 
discusses an alternative to the traditional approach described in the 
original guidelines for estimating acceptable daily intakes (ADIs) for 
carcinogenic and other hazards, such as neurotoxicological or 
reproductive/developmental hazards. The second supplement is guidance 
for the analysis of uncertainty and variability, including use of 
probabilistic risk assessment methodology, which is most relevant to 
exposure assessment.
    Like the 1992 guidelines, the supplemental guidance documents are 
not mandatory. Rather, they describe methods that CPSC staff and 
manufacturers may use to evaluate chronic hazards. The guidelines are 
intended to assist manufacturers in complying with the requirements of 
the FHSA and to facilitate the use of reliable risk assessment 
methodologies by both manufacturers and CPSC staff.

II. Response to Comments

    In response to the Commission's August 2023 Notice of Availability 
of the proposed supplemental guidance, the Commission received two 
comments. The commenters were the National Center for Health Research 
(NCHR) and one individual, Albert Donnay. They had questions about the 
timing of the release of the guidance, technical details of benchmark 
dose modeling, how to determine risk assessment approaches in the 
context of the guidance, and the citation of references after the 2008 
peer review of the supplemental guidance.
    Comment 1: NCHR noted that time has passed since a draft of the 
Supplemental Guidance was peer reviewed in 2008.
    Response 1: Although the Supplemental Guidance might have been 
finalized earlier, the methods and approaches described in the Chronic 
Hazard Guidelines and the Supplemental Guidance are neither mandatory 
nor proscriptive. Publication of the Supplemental Guidance does not 
change the Commission's substantive policies. As before, risk assessors 
are encouraged to use modern and applicable approaches to identify and 
quantify consumer product chemical hazards and risks, provided that 
methods and assumptions are documented, scientifically defensible, and 
supported by appropriate data.
    Comment 2: NCHR questioned whether it is appropriate to recommend 
using linear modeling of benchmark dose assessment for all carcinogens 
and non-carcinogens.
    Response 2: Linear dose-response modeling describes a constant 
proportional increase in a biological response (e.g., toxicity) as the 
dose or exposure level increases and is often used for low dose cancer 
risk assessments. Contrary to this comment, the supplemental guidance 
does not recommend linear modeling for all carcinogens and 
noncarcinogens. For non-cancer endpoints, the supplemental guidance 
specifically states that ``a non-linear dose response is generally 
presumed. . . .'' On the other hand, for cancer risk, the Commission 
prefers linear extrapolation to the background level from the BMD as a 
point of departure (PoD). However, the guidance also describes that a 
non-linear dose response with use of uncertainty factors may be used if 
there is convincing evidence that the dose response is non-linear at 
low doses. The preference for the linear assumption is based on 
theoretical considerations of carcinogenicity, as well as modeling 
considerations, which are described in detail in the Chronic Hazard 
Guidelines and the Supplemental Guidance. The supplemental guidance 
also states that risk assessors may use methods other than those 
described in the guidelines, provided that their methods and 
assumptions are documented, scientifically defensible, and supported by 
appropriate data.
    Comment 3: NCHR requested more specific guidance as to the 
conditions under which it would be acceptable to deviate from the 
assessment methodology outlined in the guidance.
    Response 3: CPSC's reference to the use of professional judgment is 
based on its expectation that the risk assessor has the training, 
expertise, and experience to analyze datasets using the tools and 
approaches that are most appropriate and relevant to meet the needs and 
requirements for each assessment. The Commission understands that a 
variety of tools, models, and methods currently exist, and anticipates 
further advancements in this science. Thus, the supplemental guidance 
reiterates that expertise and professional judgment are required when 
applying the guidelines and emphasizes that the guidelines cannot be 
applied mechanically.
    Comment 4: Albert Donnay asked when these supplements were most 
recently revised, what contractor(s) contributed to the latest 
revisions if they were not done solely by staff, and how many 
independent scientists with expertise in either BMD or PRA reviewed the 
post-2008 revisions before they were published in the FR.
    Response 4: After the peer review of the supplements conducted in 
2008, CPSC staff revised and updated the proposed supplements to 
incorporate discussion of more recently released

[[Page 30328]]

tools, such as benchmark dose software packages and supporting guidance 
documents from the U.S. Environmental Protection Agency (EPA) and the 
Dutch National Institute for Public Health and the Environment (RIVM). 
In addition, CPSC staff updated the references in the draft 
supplemental guidance to include literature published after 2008 and 
assessed that the more recent literature did not indicate a need for 
revision of the draft supplemental guidance or for additional 
independent review. These updates were performed by CPSC staff without 
participation of contractors.
    Having considered the comments, the Commission is finalizing the 
guidance as proposed, without changes. The Final Supplemental Guidance 
for the Use of Benchmark Dose Methodology in Risk Assessment and Final 
Supplemental Guidance for the Analysis of Uncertainty and Variability 
in Risk Assessment are stated in sections III and IV.

III. Final Supplemental Guidance for the Use of Benchmark Dose 
Methodology in Risk Assessment

A. Background

    In 1992, the U.S. Consumer Product Safety Commission (CPSC) issued 
guidelines for assessing chronic hazards under the Federal Hazardous 
Substances Act (FHSA) and the Labeling of Hazardous Art Materials Act 
(LHAMA), including carcinogenicity, neurotoxicity, reproductive/
developmental toxicity, exposure, bioavailability, risk assessment, and 
acceptable risk (CPSC 1992). 57 FR 46626. The chronic hazard 
guidelines, which are not mandatory for CPSC or stakeholders, are 
intended as an aid to manufacturers in making their determination of 
whether a product is a hazardous substance due to chronic toxicity, and 
thus would require labeling under the FHSA. The guidelines describe 
methods that CPSC staff use to assess chronic hazards under the FHSA. 
Furthermore, the guidelines are intended to be sufficiently flexible to 
incorporate the latest scientific information, such as advances in risk 
assessment methodology. Risk assessors may deviate from the default 
assumptions described in the guidelines, provided that their methods 
and assumptions are documented, scientifically defensible, and 
supported by appropriate data. However, given that the guidelines 
represent an available set of best practices, risk assessors are 
encouraged to use the information and approaches outlined therein where 
appropriate, and other methods will be reviewed by staff to determine 
acceptability.
    In the years since the guidelines were issued, there have been 
numerous advances in the basic science underlying the guidelines, such 
as the use of alternative methods to elucidate mechanisms of 
carcinogenicity and toxicity. There also have been several changes in 
the practice of risk assessment, such as in the assessment of risks to 
children, as well as wider acceptance and use of risk assessment 
methods such as the benchmark dose approach and probabilistic exposure 
assessment. Therefore, CPSC staff-initiated reviews of the existing 
chronic hazard guidelines and is recommending additions or changes, as 
appropriate. The purpose of this document is to describe supplemental 
guidance for the application of the benchmark dose approach in risk 
assessment.
    The current scientific knowledge regarding the risk assessment of 
chronic hazards is such that the guidelines cannot be applied 
mechanically (CPSC 1992, section VI.A.2, page 46633). Rather, 
considerable expertise and professional judgment are required to apply 
the guidelines properly. Furthermore, the volume of scientific 
literature on chronic hazard risk assessment, in general, and the 
benchmark dose, in particular, is extensive. Therefore, the discussion 
and guidance described below are not intended to explain how to perform 
chronic hazard risk assessments using the methods described. The 
guidelines assume that the reader has the necessary expertise. In 
addition, the discussion presented here is necessarily brief. The risk 
assessor is referred to the literature on benchmark dose, only a 
portion of which is cited here.

B. Discussion

    The benchmark dose (BMD) approach (Crump 1984a; Crump et al. 1995) 
is an alternative to the traditional method of deriving acceptable 
daily intake (ADI) \2\ levels by using no observed adverse effect 
levels (NOAELs) \3\ and lowest observed adverse effect levels (LOAELs). 
The BMD may be used for both cancer and non-cancer endpoints, quantal 
or continuous data, and animal or human data. The BMD is an estimate of 
the dose level for a particular response. For example, the 
BMD10 is the best estimate of the dose at an excess risk 
(risk over background) of 10%, and the BMDL10 is the lower 
confidence limit (LCL) of the BMD10. The benchmark response 
(BMR) level is the response level selected for deriving an ADI level or 
cancer unit risk (slope factor).\4\ The BMR is within or near the 
observable range of the bioassay used to derive the ADI or unit risk. 
Typically, selected BMR's range from 1% to 10% excess risk. To derive 
an ADI for non-cancer endpoints, the BMD is divided by the same 
uncertainty (safety) factors that are normally applied to the NOAEL. 
For cancer risk, the BMD is used as a ``point of departure'' (PoD) for 
linear extrapolation to the background level (EPA 2005). However, 
uncertainty factors may be applied for cancer risk if there is 
convincing evidence for a non-linear dose response at low doses.
---------------------------------------------------------------------------

    \2\ The ADI is an estimate of the amount of a chemical a person 
can be exposed to on a daily basis over an extended period of time 
(up to a lifetime) with a negligible risk of suffering deleterious 
effects. The ADI is roughly equivalent to a ``reference dose'' or 
``tolerable daily intake.''
    \3\ In the chronic hazard guidelines, ``NOEL'' is used 
synonymously with ``NOAEL,'' because only adverse effects are 
relevant under the FHSA.
    \4\ The term ``unit risk'' is used synonymously with ``slope 
factor'' (CPSC 1992).
---------------------------------------------------------------------------

1. Advantages of the BMD Approach
    The advantages of the BMD approach have been described in detail 
elsewhere (Barnes et al. 1995; Crump 1984a; Crump et al. 1995; Gaylor 
et al. 1998; EPA, 2012; Filipsson et al. 2003). For example, the NOAEL 
and LOAEL are limited to the doses tested in the bioassay. In contrast, 
the BMD is not limited to the doses tested in the bioassay. Thus, the 
BMD provides a more consistent basis for comparisons between studies 
that did not use the same dose levels.
    The true (parametric) value of the BMD is independent of the study 
design, such as the number of animals per dose group, n. However, the 
NOAEL is sensitive to n. The NOAEL is not a threshold, although it is 
frequently regarded as such. Rather, it is more appropriate to regard 
the NOAEL as a limit of detection. The incidence of adverse effects may 
be as high as 20% at the NOAEL. A given dose level may be a NOAEL in a 
study with small n if the incidence is not significantly different from 
background. However, the same dose in a larger study may be a LOAEL due 
to the increased sensitivity resulting from a larger n. The traditional 
NOAEL approach ``rewards'' studies with small n, by resulting in higher 
(i.e., less protective) NOAELs. Conversely, the traditional approach 
``penalizes'' studies with larger n, by resulting in lower (more 
protective) NOAELs. Thus, the traditional method is a disincentive to 
performing better, larger studies. In contrast, the BMD is essentially 
independent of n and,

[[Page 30329]]

therefore, does not penalize studies with a larger n.
    The BMD approach may account for variability in the bioassay. If 
the BMDL is used, larger studies tend to have smaller confidence 
intervals. Thus, larger studies are generally rewarded, because a 
smaller confidence interval leads to a higher BMDL. In contrast, poorly 
designed studies with inadequate sample size are penalized by having 
larger confidence intervals, leading to a lower BMDL.
    The BMD accounts for the slope and shape of the dose response curve 
and uses all of the dose response data from the study. In contrast, the 
NOAEL or LOAEL relies on the response at only one dose level. Thus, 
information on the slope and shape of the dose response curve is 
ignored.
    With the BMD approach, the methodology is the same regardless of 
whether a NOAEL is established. An additional uncertainty factor that 
is generally applied when using the LOAEL is not required in a BMD 
analysis, because the BMD can still be estimated even if a NOAEL has 
not been established.
    While there are several advantages to the BMD approach, the 
principal disadvantage is the added complexity of the methodology. BMD 
methods require expertise in statistics, as well as toxicology. The 
additional steps involved in the analysis also increases the number of 
decision points, such as the choice of BMD and mathematical model, 
which require professional judgment. This, in turn, increases the 
number and possibly the range of possible ADI values from a given data 
set and may lead to areas of disagreement among risk assessors.
2. BMD Methodology
    While the overall BMD approach is straightforward, there are many 
factors that must be considered in applying BMD methods in risk 
assessment, including the selection of the most appropriate endpoint 
and data set, dose response model, statistical methods, and selection 
of the BMD. Each of these factors requires knowledge of toxicology and 
risk assessment, as well as professional judgment.
a. Selection of the Endpoint and Data Set to Model
    Initially, the selection of the critical study and endpoint to 
model is similar to the traditional approach. The study should be well-
designed and executed, with an adequate number of animals and doses, 
and a statistically significant effect (CPSC 1992, sections VI.C.3.a, 
p. 46639; VIC.3.b, p. 46640; VI.D.2.a, p. 46642; and VI.D.3.b, p. 
46643). There should be a dose where there are no observed adverse 
effects, i.e., at or near the NOAEL. The selection of the critical 
endpoint is based, in part, on the judgment of the toxicologist or 
pathologist regarding the biological significance of the endpoint. When 
multiple studies, multiple endpoints, or multiple species are 
available, generally the most sensitive dose response is used (CPSC 
1992, section F.4.b.ii, p. 46656).
    It should be noted that the study with the lowest NOAEL will not 
necessarily lead to the lowest BMD, because the BMD also depends on the 
slope of the dose-response curve. Therefore, all relevant endpoints and 
studies should be modeled (Filipsson et al. 2005) to ensure that the 
lowest BMD is identified.
    Additionally, the data set must be amenable to modeling. That is, 
there should be a steadily increasing dose response that is not 
saturated at the high doses. If none of the available dose response 
models can adequately fit the data (see below), the BMD approach cannot 
be used.
b. Selection of the Dose Response Model
    The BMD approach is essentially a curve-fitting exercise. The 
choice of the dose-response model does not require any knowledge of the 
mode of action. Thus, the form of the model is not necessarily 
prescribed or dictated by any specific information about the studied 
activity, provided that it adequately describes the data. In some 
instances, however, mechanistic information may suggest a particular 
model, such as the Hill model when cooperative binding is observed.
    A variety of dose-response models have been used to estimate the 
BMD (Crump 1984a; Crump et al. 1995; EPA 2022; Filipsson et al. 2003; 
Gaylor et al. 1998). The BMD approach may be applied to either quantal 
(dichotomous) or continuous data. Incidence data, such as the number of 
animals with a certain adverse effect, are quantal. Serum enzyme or 
hormone levels are examples of continuous data. Generally, quantal and 
continuous data require different, though related, dose response 
models. Nested quantal models may be used with developmental studies to 
evaluate effects within and between litters.
    Dose response models for quantal data include linear (one-hit), 
quadratic, gamma multi-hit, Weibull, polynomial (multistage), logistic, 
log-logistic, probit, and log-probit models. These are slightly 
modified versions of the dose response models that have been used for 
cancer risk assessment (compare Crump 1984b; Zeise et al. 1987). The 
linear, quadratic, and Weibull models are essentially subsets of the 
polynomial model. Therefore, some or all of these models may yield 
similar results for certain data sets, such as when the dose response 
is linear. Dose response models for continuous data include linear, 
quadratic, linear-quadratic, polynomial, power, and Hill models. In 
addition, nested models are available for developmental studies. The 
mathematical forms of the models are described in detail elsewhere 
(Crump 1984a; Crump et al. 1995; EPA 2022; Filipsson et al. 2003; 
Gaylor et al. 1998).
    In applying the BMD approach to non-cancer endpoints, the dose 
response models are not used for low-dose extrapolation. Thus, in 
contrast to cancer risk assessment, there is no need to consider the 
shape of the curve at low doses. Therefore, the choice of dose response 
model depends, in large part, on the goodness of fit. That is, the 
model (or models) selected must adequately describe the data. A model 
is generally rejected if the probability based on chi-square is less 
than 0.05. In other words, if the probability that the deviation of the 
data from the model is due to random variability is less than 0.05, the 
model does not adequately describe the data. Depending on the data set, 
multiple models may provide a similar global fit to the data. In this 
case, the local fit in the low-dose range, that is, the doses nearest 
the BMR, may be considered. In practice, different models often result 
in roughly similar BMDs, provided that they adequately describe the 
data. In any case, the results from different models and the choice of 
model should be discussed.
    In some cases, it may be necessary to exclude high dose data from 
the model fitting procedure, to improve the goodness of fit. Data at 
the highest doses of a multiple dose bioassay may be considered to be 
less informative for the purpose of low dose extrapolation, especially 
in cases where the responses plateau at the high doses. Therefore, high 
dose groups may be systematically eliminated until the fit is 
acceptable (Anderson 1983).
    In other cases, such as when a non-monotonic dose response is 
observed, none of the dose response models may be able to fit the data 
adequately. When this occurs, the BMD approach should not be used. 
While the NOAEL/LOAEL approach could still be applied, the quality of 
the study should be given careful consideration. It may not be 
appropriate to derive an ADI by any method from such a data set.
    The steps for estimating the BMD may be summarized as follows:

[[Page 30330]]

     Select the bioassay(s) and endpoint(s) to model.
     Determine whether the data are quantal or continuous.
     Fit the bioassay data set(s) to several dose response 
models and determine the goodness of fit. Calculate multiple BMDs, 
including maximum likelihood estimates (MLEs) of risk and confidence 
limits. Graph the results.
     Select which model to use for determining the ADI. 
Generally, the model giving the best fit is used. If multiple models 
fit the data well, the local fit near the BMR may be considered. In 
some cases, the choice of model may be based on mechanistic 
considerations. If no model fits the data adequately, the BMD approach 
should not be used.
     If multiple endpoints or bioassays are modeled, select 
which to use for determining the ADI. The most sensitive dose response 
is generally used (CPSC 1992, section F.4.b.ii, page 46656). Other 
factors, such as severity of the effect may also be considered.
     Select which BMD (BMR) to use for deriving the ADI.
     Discuss and explain all of the decision points in the 
preceding steps.
c. Statistical Methods
    Various types of software may be used to estimate the BMD/BMDL. The 
U.S. Environmental Protection Agency (EPA) has developed Benchmark Dose 
Software (BMDS) specifically for this purpose (EPA 2022). The BMDS and 
associated documentation are in the public domain and may be downloaded 
from the EPA website. Software is also available from the Netherlands 
Ministry of the Environment (RIVM 2021) and Shao and Shapiro (2018). 
Various other statistical software packages (e.g., SAS, and R) may also 
be used. Likelihood methods are generally preferred for estimating the 
BMD and confidence limits (Crump 1984a; Crump and Howe 1985; Crump et 
al. 1995; Gaylor et al. 1998; EPA 2001). Goodness of fit is typically 
based on the chi-square distribution.
    As with cancer risk assessment, CPSC staff prefers to use extra 
risk, rather than additional risk, as a measure of the risk over 
background. Extra risk applies Abbott's correction, so that animals 
which already have a given lesion from background processes are not 
considered at risk for an exposure-induced lesion of the same type. The 
numerical difference between extra risk and additional risk is small, 
provided that the background risk is sufficiently low (<0.25). Extra 
risk (Crump and Howe 1985) is defined by:
[GRAPHIC] [TIFF OMITTED] TN23AP24.039

where:

PE is the extra risk, PD is the risk at dose 
D, and P0 is the background dose.

    Additional risk is defined by:

    [GRAPHIC] [TIFF OMITTED] TN23AP24.041
    
where:

PA is the additional risk.
d. Selection of the Benchmark Dose (BMD)--Quantal Data
    The ADI is the dose at which the risk of an adverse effect is 
considered negligible. Because such risks cannot be directly measured, 
this requires assumptions about the shape of the dose response curve in 
the low dose region. For cancer, there are theoretical reasons for 
assuming a linear response at low dose, such as the probability that a 
given chemical will interact with background processes or other 
chemicals (CPSC 1992, VI.F.3.b.ii, page 46654). For non-cancer 
endpoints, a non-linear dose response is generally presumed, although 
the shape and slope of this curve outside of the observable range is 
unknown.
    The selection of the BMD has been based on the following 
considerations: (i) The BMD should be within or near the observable 
range of the bioassay. (ii) It is roughly the dose at which a 
statistically significant effect may be observed in the bioassay (Crump 
et al. 1995). Thus, BMD's of 5% to 10% over background are typically 
used for quantal data, assuming that there is an adequate number of 
animals and the background level is not exceptionally high. (iii) The 
BMD approach is an alternative to deriving the ADI from a NOAEL. The 
BMD has generally been selected to approximate the NOAEL (Crump et al. 
1995). Thus, the study selected for estimating the BMD should include a 
dose at or near the NOAEL. Other factors, such as the shape of the dose 
response curve or the study design (e.g., CPSC 2001, 2002), may be 
considered on a case-by-case basis. For example, it may be desirable to 
select a BMD that is reflective of nonlinearity or an inflection point 
in the dose response curve (Murrell et al. 1998).
    It is important to keep in mind that the selection of a BMD is part 
of the overall risk assessment process, which includes the selection of 
the critical endpoint and uncertainty factors, among other things. The 
overall process is equally as important as the individual steps. For 
example, the risk assessor might consider applying different 
uncertainty factors, depending on the BMD selected. That is, 
consideration could be given to larger or additional uncertainty 
factors if the BMD is higher than is typical, or to smaller uncertainty 
factors if the BMD is exceptionally low.
    Numerous authors (Barnes et al. 1995; Crump 1984a; Filipsson et al. 
2003) and the EPA (EPA 2005) generally recommend using the 95% lower 
confidence limit (LCL) of the benchmark, typically the 
BMDL05 or BMDL10. This generally satisfies the 
criteria listed above. In a typical bioassay, the LCL is within or near 
the observable range, it is near the lowest detectable response, and it 
is roughly equivalent to the NOAEL. Using the LCL takes into account 
the uncertainty in the bioassay and tends to reward larger or better 
studies, which generally have narrower confidence intervals. On the 
other hand, it has been argued that using the LCL rather than the best 
estimate (maximum likelihood estimate or MLE) leads to a BMD that may 
depend more on experimental uncertainty than on the dose response 
itself (Murrell et al. 1998). Thus, using the LCL tends to defeat one 
of the principal advantages of the BMD approach, which is to make use 
of the

[[Page 30331]]

shape and slope of the dose-response curve in the analysis.
    While the choice of the BMD should be made on a case-by-case basis, 
it is desirable to have a default value for the purpose of consistency 
across different chemicals, endpoints, and risk assessors. However, 
even if the default value is used, the risk assessor must evaluate 
whether the default is appropriate in a given case, using the criteria 
described above. Risk assessors have most frequently used 
BMDL05 or BMDL10 to derive ADIs (or RfDs) (see 
above). The Chronic Hazard Advisory Panel (CHAP) convened by CPSC (CPSC 
2001) and CPSC staff (CPSC 2002) used the BMD05 to set an 
ADI level for diisononyl phthalate. Health Canada also uses the 
BMD05 to set tolerable intake levels. One advantage of using 
the MLE is that it is more reflective of the shape of the dose response 
than the LCL (Murrell et al. 1998).
    For cancer risk assessment, CPSC prefers to use the MLE risk (see 
below). However, as currently applied, the ADI is not regarded as a 
numerical estimate of risk, as is the case for cancer risk. Rather, it 
is regarded as a regulatory threshold, that is, a ``negligible risk 
level'' or ``virtually safe dose.'' Therefore, the reasons for using 
the MLE to estimate cancer risk do not necessarily apply to ADIs. This 
conclusion may change in the future, if true risk-based approaches are 
applied to non-cancer endpoints.
    At the present time it seems reasonable to use the BMD05 
(i.e., the MLE) rather than the BMDL05 (i.e., the LCL) as a 
default value, subject to the limitations discussed above. This is 
consistent with the CPSC approach to estimating cancer risk and with 
previous CPSC applications of the BMD approach. In addition, the MLE 
better reflects the shape of the dose response, as compared to the LCL.
e. Selection of the Benchmark Dose (BMD)--Continuous Data
    For continuous data, the BMD value is generally a level that is 
considered ``adverse.'' This is a matter of professional judgment by 
health scientists, such as toxicologists and pathologists, and must be 
determined on a case-by-case basis. As discussed in the previous 
section on ``Selection of the Benchmark Dose (BMD)--Quantal Data,'', 
the MLE value is preferred for risk assessment. In instances where 
there is no consensus on what constitutes an adverse effect, some risk 
assessors have used a relative change in the endpoint, such as a change 
of one standard deviation.
3. Cancer Risk Assessment
    The multistage model (Crump 1984b) has been preferred by most 
federal agencies for cancer risk assessment. The multistage model is 
defined by:

[GRAPHIC] [TIFF OMITTED] TN23AP24.042

where:

D, dose; PD, cancer risk at dose D; and q0 . . 
. q9, parameters to be fitted by the model.

    The EPA has preferred to use the upper confidence limit (UCL) of 
the estimated risk, while CPSC staff uses the MLE risk, unless the 
linear term (q1) is zero. When q1 is zero, the 
UCL risk is used to ensure linearity at low doses (CPSC 1992, 
VI.F.3.b.ii, page 46654).
    EPA began to use the BMD approach for cancer risk assessment in 
place of the multistage model in 2005 (EPA 2005). BMD is the preferred 
method for dose response assessment at EPA and other agencies (Allen et 
al. 2011). The default procedure is to use the BMR as a point of 
departure (PoD) for linear extrapolation to the background level. 
Uncertainty factors may be applied if there is sufficient reason to 
rule out a linear dose response at low doses. This procedure is 
analogous to the Mantel-Bryan procedure (Mantel & Bryan 1961; see also 
Gaylor & Kodell 1980) that was commonly used before the multistage 
model became available.
    The BMD approach described by EPA is consistent with the default 
procedures used by CPSC staff under the guidelines. The primary concern 
of CPSC staff is that linear extrapolation should remain the default 
procedure for guidelines purposes. The results from using the BMD 
methodology and the multistage model are not substantially different 
when linear extrapolation is assumed. In general, a non-linear dose 
response with use of uncertainty factors should be used only if there 
is convincing evidence that the dose response is non-linear at low 
doses. In addition, the BMD approach offers certain advantages over the 
multistage model as applied by CPSC staff. While staff prefers to use 
the MLE estimate of cancer risk, it is necessary to use the UCL risk in 
cases where the linear term (q1) is zero. By using the BMD 
approach, the MLE risk can be used in all cases. Thus, the process is 
simplified. In addition, staff use the BMD approach for non-cancer 
endpoints, BMD methods are used by EPA and other agencies for both 
cancer and non-cancer risk assessment, and the software is widely 
available.
    The practice of the CPSC Directorate for Health Sciences (HS) is to 
present the best estimate of risk, rather than the upper bound, to risk 
managers. Thus, HS prefers the MLE of risk in cancer risk assessments 
(CPSC 1992, section VI.F.3.b.iii). Presenting the best estimate of risk 
depends on a number of considerations: (i) CPSC does not routinely 
define ``safe'' levels, as is frequently done by other agencies such as 
the Food and Drug Administration (FDA) and EPA. Rather, the need for 
CPSC actions based on unsafe levels are typically determined on a case-
by-case basis. (ii) For typical cancer bioassays in animals, the 
difference between the MLE and 95% upper confidence limit (UCL) \5\ is 
generally small, about 2- to 3-fold. (iii) The overall risk assessment 
process is designed to include assumptions that tend to err on the side 
of safety when data are lacking for a particular part of the 
assessment. Thus, there is always a possibility of compounding safety 
assumptions which could result in some cases in unrealistic estimates. 
Therefore, the use of the MLE rather than the UCL generally has a small 
effect on numerical estimates.
---------------------------------------------------------------------------

    \5\ The UCL risk corresponds to the LCL dose.
---------------------------------------------------------------------------

    Therefore, the BMD approach with linear extrapolation and based on 
the MLE risk generally will be the default procedure for cancer risk 
assessments performed by CPSC staff. To further simplify the process, 
the multistage (polynomial) model generally will be the default model 
for cancer risk. However, other models that adequately describe the 
data may be used, as described above for non-cancer endpoints. While 
the choice of a PoD is not critical, the default will be the 
BMD05 (see above). Although the BMD approach will be the 
default procedure, the multistage model, as described above, can still 
be used. Risk assessors may deviate from the default assumptions 
described in the guidelines, provided that their methods and 
assumptions are documented, scientifically defensible, and supported by 
appropriate data (CPSC 1992, section VI.A.2).

[[Page 30332]]

    The following practices are recommended when applying benchmark 
dose methodology:
     The BMD approach is generally the preferred method for 
setting ADI levels for non-cancer endpoints, provided that adequate 
dose response data are available.
     Appropriate dose response models and statistical methods 
have been described in detail elsewhere (Crump 1984a; Crump et al. 
1995). Public domain software is available from EPA (EPA 2022).
     The BMD response level (BMR) used to calculate the ADI 
will be determined on a case-by-case basis. A range of BMR's, including 
best estimates and lower confidence limits, should be considered.
     As a default, CPSC staff will use the maximum likelihood 
estimate of the dose at which the extra risk is 5% (BMD05). 
The same uncertainty factors currently applied to the NOAEL will be 
applied to the BMD.
     Several dose response models should be considered. 
Generally, the model that best describes the observed dose response 
data will be selected to derive the ADI. In addition, the ADI will 
generally be based on the combination of dose response model, endpoint, 
and study that lead to the lowest ADI.
     Risk assessors may deviate from the default assumptions 
described in the guidelines, provided that their methods and 
assumptions are documented, scientifically defensible, and supported by 
appropriate data (CPSC 1992, section VI.A.2). While the BMD approach is 
typically preferred, the traditional method based on NOAELs/LOAELs may 
still be used.
    In addition, the BMD approach with linear extrapolation and based 
on the MLE risk will be the default procedure for cancer risk 
assessments performed by CPSC staff. The multistage (polynomial) model 
will be the default model for cancer risk. However, other models that 
adequately describe the data may be used, as described above for non-
cancer endpoints. While the choice of a PoD is not critical, the 
default will be the BMD05. Linear extrapolation from the PoD 
generally will be used unless there is convincing evidence that the 
dose response will be non-linear at low doses (CPSC 1992, VI.F.3.b.ii, 
page 46654). In cases where a non-linear dose response is justified, 
uncertainty factors may be applied as described for non-cancer 
endpoints. Although the BMD approach will be the preferred procedure, 
the multistage model, as traditionally applied by CPSC, can still be 
used.

C. Summary

1. Estimation of the Acceptable Daily Intake for Non-Cancer Endpoints
    The following supplements the guidance on estimating acceptable 
daily intakes (ADIs) in the CPSC Chronic Hazard Guidelines at 57 FR 
46656 (Oct. 9, 1992) in section VI.F.4.b.1.ii. This does not supersede 
the 1992 guidance; rather, it provides guidance on the use of newer 
methods for estimating ADIs.
    Traditionally, CPSC staff derived acceptable daily intake (ADI) 
levels for non-cancer endpoints by applying safety factors (uncertainty 
factors) to the no-observed-effect level (NOAEL) or lowest-observed-
effect-level (LOAEL). However, the benchmark dose (BMD) approach is now 
generally preferred over the traditional method. The benchmark dose is 
an estimate of the dose at a certain risk level. The BMD is estimated 
from a dose-response model. The advantages of the BMD approach and 
methods for estimating the BMD are described elsewhere (Barnes et al. 
1995; Crump 1984; Crump et al. 1995; EPA 2012; Filipsson et al. 2003; 
Gaylor et al. 1998). Software for estimating the BMD is available from 
the U.S. EPA (EPA 2022) and other sources. In estimating the BMD, the 
risk assessor should consider the following points: (a) The dose-
response model must provide an adequate fit to the data; the BMD 
approach may not be appropriate for all data sets. (b) Alternative dose 
response models should be considered, and the choice of model to derive 
the ADI explained. (c) Alternative endpoints and studies should also be 
considered, as appropriate. (d) A range of BMD response levels, 
including best estimates and confidence intervals should be evaluated. 
(e) Generally, different methods are required for dichotomous and 
continuous data.
    The BMD selected to derive the ADI (BMD response level) is 
determined on a case-by-case basis. The BMD response level (BMR) must 
be within or near the range of experimental dose levels. As a default, 
for dichotomous (i.e., incidence) data, the BMR will be the maximum 
likelihood estimate of the dose associated with an extra risk (risk 
over background) of 5% (BMD05). For continuous data, (e.g., 
enzyme or hormone levels), the BMD is generally based on the level 
considered to be an adverse effect. The default safety (uncertainty) 
factors described above (10-fold for human data and 100-fold for animal 
data) are applied to the BMD CPSC 1992, section VI.F.4.b.1.ii; Haber et 
al. 2018). Thus, the ADI is generally 100-fold lower than a BMD based 
on animal data. An additional uncertainty factor for ADIs based on a 
LOEL is not needed. While the BMD approach is preferred, the 
traditional method of applying safety factors to the NOAEL or LOAEL may 
still be used.
2. Estimation of Cancer Risk
    The following is a supplement to the CPSC Chronic Hazard Guidelines 
at 57 FR 46654 (Oct. 9, 1992), section VI.F.3.b.ii.
    Traditionally, CPSC staff estimated cancer unit risks (slope 
factors) using the multistage model (Global83). The maximum likelihood 
estimate (MLE) of risk was used unless the linear term (q1) 
was equal to zero; in this case, the upper confidence limit of risk was 
used. However, the benchmark dose (BMD) approach with linear 
extrapolation based on the MLE risk is now generally preferred over the 
traditional method. The multistage (polynomial) model will be the 
default model for cancer risk. However, other models that adequately 
describe the data may be used, as described above for non-cancer 
endpoints. The choice of a BMD response level (BMR) or point-of-
departure (PoD) will be made on a case-by-case basis. In general, the 
default PoD will be the MLE estimate of the dose associated with an 
extra risk (risk over background) of 5% (BMD05). Linear 
extrapolation from the PoD will be used unless there is convincing 
evidence that the dose response will be non-linear at low doses. In 
cases where a non-linear dose response is justified, uncertainty 
factors may be applied as described for non-cancer endpoints. Although 
the BMD approach generally is preferred under the guidelines, the 
traditional CPSC approach based on the multistage model may still be 
used.

D. References

Allen JA, Gift JS, Zhao QJ (2011) Introduction to benchmark dose 
methods and U.S. EPA's benchmark dose software (BMDS) version 2.1.1. 
Toxicology and Applied Pharmacology 254: 181-191.
Anderson EL (1983) Quantitative approaches in use to assess 
carcinogenic risk. Risk Analysis, 3: 277 295.
Barnes DG, Daston GP, Evans JS, Jarabek AM, Kavlock RJ, Kimmel CA, 
Park C, Spitzer HL (1995) Benchmark Dose Workshop: criteria for use 
of a benchmark dose to estimate a reference dose. Regulatory 
Toxicology and Pharmacology 21: 296-306.
Consumer Product Safety Commission (CPSC) (1992) Labeling 
requirements for art materials presenting chronic hazards; 
guidelines for determining chronic toxicity of products subject to 
the FHSA; supplementary definition of ``toxic'' under the Federal 
Hazardous Substances Act; final rules. Federal Register 57: 46626-
46674. October 9, 1992. https://

[[Page 30333]]

www.cpsc.gov/s3fs-public/pdfs/blk_pdf_chronichazardguidelines.pdf.
Consumer Product Safety Commission (CPSC) (2001) Chronic Hazard 
Advisory Panel on Diisononyl Phthalate (DINP). U.S. Consumer Product 
Safety Commission, Bethesda, MD 20814. June 2001. http://www.cpsc.gov/library/foia/foia01/os/dinp.pdf.
Consumer Product Safety Commission (CPSC) (2002) Updated risk 
assessment of oral exposure to diisononyl phthalate (DINP) in 
children's products. U.S. Consumer Product Safety Commission, 
Bethesda, MD 20814. August 2002. http://www.cpsc.gov/library/foia/foia02/brief/briefing.html (TAB L).
Crump KS (1984a) A new method for determining allowable daily 
intakes. Fundamental and Applied Toxicology 4: 854-871.
Crump KS (1984b) An improved procedure for low-dose carcinogenic 
risk assessment from animal data. Journal of Environmental 
Pathology, Toxicology and Oncology 5: 339-348.
Crump KS, Allen BA, Faustman E (1995) The Use of the Benchmark Dose 
Approach in Health Risk Assessment. Risk Assessment Forum, U.S. 
Environmental Protection Agency, Washington, DC 20460. February 
1995. EPA/630/R-94/007. https://www.epa.gov/nscep.
Crump KS, Howe RB (1985) A review of methods for calculating 
statistical confidence limits in low dose extrapolation. In 
``Toxicological Risk Assessment,'' Volume I. Clayson DB, Krewski D, 
Munro I, editors. CRC Press, Boca Raton, FL. Pages 187-203.
Environmental Protection Agency (EPA) (2022) Benchmark Dose Tools. 
U.S. Environmental Protection Agency, Washington, DC 20460. https://www.epa.gov/bmds. Accessed January 4, 2022.
Environmental Protection Agency (EPA) (2012) Benchmark Dose 
Technical Guidance Document--External Review Draft. Risk Assessment 
Forum, U.S. Environmental Protection Agency, Washington, DC 20460. 
June 2012. EPA/630/R-12/001. https://www.epa.gov/sites/default/files/2015-01/documents/benchmark_dose_guidance.pdf.
Environmental Protection Agency (EPA) (2005) Guidelines for 
Carcinogen Risk Assessment. Risk Assessment Forum, U.S. 
Environmental Protection Agency, Washington, DC 20460. March 2005. 
EPA/630/P-03/001B. https://www.epa.gov/sites/default/files/2013-09/documents/cancer_guidelines_final_3-25-05.pdf https://www.epa.gov/risk/guidelines-carcinogen-risk-assessment.
Filipsson AF, Sand S, Nilsson J, Victorin K (2003) The benchmark 
dose method--review of available models, and recommendations for 
application in health risk assessment. Critical Reviews in 
Toxicology 33: 505-542.
Gaylor DW, Kodell RL (1980). Linear interpolation algorithm for low 
dose risk assessment of toxic substances. Journal of Environmental 
Pathology and Toxicology 4: 305-12.
Gaylor D, Ryan L, Krewski D, Zhu Y (1998) Procedures for calculating 
benchmark doses for health risk assessment. Regulatory Toxicology 
and Pharmacology 28: 150-164.
Haber LT, Dourson ML, Allen BC, Hertzberg RC, Parker A, Vincent MJ 
(2018) Benchmark dose (BMD) modeling: current practice, issues, and 
challenges. Critical Reviews in Toxicology Volume 48, 2018--Issue 5.
Mantel N, Bryan WR (1961) ``Safety'' testing of carcinogenic agents. 
Journal of the National Cancer Institute 27: 455-70.
Murrell JA, Portier CJ, Morris RW (1998) Characterizing dose-
response I: critical assessment of the benchmark dose concept. Risk 
Analysis 18: 13-26.
RIVM (2021) PROAST. National Institute for Public Health and the 
Environment (RIVM), The Netherlands. September 2021. https://www.rivm.nl/en/proast.
Shao K, Shapiro RJ (2018) A Web-Based System for Bayesian Benchmark 
Dose Estimation. Environmental Health Perspectives 126(1): 017002. 
https://ehp.niehs.nih.gov/doi/10.1289/EHP1289.
Zeise L, Wilson R, Crouch EAC (1987) Dose-response relationships for 
carcinogens: a review. Environmental Health Perspectives 73: 259-
308.

IV. Final Supplemental Guidance for the Analysis of Uncertainty and 
Variability in Risk Assessment

A. Background

    In 1992, the U.S. Consumer Product Safety Commission (CPSC) issued 
guidelines for assessing chronic hazards under the Federal Hazardous 
Substances Act (FHSA), including carcinogenicity, neurotoxicity, 
reproductive/developmental toxicity, exposure, bioavailability, risk 
assessment, and acceptable risk. The guidelines are detailed in a 
Federal Register notice. 57 FR 46626 (Oct. 9, 1992).
    The chronic hazard guidelines are intended as an aid to 
manufacturers in making their determination of whether a product is a 
hazardous substance due to chronic toxicity, and thus would require 
labeling under the FHSA. The guidelines are not mandatory. The 
guidelines describe standard methods CPSC staff may use to assess 
chronic hazards under the FHSA. The guidelines are intended to be 
sufficiently flexible to incorporate the latest scientific information, 
such as advances in risk assessment methodology. Therefore, CPSC staff 
initiated reviews of the existing guidelines and is recommending 
additions or changes, as appropriate. The purpose of this document is 
to describe supplemental guidance for the analysis of uncertainty and 
variability in risk assessment, including the use of probabilistic 
techniques.

B. Discussion

    In toxicological risk assessment, uncertainty is the term used to 
describe the lack of knowledge in the underlying science, such as when 
few measurements of the particular subject have been made. Uncertainty 
may also be associated with the choice of mathematical model used to 
estimate exposure or risk. Variability refers to inherent differences 
due to heterogeneity or diversity in the population or exposure 
variable, such as body weight of people in the exposed population. 
Variability is generally not reducible by improved measurement or 
further study (EPA 1997, 2014).
    The theory and techniques of exposure assessment have been 
discussed in detail elsewhere (CPSC 1992; EPA 2014, 2019; Paustenbach 
2002). Exposure may be measured directly, but, in general, an exposure 
assessment is often based on a mathematical model that combines several 
variables describing the factors that influence exposure. For example, 
an assessment of exposure to a chemical released into the air during 
use of a product will include information about the emission rate into 
the air, the resulting concentration of the chemical in the air, the 
amount of time a person using the product or spent living, working, or 
playing in the area, and the amount of air a person breathes during the 
exposure. For a given exposure scenario, the output of an exposure 
assessment is typically an estimate of the amount of chemical that 
comes into contact with the body, usually expressed per unit of body 
weight per day during a defined period of time or over a lifetime, 
although exposure may be defined in other terms.
    For carcinogens, ``risk'' is the product of the exposure estimate 
and the dose-response value, i.e., the numerical representation of 
cancer risk per unit of daily exposure. For non-carcinogens, the 
exposure estimate is compared with the ``acceptable daily intake'' 
(ADI), which is the level of exposure at which we expect humans not to 
experience harmful health effects. Although there is no numerical 
estimate of ``risk'' in this latter case, one may calculate the hazard 
index (HI), which is the ratio of the estimated exposure to the ADI (HI 
greater than one means that the exposure may be hazardous; HI less than 
one represents negligible risk).
    There is no single, correct way to conduct an exposure or risk 
assessment for purposes of evaluating chronic hazards under the Federal 
Hazardous Substances Act (FHSA) or the Labeling of Hazardous Art 
Materials Act

[[Page 30334]]

(LHAMA). There are, however, important issues and concerns that are 
commonly encountered in risk assessment that should be considered 
regardless of the specific risk assessment approach. Because risk 
assessment is a rapidly advancing field, the discussions here should be 
supplemented with other information from the scientific literature, 
texts, and government agency guidance, as scientifically appropriate.
    In most cases, the risk assessor will consider uncertainty and 
variability in the assessment and, at a minimum, include a discussion 
of the effect of uncertainty and variability on the final risk 
estimates. The discussion may be qualitative or it may include 
quantitative estimates of uncertainty and variability. Variability and 
uncertainty are distinct issues and should be considered separately in 
each analysis using appropriate statistical techniques, such as two-
dimensional probabilistic analyses (Cullen and Frey 1999). In practice, 
however, increasingly complex analyses may not be warranted for every 
situation, as discussed below. In addition, the available data may not 
be sufficient to distinguish between variability and uncertainty or to 
allow statistical consideration of both issues.
    Risk assessors may take one of two general approaches to conduct 
risk assessments: deterministic or probabilistic (stochastic) modeling. 
Of these, probabilistic techniques explicitly include quantification of 
uncertainty and variability.
    Risk analyses have long been grounded on deterministic approaches. 
Probabilistic risk assessments have been used for many years in 
predicting accidents and systems failures, and in weather forecasting. 
Over time, probabilistic approaches have been applied to ecological and 
human health risk assessments (Kendall et al., 2001).
    Deterministic and probabilistic modeling are both valid 
mathematical approaches for estimating risk. The key difference between 
these approaches is that deterministic modeling enters point estimates 
(i.e., single values) for the model's inputs while probabilistic 
modeling uses probability distributions for some or all inputs in 
conjunction with statistical techniques such as Monte Carlo analysis. 
Consequently, the output of a deterministic assessment is a point 
estimate of the exposure or risk for the exposed individual or 
population. A probabilistic approach results in a distribution of 
exposure or risk estimates, which may provide additional information 
about the variability in the exposure of interest and the uncertainty 
in the analysis or of the true, but unknown risk.
    Exposure and risk assessments are conducted for many different 
reasons, such as to answer specific questions about exposure scenarios, 
inform decision-making, and explore options. The ultimate application 
of the assessment will help determine the methodological approaches and 
techniques to be used. The choice of approach may be based on 
considerations of the available scientific information, institutional 
policies, time and resources available, or social implications.
    Risk assessments may be iterative, e.g., subject to collection of 
new data or refinement of existing data. Assessments may be conducted 
in a tiered approach, in which each analysis is based on the knowledge 
and resources available to the risk assessor and the needs of decision-
makers and stakeholders. In general, risk analysts will work from the 
simple to the complex until, for example, the problem has been 
sufficiently characterized so that risk managers may proceed with 
decision-making and initiate any actions required to manage the hazard. 
An initial analysis may be conducted to determine whether a given 
exposure scenario is associated with relatively high or relatively low 
risk. For example, protective assumptions are sometimes used initially 
to characterize the level of risk. If such an assessment indicates a 
relatively high risk, the analyst may choose to collect more data or 
conduct a more complex assessment in order to verify the result before 
actions are taken. An initial analysis may also be used to identify 
insignificant exposure pathways that do not require further 
consideration.
    In many cases, deterministic techniques may be more desirable than 
probabilistic methods, particularly for such early analyses that are 
often under time and resource constraints, because probabilistic 
methods can be more complex, time-consuming, and costly. On the other 
hand, risk managers may find that more sophisticated techniques, 
including probabilistic methods, are valuable in providing certain 
detailed information about the risks in the exposed population, to 
explore the uncertainty in the true, but unknown risk to an individual, 
or for systematically analyzing variability, uncertainty, pathways of 
exposure, or alternative models. The risk assessor and risk manager 
must consider the utility of the risk assessment result and determine 
the value added by each assessment choice that increases the time, 
cost, and complexity of the assessment.
    Ultimately, a risk assessment is conducted to gain insight into the 
exposures and risks associated with a given scenario. See section VI.F. 
of the guidelines (CPSC 1992). Each assessment should be approached on 
a case-by-case basis, consistent with the requirements of the risk 
assessor and risk manager. Regardless of the risk analysis approach, 
the quality of the assessment depends on the quality and availability 
of relevant data.
    In general, for a given body of knowledge, a deterministic 
assessment that is based predominantly on central tendency values for 
each of the input variables (e.g., a best estimate of the available 
data, such as a mean or median), may provide results similar to a 
probabilistic assessment that is based on the same underlying 
information. However, risk analysts must be aware of the effects of 
decisions regarding the use of the available data and assumptions. For 
example, a deterministic analysis that uses multiple protective values 
rather than central values may lead to unintentionally precautious 
results, i.e., compounding safety factors. In addition, for a 
distribution of data that is skewed to the right, the mean will be 
represented by a value in the right tail and could be considerably 
larger than the median. In such a case, the mean could also be 
considered a protective value.
    The primary advantage of a probabilistic approach is the generation 
of information on the distribution of exposure and risk in a 
population, in addition to estimates of the average exposure and risk. 
This provides information on the range of exposures, including highly 
exposed individuals. However, the risk analyst must consider that 
sparse data or a poorly fitting distribution to the data for one or 
more model inputs could lead to inappropriate conclusions about the 
resulting distribution, particularly at the tails of the distribution, 
which may be most sensitive to deficiencies in the data. Further, a 
probabilistic model may be sensitive to correlations between input 
variables (e.g., body weight and body surface area). Discussion of the 
presence of correlations and dependence among variables and their 
effects on the output should be included in the assessment.
    Another advantage of probabilistic techniques is the ability to 
derive confidence intervals for exposure estimates. Thus, in addition 
to estimating the mean, median, and 95th percentiles of exposure, one 
may also estimate confidence intervals for these

[[Page 30335]]

estimates, expressed as X  Y, which provides a measure of 
uncertainty in the estimated exposure. It also gives the risk assessor 
and risk manager information on the reliability of exposure estimates. 
Typically, the confidence intervals will be larger in the tails of the 
distribution, i.e., confidence intervals for the 95th or 99th 
percentile of the distribution may be larger than the confidence 
interval about the mean. Therefore, whenever possible, methodology that 
permits the estimation of confidence intervals should be applied.
    Currently, probabilistic techniques are used primarily in 
estimating exposure, while single point estimates are derived to 
describe the dose-response (i.e., unit risk for carcinogens; ADI for 
non-carcinogens). The application of probabilistic methods to deriving 
unit risks and ADIs is not presently in widespread use, although this 
has been encouraged by the National Research Council (NRC 2009).
    A distinct issue, but related to analysis of uncertainty, is 
sensitivity analysis. Sensitivity analysis is used to identify 
variables that have the largest effect on the assessment output, and 
general approaches and statistical techniques have been developed for 
both deterministic and probabilistic analyses. It is often useful to 
know if small changes in the values for some variables result in 
relatively large changes in the output. For example, such an analysis 
may be used to identify areas of research that could improve future 
risk assessments. Sensitivity analysis may also be used to focus on 
specific subpopulations or exposure scenarios or to identify the most 
important routes of exposure.
    Such techniques also are useful for providing additional 
information in a deterministic assessment. That is, a separate 
sensitivity analysis can be used in conjunction with a deterministic 
approach to characterize the range of the most likely estimates of 
exposure and risk (e.g., one technique is to vary key input variables, 
one at a time, throughout their reasonable range of values, while 
holding other inputs constant).
    Recent exposure and risk assessments conducted by CPSC staff have 
used both deterministic and probabilistic methods based on the factors 
discussed above. For example, staff used probabilistic techniques to 
estimate the exposure and risk from oral intake of diisononyl phthalate 
by children from mouthing soft plastic toys and other objects, based on 
the strength of the available data (Babich 2002; Babich et al. 2004; 
Babich et al. 2020; Greene 2002). Yet staff used a deterministic 
approach with a separate uncertainty analysis to assess children's 
exposure to arsenic from wooden playground equipment treated with 
chromated copper arsenate (Hatlelid 2003), because staff concluded that 
the data for several key input variables were insufficient to support a 
probabilistic analysis. In this case, mainly central tendency values 
were used to estimate the exposure, and a separate uncertainty analysis 
provided additional information about the likely range of exposure.
    Section VI.F.4.b.i. of the guidelines (CPSC 1992) states that a 
carcinogenic risk of one per million or less is the appropriate level 
for defining acceptable risk; i.e., when exposure to an agent occurs, 
the exposed individual has an estimated excess risk of one chance in a 
million of developing cancer during his/her lifetime. In a 
deterministic analysis, one per million is compared directly with the 
risk value that results from the analysis. Section VI.F.1.d. of the 
guidelines also states that in most cases the best estimate of 
exposure, rather than a protective estimate, is acceptable.
    Probabilistic analyses, however, result in distributions of 
exposure and risk. While there are no generally accepted guidelines for 
interpretation of results from probabilistic analyses for carcinogens, 
this topic has received attention (Burmaster 1996; Thompson 2002; NRC 
2009). Thompson cautioned against setting ``bright-line'' criteria for 
use in any context, and Burmaster also argued that the risk manager 
must consider all the characteristics of the distribution resulting 
from the probabilistic assessment and not just a single point or 
summary statistic. As an example of how one might evaluate 
probabilistic results, Burmaster suggested that one might consider the 
skewness of the resulting risk distribution; whether the median of the 
distribution exceeds the one per million acceptable risk level; whether 
the mean exceeds one per one hundred thousand; and whether the 95th 
percentile exceeds one per ten thousand.
    CPSC staff agrees that it generally is appropriate to consider all 
of the characteristics of the risk distribution (e.g., the mean, 
median, and upper bounds values and the shape of the distribution) in 
judging whether or not the results represent an acceptable risk. 
Because of the complexity of probabilistic analyses and the diversity 
of possible probabilistic risk assessment results, staff assesses that 
it would be difficult to impose a rigid procedure for interpreting the 
results of probabilistic assessments. Staff recommends, however, that 
the one per million acceptable risk level for carcinogens currently 
defined in the guidelines generally should also serve as a guide for 
interpreting probabilistic risk assessment results. Because staff 
generally uses best estimates for exposure rather than upper bounds, 
staff assesses that interpretation of probabilistic results should be 
based in part on the relationship of the central tendency estimate of 
the resulting distribution to the one per million acceptable risk 
level. However, upper bound estimates of exposure (e.g., 95th 
percentile) may provide useful information for highly exposed 
individuals.
    Section VI.F.4.b.ii. (CPSC 1992) specifies a process for evaluating 
the acceptable daily intake (ADI) for neurotoxicological and 
developmental/reproductive agents. Staff uses these guidelines for 
other non-cancer effects, as well. The use of the ADI in a 
deterministic assessment is straightforward--the estimated exposure is 
compared with the ADI. As is the case with cancer risk assessment, 
there are no standard guidelines for interpretation of results from 
probabilistic analyses of non-cancer effects. Following the reasoning 
for cancer assessments given above, staff recommends that 
interpretation of probabilistic results for non-cancer effects should 
be based in part on comparing the central tendency estimate of the 
outcome to the acceptable daily intake, similar to the case for 
deterministic assessments. However, upper bound estimates of exposure 
(e.g., 95th percentile) may provide useful information for highly 
exposed individuals.
    Because the guidelines are not binding rules, they are meant to be 
flexible and amenable to expert judgment, as well as continuing 
scientific advances. The guidance for interpretation of both cancer and 
non-cancer exposure and risk are intended to facilitate the assessment 
process, but in practice, risk assessors and risk managers will 
consider the specific information in each case in defining acceptable 
exposure and risk.

C. Summary

    The following supplements the guidance on exposure assessment in 
the CPSC Chronic Hazard Guidelines at 57 FR 46644 (Oct. 9, 1992) in 
section VI.F.1. It does not supersede the 1992 guidance; rather, it 
provides guidance on the use of probabilistic methods as an alternative 
method for exposure assessment.
    Risk assessments may incorporate uncertainty (the lack of knowledge 
in the underlying science or in the choice

[[Page 30336]]

of mathematical model) and variability (inherent differences due to 
heterogeneity or diversity in the population or exposure variable). The 
discussion may be qualitative or include quantitative estimates of 
uncertainty and variability. While variability and uncertainty are 
distinct issues and should be considered separately in each analysis, 
in practice, the available data may not be sufficient to distinguish 
between them.
    Risk assessments may be based on deterministic or probabilistic 
modeling. Probabilistic modeling uses probability distributions for 
some or all inputs in conjunction with statistical techniques such as 
Monte Carlo analysis, and results in a distribution of exposure or risk 
estimates, providing quantification of uncertainty and variability. 
Deterministic modeling enters point estimates for the model's inputs 
and results in a point estimate of the exposure or risk. Separate 
uncertainty analysis may be used with a deterministic approach to 
characterize the range of the most likely exposure and risk.
    Because exposure and risk assessments are conducted for different 
reasons, the ultimate use of the assessment results will help determine 
the methodological approaches and techniques to be used. The choice of 
approach may be based on considerations of the available scientific 
information, institutional policies, available time and resources, and 
limitations of the methods. For example, deterministic techniques may 
be appropriate for initial analyses that are often under time and 
resource constraints; however, the use of multiple protective values in 
a deterministic analysis may lead to unintentionally protective 
results, i.e., compounding safety factors. A probabilistic assessment 
may be used to generate information on the distribution of exposure and 
risk in a population or to explore the uncertainty in the true, but 
unknown risk to an individual, but the risk assessor must consider that 
sparse data or poorly fitting distributions to the data for one or more 
model inputs could lead to inappropriate conclusions about the results, 
particularly at the tails of the distribution, which may be most 
sensitive to deficiencies in the data. A probabilistic model may be 
sensitive to correlations between input variables; the presence of 
correlations and dependence among variables and their effects on the 
output should be considered.
    A carcinogenic risk of one per million or less is the guidelines' 
default level for defining acceptable risk (16 CFR 1500.135(d)(4)(i)). 
In a deterministic analysis, one per million is compared directly with 
the risk value that results from the analysis. Interpretation of 
probabilistic results should be based in part on the relationship of 
the central tendency estimate (e.g., mean or median, as appropriate for 
the specific distribution) to the one per million acceptable risk 
level, but all characteristics of the resulting distribution should be 
considered.
    For assessment of non-carcinogens in a deterministic assessment, 
the exposure estimate is compared directly with the ADI, or the hazard 
index (HI) is calculated as the ratio of the estimated exposure to the 
ADI (HI greater than one means that the exposure may be hazardous; HI 
less than one represents negligible risk). Probabilistic results should 
be interpreted in part by comparing the central tendency estimate to 
the acceptable daily intake, but all characteristics of the resulting 
distribution should be considered.
    The guidance for interpretation of both cancer and non-cancer 
exposure and risk are intended to facilitate the assessment process, 
but in practice, risk assessors and risk managers will consider the 
specific information in each case in defining acceptable exposure and 
risk.

D. References

Babich MA. 2002. Updated risk assessment of oral exposure to 
diisononyl phthalate (DINP) in children's products. In: Response to 
Petition HP 99-1. Request to Ban PVC in Toys and Other Products 
intended for Children Five Years of Age and Under. U.S. Consumer 
Product Safety Commission. Washington, DC 20207. August 2002. http://www.cpsc.gov/library/foia/foia02/brief/briefing.html (TAB L).
Babich MA, Greene MA, Chen S, Porter WK, Kiss CT, Smith TP, Wind ML. 
2004. Risk assessment of oral exposure to diisononyl phthalate from 
children's products. Regulatory Toxicology and Pharmacology 40: 151-
167.
Babich MA, Bevington C, Dreyfus M (2020) Plasticizer migration from 
children's toys, child care articles, art materials, and school 
supplies. Regulatory Toxicology and Pharmacology 111: 104574.
Burmaster DE. 1996. Benefits and Costs of Using Probabilistic 
Techniques in Human Health Risk Assessments--with an Emphasis on 
Site-Specific Risk Assessments. Human and Ecological Risk Assessment 
2(1): 35-43.
Consumer Product Safety Commission (CPSC). 1992. Labeling 
requirements for art materials presenting chronic hazards; 
guidelines for determining chronic toxicity of products subject to 
the FHSA; supplementary definition of ``toxic'' under the Federal 
Hazardous Substances Act; final rules. 57 FR: 46626-46674 (9 October 
1992). https://www.cpsc.gov/s3fs-public/pdfs/blk_pdf_chronichazardguidelines.pdf.
Cullen AC and Frey HC. 1999. Probabilistic Techniques in Exposure 
Assessment: A Handbook for Dealing with Variability and Uncertainty 
in Models and Inputs. New York: Plenum Press.
Greene M. 2002. Oral DINP Intake Among Young Children. In: Response 
to Petition HP 99-1. Request to Ban PVC in Toys and Other Products 
intended for Children Fiver Years of Age and Under. U.S. Consumer 
Product Safety Commission. Washington, DC 20207. August 2002. http://www.cpsc.gov/library/foia/foia02/brief/briefing.html (TAB K).
Hatlelid KM. 2003. Cancer risk assessment for arsenic exposure from 
CCA-treated wood playground structures. In: Re: Petition HP 01-3. 
Request to Ban Chromated Copper Arsenate (CCA)-Treated Wood in 
Playground Equipment. U.S. Consumer Product Safety Commission. 
Washington, DC 20207. February 2003.
Kendall RJ, Anderson TA, Baker RJ, Bens CM, Carr JA, Chiodo LA, Cobb 
III GP, Dickerson, RL, Dixon, KR, Frame LT, Hooper MJ, Martin CF, 
McMurry ST, Patino R, Smith EE, Theodorakis CW. 2001. Ecotoxicology. 
In, Casarett & Doull's Toxicology: The Basic Science of Poisons. CD 
Klaassen, Ed. New York: McGraw-Hill.
Morgan MG and Henrion M. 1990. Uncertainty: A Guide to Dealing with 
Uncertainty in Quantitative Risk and Policy Analysis. New York: 
Cambridge University Press.
National Research Council (NRC). 1983. Risk Assessment in the 
Federal Government: Managing the Process. Washington, DC: National 
Academy Press.

Alberta E. Mills,
Secretary, Consumer Product Safety Commission.
[FR Doc. 2024-08604 Filed 4-22-24; 8:45 am]
BILLING CODE 6355-01-P