This report contains the collective views of an international group of experts and

does not necessarily represent the decisions or the stated policy of the United
Nations Environment Programme, the International Labour Organization or the
World Health Organization.

Environmental Health Criteria 239

PRINCIPLES FOR MODELLING

DOSE–RESPONSE FOR THE RISK
ASSESSMENT OF CHEMICALS

First draft prepared by the WHO Task Group on Environmental Health

Criteria on Principles for Modelling Dose–Response for the Risk
Assessment of Chemicals

Published under the joint sponsorship of the United Nations

Environment Programme, the International Labour Organization
and the World Health Organization, and produced within the
framework of the Inter-Organization Programme for the Sound
Management of Chemicals.
The International Programme on Chemical Safety (IPCS), established in 1980, is a
joint venture of the United Nations Environment Programme (UNEP), the International
Labour Organization (ILO) and the World Health Organization (WHO). The overall objec-
tives of the IPCS are to establish the scientific basis for assessment of the risk to human
health and the environment from exposure to chemicals, through international peer review
processes, as a prerequisite for the promotion of chemical safety, and to provide technical
assistance in strengthening national capacities for the sound management of chemicals.
The Inter-Organization Programme for the Sound Management of Chemicals
(IOMC) was established in 1995 by UNEP, ILO, the Food and Agriculture Organization
of the United Nations, WHO, the United Nations Industrial Development Organization, the
United Nations Institute for Training and Research and the Organisation for Economic Co-
operation and Development (Participating Organizations), following recommendations
made by the 1992 UN Conference on Environment and Development to strengthen coop-
eration and increase coordination in the field of chemical safety. The purpose of the IOMC
is to promote coordination of the policies and activities pursued by the Participating
Organizations, jointly or separately, to achieve the sound management of chemicals in
relation to human health and the environment.

WHO Library Cataloguing-in-Publication Data

Principles for modelling dose-response for the risk assessment of chemicals.

(Environmental health criteria ; 239)

1.Chemicals. 2.Dose-response relationship, Drug. 3.Dose-response relationship,

Radiation. 4.Risk assessment. 5.Environmental exposure. I.World Health
Organization. II.Inter-Organization Programme for the Sound Management of
Chemicals. III.Series.

ISBN 978 92 4 157239 2 (NLM classification: QV 38)

ISSN 0250-863X

© World Health Organization 2009

All rights reserved. Publications of the World Health Organization can be obtained
from WHO Press, World Health Organization, 20 Avenue Appia, 1211 Geneva 27,
Switzerland (tel.: +41 22 791 3264; fax: +41 22 791 4857; e-mail: [email protected]).
Requests for permission to reproduce or translate WHO publications – whether for sale or
for noncommercial distribution – should be addressed to WHO Press, at the above address
(fax: +41 22 791 4806; e-mail: [email protected]).
The designations employed and the presentation of the material in this publication do
not imply the expression of any opinion whatsoever on the part of the World Health
Organization concerning the legal status of any country, territory, city or area or of its
authorities, or concerning the delimitation of its frontiers or boundaries. Dotted lines on
maps represent approximate border lines for which there may not yet be full agreement.
The mention of specific companies or of certain manufacturers’ products does not
imply that they are endorsed or recommended by the World Health Organization in
preference to others of a similar nature that are not mentioned. Errors and omissions
excepted, the names of proprietary products are distinguished by initial capital letters.
All reasonable precautions have been taken by the World Health Organization to verify
the information contained in this publication. However, the published material is being
distributed without warranty of any kind, either expressed or implied. The responsibility
for the interpretation and use of the material lies with the reader. In no event shall the
World Health Organization be liable for damages arising from its use.
The named authors alone are responsible for the views expressed in this publication.
This document was technically and linguistically edited by Marla Sheffer, Ottawa,
Canada, and printed by Wissenchaftliche Verlagsgesellschaft mbH, Stuttgart, Germany.
CONTENTS

ENVIRONMENTAL HEALTH CRITERIA ON

PRINCIPLES FOR MODELLING DOSE–RESPONSE FOR
THE RISK ASSESSMENT OF CHEMICALS

PREAMBLE ix

PREFACE xvii

ACRONYMS AND ABBREVIATIONS xix

1. SUMMARY, CONCLUSIONS, AND

RECOMMENDATIONS 1

1.1 Summary 1
1.2 Conclusions 5
1.3 Recommendations 6

2. INTRODUCTION 9

2.1 Background 10
2.2 Scope 10

3. RISK ANALYSIS 13

3.1 Decision paradigms 13

3.2 Risk analysis paradigms 13
3.3 Motivations and considerations for producing a
formal risk assessment 15
3.3.1 Transparency and justification 15
3.3.2 Public health and individual health 16
3.3.3 Quantification and computation 17
3.3.4 Cost of assessment 17
3.4 Risk assessment 17
3.4.1 Problem formulation 18
3.4.1.1 Defining the question 19
3.4.1.2 Prior knowledge 19
3.4.1.3 Desired outcomes 19
3.4.2 Risk assessment outcomes 20

iii
EHC 239: Principles for Modelling Dose–Response

4. DOSE–RESPONSE MODELLING: BASIC CONCEPTS 22

4.1 Introduction 22
4.2 What is dose? 23
4.3 What is response? 24
4.4 What is a model? 25
4.5 What is dose–response modelling? 27
4.6 Risk versus safety in dose–response modelling 31
4.7 Summary 33

5. DOSE–RESPONSE MODELLING: WHY AND WHEN

TO USE IT 34

5.1 Historical perspectives 34

5.1.1 The no-observed-adverse-effect level approach
to acceptable/tolerable daily intake 35
5.1.2 The benchmark dose approach to
acceptable/tolerable daily intake 38
5.2 Points of consideration 40
5.2.1 General aspects of definition 41
5.2.2 Estimation procedure 42
5.2.3 Uncertainty 43
5.2.4 Study design 44
5.2.5 Biological information 46
5.2.6 Comparison of experimental results 46
5.2.7 Risk management perspectives 47
5.3 Implementation issues 47
5.4 Summary 48

6. PRINCIPLES OF DOSE–RESPONSE MODELLING 49

6.1 Data 49
6.1.1 Selection of data 49
6.1.2 Data types 49
6.2 Models and distributions 51
6.2.1 Dose–response models 51
6.2.1.1 Continuous dose–response models 51
6.2.1.2 Quantal dose–response models 53
6.2.1.3 Thresholds 53
6.2.1.4 Severity (degree of effect) 55
6.2.1.5 Modelling with covariates 57

iv
 6.2.1.6 Biologically based dose–response
models 57
6.2.2 Statistical distributions 59
6.2.2.1 Continuous distributions 59
6.2.2.2 Discrete distributions 60
6.3 Model fitting and estimation of parameters 61
6.3.1 Criterion function 62
6.3.2 Search algorithms 62
6.4 Model comparison 63
6.5 Representing uncertainty 65
6.5.1 Sampling error 65
6.5.2 Study error 66
6.5.3 Model error 66
6.6 Benchmark dose and benchmark response selection 73
6.7 Summary 76

7. COMMUNICATING THE RESULTS OF

DOSE–RESPONSE MODELLING 78

7.1 Introduction 78
7.2 Incorporation of the outputs of dose–response
modelling into risk assessments 80
7.3 Derivation of health-based guidance values 81
7.4 Estimation of the margin of exposure 82
7.5 Quantitative estimations of the magnitude of
the risk at levels of human exposure 83
7.6 Presentation of results 84
7.6.1 Tables 84
7.6.2 Graphs 84
7.7 Risk assessment context and questions 89
7.8 Synopsis of approach to modelling 89
7.8.1 Data sets 90
7.8.2 Uncertainty 90
7.9 Explaining/interpreting the output of the
dose–response analysis 91
7.9.1 Outputs in the observable biological range 91
7.9.1.1 Health-based guidance values 92
7.9.1.2 Margin of exposure 92
7.9.2 Outputs outside the observable biological
range 94
7.9.2.1 Prediction of risks at specified
exposure levels 95

v
EHC 239: Principles for Modelling Dose–Response

7.9.2.2 Prediction of exposure levels

producing specified risk levels 96
7.9.2.3 Uncertainty analyses 96
7.10 Issues for risk managers 97
7.10.1 Risk assessment issues 97
7.10.1.1 Population versus individual effects 97
7.10.1.2 Risk characterization 98
7.10.2 Risk management issues 98
7.10.2.1 Risk management options 98
7.10.2.2 Cost–benefit and risk–benefit
analyses 99
7.10.2.3 Acceptable level of risk 99

8. CONCLUSIONS AND RECOMMENDATIONS 101

8.1 Conclusions 101

8.2 Recommendations 102

REFERENCES 103

ANNEX 1: TERMINOLOGY 111

RESUME, CONCLUSIONS ET RECOMMANDATIONS 123

RESUMEN, CONCLUSIONES Y RECOMENDACIONES 130

vi
 NOTE TO READERS OF THE CRITERIA MONOGRAPHS

Every effort has been made to present information in the

criteria monographs as accurately as possible without unduly
delaying their publication. In the interest of all users of the
Environmental Health Criteria monographs, readers are requested to
communicate any errors that may have occurred to the Director of
the International Programme on Chemical Safety, World Health
Organization, Geneva, Switzerland, in order that they may be
included in corrigenda.

vii
 Environmental Health Criteria

PREAMBLE

Objectives

In 1973, the WHO Environmental Health Criteria Programme

was initiated with the following objectives:

(i) to assess information on the relationship between exposure to

environmental pollutants and human health, and to provide
guidelines for setting exposure limits;
(ii) to identify new or potential pollutants;
(iii) to identify gaps in knowledge concerning the health effects of
pollutants;
(iv) to promote the harmonization of toxicological and epidemio-
logical methods in order to have internationally comparable
results.

The first Environmental Health Criteria (EHC) monograph, on

mercury, was published in 1976, and since that time an ever-
increasing number of assessments of chemicals and of physical
effects have been produced. In addition, many EHC monographs
have been devoted to evaluating toxicological methodology, e.g. for
genetic, neurotoxic, teratogenic, and nephrotoxic effects. Other
publications have been concerned with epidemiological guidelines,
evaluation of short-term tests for carcinogens, biomarkers, effects
on the elderly, and so forth.

Since its inauguration, the EHC Programme has widened its

scope, and the importance of environmental effects, in addition to
health effects, has been increasingly emphasized in the total
evaluation of chemicals.

The original impetus for the Programme came from World

Health Assembly resolutions and the recommendations of the 1972
UN Conference on the Human Environment. Subsequently, the
work became an integral part of the International Programme on
Chemical Safety (IPCS), a cooperative programme of WHO, ILO,
and UNEP. In this manner, with the strong support of the new

ix
EHC 239: Principles for Modelling Dose–Response

partners, the importance of occupational health and environmental

effects was fully recognized. The EHC monographs have become
widely established, used, and recognized throughout the world.

The recommendations of the 1992 UN Conference on Environ-

ment and Development and the subsequent establishment of the
Intergovernmental Forum on Chemical Safety with the priorities for
action in the six programme areas of Chapter 19, Agenda 21, all
lend further weight to the need for EHC assessments of the risks of
chemicals.

Scope

Two different types of EHC documents are available: 1) on

specific chemicals or groups of related chemicals; and 2) on risk
assessment methodologies. The criteria monographs are intended to
provide critical reviews on the effect on human health and the
environment of chemicals and of combinations of chemicals and
physical and biological agents and risk assessment methodologies.
As such, they include and review studies that are of direct relevance
for evaluations. However, they do not describe every study carried
out. Worldwide data are used and are quoted from original studies,
not from abstracts or reviews. Both published and unpublished
reports are considered, and it is incumbent on the authors to assess
all the articles cited in the references. Preference is always given to
published data. Unpublished data are used only when relevant
published data are absent or when they are pivotal to the risk
assessment. A detailed policy statement is available that describes
the procedures used for unpublished proprietary data so that this
information can be used in the evaluation without compromising its
confidential nature (WHO (1990) Revised Guidelines for the
Preparation of Environmental Health Criteria Monographs.
PCS/90.69, Geneva, World Health Organization).

In the evaluation of human health risks, sound human data,

whenever available, are preferred to animal data. Animal and in
vitro studies provide support and are used mainly to supply
evidence missing from human studies. It is mandatory that research
on human subjects is conducted in full accord with ethical
principles, including the provisions of the Helsinki Declaration.

x
 The EHC monographs are intended to assist national and
international authorities in making risk assessments and subsequent
risk management decisions. They represent a thorough evaluation of
risks and are not, in any sense, recommendations for regulation or
standard setting. These latter are the exclusive purview of national
and regional governments.

Procedures

The following procedures were followed in the development

and publication of this EHC. A designated IPCS Staff Member (Dr
Sam Page and subsequently Dr A. Tritscher), responsible for the
scientific content of the document, served as the Responsible
Officer (RO). The IPCS editor was responsible for layout and
language.

The WHO Planning Group for the IPCS Harmonization Project

on Dose–Response Modelling met on 10 October 2002 in Geneva to
develop an outline and proposed time frame for the project. A first
draft working paper, including contributions from several additional
authors, was prepared by Drs C. Carrington and M. Bolger and
distributed to the Task Group prior to the Task Group meeting,
which was held from 13 to 17 September 2004. The first draft
working paper was revised during the Task Group meeting and
during a subsequent internal Task Group Internet forum. This
revised draft was available on the IPCS web site for external review
and comment. Comments received are available on request from the
WHO Secretariat. They were reviewed by the Task Group, and
necessary additions and revisions to the document were made.

The Task Group members serve as individual scientists, not as

representatives of any organization, government, or industry. All
individuals who as authors, consultants, or advisers participate in
the preparation of the EHC monograph must, in addition to serving
in their personal capacity as scientists, inform the WHO Secretariat
if at any time a conflict of interest, whether actual or potential,
could be perceived in their work. They are required to sign a
declaration of interest statement. The Chairpersons of Task Groups
are briefed on their role and responsibility in ensuring that these
rules are followed. Such a procedure ensures the transparency and
probity of the process. Their function is to evaluate the accuracy,
significance, and relevance of the information in the document. A

xi
EHC 239: Principles for Modelling Dose–Response

summary and recommendations for further research and improved

safety aspects are also required. The composition of the Task Group
is dictated by the range of expertise required for the subject of the
meeting and, where possible, by the need for a balanced
geographical distribution.

xii
WHO PLANNING GROUP FOR THE IPCS
HARMONIZATION PROJECT ON DOSE–RESPONSE
MODELLING

Members

Dr P.M. Bolger, Food and Drug Administration, College Park, MD,

United States of America (USA)

Professor E. Dybing, Norwegian Institute of Public Health, Oslo,

Norway

Dr L. Edler, German Cancer Research Center, Heidelberg, Germany

Dr M. Hartley, National Industrial Chemicals Notification and

Assessment Scheme (NICNAS), Sydney, Australia

Dr A. Knaap, National Institute of Public Health and the

Environment (RIVM), Bilthoven, Netherlands

Dr C. Portier, National Institute of Environmental Health Sciences,

Research Triangle Park, NC, USA

Professor A. Renwick, University of Southampton, Southampton,

United Kingdom

Secretariat

Dr T. Damstra, International Programme on Chemical Safety,

World Health Organization, Research Triangle Park, NC, USA

Dr S. Page, International Programme on Chemical Safety, World

Health Organization, Geneva, Switzerland

Ms C. Sonich-Mullin, International Programme on Chemical Safety,

World Health Organization, Cincinnati, OH, USA

Ms C. Vickers, International Programme on Chemical Safety,

World Health Organization, Geneva, Switzerland

xiii
EHC 239: Principles for Modelling Dose–Response

WHO TASK GROUP ON ENVIRONMENTAL HEALTH

CRITERIA ON PRINCIPLES FOR MODELLING
DOSE–RESPONSE FOR THE RISK ASSESSMENT
OF CHEMICALS

Dr S. Page and Dr A. Tritscher, IPCS, served as the

Responsible Officer (RO) and were responsible for the preparation
of the final document and for its overall scientific content. Marla
Sheffer, Ottawa, Canada, was the IPCS editor responsible for layout
and language.

***

Risk assessment activities of IPCS are supported financially by

the Department of Health and Department for Environment, Food &
Rural Affairs, Food Standards Agency, United Kingdom;
Environmental Protection Agency, Food and Drug Administration,
and National Institute of Environmental Health Sciences, USA;
European Commission; German Federal Ministry of Environment,
Nature Conservation and Nuclear Safety; Health Canada; Japanese
Ministry of Health, Labour and Welfare; and Swiss Agency for
Environment, Forests and Landscape.

***

Task Group members

Dr P.M. Bolger, Food and Drug Administration, College Park, MD,

USA

Professor A. Boobis, Imperial College London, London, United

Kingdom

Dr C. Carrington, Food and Drug Administration, College Park,

MD, USA

Dr V. Cogliano, International Agency for Research on Cancer,

Lyon, France

Professor E. Dybing, Norwegian Institute of Public Health, Oslo,

Norway

xvi
Dr L. Edler, German Cancer Research Center, Heidelberg, Germany

Professor E. Faustman, University of Washington, Seattle, WA,

USA

Dr M. Healy, Food Standards Australia New Zealand, Canberra,

Australia

Mr J. Howlett, Food and Agriculture Organization of the United

Nations, Rome, Italy

Dr J. Kleinberg, European Food Safety Authority, Brussels,

Belgium

Dr A. Knaap, National Institute of Public Health and the

Environment (RIVM), Bilthoven, Netherlands

Dr J. Larsen, Danish Institute of Food and Veterinary Research,

Solberg, Denmark

Dr D. Lovell, University of Surrey, Surrey, United Kingdom

Dr C. Portier, National Institute of Environmental Health Sciences,

Research Triangle Park, NC, USA

Professor A. Renwick, University of Southampton, Southampton,

United Kingdom

Professor T. Sanner, Institute for Cancer Research, Oslo, Norway

Dr J. Schlatter, Swiss Federal Office of Public Health, Zurich,

Switzerland

Dr R. Setzer, Jr, Environmental Protection Agency, Research

Triangle Park, NC, USA

Professor W. Slob, National Institute of Public Health and the

Environment (RIVM), Bilthoven, Netherlands

Dr A. Wadge, Food Standards Agency, London, United Kingdom

xv
EHC 239: Principles for Modelling Dose–Response

Professor G. Williams, New York Medical College, Valhalla, NY,

USA

Secretariat

Dr S. Page, International Programme on Chemical Safety, World

Health Organization, Geneva, Switzerland

Dr A. Tritscher, International Programme on Chemical Safety,

World Health Organization, Geneva, Switzerland

Other contributors

Professor E. Calabrese, University of Massachusetts, Amherst, MA,

USA

Dr I. Dewhurst, Pesticides Safety Directorate, York, United

Kingdom

Dr D. Gaylor, Gaylor and Associates, Eureka Springs, AR, USA

Professor P. Grandjean, Environmental Medicine, University of

Southern Denmark, Odense, Denmark

Dr I. Mangelsdorf, Fraunhofer Institute of Toxicology and

Experimental Medicine, Hanover, Germany

Dr S. Sand, Institute of Environmental Medicine, Karolinska

Institutet, Stockholm, Sweden

Dr K. Victorin, Institute of Environmental Medicine, Karolinska

Institutet, Stockholm, Sweden

xvi
PREFACE

The International Programme on Chemical Safety (IPCS) was

initiated in 1980 as a collaborative programme of the United
Nations Environment Programme (UNEP), the International Labour
Organization (ILO), and the World Health Organization (WHO).
One of the major objectives of IPCS is to improve scientific
methodologies for assessing the effects of chemicals on human
health and the environment. As part of this effort, IPCS publishes a
series of monographs, called Environmental Health Criteria (EHC)
documents, that evaluate the scientific principles underlying
methodologies and strategies to assess risks from exposure to
chemicals.

This EHC is part of the ongoing review of the underlying

scientific bases for decision-making in chemical risk assessment by
IPCS. It involves specific consideration of the area of dose–
response assessment in the evaluation of information from
toxicological studies in animals and from human clinical and
epidemiological studies. It covers toxicants with threshold effects
and those for which there may be no practical threshold, such as
substances that are genotoxic and carcinogenic. The discussions are
concerned with that subset of cause–effect relationships commonly
referred to as dose–response models, which are typically used to
characterize the biological effects of intentional (e.g. drugs and
nutrients) and unintentional (e.g. contaminants) exposure to
chemicals.

This EHC is intended primarily to provide descriptive guidance

for risk assessors in using dose–response modelling in hazard
characterization. It will also provide mathematical modellers with
an appreciation of issues to be considered when modelling in the
context of the risk assessment process. Risk managers will be able
to obtain a general understanding of the applications and limitations
of dose–response modelling. For both risk assessors and risk
managers, some considerations for communicating the results of
risk assessments that use dose–response modelling are presented.

The efforts of all who helped in the preparation, review, and

finalization of the monograph are gratefully acknowledged. Special
thanks are due to Health Canada, the Ministry of Health of Japan,
the United Kingdom Food Standards Agency and the United States
xvii
EHC 239: Principles for Modelling Dose–Response

National Institute of Environmental Health Sciences for their

financial support of the project.

xviii
 ACRONYMS AND ABBREVIATIONS

ADI acceptable daily intake

AIC Akaike’s information criterion
ALT alanine aminotransferase
BMD benchmark dose
BMD10 benchmark dose at 10% risk
BMDL lower confidence limit on the benchmark dose
BMDS Benchmark Dose Software (United States
Environmental Protection Agency)
BMR benchmark response
CDF cumulative distribution function
CSAF chemical-specific adjustment factor
DDT dichlorodiphenyltrichloroethane
DNA deoxyribonucleic acid
DRM dose–response modelling
EHC Environmental Health Criteria
ED10 effective dose for a 10% risk
EPI exposure potency index
F Frequency
FAO Food and Agriculture Organization of the United
Nations
fx dose–response function
f(x) dose–response function
IPCS International Programme on Chemical Safety
JECFA Joint FAO/WHO Expert Committee on Food
Additives
LD50 median lethal dose
LED10 lower bound on the effective dose resulting in a 10%
increase in risk (ED10)
LL log-likelihood
LOAEL lowest-observed-adverse-effect level
MOE margin of exposure
NOAEL no-observed-adverse-effect level
NOEL no-observed-effect level
RfD reference dose
SD standard deviation

xix
EHC 239: Principles for Modelling Dose–Response

T25 chronic daily dose that gives 25% of the animals

tumours above background at a specific tissue site
TDI tolerable daily intake
UF uncertainty factor
WHO World Health Organization

xx
 1. SUMMARY, CONCLUSIONS, AND
RECOMMENDATIONS

1.1 Summary

Dose–response modelling (DRM), for use in quantitative risk

assessment and ultimately for informing public health decisions
about chemical exposures, can be described as a six-step process.
The first four steps—data selection, model selection, statistical
linkage, and parameter estimation—constitute dose–response
analysis. These steps relate to the process through which a
mathematical description of the data is obtained in order to evaluate
predicted responses for known doses or to obtain dose estimates
when a chosen response is of interest. The fifth step involves the
integration of the results of the dose–response analysis with
estimates of exposure for the purposes of guiding public health
decisions. The final step, which can optionally be applied earlier in
DRM, involves an assessment of the quality of the dose–response
analysis and the sensitivity of model predictions to the assumptions
used in the analysis.

The characterization of dose–response relationships in animal

and human studies has been a major component of hazard
characterization and has been used in the extrapolation of
incidences of adverse effects in the range of human exposure levels.
Over the years, a variety of methods have been developed to
accommodate such relationships, for improving extrapolation to low
doses and deriving health-based guidance values, such as acceptable
daily intakes (ADIs), tolerable daily intakes (TDIs), and reference
doses (RfDs). DRM may prove useful in risk assessments for
making better use of available data and for providing tools to
evaluate the quality of data and the ensuing uncertainties in dose–
response estimates.

In general, DRM estimates are based on data from the entire

dose–response curve for the critical effect. The standard no-
observed-adverse-effect level (NOAEL) approach can be regarded
as a special, simplified case of dose–response analysis, as it
identifies a single dose that is assumed to be without an appreciable

1
EHC 239: Principles for Modelling Dose–Response

adverse effect. The dose–response model reflects the characteristics

of the dose–response curve, particularly in providing estimates of
the slope. In the case of a regression framework, it provides
standard error and confidence intervals for the model parameters. A
disadvantage of using the NOAEL approach is that it is not possible
to quantify the degree of variability and uncertainty that may be
present, whereas other dose–response models can facilitate the
analysis of sensitivity and uncertainty. Consideration of a dose–
response model can optimize study design and clarify the need for
additional studies. The NOAEL approach incorporates biological
information through the application of “expert” but subjective
judgement. Full DRM has the potential for a more “science-rich”
analysis through the more formal quantitative inclusion of, for
example, factors and covariates into the models. Estimates derived
from DRM enhance the ability to compare quantitatively different
experiments, effects, and compounds within a common framework.
DRM can enhance risk and safety assessments as well as provide
opportunities to consider the likelihood of effects outside the
observable range.

The choice of the models to be used depends upon the type of

data. The models should include a model for dose–response and a
model for the variability of the data. Once models are fit to a data
set, the degree to which they individually describe the data can be
evaluated using goodness-of-fit measures. In addition, their ability
to describe the data with respect to each other may be compared.
Uncertainties about the inferences that result from such models fall
into four main categories: statistical uncertainty of inferences due to
variability among the responses of experimental subjects,
experimental errors (e.g. imperfect randomization, dosing errors,
unfavourable dose location), variability among experiments due to
unavoidable differences in experimental execution, and uncertainty
due to the fact that the “true model” for the data is unknown. Dose–
response analysis needs to address all four sources of variability and
uncertainty whenever possible.

One particularly important application of DRM is the

calculation of benchmark doses (BMDs). These are doses at which
it is inferred that a particular level of response would occur. When
appropriate data are available, BMDs are an alternative to the
NOAEL approach in the calculation of health-based guidance
values. When extrapolation is necessary, the uncertainty associated

2
 Summary, Conclusions, and Recommendations

with a prediction should be represented. Here it is especially

important to include model uncertainty.

Full DRM offers the potential to provide additional information

for the risk manager. The output of DRM should be directed
towards addressing specific questions about the likelihood of
adverse health effects. It can be presented in three principal ways.
Firstly, it can be used for the establishment of a health-based
guidance value, such as an ADI, TDI, or RfD, analogous to current
procedures based on a NOAEL or lowest-observed-adverse-effect
level (LOAEL). DRM can be a more scientifically robust method
for determining health-based guidance values. Secondly, the output
from DRM can be used in risk management to estimate a margin of
exposure (MOE), by calculation of the ratio of the dose
corresponding to a given limit of response to a human exposure
level. Thirdly, based on the modelled dose–response relationship,
the output can be a quantitative estimation of the magnitude of the
risk/health effect at the level of human exposure, with the generally
accepted assumption that the uncertainty factors used cover the
uncertainties about differences in sensitivity between individuals
and species. DRM can provide better information on the likelihood
of effects at low doses that are below the levels observed in
biological systems and can also provide better estimates of the
statistical uncertainties surrounding estimates of likely effects.

Two factors that can impact the type of outputs from DRM
exercises and that may be of importance to the risk manager are
multiple data sets and uncertainties. DRM can be used with
exposure data to identify subpopulations at risk. DRM can also be
used to assist risk managers in determining priorities and evaluating
the consequences of proposed interventions aimed at reducing the
risk. For risk communication, the use of DRM techniques offers
opportunities and challenges. DRM evaluations can produce
information in several formats, including dose–response functions
that allow, along with estimates of exposure, the prediction of risks
at specified exposure levels and functions that allow the estimation
of exposure levels resulting in specified risks. This includes
estimates of the possible risk at intakes above a health-based
guidance value, such as an ADI. DRM evaluations also offer
approaches to compare competing risks or benefits and provide a
focus on uncertainties that can influence the predicted risk.

3
EHC 239: Principles for Modelling Dose–Response

However, unless the situation of risk is viewed at the population

level, there is the risk communication problem that while explaining
the level of risk in those circumstances where there is no safe level
of exposure, some percentage of the population will be predicted to
experience some effect deemed to be adverse. It must be recognized
that the use of DRM requires a certain quantity and quality of data,
as well as specific expertise.

The potential “ongoing” use of the estimates from DRM can,

from a risk management perspective, give an improved
characterization for decision-making by:

x providing information about what happens above the health-

based guidance value (magnitude and types of health impacts);
x showing benefits from different regulatory actions;
x giving the decision-maker a “more-than-one-point” apprecia-
tion of the data;
x promoting consistency in decisions, if appropriate adjustments
are made for differences in effect, effect level, species, and
study design; and
x allowing for an iterative interaction between the risk assessor
and risk manager on a continuous and ongoing basis.

The use of DRM and probabilistic assessment techniques to

quantitatively describe variability and uncertainty brings new
challenges in risk communication. Some of these challenges are:

x explaining that some percentage of the population is predicted

to exceed the safety level and/or experience an adverse effect;
x explaining the level of risk in those circumstances where there
is assumed to be no safe level of exposure;
x comparing competing risks or benefits;
x providing a focus on uncertainties that influence the predicted
risk; and
x explaining that a risk estimate pertains to what may occur at a
population level, rather than the individual level, and noting
that this is also the case for the ADI/TDI approach.

4
 Summary, Conclusions, and Recommendations

1.2 Conclusions

x Full DRM can be considered a more sophisticated or robust

alternative to the NOAEL approach in all cases where suitable
dose–response data are available (e.g. several dose groups with
different response levels).

x For quantal dose–response data, the interest is often in low

response (incidence) levels. This may call for low-dose
extrapolation by several orders of magnitude (e.g. for tumour
incidences). However, equally plausible dose–response risk
models may result in highly divergent low estimates. A
currently applied approach is to estimate a BMD10 (dose at 10%
risk) and linearly extrapolate from that point downwards, as a
conservative approach. Another option, currently under
development, is to apply a Bayesian approach that considers the
various models all together.

x For continuous dose–response data, two approaches of DRM

exist. One is to transform the continuous data into quantal data.
The other is to consider continuous dose–response data as
information on the severity of the effect and therefore as a
function of dose. In the latter approach, measurable changes of
effect are often close to response levels considered as adverse
(e.g. 10% inhibition of cholinesterase), and the low-dose
extrapolation problem is minor or non-existent.

x For the purpose of deriving an ADI, TDI, or RfD, DRM may be

used for deriving a BMD, to be used as a point of departure in
the same way as the NOAEL is used (i.e. the same uncertainty
factors would be applied to the BMD as to the NOAEL).

x DRM may also be used for estimating risks at a given (human)

exposure level. For risks in terms of incidences (quantal data),
this may involve low-dose extrapolation.

x DRM exercises can provide information on uncertainties

associated with the data and identify factors contributing to
uncertainties in risk estimates.

5
 EHC 239: Principles for Modelling Dose–Response

x Application of DRM for all end-points can be cost prohibitive,

so it is efficient to pre-select the apparently more sensitive end-
points. In some cases, however, it is not easy to identify the
most sensitive end-points by visual inspection, so all of the
end-points may need to be modelled.

x The BMD and the lower confidence limit of the BMD (BMDL)
should always be reported, so that the quality of the data and
the model fit are clear and potencies can be compared on the
basis of the BMD.

x The output of the different models used in DRM should be

presented.

1.3 Recommendations

x Toxicity testing protocols (e.g. Organisation for Economic Co-

operation and Development guidelines) should be reviewed for
optimization for BMD and other DRM approaches, including
optimal designs for the number of animals and number of doses
for different dose–response curves. Additional research is
needed for the development of optimal study designs. Guidance
should be developed for combining existing studies with a view
to DRM.

x Better guidance needs to be developed for combined analysis of

different data sets for more precisely estimating BMDs.

x Better understanding of when and how to use the benchmark

response (BMR) needs to be developed.

x Better understanding of the shape of the dose–response curve at

low doses needs to be developed. Additional research is needed
to determine the biological basis for extrapolation (e.g. by
using biomarkers, tumour precursors, genetically modified
animals, and toxicokinetics for target dose estimation).

x Improved guidance needs to be developed for risk

communication based on the results of DRM and probabilistic
assessment techniques. This should include communication of

6
 Summary, Conclusions, and Recommendations

the types of uncertainty and the relation to statistical variability,

imprecision, and the use of confidence intervals.

x The use of DRM should be reviewed and additional general

principles for its use developed when more experience becomes
available.

7
 2. INTRODUCTION

The International Programme on Chemical Safety (IPCS) and

other public health organizations have recognized the importance of
the harmonization of procedures to enhance the quality of risk
assessments, to improve the transparency of the risk assessment
process, and to facilitate risk communication.

Public health decisions on the plausible risks of chemical

exposures can include several possible outcomes. The ultimate goal
is to implement a risk management action that will produce the
desired reduction of risk. Among the first objectives of a risk
assessment is the determination of the presence or absence of a
cause–effect relationship. If there is sufficient plausibility for the
presence of such a relationship, then dose–response information is
needed and will be subject to an analysis of a dose–response
relationship.

Extrapolation is a fundamental problem in the quantitative

health risk assessment of exposures to chemicals that produce
toxicity in experimental systems. Adverse health effects of
chemicals are, in the absence of human data, typically evaluated in
laboratory animals at significantly higher doses than the levels to
which humans may be exposed. Also, for certain substances for
which the exposure can be controlled, such as food additives and
residues of pesticides and veterinary drugs, the quantification of the
risk above the level of exposure that has been assessed to be safe
(e.g. the acceptable daily intake [ADI]) can be difficult. This is
particularly true in cases of temporary excursions above an ADI.

The use of mathematical and statistical approaches in hazard

characterization is increasing. Although dose–response models have
been available for some time, their use has been somewhat limited
because of a lack of either appropriate scientific information or
agreed-upon approaches and methods for how to obtain and use
available dose–response information appropriately. Dose–response
modelling (DRM) involves a number of choices based upon
scientific experience, data availability, and mathematical tractability
and can take on many different forms and be used in many different

9
 EHC 239: Principles for Modelling Dose–Response

ways. A recent review of the available quantitative approaches for

hazard characterization noted that mathematical modelling of the
dose–response relationship could improve the risk assessment
process (Edler et al., 2002).

Dose–response models may improve and generate more reliable

predictions, but they can never be proved to be completely correct.
Therefore, it is necessary to rely on scientific judgement to
determine the utility of risk predictions from DRM in making public
health decisions. It is important to remember that risk numbers
derived from DRM can be misleading for a variety of reasons; like
any other tool used in science, DRM needs to be utilized in a
broader context of all of the available scientific knowledge.
Although mathematical and statistical rigor are important factors in
risk assessment, the final standard that prevails remains biological
plausibility. It is this inherent uncertainty and its communication for
which modelling and quantitative risk assessment can be
particularly valuable.

2.1 Background

This Environmental Health Criteria report (EHC) is intended

primarily to provide descriptive guidance for risk assessors in using
DRM in hazard characterization. It will also provide mathematical
modellers an appreciation of the issues to be considered when
modelling in the context of the risk assessment process. Risk
managers will be able to obtain a general understanding of the
applications and limitations of DRM. For both risk assessors and
risk managers, some considerations for communicating the results
of risk assessments that use DRM are presented.

2.2 Scope

This EHC is part of the ongoing review of the underlying

scientific bases for decision-making in chemical risk assessment by
IPCS. It involves specific consideration of the area of dose–
response assessment in the evaluation of information from
toxicological studies in animals and from human clinical and
epidemiological studies; it does not include consideration of other
aspects of quantitative risk assessment, such as physiologically
based modelling. It covers toxicants with threshold effects and those
for which there may be no practical threshold, such as substances
10
 Introduction

that are genotoxic and carcinogenic. The discussions are concerned

with that subset of cause–effect relationships commonly referred to
as dose–response models, which are typically used to characterize
the biological effects of intentional (e.g. drugs and nutrients) and
unintentional (e.g. contaminants) exposure to chemicals. Dose–
response models are also commonly used in microbiological risk
assessments (e.g. WHO, 2004a).

This document focuses primarily on experimental animal

studies. In DRM of human epidemiological data, several important
issues should be considered:

x Impact of imprecision of the dose estimate. This issue differs

substantially from the situation with experimental animal
studies. In observational studies, where the dose is not a matter
of design, this imprecision is likely to be substantial.

x Absence of a true control group. In many observational studies,

there may not be any subjects who are completely free from
exposure. The response at zero exposure cannot be observed
and has to be estimated.

x Shape of the dose–response curve at low doses. The shape may

depend on both the outcome parameter and the toxicant. For
most contaminants, insufficient information is available, and
the impact on uncertainties must therefore be considered.

x Confounder adjustment. In epidemiological studies, confounder

adjustment must be included. Decisions therefore need to be
made as to which confounders to include in the DRM.

x Meta-analysis. If more than one study is available, a meta-

analysis can provide improved information on the dose–
response models.

Many of the considerations in this EHC are also relevant to

ecotoxicological studies.

This EHC is intended to provide guidance in a number of areas

relevant to DRM. Initially, there is a discussion of the risk analysis
paradigm (chapter 3) and the basic concepts of DRM (chapter 4). In

11
EHC 239: Principles for Modelling Dose–Response

chapter 5, the use of DRM is described, including comparing the

no-observed-adverse-effect level (NOAEL) approach with the
benchmark dose (BMD) modelling method. Chapter 6 provides the
principles of DRM, including data considerations, model
descriptions, model fitting and parameter estimation, model
comparisons, and uncertainty. This chapter also includes discussion
of BMD approaches. Chapter 7 discusses the provision of scientific
advice by risk assessors to the risk managers. This chapter includes
an explanation of the output of the dose–response analysis and the
strengths and weaknesses of DRM. The final chapter, chapter 8,
summarizes the conclusions of the EHC and provides
recommendations for future research.

There is only limited treatment of the mathematical and

statistical considerations for DRM. References and links are
provided for more in-depth treatments, modelling tools, and
examples.

12
 3. RISK ANALYSIS

3.1 Decision paradigms

A risk analysis decision paradigm is a formal representation of

a process that distinguishes the scientific bases from the risk
management objectives and generally contains a component in
which the probability of harm is estimated. This component of the
decision paradigm is referred to as the risk assessment. As a
probability calculation, a risk assessment will include both a
statement of the objective under consideration (i.e. the harm) and
the basis for the assertion that the harm may occur (i.e. the
probability).

3.2 Risk analysis paradigms

The first risk analysis paradigm for public health was proposed
by the National Academy of Sciences of the United States of
America (NRC, 1983) and focused on assessing the risk of cancer
from exposure to chemicals in food. The decision process was
divided into three major steps: research, risk assessment, and risk
management. The risk assessment process was further divided into
hazard identification, dose–response assessment, exposure
assessment, and risk characterization. Risk management is the
decision-making process involving the consideration of political,
social, economic, and technical factors with relevant risk assessment
information relating to a hazard so as to develop, analyse, select,
and implement appropriate risk mitigation options. Risk
management comprises three elements: risk evaluation, emission
and exposure control, and risk monitoring.

In the National Academy of Sciences paradigm, the principal

steps were considered to be sequential, with the decision process
commencing with research and concluding with the decision. A
drawback of this sequential concept is an absence of the recognition
of the influence that the risk analysis might have on data collection
or of the impact that political, social, and economic objectives may
have on the need to identify the hazard.

13
EHC 239: Principles for Modelling Dose–Response

More recent examinations of risk assessment/analysis

methodology have paid much closer attention to the influence of
risk management on the risk assessment process (NRC, 1994, 1996;
Presidential Commission, 1997; Renwick et al., 2003). Rather than
insist that management be insulated from the risk assessment
process for the sake of preserving scientific objectivity, it is
acknowledged that risk management should interact with risk
assessment for the scope of the analysis, particularly in problem
formulation. The focus on this interaction leads to the notion that
the relationship between risk assessment and risk management is an
interactive, often iterative, and circular process (see Figure 1).

5HVHDUFK
3ULRU
.QRZOHGJH

3UREOHP
)RUPXODWLRQ

5LVN 5LVN
$VVHVVPHQW 0DQDJHPHQW

Fig. 1. Interactions of risk assessment with risk management.

As a general rule, formal risk assessments are preceded by

preliminary risk assessments. These are usually subjective and
informal and may be initiated from inside or outside the risk
assessment and scientific communities. A key consideration of these
preliminary risk assessments is whether or not a formal risk
assessment is necessary. The transition process from preliminary
assessments to formal risk assessments has been described as
problem formulation (Renwick et al., 2003). It is an iterative
process that facilitates the critical interface between risk assessment
and risk management. Risk communication, with stakeholder
involvement, is particularly essential during the problem
formulation.

14
 Risk Analysis

As the risk analysis paradigm evolved, the need for risk

communication as an integral part was recognized (see Figure 2).
Risk communication not only is the interactive exchange of
information and opinions among risk assessors and risk managers,
but necessarily includes all interested parties. The issues of risk,
risk-related factors, and risk perceptions should involve interactive
exchange throughout the risk assessment process. The
communication of the results of the risk assessment as the basis of
the risk management decisions demands transparency and
appreciation for the uncertainties involved.

5LVN

5LVN 5LVN
$VVHVVPHQW 0DQDJHPHQW

&RPPXQLFDWLRQ

Fig. 2. The risk analysis paradigm.

3.3 Motivations and considerations for producing a

formal risk assessment

There are several different reasons for preparing a formal risk

assessment. The relative importance of these different motivations
may influence the scope or the methodology used.

3.3.1 Transparency and justification

A major function of formal risk assessment is to serve as a

transparent justification of a public health decision, whereby each
step and assumption are clearly described. A key reason for
undertaking such an assessment is to separate clearly scientific
knowledge from values. Formal risk assessments are almost always

15
 EHC 239: Principles for Modelling Dose–Response

performed for notable public health issues where there is a wider

interest in the political, social, and economic consequences of such
assessments. Identifying the public health objectives before the
technical analysis will allow participation in the debate of the other
issues involved without necessarily requiring involvement in the
scientific discussion. There may be areas of a risk assessment that
can be obscure to someone not privy to their development. As a
result, transparency is often audience dependent, relative to the level
of comprehension and involvement. Since less can be taken for
granted, the extent of the explanation required will increase as the
audience broadens and its level of interest increases and
sophistication decreases. Producing records of an assessment with
varying degrees of technical detail may be a useful objective.

The World Trade Organization, under the Agreement on the

Application of Sanitary and Phytosanitary Measures, has recognized
the importance of harmonized science-based risk assessments. The
World Trade Organization has specifically cited the standards,
guidelines, and recommendations of the Codex Alimentarius
Commission as reflecting international consensus regarding the
requirements to protect human health from foodborne hazards. The
Codex Alimentarius Commission has formally adopted the risk
analysis paradigm in its decision-making (Codex Alimentarius
Commission, 2003). Other organizations have also adopted this
paradigm (European Commission, 2000).

3.3.2 Public health and individual health

A public health risk assessment is concerned with a population.

The behavioural, environmental, or biological characteristics will
vary among individuals in the population of concern. This variation
is considered in probabilistic approaches and determines the
statistical nature of health risk measures and conclusions made on
populations. A risk assessment may need to describe or model these
individual characteristics to produce a prediction of what might be
expected to happen in the population. Specifying the population
with which the risk assessment is concerned may be an important
part of the problem formulation. In a public policy setting, the
population will generally be defined by the risk managers, often in
view of social, economic, and other considerations.

16
 Risk Analysis

3.3.3 Quantification and computation

Public health issues often involve matters of degree,

particularly in regard to level of exposure and risk, and may be
defined by measures of quantity or statistical rates. If an uncertainty
analysis is conducted, knowledge may be quantified as a matter of
degree. Although judging matters of degree does not require the use
of numbers, communication of degree does. Quantitative risk
assessment approaches, including DRM, can be valuable in
providing information to address these issues.

Formal risk assessments often involve the interaction of

multiple quantitative measures that may lead to extensive and
complicated calculations. Particularly in DRM, mathematical and
statistical considerations are often complex. Although computers
can carry out these calculations more accurately and quickly,
knowledge of the scientific basis and experience with the
applications of DRM are essential in order to avoid misinterpreting
and incorrectly communicating the outcomes.

3.3.4 Cost of assessment

Risk assessments take time and effort to develop. The time and
effort required will increase with the complexity of the problem and
often with the degree of transparency that is required. The level of
scientific detail addressed by the models and the level of
documentation needed may vary with the nature and magnitude of
the motivations for producing the risk assessment in the first place.
In order to tailor the risk assessment to the decision problem, it may
be desirable to develop the risk assessment by an iterative process
that commences with the simplest possible statement of the problem
and becomes more complicated as the risk assessment is developed.

3.4 Risk assessment

The risk assessment paradigm, incorporating problem

formulation, is illustrated in Figure 3 (based on Renwick et al.,
2003).

17
 EHC 239: Principles for Modelling Dose–Response

3UREOHP)RUPXODWLRQ

+D]DUG,GHQWLILFDWLRQ
,GHQWLILFDWLRQRIWKHW\SHDQGQDWXUHRIDGYHUVHKHDOWK
HIIHFWV
KXPDQVWXGLHV
DQLPDOEDVHGWR[LFRORJ\VWXGLHV
LQYLWURWR[LFRORJ\VWXGLHV
VWUXFWXUH±DFWLYLW\VWXGLHV

([SRVXUH$VVHVVPHQW
(YDOXDWLRQRIFRQFHQWUDWLRQRUDPRXQWRID
SDUWLFXODUDJHQWWKDWUHDFKHVDWDUJHW
SRSXODWLRQ
PDJQLWXGH
IUHTXHQF\
GXUDWLRQ
URXWH
H[WHQW
+D]DUG&KDUDFWHUL]DWLRQ
4XDOLWDWLYHRUTXDQWLWDWLYHGHVFULSWLRQRI
LQKHUHQWSURSHUWLHVRIDQDJHQWKDYLQJWKH
SRWHQWLDOWRFDXVHDGYHUVHHIIHFWV
VHOHFWLRQRIFULWLFDOGDWDVHW
PRGHPHFKDQLVP V RIDFWLRQ
NLQHWLFYDULDELOLW\
G\QDPLFYDULDELOLW\
GRVH±UHVSRQVHIRUFULWLFDOHIIHFW
LGHQWLILFDWLRQRIVWDUWLQJSRLQW

5LVN&KDUDFWHUL]DWLRQ
$GYLFHIRUGHFLVLRQPDNLQJ
SUREDELOLW\RIRFFXUUHQFH
VHYHULW\
JLYHQSRSXODWLRQ
DWWHQGDQWXQFHUWDLQWLHV

Fig. 3. Risk assessment (adapted from Renwick et al., 2003).

3.4.1 Problem formulation

Problem formulation is the initial phase in a risk assessment

that determines if a detailed risk assessment is necessary and, if so,
whether it is possible. Further, it serves as the transition from an
informal risk assessment to a formal risk assessment. Problem
formulation requires at least some preliminary consideration of the
hazard identification, hazard characterization, and exposure
assessment and usually proceeds in iterative stages. The output is a
plan for the risk assessment process, which can be changed as the
risk assessment progresses.

18
 Risk Analysis

3.4.1.1 Defining the question

Among additional considerations are those that address who

should be involved in the risk assessment and risk management
processes. The transparency of a risk assessment will depend on
how well these are described. It is not necessary to establish beyond
all doubt that there is a cause–effect relationship in order to conduct
a risk assessment. The suspicion that there may be such a
relationship is sufficient. The consideration of the evidence for or
against the supposition is often an integral part of the analysis.
Identifying the problem may be politically controversial. That is, it
may constitute a risk management issue that must be resolved
before the risk assessment may be used as the justification for a
decision. Non-scientific controversy may be diverted from the risk
assessment by separating the valuation of the effect from the risk
assessment per se (i.e. the risk assessment may be used as part of a
cost–benefit analysis, but the cost–benefit analysis is not part of the
risk assessment). Predicting the occurrence of an event is not part of
an expression of the level of public health concern. However,
suggesting that the problem is big enough to merit a formal risk
assessment does imply that the risk may be of some significance.

3.4.1.2 Prior knowledge

Organizing information regarding public health issues that may

involve many details and complex cause–effect relationships may
benefit from the methodical collection and evaluation of prior
knowledge of the agent, exposure to the agent, and possible
biological effect(s) resulting from exposure to the agent. This is
essential for determining the feasibility of a detailed assessment.
Prior knowledge is also important for prioritizing and directing the
risk assessment. Organization of information may also instigate and
support specialization; different experts may produce or oversee
different parts of the risk assessment. This information may in turn
influence the conception of the problem (where management
specifies the objective of the analysis) and also may influence
additional research that may be needed.

3.4.1.3 Desired outcomes

The desired outcomes of the problem formulation are:

19
 EHC 239: Principles for Modelling Dose–Response

x explicit questions to be answered in the risk characterization to

meet the needs of the risk manager;
x determination of the resources that are needed and available;
and
x time frame for completing the assessment.

3.4.2 Risk assessment outcomes

The advice to risk managers that is formulated in the risk

characterization may be qualitative or quantitative. Quantitative
advice includes:

x health-based guidance values;

x estimates of the risks at different levels of exposure;
x exposure-based estimates used with low levels of exposure (e.g.
threshold of toxicological concern); and
x risks at minimum and maximum intakes (e.g. nutrients).

Qualitative advice includes:

x statements/evidence that the agent is of no toxicological

concern owing to the absence of toxicity even at high exposure
levels (e.g. ADI “not specified”);
x statements/evidence that the agent is “safe” in the context of a
specified use; and
x recommendations for avoidance, minimization, or reduction of
exposure.

Risk characterization should include all key assumptions and a

clear explanation of the uncertainties in the risk assessment. It
should also include information on susceptible subpopulations,
including those with greater potential exposure and/or specific
physiological conditions or genetic factors. At present, this is
limited, and generic approaches have to be used (e.g. 10 × 10
uncertainty factors for interspecies differences and human
variability). The advice to risk managers can be in the form of a
comparison of the relative risks resulting from choosing different
risk management options.

The risk assessment that is produced is followed by either a risk

management decision or a request for further analysis, which may

20
 Risk Analysis

influence the further research that is conducted. In one sense, the

risk assessment process may never end. However, from a risk
management standpoint, there is usually some imperative and
timeline that conclude the process. Therefore, in another sense, the
risk assessment ends when the risk management decision is made.
The record produced by a risk assessment stands as a justification
for a decision at the time the decision is made. However, with
additional information, such as that which can reduce the
uncertainties identified in the risk assessment, the risk
assessment/analysis may be reopened. It is also possible that
additional information can increase uncertainty.

21
 4. DOSE–RESPONSE MODELLING: BASIC
CONCEPTS

4.1 Introduction

Toxicology is the science of identifying and quantifying

harmful or adverse effects of chemical and physical agents in the
human environment. This can be accomplished by observations in
humans (i.e. epidemiological and clinical studies), experimental
studies using animal models (i.e. in vivo bioassays), or cellular and
molecular studies. All these approaches have firmly established the
principle of dose–response. Accordingly, dose–response toxicities
of chemicals can be and have been expressed quantitatively (e.g. the
median lethal dose, or LD50).

However, scientific data alone are not sufficient to allow a

decision to be made regarding the potential toxicity of chemicals
and other agents that humans encounter; it is the analysis and
interpretation of these data that lead to a scientifically supported
decision regarding potential health effects. Many analytical
processes have been developed to address the evaluation of the
toxicities of chemicals, ranging from very simple approaches based
solely upon the identification of the possibility of a hazard (NTP,
2002; Cogliano et al., 2004; USEPA, 2005) to much more
complicated approaches incorporating biological mechanisms,
complicated mathematical models, bioavailability in humans, and
direct predictions of chemically induced changes in disease
incidence in the affected human population (Portier & Kohn, 1996;
Kim et al., 2002). All of these methods have two basic steps in
common: analysis of the dose–response information and
implementation of the results of that analysis to formulate a
conclusion. The combined two-step approach will be referred to as
DRM.

This chapter describes the elements that embody DRM. Most of

the information presented is found in more extensive detail in other
chapters of this guidance document. This chapter sets the stage for
discussion of dose/exposure–response modelling by briefly

22
 Dose–Response Modelling: Basic Concepts

answering the questions: What is dose? What is response? What is a

model? It then goes on to introduce the reader to the types of data
and information that may have an impact on the development of
dose–response models.

4.2 What is dose?

It is critical when performing dose–response analyses to have a

clear concept of what is meant by “dose” and how it applies to the
response. There are three basic types of “dose” that arise from
scientific investigations: the administered or external dose, the
internal (absorbed) dose, and the target or tissue dose. External dose
denotes the amount of a chemical or other agent administered to an
experimental animal or human in a controlled experimental setting
by some specific route at some specific frequency. In the
terminology used by the Joint FAO/WHO Expert Committee on
Food Additives (JECFA), intake (dietary exposure) refers to
external dose. Internal dose is the amount determined by
toxicokinetics to be systemically available. It is a consequence of
absorption, distribution, metabolism, and excretion of the chemical.
The tissue dose is the amount that is distributed to and present in a
specific tissue of interest. The three are, of course, related, and each
can be used to express dose–response.

Two other parameters are important: the dose frequency and

duration of dosing. Dosing can be acute, subchronic, or chronic. For
simplicity, the term dose in DRM will be used as an inclusive term
referring to all three forms of dose described above. In general,
units of dose should reflect the magnitude, frequency, and duration
over which it applies. Dose can be expressed in a multitude of
metrics. Some of these metrics include daily intake (e.g. ng/kg body
weight per day), total body burden (e.g. ng/kg body weight), body
burden averaged over a given period of time, or tissue concentration
(ng/kg).

For humans, where dosing of xenobiotics is not intentional, the

term exposure is used for the external dose. In epidemiological
studies, exposure is rarely known, and best estimates are made
using several assumptions and/or biomonitoring of tissue (usually
blood) concentrations at very few time points, often many years

23
 EHC 239: Principles for Modelling Dose–Response

after what is believed to be the period of first/highest exposure.

Sometimes, when laboratory animals are used for DRM, the dose
used in the animal study is transformed to an equivalent human
exposure prior to modelling. Exposure assessment is the qualitative
and/or quantitative evaluation of the likely intake of chemical
agents via food, as well as exposure from other sources, if relevant
(WHO, 1997). In this situation, models of exposure linked to
response data may be used to develop a dose–response model.
However, limited knowledge of the events controlling absorption
and tissue distribution (especially in humans at low levels of
exposure), metabolism, and excretion and the other molecular and
biochemical processes that ultimately lead to particular responses
contribute to the uncertainty in these analyses.

4.3 What is response?

Response, in this context, generally relates to an observation or

effect seen in a laboratory cell culture, an animal, or a human
following exposure. These end-points cover a broad range of
observations, from early responses, such as biochemical alterations,
to more complicated responses, such as cancer and developmental
defects. Responses can be either adaptive or adverse (e.g. Williams
& Iatropoulos, 2002). The latter are defined as a change in the
morphology, physiology, growth, development, reproduction, or
lifespan of an organism or subsystem (e.g. subpopulation of cells)
that results in an impairment of functional capacity, an impairment
of the capacity to compensate for additional stress, or an increase in
susceptibility to other influences. These are critical responses that
are likely to underlie an adverse health effect in humans. The
responses are sometimes species and/or tissue specific and have
different degrees of variation across individuals. Nevertheless, there
is some commonality across species, and there are known linkages
between some responses (e.g. DNA damage is a precursor for
mutations). DRM can address each response, provide insight into
their quantitative similarity across species and tissues, and link
responses in a mechanistically reasonable manner.

Response is generally considered to vary across experimental

units (animals, humans, cell cultures) in the same dose group in a
random fashion. This random variation is usually assumed to follow
some statistical distribution describing the frequency of any given
24
 Dose–Response Modelling: Basic Concepts

response for a population. In general, statistical distributions are

characterized by their central tendency (usually the mean or average
value) and their effective range (usually based on the standard
deviation). Most responses of interest in the context of dose–
response assessment fall into one of four basic categories:

x Quantal responses. Quantal responses generally relate to the

number of experimental units responding in a given period of
time (e.g. the proportion of animals with a tumour in a cancer
bioassay).

x Counts. Count data generally relate to a discrete number of

items measured in a single experimental unit (e.g. number of
papillomas on the skin).

x Continuous measures. Continuous measures generally take on

any value in a defined range (e.g. body weight).

x Ordered categorical measures. Ordinal categorical measures

generally take on one value from a small set of ordered values
(e.g. tumour severity grades).

Sometimes it is useful to convert continuous data into proportions

(e.g. number of animals outside a clinically relevant range for an
immune system marker) or categories (e.g. measured degree of liver
necrosis converted to minimal, moderate, or extensive).

For each of these different data types, there will be some

differences in how they will be handled for DRM; as a general rule,
however, the goal of DRM is to describe the mean and variance of
the response as a function of exposure and/or time.

4.4 What is a model?

Dose–response models are mathematical models used to

characterize the relationship between dose and response for a given
set of scientific data. Mathematical models consist of three basic
components: assumptions used to derive the model, a functional
form for the model, and parameters that are components of the

25
 EHC 239: Principles for Modelling Dose–Response

functional form. For example, the simplest dose–response model is

a linear model to describe a continuous response (see Figure 4).



‡  $ATA

0OINT
2ESPONSE
2$ 
$OSE 2ESPONSE -ODEL
2$ D E $ WHERE
 D AND E



!DDED 2ESPONSE 
[ 

-ODEL
0REDICTIONS 
    

$OSE $OSE
$

Fig. 4. Dose–response illustration displaying a linear model fit to continuous

data for which prediction of the dose associated with an added response of 5
units (not designated) is a dose of 1 unit (not designated).

For this model, the key components are:

x Assumptions: Mean added response is proportional to dose.

x Functional form: R(D) = Į + ȕ·D, where R(D) is the mean
response as a function of dose, denoted D.
x Parameters: Į is a parameter describing the mean response in
the control (unexposed) group, and ȕ is a parameter
describing the mean change in response per unit dose.

Dose–response models range from very simple models, such as

the linear model described above, to extremely complicated models
for which the eventual functional form cannot easily be expressed
as a single equation (e.g. biologically based dose–response models).
Models can also be linked, meaning that one model could describe
part of the dose–response process while another describes the
remainder of the process. For example, in most cases for chemical
carcinogenesis, cancer risk is more closely linked to tissue

26
 Dose–Response Modelling: Basic Concepts

concentration than to administered dose. Given data on dose, tissue

concentration, and tumour response, one can use a toxicokinetic
model to relate dose to tissue concentration and use a multistage
cancer model to relate tissue concentration to response. The two
models combined are needed to describe the dose–response
relationship.

Dose–response models may incorporate other information into

the model form. Age and time-on-study are commonly used in
DRM, but other factors, such as species/strain/human ethnicity, sex,
body weight, etc., have also been used to expand the utility of dose–
response models.

4.5 What is dose–response modelling?

DRM can be described by six basic steps, with a variety of

options at each step (Table 1). The first four steps, which will be
referred to as dose–response analysis, are aimed at the analysis of
the data available for DRM. Dose–response analysis provides the
linkage of a model to dose–response data for the purposes of
predicting response to a given dose or predicting dose from a given
response. The last two steps deal with implementation and
evaluation of the analysis results.

Table 1. Basic steps in dose–response modelling

Section links
Step Description Options for chapter 6
1. Data Determine the response to End-point, data quality, 6.1
selection be modelled, and select sample size, data utility,
appropriate data data availability
2. Model Choose the type of model End-point, data availability, 6.2.1
selection to be applied to the data purpose
3. Statistical Assumes that statistical End-point, data type, model 6.2.2
linkage distributions describe the choice, software availability
response
4. Parameter Combine the first three Linkage function, software 6.3
estimation steps in an appropriate availability, variance
computer program to
obtain estimates of the
model parameters
5. Implemen- Use the estimated model Outputs, target selection, 6.3
tation parameters and the model model predictions, BMD,

27
 EHC 239: Principles for Modelling Dose–Response

Section links
Step Description Options for chapter 6
formula to predict direct extrapolation
response/dose as needed
6. Evaluation Examine the sensitivity of Model comparison, 6.4, 6.5, 6.6
the resulting predictions to uncertainty
the assumptions used in
the analysis

Step 1 involves selection of the appropriate data for DRM. The

type of data available can have a marked impact on the complexity
of the model that can be used. For example, whereas two points can
be used to identify the slope of a line, it takes at least three points to
identify the shape of a more complex dose–response relationship
(e.g. straight line versus two connected lines). The issue of whether
there are enough data to support a given model is quite complex
(Portier, 1994) and is discussed in greater detail in section 6.1. In
general, the data can restrict the type of model that can be used.

The second step is then to choose an appropriate model. Many

choices exist for modelling dose–response data, and examples of
some of the possible choices are presented in chapter 6. These
models have been generally divided into two categories: empirical
and biologically based models. Empirical models generally refer to
functional forms for which there is limited mechanistic justification
(e.g. the linear model above). Most of the DRM that has been done
to date has focused on the use of empirical models. Biologically
based models generally have functional forms that are derived from
some basic principles about the onset and progression of disease in
a biological system. These models are generally functionally
complicated and require that experience in mathematics, statistics,
and computer science be linked to experience with biological
mechanisms. Mechanistic models also generally have greater data
needs than do empirical models.

The third step requires the choice of a statistical linkage

between the data and the model. The most common linkage method
is to assume a statistical distribution for the response and use that
distribution to derive a mathematical function describing the quality
of the fit of the model to the data. However, a considerable amount
of DRM has been done by simpler linkage functions, such as
drawing a straight line through the data points. The advantage of

28
 Dose–Response Modelling: Basic Concepts

choosing a formal statistical linkage is the ability to test hypotheses

and derive confidence intervals for model predictions.

In DRM, fitting the model to the data is the fourth step. Since
the primary components of a model are the parameters that define
the model, curve fitting simply involves choosing values for the
parameters in the model. If a formal statistical linkage has been
developed for linking the data to the model, then the parameters are
chosen such that they “optimize” the value of the linkage function.
For example, a common choice is to link the data to the model using
2
the squared distance, denoted [R(di) • oij] , between the predicted
value from the model, denoted R(di), and the observed value,
denoted oij. These squared differences can be summed across all
data points, and model parameters are chosen to minimize this sum;
this is the common least-squares algorithm. Simpler methods can
also be used to estimate model parameters. For example, by
drawing a line through the data points, the parameters in the linear
model can be estimated directly, since the value of can be
estimated as the point where the line crosses the y-axis (zero dose)
and the value of can be estimated by calculating the slope of the
drawn line. Formal optimization is a better choice for modelling
than ad hoc procedures, which often do not meet the criterion of
transparency.

The fifth step in DRM is to make the inferences necessary to

develop measures to protect public health. In its simplest form, a
dose–response model allows the prediction of the response if the
dose is known and the calculation of the dose if the aim is to target
a specific level of response. In addition, implementation of the
dose–response analysis (steps 1–4) also encompasses the
extrapolation of results from the specific responses seen for the
experiment being modelled to other exposure scenarios and other
doses. This step can also involve an extrapolation from a laboratory
species to humans. Usually, when making a prediction, the
emphasis is on the change in response seen in the treated animals
compared with the response seen in the controls. The different types
of data (quantal, count, continuous, categorical) require different
methods for predicting changes in response beyond the normal
response. In general, the targets used for additional response fall
into the categories of added response (simply subtract control

29
EHC 239: Principles for Modelling Dose–Response

response), relative response (fold change relative to control

response), and extra response (added response scaled to range from
zero to the maximum possible response). Each of these choices can
impact the final decision, so care should be taken to understand why
a specific choice is made. Figure 4 illustrates some of the basic
components of DRM for the simple linear model case and added
response.

Measures used by public health agencies to prevent excess

exposure to a hazardous agent generally fall into the categories of
direct banning or limiting exposure. DRM could inform both
choices, although its major impact is in the area of limiting
exposure. Several methods on how to use DRM in this context have
been proposed. The simplest is to use the predicted model to find
the dose associated with a negligible (e.g. one in a million) or zero
response over control. In general, this results in extrapolation far
beyond the range of the data, which creates a great deal of
uncertainty. A second approach is to use the dose–response model
to identify a dose with a known response at or slightly below the
observable range (the limit of scientific certainty) and use other
models to get into a range where the response is assumed to be
virtually unchanged relative to the control response. In this
approach, a functional model structure can be used, such as a
straight line, or something simpler, such as uncertainty factors
(UFs), to identify a safe level of exposure. All of these options are
discussed in chapter 6.

The basic steps in DRM shown in Table 1 can be repeated to

consider other options in the process in order to understand the
impact of choices on the predictions from DRM. This final step
(step 6) in DRM is aimed at understanding the sensitivity of the
analysis to specific choices and judging the overall quality of the
final predictions. The simplest way to evaluate sensitivity is by
considering several choices and determining if the results
dramatically change. Depending on the degree of difference
between choices, there could be value in performing a formal
analysis of the quality of the fit of the model to the data. Other
methods can also be used to assess the impact of choices used in the
modelling on the eventual outcome, such as uncertainty analysis
and Bayesian mixing. In some cases, step 6 is performed before step
5, with a focus on the assumptions used in the dose–response
30
 Dose–Response Modelling: Basic Concepts

analysis, and/or after step 5, with a focus on the assumptions used

for implementation. These steps are further described in chapter 6.

4.6 Risk versus safety in dose–response modelling

Risk as used in this discussion is the direct estimation of the

likelihood or degree of an event or its prevalence in a human
population as a function of exposure. Given sufficient data in
humans in the range of exposure where there is concern, it is
possible to obtain scientifically supported estimates of risk. In most
cases, the data used to develop dose–response models are not from
studies in humans in the range of exposures that humans generally
encounter. The most common type of data used for DRM comes
from experiments in laboratory animals, generally at administered
doses significantly exceeding the exposures that humans encounter.
Even when human data are available and suitable for dose–response
analysis, they are generally from selected populations, such as
workers in occupational settings, whose exposures differ from those
of the general population.

Thus, in many cases, dose–response analyses need to be

extrapolated from an observable region where scientific support is
available to a region where scientific support is weaker or non-
existent. For dose–response analyses based on human studies, the
extrapolation is generally a downward extrapolation to different
exposure levels, but extrapolations can also be to different life
stages (e.g. fetus, child) or to different populations with different
environmental factors that might affect exposure (e.g. dietary
differences). For dose–response analyses based upon laboratory
data using animals, there is the additional problem of extrapolating
from animals to humans.

Most of the methods used to implement the results of a dose–

response analysis (step 5) address these extrapolation issues. The
methods that have been used for extrapolation are diverse and
sometimes contentious, with different countries, and even different
agencies within a given country, using different approaches. The
strategies used for extrapolation basically fall into two categories:
those aimed at using estimates of risk for exposures outside of the

31
EHC 239: Principles for Modelling Dose–Response

range of the data used in the dose–response analysis, and those

aimed at establishing safety without using an estimate of risk.

Estimates of risk and the dose associated with that risk

generally require extrapolation from the data on responses and
doses to a lower dose range. These extrapolations can be done using
the model (step 2) that was fit (step 4) to the data (direct estimation)
or a different model, usually a line, extending from the lowest dose
to a point of zero risk. The latter approach is generally envisioned to
be conservative, assuming that the true risk is less than would be
estimated by this second model at all doses below the dose for
which scientific support is clear. In contrast, methods used to
establish safety for a given dose without presenting an estimate of
risk rely upon the concept that a dose that is sufficiently distant
from the lowest dose associated with the observable range will be
safe. This is generally done using uncertainty factors that have been
developed over years of experience. In some cases where the
general human exposure is estimated, however, the difference
between the estimated exposure and the dose at the lowest edge of
scientific support is used (margin of exposure, or MOE).

Regardless of how dose–response analysis is performed,

additional methods are employed to extrapolate to humans. These
methods are also varied, ranging from the use of additional
uncertainty factors to more complicated modelling schemes based
upon differences in toxicokinetics and toxicodynamics between
humans and animals.

The term “risk assessment” is generally used to describe the

entire process of making a public health decision regarding a
specific chemical or agent. However, risk assessment can be
defined further to differentiate between analyses aimed at
establishing safety (as defined above) and analyses aimed at
estimating risks. In this case, “safety assessment” would refer to the
decision process aimed at establishing safety, whereas “risk
assessment” would refer to assessments aimed at estimating risks
that are part of a larger decision process. Safety assessments are
more often used in cases where exposure can be controlled, such as
for food additives and residues of pesticides and veterinary drugs in
foods.

32
 Dose–Response Modelling: Basic Concepts

4.7 Summary

DRM, as used for informing public health decisions about

chemical exposures, is a six-step process. The first four steps
constitute dose–response analysis and relate to the process through
which a mathematical description of the data is obtained in order to
evaluate predicted responses for known doses or to obtain dose
estimates when a chosen response is of interest. The fifth step
involves the implementation of the results of the dose–response
analysis for the purposes of guiding a public health decision. The
final step, which can optionally be applied earlier in DRM, involves
an assessment of the quality of the dose–response analysis and the
sensitivity of model predictions to the assumptions used in the
analysis. DRM, because it involves a large number of choices based
upon scientific experience, can take on many different forms and be
used in many different ways. The remaining chapters of this report
focus on the range of choices available for each step in the process
and some guidance to be used in making these choices.

33
 5. DOSE–RESPONSE MODELLING: WHY AND
WHEN TO USE IT

Dose–response analysis is a major part of the hazard

characterization within the risk assessment paradigm and has been
used in the past for both the characterization of dose–response
relationships observed in animal bioassays as well as the low-dose
extrapolation of incidences of adverse effects to the range of human
exposure levels. Dose–response analysis includes the use of the
NOAEL (pairwise testing) for deriving health-based guidance
values such as the ADI and the use of DRM (fitting functions).

5.1 Historical perspectives

It has always been a challenge to extrapolate from effects

observed in experimental animal bioassays to potential effects in
humans in order to protect humans from potentially harmful
chemical exposures. A variety of approaches have been developed.

The prototype chemical safety assessment uses the ADI

methodology, which was introduced by Lehman & Fitzhugh (1954)
and has come to be widely employed for the derivation of health-
based guidance values (IPCS, 1987). The ADI was originally
devised as a procedure for the regulatory approval of food additives.
Since food additives are deliberately added, the process often
defines what the regulatory agency is willing to accept as a legal
standard of safety. The same methodology is used to derive health-
based guidance values for chemical contaminants. However,
because “acceptable” was deemed to be an inappropriate term for
chemical contaminants, the term “tolerable” was used instead (i.e.
tolerable daily intake, or TDI). Comparable terms that have been
used are provisional maximum tolerable daily intake (IPCS, 1987)
and reference dose (RfD) (Barnes & Dourson, 1988). Other similar
methods exist for different types of exposures, such as for
compounds with accumulating properties—for example, provisional
maximum tolerable weekly intake or provisional maximum
tolerable monthly intake (IPCS, 1987; WHO, 2002).

34
 Dose–Response Modelling: Why and When to Use It

5.1.1 The no-observed-adverse-effect level approach to

acceptable/tolerable daily intake

Calculation of the ADI based on the NOAEL approach for the

case of quantal data is summarized in Table 2.

Table 2. NOAEL-derived ADI for the case of quantal data

Step NOAEL-derived ADI

1. Data Sufficient sample sizes, at least one dose with “no” effect
selection and one dose with effect. Relevant end-points in a relevant
species are important for any approach.
2. Model Statistical method
selection

­
° LIUHVSRQVHDWGRVH'
° LVQRWVLJQLILFDQWO\GLIIHUHQW
° IURPFRQWUROUHVSRQVH
5' ®
° LIUHVSRQVHDWGRVH'
° LVVLJQLILFDQWO\GLIIHUHQW
°
¯ IURPFRQWUROUHVSRQVH

3. Statistical Pairwise statistical tests between dose groups and control

linkage group.
4. Parameter Assessment of point of departure
estimation
12$(/ '12$(/
ZKHUH5 '  IRUDOO'”'DQG
12$(/

5 '  IRUDOO'!'12$(/

This procedure presupposes that all doses below the

NOAEL are non-significant and all doses above the LOAEL
are significant. This is often not the case.
5. Implemen-
tation 12$(/
$',
8)V

where UF is uncertainty factor.

6. Evaluation Statistical power analysis should be performed to check if
the test was sensitive enough to detect relevant effects.

Selecting the data needed to calculate the ADI based on the

NOAEL approach (step 1) is similar to choosing the data to be used

35
 EHC 239: Principles for Modelling Dose–Response

for more complicated modelling; the better data sets have an

appropriate number of relevant doses, sufficient sample sizes, and
relevant end-points in a relevant species. The next step in
calculating an ADI is to determine the NOAEL, which is the highest
concentration or dose of a chemical, found by experiment or
observation, that causes no detectable adverse effect, as defined
above. This includes a statistical method (step 2), statistical linkage
(step 3), and a method of assessment of a point of departure (step 4)
that describes the identification of the NOAEL. Consider a response
procedure, R(D), of the form:

­ LIUHVSRQVHDWGRVH'LVQRWVLJQLILFDQWO\GLIIHUHQWIURPFRQWUROUHVSRQVH
5' ®
¯ LIUHVSRQVHDWGRVH'LVVLJQLILFDQWO\GLIIHUHQWIURPFRQWUROUHVSRQVH

The statistical linkage (step 3) between this procedure and the

data is represented by the statistical test used to determine if a
response at any given dose is different from the control response.
When the response is non-significant, we simply act as if the effect
were in fact zero. Obviously, we cannot conclude that the effect
actually is zero. When the NOAEL approach is chosen, the
statistical test is used to decide upon the existence of a statistically
significant increase (e.g. at the 5% level) over background (e.g. the
control group) for each dose level separately. The selection of the
NOAEL (step 4) is then achieved by choosing the largest dose,
DNOAEL, for which all smaller doses have R(D) = 0 and all larger
doses have R(D) = 1. Mathematically, this assessment can be
written as:

12$(/ '12$(/ ZKHUH5 '  IRUDOO'”'DQG

12$(/

5 '  IRUDOO'!' 12$(/

This procedure presupposes that all doses below the NOAEL are
non-significant and all doses above the LOAEL are significant. This
is not always the case.

The ADI methodology specifies that an acceptable dose of a

chemical may be calculated by dividing the NOAEL by appropriate
uncertainty factors (also called safety factors). Uncertainty factors
are default factors used to account for both uncertainty and
variability.

36
 Dose–Response Modelling: Why and When to Use It

Historically, an uncertainty factor of 100-fold has been used to

convert the NOAEL from an animal study into a health-based
guidance value (Lehman & Fitzhugh, 1954; Dourson & Stara, 1983;
IPCS, 1987). Additional uncertainty factors may be used to allow
for database deficiencies, such as the absence of a chronic study
(IPCS, 1994). The default 100-fold uncertainty factor may be seen
to represent the product of two separate 10-fold factors that allow
for interspecies differences and human variability (IPCS, 1987;
Renwick & Lazarus, 1998). The recognition that the original 100-
fold uncertainty factor could be considered to represent two 10-fold
factors allowed some flexibility, because different factors could be
applied to the NOAEL from a study in humans and from a study in
animals. The concept of chemical-specific adjustment factors
(CSAFs) (IPCS, 1994, 2005) was introduced to allow appropriate
data on species differences and/or human variability in either
toxicokinetics (fate of the chemical in the body) or toxicodynamics
(actions of the chemical on the body) to modify the relevant default
10-fold uncertainty factor. The strategy used by WHO/IPCS in the
NOAEL/ADI approach involves replacing the original 100-fold
uncertainty factor with CSAFs where there are adequate data (IPCS,
1994, 2005).

Regardless of the quantities chosen for the uncertainty factor,

the prediction (step 5) of the ADI from NOAEL-based DRM is
given by the equation:

12$(/
$',
8)V

Step 6 can be extended to the evaluation of the sensitivity of the

ADI to the assumed values of the uncertainty factors.

Some scientists have raised concerns regarding the use of the

NOAEL to determine an ADI. The greatest concern is that the
NOAEL tends to yield lower ADIs for chemicals for which there
are more or better data. Therefore, stakeholders using usually more
costly, better data are “punished” (Crump, 1984; Dourson et al.,
1985; Kimmel & Gaylor, 1988; Barnes et al., 1995; Slob & Pieters,
1998).

37
 EHC 239: Principles for Modelling Dose–Response

5.1.2 The benchmark dose approach to acceptable/tolerable daily

intake

The BMD concept was introduced as an alternative to the

NOAEL approach (Crump, 1984; Kimmel & Gaylor, 1988). The
BMD method has a number of advantages, including the possibility
to extrapolate outside the experimental dose range and respond
appropriately to sample size and the associated uncertainty.

Calculation of the ADI based on the BMD approach is

summarized in Table 3 for the case of quantal data. A generic form
of the BMD and benchmark dose lower confidence limit (BMDL) is
presented in this table. In this document, a variety of response
levels, such as 1%, 5%, and 10%, will be discussed.

Table 3. BMD-derived ADI (Weibull model) for the case of quantal data

Step BMD-derived ADI

1. Data selection Sufficient number of doses with different response
levels and a sufficient number of total subjects.
2. Model selection Fit dose–response model (e.g. Weibull model).
3. Statistical linkage Predicted fractions are linked to observed fractions,
and their “distance” is minimized by optimizing
some fit criteria function (e.g. likelihood function
based on assumed distribution).
4. Parameter Choose an appropriate response, p, in the range of
estimation experimental response. Estimate BMDLp, the 95%
lower confidence bound on the BMDp, where

5 %0'P í5 
S
í5 

5. Implementation
%0'/ S
$',
8)V

6. Evaluation Sensitivity of BMD to model choice can be checked

by fitting various models.

In choosing the data (step 1) for BMD modelling, the same

basic considerations apply as for the NOAEL method. In addition,
studies showing a graded monotonic response with a significant
dose-related trend work best. This is generally true for all DRM
analyses.

38
 Dose–Response Modelling: Why and When to Use It

Choosing a model (step 2) for the BMD method is dependent

upon the types of data available and the characteristics of the
response being modelled. Complicated models will require a larger
number of dose groups than simpler models. Several models have
been proposed for each type of data. In the United States
Environmental Protection Agency’s Benchmark Dose Software
(BMDS) program, a number of routinely used models are cited
(http://www.epa.gov/ncea/bmds/). As an example, assuming the
availability of data that represent the proportion of animals
responding to a given exposure with an adverse effect (e.g. cancer)
from each dose group, one model choice could be the Weibull
model, which has the form:

J
í Eî'
5 '  D í
D íH

where Į is the proportion responding in the unexposed group, ȕ

describes the increase in probability of adverse effect per unit dose,
and Ȗ describes the shape of the dose–response curve (e.g. Ȗ >> 1
implies threshold-like behaviour; Ȗ = 1 implies log-linear
behaviour).

The statistical linkage (step 3) between the data and the model
can assume a number of different forms, as described previously
(section 4.5) and in section 6.2. For quantal data, it is appropriate to
assume that the data are binomially distributed for each dose group.
Estimating model parameters (step 4) for the BMD method can also
be based upon a variety of different methods. For the Weibull
example, one routinely used approach would be to choose the
parameters that maximize the binomial-based log-likelihood.

The concept of the BMD comes from the idea that it is

desirable to use a dose–response model to capture the general
pattern of response for all dose groups in the experimental data set,
but there was some dose, the BMD, below which predictions would
be tenuous. This BMD can be selected in a number of ways (e.g.
Barnes et al., 1995; Murrell et al., 1998), but the most common way
is to choose an excess response, the benchmark response, or BMR
(p), below which there was insufficient support from the data. A
common choice for BMR is p = 10%. Once the BMR (p) is selected,
the BMD, specifically denoted BMDp, is calculated according to the
following equation, if the extra risk formula is used:

39
 EHC 239: Principles for Modelling Dose–Response

5 %0' S í 5 
S
í 5 

Empirical investigations showed for a large and representative

set of compounds that the 95% statistical lower bound on the
estimated BMD may be regarded as an analogue to a NOAEL, and
substituting one with the other would result in similar ADIs
(Crump, 1984; Barnes et al., 1995). As with all aspects of
modelling, many choices exist for calculating confidence bounds,
and these are discussed further in chapter 6.

Having chosen a method for estimating a 95% statistical lower

bound on BMDp, which can be called BMDLp, the ADI can be
calculated as follows:

%0'/ S
$',
8)V

In this calculation, the values of the uncertainty factors could

be the same as those used for the NOAEL or adjusted to account for
a slightly different interpretation for the BMDLp relative to the
NOAEL (Renwick et al., 2003).

The BMD method includes the determination of the response at

a given dose, the dose at a given response, and their confidence
limits. Using extrapolation of the dose–response model below the
biologically observable dose range, the response at specified
(lower) dose levels can be estimated as well as the dose
corresponding to a specific response level.

5.2 Points of consideration

The use of DRM in general for hazard characterization is

possible when a sufficient amount of dose–response information is
available, either from an experimental animal bioassay or from a
human study (epidemiological study or clinical trial). As shown in
the previous section, the BMD can be considered as an alternative
point of departure for deriving an ADI in those situations where a
NOAEL would have been used as a point of departure in current
procedures. In addition, DRM may be helpful in those situations

40
 Dose–Response Modelling: Why and When to Use It

where there is a need for low-dose extrapolation (e.g. substances

that are genotoxic and carcinogenic). It should be noted, however,
that extrapolation from a single model that fits the data in the
observed range cannot be justified, since other models fitting the
data equally well may result in substantially different estimates of
low-dose risk. Bayesian approaches are currently under
development, which take into account both statistical uncertainties
in the data and model uncertainty (see section 6.5). In practice,
linear extrapolation from a BMD10 (or ED10, effective dose for a
10% risk, approximately equal to BMD10) is often applied as a
simple method for low-dose extrapolation. This is considered a
conservative approach. As another application, DRM may be used
to estimate risks at any given (human) exposure level. Since human
exposure levels are usually lower than the doses in the observed
range in animal studies, methods for low-dose extrapolation may
also be needed in this application.

5.2.1 General aspects of definition

The NOAEL is a parameter derived directly from the observed

dose–response data and is defined as the highest administered dose
at which the effect is still not significantly different from that at
dose 0 (see section 5.1). The NOAEL is based on a multiple test
procedure performed along the applied dose series. It lacks further
detailed statistical properties compared with a parameter of a dose–
response model, for which the precision of the estimate can be
quantified.

The dependence of the NOAEL on the statistical significance

test, however, tends to penalize chemicals for which there are more
or better data by giving a higher estimate for those chemicals with
less precise data. This problem does not occur in DRM. In fact, the
opposite relationship holds: it penalizes studies with few or poor
data.

The NOAEL approach can be formally considered a

dichotomous procedure, where no effect is assumed to be present
below the NOAEL and where an expression of the critical effect is
present above the NOAEL (see section 5.1 and Table 2). Given the
typical animal studies used in toxicology, the effect size that can be
detected by a statistical test may be larger than 10% (additional

41
 EHC 239: Principles for Modelling Dose–Response

risk). Therefore, the NOAEL may be expected to be a dose at which

the effect is in reality somewhere between 0% and 10% or more. In
contrast, the BMD is a dose for which the size of the effect has been
predefined, and thus it is under the control of the risk assessor.
Furthermore, while the dichotomy of the NOAEL approach does
not provide quantitative information about risk above the ADI, such
information might be derived from fitted dose–response models,
where such dichotomy does not exist.

In general, DRM estimates are based on data from the entire

dose–response curve for the critical effect. The standard NOAEL
approach can be regarded as a special, simplified case of dose–
response analysis, as it identifies a single dose that is assumed to be
without an appreciable adverse effect. The dose–response model
thus reflects the characteristics of the dose–response curve,
particularly in providing estimates of slope. In the case of a
regression framework, it provides standard error and confidence
intervals for the model parameters.

5.2.2 Estimation procedure

NOAELs are restricted by the set of doses used in the specific

studies. An important consequence is that the NOAEL may be
either below or above the threshold it aims to approximate,
assuming one exists. When the true threshold is higher than the
NOAEL, the distance between the two can be expected to be limited
(related to the dose spacing used). However, when the true
threshold is lower than the NOAEL, the distance between the two is
unlimited: the true threshold could be anywhere between zero and
the NOAEL.

The actual value of the NOAEL depends strongly on the

following characteristics of the study design:

x Group size. The power to detect a NOAEL at some dose level

is directly dependent on the sample sizes chosen at those dose
levels (Gaylor, 1989). The larger the group size, the smaller the
potential true effect size at the NOAEL.

x Dose location. Since the NOAEL is an applied dose that did

not show significant effects, while the next higher dose applied
did show significant effects, the NOAEL can only be one of the

42
 Dose–Response Modelling: Why and When to Use It

doses actually applied in the study. A particularly disturbing

disadvantage of the NOAEL approach is that in some cases no
NOAEL can be assessed because the lowest applied dose
showed effects.

x Experimental variation. Larger experimental variation between

subjects will result in lower statistical power and, hence, higher
NOAELs. In quantal data, this phenomenon is somewhat
hidden, but in continuous data, it is directly visible: it is
reflected by the scatter in the data per dose group. This
experimental variation comprises various things: biological
(e.g. genetic) variation between subjects, variation in
experimental conditions (e.g. time of feeding, location in
experimental room, time of section or interim measurements),
and measurement errors.

DRM-derived estimates are based on interpolation, and these

estimates are not restricted to the actually applied doses. DRM can
also be used on a study where no NOAEL (only a LOAEL) can be
defined, so in this situation another study may be unnecessary. A
comparison of different models can be useful. When multiple
models are fit to the same data and produce widely varying BMD
estimates, caution should be used in interpreting the results, as this
could indicate insufficient data for modelling (see chapter 6).

It should be noted that in comparison with the NOAEL

approach, implementation of the full DRM approach may lead to
differences between the NOAEL and the BMD in individual data
sets. However, on average, BMDs that represent the lower
confidence interval on the dosage giving a BMR of 5% or 10% tend
to be quite similar to the NOAELs (Allen et al., 1994). Therefore, in
data sets where DRM cannot be applied, the NOAEL may serve as a
reasonable surrogate of the BMD.

5.2.3 Uncertainty

A modelling approach facilitates both sensitivity and

uncertainty analyses. Uncertainty (see section 6.5) can be expressed
numerically when the doses and responses are linked by a model.
Such numerical analyses can also be subject to sensitivity analyses,
to test the contribution of different aspects of the database or of

43
 EHC 239: Principles for Modelling Dose–Response

model characteristics to the overall uncertainty. The uncertainty in

the risk estimates that arises from aspects of study design, such as
dose spacing, sample size, and biological variability, can be
assessed in a dose–response model. While uncertainty factors are
amenable to uncertainty analysis (Slob & Pieters, 1998), the
threshold procedure of the NOAEL is not readily amenable to
quantitative estimation of uncertainty or to a sensitivity analysis.

A disadvantage of using the threshold procedure of the

NOAEL for the estimation of a point of departure (“starting point”)
for formulating advice to risk managers is that it is not possible to
quantify the degree of variability and uncertainty that may be
present. The NOAEL is assumed to be a dose without biologically
significant effects. This assumption is more likely to be valid in
toxicological studies with larger sample sizes.

5.2.4 Study design

A design optimal for the NOAEL approach could limit the use
of DRM, and vice versa. While the NOAEL approach requires
sufficient sample sizes within dose groups (to warrant statistical
power), the DRM approach requires a sufficient number of dose
groups (to warrant a description of the whole dose–response).
Given the restrictions on the total number of animals used in a
single study, these two requirements may not be compatible.

An important point to bear in mind is that DRM can be used on

studies carried out in the past and based on the traditional designs
(with three dose groups and a control). Some have argued that
optimal designs for dose–response models may have the advantage
for animal welfare that fewer animals could be used (Slob et al.,
2005).

While DRM provides uncertain estimates when the number of

dose groups is too small, the determination of both the
BMD/BMDL and the NOAEL may prove inadequate at different
points when the number of animals per dose group is too small. For
example, when the critical effect is seen in a larger experimental
animal, such as the dog, with few animals per dose group, the
NOAEL may be high owing to the insensitivity of the test. The
BMD/BMDL approach, however, can be used to evaluate sparse
dose–response data and quantify the inherent uncertainty. However,

44
 Dose–Response Modelling: Why and When to Use It

even here, where an apparent dose–response relationship in the data

remains, the BMD/BMDL may also provide very uncertain
estimates. Therefore, a typical four-dose study with a few animals
per dose may in practice be unreliable whatever method, NOAEL or
BMD, is applied. However, the advantage of DRM is that this
uncertainty is made visible, whereas in the NOAEL approach it
remains hidden.

DRM reduces the need for more experiments when a small

degree of extrapolation is needed (e.g. when the doses used are near
the human exposure level). In contrast, the NOAEL approach may
require further experiments where no clear NOAEL (or LOAEL)
can be identified. This can be illustrated by the study of Allen et al.
(1996) on developmental risk assessment in rats exposed to boric
acid in their diet. This study failed to establish a NOAEL; however,
the BMD approach could have been applied, thereby avoiding the
need for repeat studies (see also section 5.2.2 above). Distributing
the total number of animals over more dose groups does not result
in poorer performance, despite the smaller number of animals per
dose group, as shown by Slob et al. (2005). The above example of
Allen et al. (1996) suggests that the BMD approach provides a
reasonable basis for appropriately comparing and combining
studies, as opposed to ad hoc combinations of study results.

A major advantage of DRM is the ability to estimate risks

within the observable range of effects. In animal studies, it is
possible to estimate risk over the full range of doses used.
Estimation of risks outside the observable range will be more and
more unreliable when risks get smaller and smaller (Murrell et al.,
1998). Some studies (e.g. Sand et al., 2002) have investigated the
effect of model dependence at different response levels.

Some experts have argued that extrapolation to risk levels

outside the observable range might be warranted when there are
indications that the same toxicological mechanism is active in both
the extrapolation region and the experimental region of the model
fit. However, mathematical models that adequately describe the full
complexity of the mechanisms involved are very rare; even then, it
needs to be additionally assumed that the parameters estimated from
the data (i.e. the observable range) are adequate for the low-dose

45
 EHC 239: Principles for Modelling Dose–Response

range and have sufficient precision to make the prediction (see also
section 6.5.3).

5.2.5 Biological information

The NOAEL approach incorporates biological information

through the application of “expert” but subjective judgement. Full
DRM has the potential for a more “science-rich” analysis through
the more formal quantitative inclusion of factors/covariates into the
models, in the case of both human epidemiological and animal data.

Such an approach can lead to more certain estimates, centred

on a toxicologically based concept of estimating the dose–response
relationship on the basis of all available biological knowledge,
using empirical data and applying statistical inference. More
complicated models can be developed on the basis of toxicokinetics
and toxicodynamics.

5.2.6 Comparison of experimental results

NOAELs derive from an algorithmic analysis of the results of a

single experiment. Meta-analysis on data such as NOAELs across a
range of studies on a specific chemical is possible, such as when
data are insufficient to build a dose–response model, but may be
limited by the statistical properties of the NOAEL estimates.

Estimates derived from full DRM, however, enhance the ability

to compare different experiments, effects, and compounds using a
common framework. The estimates obtained may provide a test of
consistency among different studies that may use different dose
levels. DRM methodology can be used to describe dose–response
relationships in different studies (e.g. rat and mouse, chronic and
subchronic exposure, healthy and diseased animals) if suitable data
sets exist.

Rules for combining studies, however, need to be developed.

Descriptions of the dose–response on the same end-points in
different studies may be integrated to provide a cohesive picture of
the chemical’s toxicity. The values obtained using DRM may result
in estimates for each end-point on the basis of biological and
functional relevance.

46
 Dose–Response Modelling: Why and When to Use It

5.2.7 Risk management perspectives

The potential use of the estimates from DRM can, from a risk
management perspective, give an improved characterization for
decision-making by:

x giving the decision-maker a “more-than-one-point” apprecia-

tion of the data;
x providing information about what happens above the safety
level (magnitude and types of health impacts);
x quantifying benefits in risk reduction from different regulatory
actions;
x promoting consistency in decisions, if appropriate adjustments
are made for differences in effect, effect level, species, and
study design; and
x allowing for an iterative interaction between the risk assessor
and risk manager on a continuous and ongoing basis.

5.3 Implementation issues

In the case of the BMD, there are a number of decisions to be

made in applying the method and determining a BMDL: for
example, which mathematical model to use; what degree of
confidence to use in calculating confidence limits; what response
level to predetermine as the BMR (e.g. BMR = 1%, 5%, or 10%
incidence of an effect, or a 5% or 10% change in a continuous end-
point, such as body weight or red blood cell counts). It is often not
clear what response level (BMR) can be considered as non-adverse.
For example, should a 5% decrease in red blood cell counts be
considered as adverse, or should a smaller (or larger) change be
chosen? Should up to a 5% increased incidence in hepatocellular
hypertrophy be considered as acceptable in an animal study, or is a
maximum of 10% increase adequate? These and other choices need
additional discussion among toxicologists and clinicians. Although
an explicit statement on the BMR is an improvement compared with
the generally unknown response level associated with a NOAEL,
choices of a BMR need consensus building.

47
 EHC 239: Principles for Modelling Dose–Response

5.4 Summary

The characterization of dose–response relationships in animal

and human studies has been a major component of hazard
characterization. Over the years, a variety of methods have been
developed to accommodate such relationships. DRM may be
regarded as the most adequate approach for analysing dose–
response data, provided that a suitable data set of animal or human
dose–response data is available.

The standard NOAEL approach identifies a single dose that is

assumed to be without appreciable effect, whereas the BMD is
based on data from the entire dose–response curve, estimated for
the critical effect. Although the effect at the NOAEL is assumed to
be zero, it will be non-zero in many cases, although to what extent
remains unknown. The size of the effect at the BMD is made
explicit and, as far as possible, is based on toxicological knowledge.
While the uncertainty in a NOAEL cannot be quantified, the
uncertainty in a BMD can be quantified by a confidence interval.
The use of DRM may call for different guidelines for optimal study
designs, as the number of dose groups should be sufficiently large.
Distributing the total number of animals over more dose groups
may be done without loss of precision. DRM can more effectively
compare different experiments, effects, and compounds. While risks
above the ADI based on a NOAEL cannot be quantified, such may
be possible for exposures exceeding the ADI based on DRM. For
estimating risks below the observable range, extrapolation based on
a single fitted model is unwarranted. Here, linear extrapolation may
be considered as a conservative approach. Currently, more
advanced methods are being developed (e.g. Bayesian approaches)
for low-dose extrapolation based on DRM.

48
 6. PRINCIPLES OF DOSE–RESPONSE MODELLING

6.1 Data

6.1.1 Selection of data

When considering which data to use from a set of available

toxicity studies on a particular compound, it may not be effective to
do a dose–response analysis for each observed end-point in each
study. As a first step, one may omit studies that have obviously
larger NOAELs compared with the other studies. In this way, one
may, for example, select for a given type of toxic response (e.g.
chronic, developmental) for the most sensitive species. For a given
study, many end-points may have been measured. End-points not
showing a clear dose–response on visual inspection can be omitted.
Then, based on the toxicological impact together with the apparent
magnitude of the response, a selection of end-points can be made as
candidates for modelling. It would be very helpful if submitted
studies included an annex with plots (in addition to tables) of
observed data points for each end-point, possibly with fitted curves
to the plots, to enhance the process of selecting end-points. At a
minimum, these should be included for end-points showing evident
effects.

After selecting the potentially relevant end-points, one must

decide whether each dose–response data set is actually amenable for
a dose–response analysis. Generally, it is desirable to have at least
three or four different doses (including controls). In addition, the
associated effect levels need to be different from each other; it is
preferable to have at least three different response levels.

6.1.2 Data types

There are various types of response data, and these can be

categorized in various ways. The main distinction relevant for
effects is that between quantal and continuous data. Quantal data
relate to an effect that is observed or not in each individual subject
(laboratory animal or human). Hence, for each dose, the number of
subjects responding out of the number of subjects available is

49
EHC 239: Principles for Modelling Dose–Response

reported. In continuous data, a quantitative measurement is

associated with each individual subject. As an intermediate type of
data, ordinal data reflect (ordered) severity categories—that is, they
are qualitative data but with a rank order (e.g. histopathological
severity data). When the categories are non-ordered, they are called
categorical data, but these are rare for response data. Finally, count
data form another class of data (i.e. discrete data), but in practice
they can often be treated as continuous data (see also section 4.3).

Although the type of data is important for statistical reasons

(see section 6.2.2 on distributions), the distinction between quantal
and continuous data also has a crucial impact on interpretation of
results and their ensuing use in risk assessment. In the case of
quantal dose–response data, information on the change of incidence
with dose is available at one particular degree of effect. For
example, the incidence of cleft palate may increase as dose
increases, but under the categories “no cleft palate” and “cleft
palate”, there is no information about the degree of the effect. In
ordinal and continuous data, in contrast, information on both the
degree of effect and the incidence is available as a function of dose.
So, for example, cleft palate might be categorized into an ordinal
variable with levels “no clefting”, “mild clefting”, “moderate
clefting”, and “severe clefting”, or it might be quantified in a
continuous variable as, for example, the fraction of closure. The
relationship between the average response and dose gives
information about how exposure changes the degree of effect. For
instance, a plot of (average) red blood cell count may show the
decrease in mean red blood cell count (i.e. the degree of the effect)
as a function of dose. By also considering the individual data points,
information on the incidence can be derived as well. For example,
an estimate of the fraction of individuals with red blood cell counts
less than some critical value can be derived.

When using animals as a model for human response, the

observed dose–response information is assumed to mimic the dose–
response in humans to some approximation. It might be argued that
this assumption is more plausible for degree of effect than for
incidence. The problem is that the observed dose–incidence
relationship for animals largely reflects the variation in the animals
used, which is highly controlled in a laboratory experiment. Hence,
it may not mimic the human variation.

50
 Principles of Dose–Response Modelling

6.2 Models and distributions

6.2.1 Dose–response models

6.2.1.1 Continuous dose–response models

The models listed in Table 4 are some of the forms that may be
used to describe the relationship between dose and the magnitude of
a response on a continuous scale in an individual. When combined
with a statistical distribution (e.g. normal or lognormal), these
equations can also be used to describe the relationship between dose
and a continuous response in a population, where the continuous
model corresponds to the central estimate.

Dose–response data are often adjusted by subtracting the

(mean) control value from each individual observation. However,
this procedure does not account for the fact that the background
response level in the controls is, like the response level in the
experimental groups, subject to sampling error. A better approach is
to account for the background response in the model with a
parameter that needs to be estimated from the data. Among the
many ways in which this can be done, the following are three of the
simplest:

1. y = a + fx(D)
2. y = a × fx(D)
3. y = fx(a + D)

where D is dose, a is the background term, and fx may be any dose–

response function. For some assessments, there may be mechanistic
information that makes one form preferable to another. For
example, the first form is preferable for modelling an influence that
produces the effect independently, the second corresponds to the
idea of normalizing the response as a fraction of the background
response, and the third reflects a contribution from another agent
acting by the same mechanism.

51
 Table 4. Continuous dose–response models

Name(s) Notes Equation for response Parameter explanations

Michaelis-Menten law of A theoretical account of enzyme- or receptor- RMax is the maximum rate of the
mass action based activity where the rate of action is a >6@ reaction, [S] is the substrate
50D[
function of the rate of association (ka) and . 0  >6@ concentration, and KM is the Michaelis-
the rate of dissociation (kd). Menten constant, which is equal to
ka/kd.
Hill equation log-logistic A modification of the Michaelis-Menten RMax is the maximum response, D is
equation that supposes that the occupation of 'Q the dose, KD is the reaction constant for
50D[ Q Q
multiple sites or receptors is required for the .'  ' the drug–receptor interaction, and n is
production of an effect. the number of (hypothetical) binding
sites.
First-order exponential If the interaction between a chemical and a RMax is the maximum response, D is
target site is irreversible, then the rate of the 50D[ í Hí U' the dose, and r is the exponential rate
reaction is determined by the rate of constant.
association (ka) only.
Į
Power Simple exponential model. = ȕD D is the dose, Į is the shape
parameter, and ȕ is the scale
parameter.
Linear Although there is usually no biological theory = mD D is the dose and m is the slope.
to suggest it, linear models are often justified
by their simplicity; linear models have but a
single parameter.
 Principles of Dose–Response Modelling

6.2.1.2 Quantal dose–response models

Quantal dose–response functions describe the relationship

between dose and the frequency of a particular outcome in a
population (see Table 5). For a group of homogeneous or nearly
identical individuals, the relationship between dose and frequency
can be described with a step function where all subjects either
respond or fail to respond at any given dose. However, because
variability is ubiquitous in living organisms, quantal dose–response
data typically show gradually increasing incidence with dose. One
interpretation of this is that individual subjects differ in tolerance to
the agent, which can be described by a statistical tolerance
distribution. Hence, any cumulative distribution function may be
used as a quantal dose–response function. Other models have been
derived from statistical assumptions about how the agent might
exert its effect in an organism, such as the gamma multi-hit model.
Background response rates should, just as in the case of
continuous data, be accounted for by incorporating an additional
parameter in the dose–response model. The two simplest ways of
doing this are:

1. y = a + (1 – a) f(x)
2. y = f(x + a)

where f(x) is any dose–response function (varying from 0 to 1). As

with continuous data, correcting the data for background response
prior to the dose–response analysis is statistically unsound. The
background response level should be estimated simultaneously with
the dose–response model and be treated in the same way as the
observed responses in the other dose groups.

6.2.1.3 Thresholds

The term “threshold” can be used in three different senses.

First, it is used in a scientific sense to indicate a level of exposure at
which no effect occurs (e.g. there is a physical stimulus, but there is
no response). Second, a threshold may be thought of as a level at
which there may or may not be an effect, but it is too small to be
observed (e.g. a NOAEL). In this case, it is the perceptual limitation
of an observer or analyst, rather than the actual subject of the
experiment, that is being described. As a third meaning, a “practical

53
 Table 5. Quantal dose–response models
Name(s) Theoretical basis Equation for Parameter explanations
frequency (F)
Step function No variability. If D < T, F = 0 D is the dose and T is the threshold parameter.
If D • T, F = 1
One-hit (single- Hit theory models employ the use of a rate D is the dose, e is Euler’s constant, Į is a location
hit) to describe the interaction between a íHí AB' parameter, and ȕ is the slope parameter.
group of causal agents (e.g. molecules)
and a group of targets (e.g. a human
population).
Gamma multi- An expansion of the one-hit model, which = ī(gamma*D, k) ī() is the incomplete gamma CDF, D is the dose,
hit is based on the notion that multiple hits or gamma is a rate parameter, and k is the number of
events are required to produce a particular hits required to produce the effect.
effect.
Probit normal A descriptive model based on a normal or = )(Į + D*ȕ) )() is the normal CDF, D is the dose, Į is a location
Gaussian distribution. parameter, and ȕ is the slope parameter.

Logistic The statistical logistic model is also a = D is the dose, Į is a location parameter, and ȕ is the
descriptive tool with no theoretical basis. Hí A í' î B slope parameter.
G
í A  B î'
Weibull A flexible descriptive model originally =H D is the dose, Į is the background parameter, ȕ is
developed to describe survival data in the slope parameter, and Ȗ is an exponent.
demography.
CDF, cumulative distribution function
 Principles of Dose–Response Modelling

threshold” is a response where the consequences are determined to

be trivial and not worth further consideration.

A threshold in the first sense may be incorporated into a model.

The introduction of a threshold parameter truncates the dose–
response relation at a threshold dose:

x Below threshold, the effective dose is zero.

x Above threshold, the effective dose is the dose minus threshold.

Threshold terms generally are difficult to estimate accurately

and have large confidence limits.

6.2.1.4 Severity (degree of effect)

The severity of toxic responses is rarely used in DRM other

than in a qualitative manner (e.g. tumour formation vs reduced
fertility). However, one may also consider severity or degree of
response in a quantitative way at the level of a single end-point. As
noted above, the dose–response of continuous end-points may be
directly interpreted as a dose-related change in degree of effect—for
example, a per cent decrease in haematocrit (Woutersen et al., 2001)
or a per cent change in body weight (see Figure 5, which represents
the dose–response relationship between body weight and exposure
to the mycotoxin deoxynivalenol) (Pieters et al., 2004). Here, a
certain degree of effect (5% reduction in body weight) is chosen for
deriving the BMD. The BMDL is then defined as the dose
associated with a particular (e.g. 5%) change in degree of effect for
that end-point.

An important advantage of defining a BMR in terms of degree

of effect based on continuous response data is that values for the
BMR that may be considered non-adverse are within or close to the
range of observations. Therefore, low-dose extrapolation may not
be needed or needed only to a small extent when continuous end-
points are considered.

55
EHC 239: Principles for Modelling Dose–Response

\ D H[S E[



%RG\ZHLJKW J
 

    

/RJRIGRVH PJNJGD\
Fig. 5. Dose–response model fitted to male (circles) and female (triangles)
body weights plotted against log dose (exposure to the mycotoxin
deoxynivalenol). The plotted marks represent the (geometric) means of about
40 mice, with 90% confidence intervals. The BMD associated with a BMR of
5% is estimated at 0.24 mg/kg body weight (log equivalent = í0.62), with a
lower confidence bound of 0.22 mg/kg body weight (log equivalent = í0.66).
The latter value can be considered as a BMDL for this end-point (adapted
from Pieters et al., 2004).

In the case of a histopathological end-point resulting in ordinal

data, a dose–response function may be fit using categorical
regression, and the BMDL associated with a particular degree of
effect (e.g. minimal or mild) may be estimated (e.g. Piersma et al.,
2000; Woutersen et al., 2001).
Categorical regression may also be applied at a higher level—
that is, in an analysis of multiple studies (Hertzberg & Miller, 1985;
Hertzberg, 1991; Hertzberg & Wymer, 1991). In this application of
categorical regression, severity categories are defined covering
disparate end-points. Most of these applications focus on estimating

56
 Principles of Dose–Response Modelling

the likelihood that a given category of severity may occur at a given

dose level.

6.2.1.5 Modelling with covariates

In some circumstances, it is desirable to include variables in

addition to an exposure variable in dose–response models. For
example, in epidemiological studies, it is common to model disease
risk in terms of not only exposure, but also age, sex, socioeconomic
status, smoking status, and other measurements that may be relevant
to the disease state. These other factors may be correlated with
exposure status because of the way in which the sample was taken.
Then, unless the proper covariates are included in a model for the
relationship between exposure and the health end-point, the effect
of exposure will be incorrectly estimated. In bioassay studies, in
which animals are randomized to treatment groups, this sort of
confounding cannot, in principle, occur, but it may be useful to
include a covariate such as sex to account for some of the variability
in a related measure (see Figure 6).

6.2.1.6 Biologically based dose–response models

While biological considerations may motivate the choice of one

or several empirical models, the level of biological detail in such
models is minimal. Thus, their credibility for interpolating and
extrapolating a data set derives mainly from their fit to the data, as
evaluated statistically. Another class of model, the biologically
based dose–response model, is much more complicated and is
explicitly designed to model the biological details that lead from
initial exposure to a toxicant to the ultimate pathological outcome.
Typically, such a model includes a physiologically based
toxicokinetic model to describe the distribution and metabolism of
the parent compound and toxic metabolites and other mechanistic,
or toxicodynamic, models that link target tissue concentration to the
ultimate response. The toxicodynamic part of the models may be
relatively simple (e.g. when the outcome is inhibition of
acetylcholinesterase in the model for chlorpyrifos; Timchalk et al.,
2002) or as complicated as a fully elaborated stochastic model for
carcinogenesis (Sherman & Portier, 1998). Such a model is really a
quantitative expression of a set of biological hypotheses and, when
rigorously tested against critical experiments, becomes a credible
tool for extrapolating from experimental results into exposure

57
 EHC 239: Principles for Modelling Dose–Response

realms that are difficult or expensive to reproduce in controlled

experiments. Such models are quite expensive to construct both in
resources and in time and thus would be expected to be developed
fully only for exposures and toxicities of the highest concern.

(\ D H[S E[


6HUXP$/7 8,





   

'RVH PJNJERG\ZHLJKWGD\

Fig. 6. Dose–response model fit to serum alanine aminotransferase (ALT)

levels observed in males (circles) and females (triangles), where sex is
treated as a covariate. In this case, the parameter a (background response
level) differs between sexes, whereas parameter b and the residual variance
(var) for the log(data) do not differ between sexes.

58
 Principles of Dose–Response Modelling

6.2.2 Statistical distributions

6.2.2.1 Continuous distributions

The normal or Gaussian distribution is symmetrical and defined

from minus to plus infinity. It has two parameters: the mean and
standard deviation, which control the location and scale of the
distribution, respectively. Because sums of large numbers of small
effects tend to be approximately normally distributed, this
distribution is often used to describe variability and the variation of
measurement error.

The lognormal distribution has two parameters: the geometric

mean and the geometric standard deviation. It can be considered as
a derivative of the normal distribution where the logarithms of the
observed or predicted values are assumed to be normally
distributed. This produces a skewed distribution on the original
scale. Another consequence of using a lognormal distribution is that
it will not generate negative values, which makes it more suitable
for describing positive-only data sets and unsuited for values with
negative values. Since many distributions are skewed and contain
only positive numbers, the lognormal distribution often provides a
good description. In addition, products of a large number of small
effects tend to be approximately lognormally distributed. Since
effects in biological measures tend to be multiplicative
(proportional) rather than additive, the lognormal distribution is
generally more suitable for biological measures.

The Weibull distribution is most commonly used to represent

the survival or “lifetime” distribution of physical systems/products
or biological systems, depending upon the context. In many
applications, there is no explicit theoretical reasoning indicating that
a Weibull distribution is appropriate or should be used, although the
distribution does have some theoretical underpinning within the
class of extreme value distributions. From a curve-fitting
standpoint, the functional form of the distribution is simply a power
transformation of the exponential model, which gives the model
more flexibility for describing data. The multi-hit model is a special
case of the Weibull model.

A more complete list of continuous distributions is given in

Evans et al. (1993).

59
 EHC 239: Principles for Modelling Dose–Response

6.2.2.2 Discrete distributions

Discrete distributions describe responses on a finite or infinite

scale, preferably count data; a special case is a response with a
dichotomous quantal outcome of 0 or 1.

A Bernoulli distribution has an outcome of 1 or 0,

corresponding to the occurrence or absence of an event that occurs
with frequency f over an infinite sequence of trials. The Bernoulli
distribution is then simply “1” with frequency f and “0” with
frequency 1 í f. The Bernoulli trial is the basis of the binomial
distribution, the definition of which subsumes the former.

The binomial distribution is defined as the distribution of a sum

of a given number of Bernoulli trials with outcome of 1 or 0,
denoting the occurrence or absence of a specified event,
respectively. In toxicological applications, the number of trials is
fixed by the experimental design, and the proportion of subjects in
which the specified event occurs is the response to be estimated. As
a result, the binomial distribution is the distribution typically used to
estimate quantal response model parameters.

The Poisson distribution is a one-parameter distribution for a

positive and discrete valued response. The domain of the response
variable is any positive integer. The distribution was originally
derived as a distribution of rare events: specifically, the number (n)
of events occurring in a sequence of Bernoulli trials where the
number of trials is large and the probability (P) of events per trial is
small. Consequently, the Poisson distribution can be used as an
approximation of the binomial distribution when n is large and P is
small. The Poisson distribution is commonly used in analyses of
epidemiological data when the study design involves prospectively
following a cohort of subjects over a time period for which the
expected incidence of adverse events is small relative to the cohort
size.

A more complete list of discrete distributions can be found in

Evans et al. (1993).

60
 Principles of Dose–Response Modelling

6.3 Model fitting and estimation of parameters

The general principles of parameter estimation and model

fitting have been discussed in chapter 4. Two basic methodologies
are available for model fitting: conventional, in which parameters
are selected to minimize or maximize an objective function, and
Bayesian, in which information in a data set is combined with prior
information about model parameters, resulting in a posterior
distribution for those parameters that reflects the degree of
uncertainty about those parameters. For historical and
computational reasons, “user-friendly” software designed for
carrying out dose–response analysis and non-linear modelling in
general has been restricted to using conventional methodologies,
whereas Bayesian methods are implemented in packages that
require more extensive programming and substantially greater
understanding of the statistical details (for further details on
Bayesian approaches, see Hasselblad & Jarabek, 1995; Gelman et
al., 2004). While such software requires substantial statistical
understanding for successful use of Bayesian methods and is thus
beyond the scope of this document, even conventional methods
require an understanding of some basic principles before outcomes
from applying the software can be properly interpreted. Some
general remarks are given below.

6.3.1 Criterion function

The general approach of fitting a model is to find parameter

values for the model that optimize the fit of the model to the data.
To that end, a criterion function is defined, reflecting the fit of the
model. The goal is to find the parameter values that optimize the
value of the criterion. For many models typically used, this can be
achieved only by an iterative “trial and error” approach (see below).

In many applications, the logarithm of the likelihood function is

used as the criterion. The likelihood directly derives from the
distribution assumed for the scatter in the data. For quantal data, the
binomial likelihood is typically used. For continuous data, the
normal likelihood is often used, be it for the observed responses
themselves or for the log-transformed responses. Note that
maximizing the likelihood function for data that are assumed to be

61
 EHC 239: Principles for Modelling Dose–Response

normally distributed is in fact equivalent to minimizing the sum of

squares.

6.3.2 Search algorithms

Computer software employs algorithms to find parameter

values that optimize the fit of the model to the data, and the user
does not need to worry about the exact nature of the calculations.
However, some basic understanding of the search process is
required in order to interpret the outcomes.

An iterative search algorithm tries to find “better” parameter

values in a process by evaluating whether the fit can be improved
by changing the parameter values through a trial and error process.
More advanced algorithms operate by evaluating the slope of the
likelihood at which the fit is improved for one or more parameter
value changes (basically using the slope to “climb the likelihood
function” as quickly as possible to find the top value). The
algorithm can start searching only when the parameters have values
to start with. Although the software often gives a reasonable first
guess for the starting values, the user may have to change these. It is
not unusual (in particular when the information in the data is hardly
sufficient to estimate the intended parameters) that the end result
depends on the starting values chosen, and the user should be aware
of that.

The algorithm keeps on varying the parameter values until

criteria for stopping are satisfied. There are two major reasons for
the algorithm to stop the searching process:

1. The algorithm has converged (e.g. it has found a clear

maximum in the log-likelihood function). In this case, the
associated parameter values can be considered as the “best”
estimates—e.g. the maximum likelihood estimate—if the
likelihood was maximized. However, it can happen that the log-
likelihood function has not one but more (local) maxima. This
means that one may get other results when running the
algorithm again, but with other start values. This can be
understood by remembering that the algorithm can only “feel”
the slope locally, so that it usually finds the optimum that is
closest to the starting point.

62
 Principles of Dose–Response Modelling

2. The algorithm has not converged (i.e. the algorithm was not
able to find a clear optimum in the likelihood function, but it
stops because the maximum number of iterations [trials] is
exceeded). This may occur when the starting values were
poorly chosen, such that the associated model would be too far
away from the data. Another reason could be that the
information in the data is poor relative to the number of
parameters to be estimated. For example, a dose–response
model with five unknown parameters cannot be estimated with
a study with four dose groups. As another example, the
variation between the observations within dose groups may be
large compared with the overall change in the dose–response.
In these cases, the likelihood function may be very flat, and the
algorithm cannot find a point where the function changes
between increasing and decreasing. The user may recognize
such situations by high correlations between parameter
estimates (i.e. changing the value of one parameter may be
compensated by another), leaving the model prediction
practically unchanged.

6.4 Model comparison

The fundamental criterion for judging a model is that the

selected model should describe the data, especially in regions of the
dose–response where inferences are needed. Most fitting methods
provide a global goodness-of-fit measure, usually providing a p-
value. These measures quantify the degree to which the model
predictions correspond to the data. Small p-values indicate a poor fit
to the data. Since it is particularly important that the data be
adequately described, it is recommended that a p-value of 0.1 be
used to compute the critical value for goodness of fit, instead of the
more conventional values of 0.05 or 0.01.

Another way to detect the form of these deviations from fit is

with graphical displays. Plots should always supplement goodness-
of-fit testing. For continuous data, it would be extremely helpful for
plots that include data points to also include a measure of dispersion
of those data points. In certain cases, the typical models used in
DRM cannot fit the observed data, such as when the data are not
monotonic or when the response rises abruptly after some lower
doses that give only the background response. In these cases,

63
EHC 239: Principles for Modelling Dose–Response

adjustments to the data (e.g. a transformation of dose) or the model

(e.g. adjustments for unrelated deaths) may be helpful.

When fitting many different models to the same data, they

generally will not all result in the same fit, and some care must be
taken in choosing which model or models will be considered. In
applying a statistical theory to this problem, one of four possible
situations may arise:

1. The models form a nested series of models in the same family,

in the sense that there is a “full” model, and other “restricted”
models are derived from that full model by setting successively
more parameters to a fixed value or, conversely, successively
incorporating more parameters into the model. Likelihood ratio
tests can be used to evaluate whether the improvement in fit
afforded by estimating additional parameters is justified. The
general form of the test is to calculate 2 × (LLfull – LLrestricted),
where LL is log-likelihood, and compare this with a critical
value from the chi-squared distribution with Pfull – Prestricted
degrees of freedom (where Px is the number of parameters
estimated in model x).

2. The models are from the same family, but do not form a nested
series. Some statistics, notably Akaike’s information criterion
(AIC is í2LL + 2P, where LL is the log-likelihood at the
maximum likelihood estimates for the parameters, and P is the
number of model degrees of freedom) can be used to compare
models (Akaike, 1973; Burnham & Anderson, 2002). In this
case, the model with the smallest AIC value is selected,
although models with similar AIC values (differing by no more
than about 4) are probably equivalent (Burnham & Anderson,
2002).

3. The models are not from the same family, but are fit using the
same assumptions about the underlying probability
distributions (e.g. all using a lognormal likelihood or all using a
normal likelihood). In this case, Burnham & Anderson (2002)
argue that AIC can still be used to identify the best model, but
this appears to be a controversial point. Sand et al. (2002) have
shown that it may be difficult to discriminate between the
commonly used quantal dose–response models based on the
AIC, which may be due to the fact that these models are quite

64
 Principles of Dose–Response Modelling

similar in their structure and include a similar number of

parameters. In general, this case is still the subject of statistical
research. At present, it will probably be adequate to use AIC to
select a model as in the previous case, recognizing that this
guidance may change.

4. Models do not use the same probability distribution. In this

case, little formal statistical guidance is available. The
plausibility of assumptions about the distribution of data needs
to be examined by looking at the distribution of individual data.
However, continuous data are often aggregated and reported as
means and standard deviations, which eliminates the possibility
of examining distributional assumptions. In these situations, the
best that can be done is to rely on past experience with the end-
points being modelled and select a reasonable probability
distribution.

6.5 Representing uncertainty

Any parameters or predictions estimated from a given model

are only point estimates and, to a larger or smaller extent, uncertain.
This uncertainty arises from at least three sources:

1. Sampling error—the sampling error arising from inferences

about a larger population from a single experiment;
2. Study error—the reality that dose–response estimates often
differ among experiments with different experimental design,
protocol, or uncontrolled circumstances; and
3. Model error—the fact that the “true” model is not known,
which results in additional uncertainty when interpolating
between doses, but even more so when extrapolating outside
the dose range containing observations.

These three sources of uncertainty are briefly discussed below.

6.5.1 Sampling error

Uncertainty arising from sampling error with a single

experiment is perhaps the easiest to evaluate and report. It may
typically be quantified by a standard error or, preferably, by a

65
 EHC 239: Principles for Modelling Dose–Response

confidence interval. Confidence intervals may be calculated in

several ways:

x plus or minus twice the parameter’s standard error (provided by

most dose–response software), which is estimated by the
second derivative of the likelihood function (Hessian or
information matrix);
x based on the profile of the log-likelihood function, using the
chi-square approximation of the log-likelihood;
x bootstrap methods (see, for example, Efron, 1987; Efron &
Tibshirani, 1993); and
x Bayesian methods, in particular if one has some preliminary
knowledge of the plausible range of the parameter.

Various studies have compared the first three methods and

concluded that the first may result in inaccurate intervals, whereas
the second and third methods give similar results (see, for example,
Moerbeek et al., 2004).

6.5.2 Study error

Uncertainty about the true value of a parameter that stems from

variability among experiments can often be handled by treating the
experiments as comprising an additional level of hierarchy, when
the experiments are very similar in design and intent (e.g. same
agent on the same end-point in the same strain and species). To
characterize uncertainty in a statistical framework, it can be
assumed that there is a population of experiments from which the
ones at hand were selected (e.g. Davidian & Giltinan, 1995). As a
result, the prediction or parameter of interest varies around a mean
value among the members of that population of experiments, and an
estimate of the mean and the degree of confidence can be derived. It
should be noted that, even if data from only one experiment are
available for analysis, this source of uncertainty still exists—it may
be possible to quantify this uncertainty by analogy.

6.5.3 Model error

The third area of uncertainty, model uncertainty, is reflected by

the question: to what extent do the data, possibly along with other
knowledge about dose–response shape, constrain the set of possible
dose–response shapes? A statistical model completely hinges on the

66
 Principles of Dose–Response Modelling

dose–response data, and the quality of the data is in fact the crucial
aspect. In the fitting process, a model tries to hit the response at the
observed doses. However, when a model is used to make
inferences, interpolation between observed doses and extrapolation
beyond the non-control doses are possible approaches. Thus, the
model must also predict the response in the non-observed dose
range. In other words, there are two aspects in evaluating the fitted
model: one should assess not only if the model succeeded in
describing the observed responses, but also if the model can be
trusted to describe the non-observed responses where it is desirable
to make inferences. The former aspect focuses on the quality of the
model, the latter on the quality of the data. The following discussion
elaborates on how to deal with the second of these two aspects (the
first was addressed in section 6.4, Model comparison).

There are two ways to evaluate whether the data provide

sufficient information to constrain the model and allow inference in
some defined range outside of the range of the data. The fitted
dose–response model should be visually inspected, to check if the
data provide sufficient information to confine the model. Here, the
question should be asked: if a curve is drawn through the data
points by hand, could that be done in disparate ways? For instance,
in the top panel of Figure 7, three curves have been drawn through
the data points, each of which might be close to the true dose–
response curve in the range between 2 and 5. In the bottom panel,
however, it is very difficult to imagine that the true dose–response
relationship would be very different from the (single) curve drawn
here in the same range.
Another way to deal with this question is by comparing the
outcomes from different fitted models. If the data do contain
sufficient information to confine the shape of the dose–response
relationship, different models fitting the data (nearly) equally well
will result in similar fits and similar inferences. As an illustration,
Figure 8 shows two different models fit to the same (continuous)
data. Owing to the good quality of the data, they result in very
similar estimated dose–response relationships. Inferences from
dose–response models bear an additional level of uncertainty in
proportion to the degree with which those inferences depend on the
model used.

67
EHC 239: Principles for Modelling Dose–Response


UHVSRQVH




    
GRVH

UHVSRQVH




    
GRVH

Fig. 7. Two data sets illustrating the idea of model uncertainty. In the top
panel, the data (either quantal or continuous) do not contain sufficient
information to confine the dose–response relationship in the range between 2
and 5: one may imagine various disparate curves that are all in agreement
with the data, and hence they all might represent the true dose–response
relationship. In the bottom panel, the data points prohibit the possibility of
drawing disparate curves between 2 and 5.

In current practice, there is a tendency to focus only on

goodness of fit, and passing a formal goodness-of-fit test is often

68
 Principles of Dose–Response Modelling

regarded as sufficient evidence that the model is acceptable. This is

unfortunate, since a goodness-of-fit test tends to be more easily
passed for data with few dose groups or when few dose-related
responses are noted and therefore non-observed responses are
important or dominate. In addition, a goodness-of-fit test assumes
that the experiment was carried out perfectly (i.e. perfectly random
with respect to all potentially relevant experimental factors and
actions). Clearly, this assumption is not realistic.

It is re-emphasized that a dose–response model, as long as it is

not based on the mechanism of action of the particular chemical,
serves only to smooth the observed dose–response relationship and
to provide for a tool to assess confidence intervals. A statistical
regression model itself has little, if any, biological meaning, and the
choice of the model is to some extent arbitrary. It is the data, much
more than the model, that should determine the dose–response
relationship and any inferences derived from it. When different
models (with similar goodness of fit and equal number of
parameters) result in different estimates, this reflects a component
of uncertainty that needs to be quantified and communicated with
the estimate.
Dose–response models that are based on the mechanism of
action of a particular chemical stand in opposition to statistical
models as described here. Such mechanistic models contain
information gleaned from biological theory and typically multiple
experiments and therefore are less sensitive to data gaps (between
dose groups). However, they do contain unknown parameters that
need to be estimated from the data and thus require the resulting
uncertainties to be quantified. Since such models are typically
complex and idiosyncratic, little further general advice can be
given, and it is suggested that professional statistical advice be
sought in such cases.
Model uncertainty is particularly relevant to the issue of low-
dose extrapolation. Here, the problem is that there may well be
several models that are consistent with the data, as shown in the top
half of Figure 9, and so give similar predictions in the range of the
data, but whose predictions diverge at the low end of the dose
range, as depicted in the lower half of Figure 9. One way to collect
and represent model uncertainty in a risk assessment is through the

69
EHC 239: Principles for Modelling Dose–Response

\ D >F F H[S E[AG @

3URDVW


YDU
D

&('
F
G


ORJON
)HWDOZHLJKW J

E


%05
%0'/


%0'/



      

'RVH PJNJEZGD\

\ D >F F H[S E[AG @

3URDVW


YDU
D


&('
F
G


ORJON
)HWDOZHLJKW J

E


%05
%0'/


%0'/



      

'RVH PJNJEZGD\

Fig. 8. Two different models (both with four parameters) fitted to the same
data set resulting in similar dose–response relationships and similar
BMD(L)s. Small circles indicate individual observations, large circles
(geometric) group means.

70
 Principles of Dose–Response Modelling

use of probability trees (Rescher, 1969; Hacking, 1976). A

probability tree is a logical construct that may be used to represent a
set of mutually exclusive propositions. For example, if the three
models depicted in the top panel of Figure 8 were equally well
supported, then each model would have a probability of 0.33. If one
model had a weight that was 6 times greater than the others, then it
would have a probability of 0.75, whereas the others would have a
probability of 0.125. Note that the probability of a model does not
depend only on the strength of evidential support; it also depends on
what other models are being considered. A model with little support
may have a high probability if all the alternatives under
consideration have even less support. Quantitative measures of
model preference may be combined to produce an overall rank or to
provide a formal measure of the weight of the evidence.
To some extent, all quantitative methods for assigning model
probabilities rely on untestable assumptions or elements of
judgement. Therefore, the simplest and most straightforward
method for assigning probabilities to models is to simply give them
all the same weight. This approach is implicit when the predictions
from different models are simply listed (e.g. Ghani at al., 2000).
Another relatively simple approach is to ask the experts to identify
plausible theories and then apply probabilities to them (Evans et al.,
1994; IPCS, 2000). These probabilities can then be updated to
incorporate additional information in the data by using Bayesian
methods. However, there are many formal techniques for assigning
weights or probabilities to models (Bozdogan, 1987; Raftery et al.,
1997). A semiformal approach may be used in which the same
criteria discussed in the section for selecting models (section 6.2.1)
may also be used to weight and assign probabilities to each
alternative model considered (e.g. Carrington & Bolger, 2000).
Model uncertainty may also be integrated with sampling error by
using bootstrapping techniques. This involves repeatedly drawing
random samples from the data set and refitting each data set with a
set of models. The best models from each bootstrap are then
retained in a probability tree to represent both parameter and model
uncertainty.

71
EHC 239: Principles for Modelling Dose–Response







([SRQHQWLDO

)UHTXHQF\

7UXQFDWHG/RJLVWLF
:HLEXOO
 *DPPD
/RJ/RJLVWLF
/RJ3URELW

'DWD








          

%HQ]R D S\UHQHGRVH PJNJERG\ZHLJKWSHUGD\


)UHTXHQF\ DGGHGULVN



([SRQHQWLDO
7UXQFDWHG/RJLVWLF
:HLEXOO
 *DPPD
/RJ/RJLVWLF
/RJ3URELW
'DWD





( ( ( ( ( ( ( ( (

%HQ]R D S\UHQHGRVH PJNJERG\ZHLJKWSHUGD\

Fig. 9. Model uncertainty in low-dose extrapolation. Different models may all

fit the data reasonably well (top), but yield highly divergent response
estimates at low doses (bottom). The data and models are taken from
Fitzgerald et al. (2004).

72
 Principles of Dose–Response Modelling

Alternatively, some people have addressed this uncertainty by

choosing a subset of the models that appear to fit the data well.
From these models, those with adequate fits are summarized with a
range and associated variance. When choosing a final value for the
BMD, these values can be aggregated by taking a mean or
geometric mean to provide a central point estimate (National Health
and Medical Research Council, 1999) or a value simply chosen
through expert judgement (WHO, 2006).

6.6 Benchmark dose and benchmark response selection

One important use of DRM is the calculation of BMDs. A

BMD is the dose at which it is inferred that a particular,
prespecified level of response would occur. The methodology was
introduced in Crump (1984) as an alternative to the use of NOAELs
and LOAELs in dose–response assessment for determining
quantities such as ADIs. The main advantages of the use of the
BMD over NOAELs and LOAELs stems from the more complete
use of dose–response data by BMD methods and from the fact that
uncertainties about the value of a BMD can be quantified using
statistical methodology. The uncertainty of a BMD may be
expressed as a confidence interval, in which case the lower end of a
one-sided 95% confidence interval is termed the BMDL, or as a full
Bayesian posterior distribution.

The BMR is the response for which the BMD is to be

calculated. There are both technical and policy aspects associated
with selecting the BMR. The technical aspects have to do with just
how the BMR is expressed; different types of end-points, such as
quantal and continuous, require different treatments. Also, in
somewhat more complicated situations, such as when covariates
have been used in the modelling, the BMD depends on the BMR
and possibly on the values of the covariates. Policy issues have to
do with just how high or low down the dose–response curve the
BMR should be. This section discusses the technical issues
surrounding the choice of BMR and some of the consequences that
need to be considered in making the policy decision about where to
set the BMR, but it does not directly address the choice of its
particular value.

73
EHC 239: Principles for Modelling Dose–Response

The way in which the BMR is expressed depends upon the kind
of response variable being modelled. For end-points with two states
(affected/not affected), the BMR is usually expressed in a way that
adjusts for background. Two equations are common. One is that of
added risk (AR):

BMRAR f ( BMD )  f (0)

where fx represents the dose–response function evaluated at dose x.

The other, which is probably most widely used, is extra risk (ER):

f ( BMD)  f (0)
BMRER
1  f (0)

where added risk is divided by the non-affected fraction of the non-

exposed population. The response at the BMDER is always smaller
than the response at the BMDAR for the same numerical value of
BMR when there is a background incidence. However, for small to
moderate background response, the difference is small.
A third equation, common in epidemiological analyses, but
applicable to animal studies as well, is relative risk (RR):

BMRRR f BMD / f 0

BMRs for continuous end-points can be expressed directly in

terms of changes in the mean response level or indirectly in terms of
the fraction of experimental animals that exceed (or drop below)
some critical level. For example, the BMD for mean adult body
weight might be selected to be the dose at which the mean body
weight drops below 90% of the body weight in controls or at which
brain acetylcholinesterase activity is inhibited by 10% relative to
control levels (this is often termed the critical effect size). One
might also specify a fixed value or fixed drop in the mean,
selecting, for example, the dose at which the mean nerve conduction
velocity drops below a fixed rate or a fixed difference from that in
unexposed individuals. For end-points that demonstrate a sigmoidal
response, as does enzyme induction, it has been suggested (Murrell
et al., 1998; see Gaylor & Aylward, 2004, for a contrary argument)
that a formulation similar to extra risk be used: for these end-points,
the authors suggest that the BMD is best characterized as the dose at

74
 Principles of Dose–Response Modelling

which the response is a specified fraction of the total dynamic range

(e.g. the difference between background and maximum possible
induction) of the response. The Gaylor & Aylward (2004) approach
considers a certain setting within the definition of the response (i.e.
a 1% change) and compares the uncertainties in the resulting BMD
with the uncertainties in BMDs estimated using the specific setting
in the “hybrid” approach. Thus, their conclusion may not hold in
general terms (e.g. considering a 5% or 10% change in response
relative to the total dynamic range).

Indirect or “hybrid” approaches have been advocated by Crump

(2002) and Gaylor and his co-authors (Gaylor & Slikker, 1994;
Kodell et al., 1995). In indirect approaches, the relationship
between the mean of a continuous variable and dose is modelled, in
the same manner as in the direct approaches. Next, a critical value
for the continuous variable is determined that is to be considered as
adverse, and an extra (or additional) risk BMR is selected for which
to calculate a BMD. It is preferable that the critical value be based
upon biological considerations, but it may otherwise be a value in
the tail of the distribution of values in the control group. As the
mean response increases, so will the fraction of subjects that exceed
the previously determined critical value. The BMD is the dose at
which the fraction exceeding the critical value corresponds to the
fraction of affected animals associated with the BMR as defined for
quantal data (e.g. BMRER).

It is possible to approximate the BMD as calculated in the

previous paragraph (Crump, 1995) for a critical value
corresponding to a “small” (e.g. 0.1–2%) risk in the control group
and extra risk in the vicinity of 10%. This BMD corresponds
approximately to the dose at which the mean of the response
variable differs from the control mean by an amount equal to the
standard deviation of the control group. This gives another way to
specify a BMR for continuous variables, based on the variability of
the animals used in the bioassays.

Both hybrid methods based on variability discussed above

require that the variability be true interindividual variability, and not
be due to large assay errors. They depend critically on the idea that
extreme quantiles of an unexposed population may be thought of as
affected in the same sense as an individual with the same value from

75
 EHC 239: Principles for Modelling Dose–Response

an exposed population. Sand et al. (2003) examined how the hybrid

approach depends on the estimate of variance. Gaylor & Slikker
(2004) discussed how different sources of variability may be
separated.

In some cases, the dose is not the only independent variable in

a dose–response model. For example, in epidemiological studies,
often many covariates that help characterize an individual and that
might influence the response variable and be incidentally associated
with the exposure variable are included in analyses in an attempt to
reduce bias in the estimates of the effects of exposure (see section
6.2.1.4). In developmental bioassays, characteristics of the dam or
the litter as a whole (e.g. number of implantation sites) may be used
as a covariate in the modelling to help explain some of the
additional variation among litters usually seen in such studies. Even
adult-only rodent bioassays are usually segregated by sex.
Typically, then, the assessor needs to decide for which values of the
covariates BMDs need to be calculated. When there are few,
discrete covariates, it may make sense to calculate a separate BMD
for each set of values (e.g. a BMD for both males and females).
When covariates are continuous (or treated as such, as in number of
implantation sites), in an animal bioassay, it is usual to pick a
typical value in the control group. However, if BMD changes with
the value of the continuous variable, a detailed analysis of the
dependency should be undertaken (e.g. modelling the BMD as a
function of that covariate). If the variable makes sense for
extrapolation to the human situation, it might be informative to
calculate the BMDs for several values of the covariate, to evaluate
the sensitivity of the BMD to the range of covariate values for
humans.

6.7 Summary

Data sets for DRM generally need to be selected to reflect the

more sensitive end-points available, just to reduce potential
workload. Models used depend upon the type of data (continuous,
ordered categorical, quantal, or counts) and include a model for
dose–response and a model for the variability of the data. Once
models are fit to a data set, the degree to which they individually
describe the data is evaluated using goodness-of-fit measures; in
addition, their ability to describe the data with respect to each other
may be compared using measures such as the AIC.

76
 Principles of Dose–Response Modelling

Uncertainty about the inferences that result from such models

fall into three main categories: statistical uncertainty of inferences
due to variability among responses in experimental subjects,
variability among experiments due to unavoidable differences in
experimental execution, and uncertainty due to the fact that
different models yield different approximations of the true dose–
response relationship. Dose–response analysis needs to address all
three sources of uncertainty whenever possible.

One particularly important application of DRM is the

calculation of BMDs, doses at which it is inferred that a particular
level of response would occur. When data are available, BMDs are
a better alternative than NOAELs or LOAELs in the calculation of
guidance values such as ADIs or TDIs. When extrapolation is
necessary, the uncertainty associated with any predictions made
should be represented. It is often especially important to include
model uncertainty.

77
 7. COMMUNICATING THE RESULTS OF DOSE–
RESPONSE MODELLING

7.1 Introduction

Risk communication has been defined as the “interactive

exchange of information about (health or environmental) risks
among risk assessors, managers, news media, interested groups and
the general public” (IPCS, 2004). Risk communication has evolved
with the rest of the risk analysis paradigm to embrace the
“interactive” nature of the processes. The transition from
monologue to reflexive dialogue in risk communication has
necessitated awareness that risk perception issues are extremely
important. The scientific, political, and social perspectives of bench
scientists, risk assessors, risk managers, media, and the public can
result in considerable misunderstandings and misinterpretations
(Garvin, 2001). The preconception that scientific and technical
knowledge and their application in risk analysis are value free and
objective has often resulted in the marginalization of insights from
other sources.

General public perception, resulting from health-based

guidance approaches and terminology such as “ADI”, “TDI”, and
“threshold”, is that there is a bright line between “safe” and
“unsafe”. These approaches are not designed to incorporate risk and
benefit dynamics and may not require or even allow an outside
audience to become engaged in the decision process. For many
considerations of chemical exposures, these dynamics do not have
to be dealt with because the outcome of the safety/risk assessment
provides a perfectly useful and acceptable answer to the risk
manager. However, there are instances where these dynamics will
need to be considered and evaluated.

The use of DRM and other probabilistic assessment techniques

to quantitatively describe variability and uncertainty brings new
challenges in risk communication. Some of these challenges are:

x explaining that a certain percentage of the population is

predicted to experience some effect;

78
 Communicating the Results of Dose–Response Modelling

x explaining the level of risk in those circumstances where there

is no safe level of exposure;
x comparing competing risks or benefits;
x providing a focus on uncertainties that are attendant to the
predicted risk; and
x explaining that the risk generally is described at the population
level, rather than the individual level, noting that this is also the
case for the ADI/TDI approach.

In addition, one of the limitations of the current health-based

guidance approach is that it gives no information about risk when
the ADI/TDI is exceeded. For example, some subpopulations may
exceed the health-based guidance value for dioxins, and the DRM
approach may provide additional information that is useful for the
risk manager and communicator.

An appreciation of the variability in most populations clearly

impacts risk communication. This is particularly true for genotoxic
carcinogens and other substances, such as lead, that are unavoidable
contaminants and may be toxic at low levels. Using a point estimate
to depict an entire population in the context of risk communication
can be misleading, because it can suggest that the risks are larger for
the entire population than they really are if upper percentile point
estimates are used, and it ignores the fact that some portion of the
population does have a somewhat higher level of risk. Becoming
involved in a public decision requires a transformation from
concern for an individual to concern for a population and thinking
about variability as an inherent part of the problem rather than just a
source of uncertainty.

In risk communication, uncertainties can facilitate dialogue.

Uncertainty analysis can inform all the parties of what is known,
what is not known, and the weight of evidence for what is only
partially understood. However, there are currently no general
criteria for the application of weight-of-evidence approaches. An
appreciation of uncertainty, including uncertainty about variability,
can lead to better consideration of the options for seeking better
information, using a value-of-information approach (Thompson,
2002). However, in risk communication, “uncertainty” can be a
double-edged sword. When the results of a probabilistic risk
assessment are presented, uncertainty is specifically described rather

79
 EHC 239: Principles for Modelling Dose–Response

than managed by the use of a default factor. Since the responsibility

for managing the uncertainty is left to the discretion of the
management process, communicating the uncertainty to the
participants in that process is very important.

The application of DRM and other probabilistic risk assessment

techniques has the potential for improving risk analysis and public
risk perception. There must be an acknowledgement of the
limitations and weaknesses of the technical knowledge in addition
to its strengths. There should also be the realization that there may
be difficulties with risk comparisons and that social perceptions can
drive precautionary considerations. There may not be agreement on
how to interpret new information or on the appropriate criteria for
making or reversing risk decisions. The critical contribution of
probabilistic approaches is that they can improve the processes of
risk assessment and risk management and thereby facilitate
communication. As a result, participation in the decision process
will be broadened.

7.2 Incorporation of the outputs of dose–response

modelling into risk assessment

The output of dose–response analysis can be used in various

ways, depending on problem formulation and the nature of the
effect modelled. An output may be presented in three principal ways
as the basis for advice on the possible health implications of human
exposure:

1. establishment of a health-based guidance value, such as an ADI

or TDI, which is a daily intake over a lifetime that is considered
to be without appreciable health risk (this would be analogous
to current procedures based on a NOAEL or LOAEL);
2. estimation of the MOE as the ratio between the dose–response
output and the estimate of human exposure; and
3. quantitative estimation of the magnitude of the risk at the level
of human exposure, derived from the modelled dose–response
relationship.

The discussion below assumes that the dose used in the dose–
response model was the external dose expressed in milligrams per
kilogram body weight. The use of internal or target organ dose
estimated by a physiologically based toxicokinetic model would
80
 Communicating the Results of Dose–Response Modelling

reduce the uncertainties of interspecies extrapolation, because

kinetics are a major source of species differences, such that a
reduced uncertainty factor would be required.

7.3 Derivation of health-based guidance values

Traditionally, a health-based guidance value for threshold

effects has been derived from a NOAEL or LOAEL divided by an
appropriate composite uncertainty factor, either default values or
CSAFs (IPCS, 2005), on the assumption that the NOAEL represents
an intake close to the threshold for the adverse effect. In practice,
the limit of detection for the incidence of adverse effects in animal
experiments depends on the sample size, and more than 100 animals
may be needed to achieve confidence intervals in the range of ±5%.

Many studies have shown that the BMDL for a 5% response is

similar to the experimental NOAEL (Allen et al., 1994). Fowles et
al. (1999) came to a somewhat different conclusion. They examined
acute inhalation lethality data and compared NOAELs with BMDs
corresponding to 1%, 5%, and 10% response incidences. Similarly
to the “quantal” parts of the results of the Allen et al. (1994) studies,
BMDLs based on 10% incidence corresponded approximately to
NOAELs. However, because the dose–response for lethality is so
steep, BMDLs for 5% and 1% incidences were very close to those
for 10% incidence. As a result, the BMDLs for a 1% incidence were
on average only about 1.6 or 3.6 times smaller than a NOAEL,
depending on whether a log-probit or Weibull model was used. This
possibly can be explained by the smaller sample sizes in these
experiments, not by the difference in end-points.

Given the uncertainty in the relationship of the NOAEL and the

threshold of the adverse effect, finding a BMR such that the
resulting BMD and BMDL correspond numerically (on average) to
a NOAEL may not be relevant and is certainly not necessary for the
application of BMD approaches. Also, the use of the BMDL to set a
health-based guidance value would need to take into account the
same uncertainties as when a NOAEL is used as the basis for
establishing an ADI/TDI.

81
 EHC 239: Principles for Modelling Dose–Response

7.4 Estimation of the margin of exposure

The normal default uncertainty factor of 100 has a long history

of use for threshold effects and can be regarded as the margin
between two points—the NOAEL or BMDL from the experimental
data and a level of human intake/exposure that would be without
appreciable health risk. Because this is based on a NOAEL or
BMDL, the ratio is equivalent to a margin of safety, and there
would be negligible risk providing that the intake was at or less than
the ADI/TDI.

In the case of adverse effects that are considered not to show a

biological threshold in their dose–response, the BMDL could not be
considered to represent an intake close to a threshold, but is simply
the confidence interval on the BMD. Consequently, the margin
between the BMDL and the estimated human intake/exposure
would not be a margin or safety and is therefore termed an MOE.
The MOE is calculated as the ratio between two experimental
estimates, the BMDL and the predicted or estimated human
intake/exposure. Calculation of an MOE does not require
extrapolation of the data beyond the range of observations (IPCS,
1999; Edler et al., 2002).

Uncertainties related to interspecies differences and human

variability, which are the basis for the usual 100-fold uncertainty
factor used in the derivation of an ADI/TDI, would be equally
applicable to an MOE based on animal data, but there would be
additional uncertainties related to the nature of the dose–response
relationship below the experimental/observable range, the impact of
genetic polymorphisms in the processes critical to the production of
a mutated cell, and the subsequent clonal expansion and progression
into a cancer. Consequently, an MOE of 100 would be inadequate
to reflect the fact that the starting point (the BMDL) cannot be
regarded as a threshold or the additional uncertainties related to the
mode of action.

Application of linear low-dose extrapolation using the BMDL

for a 5% response (see below) to estimate a one in a million lifetime
risk is equivalent to an MOE of 50 000.

82
 Communicating the Results of Dose–Response Modelling

7.5 Quantitative estimations of the magnitude of the risk

at levels of human exposure

The results of a dose–response model can be used to estimate

the possible risks at intakes/exposures above a health-based
guidance value such as the ADI and at very low levels of human
exposure or to estimate intakes/exposures associated with
predefined levels of risk, such as a one in a million lifetime risk of
cancer.

Estimation of risks of intakes/exposures above a health-based

guidance value, derived by the application of uncertainty factors to
a BMDL from a study in either animals or humans, would need to
use the slope characteristics in the dose–response model. For
example, if an intake is of concern because it is above the health-
based guidance value, then the extent of any risk could be estimated
by reference back to the modelled animal dose–response
relationship. Traditionally, an estimate of the possible risk has not
been made, and intakes above the ADI/TDI have been considered to
have eroded the uncertainty factor. However, if one assumes that
the dose–response relationship in humans has a similar shape to that
in the animal study, the ADI is set with default uncertainty factors
that will obscure any quantitative estimates of the risk above the
ADI (the risk at the ADI is assumed to be negligible). More
accurate estimates of differences in sensitivity between humans and
animals would be required for such calculations.

Estimation of risks at very low levels of human exposure or of

exposures associated with responses below the BMR requires
extrapolation outside the data used to generate the dose–response
model. Extrapolation outside the observed range—for example,
from an incidence of about 5% to one in a million—will require
extrapolation over many orders of magnitude. Low-dose
extrapolation may be undertaken using the dose–response
relationship defined by the model that was fitted to the experimental
data or by application of a standardized mathematical approach,
such as linear extrapolation, to the starting point. An advantage of
using a model is that the risk estimates can be compared across
different compounds. The major uncertainty associated with such
estimates is the biological relevance of the model in the region of
extrapolation.

83
 EHC 239: Principles for Modelling Dose–Response

7.6 Presentation of results

In a scientific or logical sense, the risk assessment is finished

when the conclusions have been drawn. However, when the
conclusions are simulation results, some distillation or condensation
is often necessary in order to make the results comprehensible.
Since there is always some danger that crucial information may be
lost, care must be exercised to ensure that the summary process
does not omit information that is important for the decision.

7.6.1 Tables

Precise communication of quantitative information requires

numbers. More numbers will portray more information than fewer
numbers, but will take longer to assimilate. Tables 6–10 give
examples of the range of options, from high to low complexity, that
may be considered, all taken from the same simulation results for
exposure. It is recommended that, in case of effects of concern or a
single effect found in several studies, all quantitative results be
summarized in a table. The risk assessor should sort out the most
relevant results and present the data to the risk manager in a clear
and understandable way.

7.6.2 Graphs

Although they may allow quick comparison, tables inherently

compare one value at a time. Graphing or visualization is in some
ways a better means of digesting the entire distribution. A one-
dimensional simulation will produce a frequency distribution (when
simulating variability) or a likelihood distribution (when
representing uncertainty). There are two ways of plotting frequency
or likelihood curves (see Figure 10). The first is to plot density
against value, which emphasizes the values that are the most
common or likely. The second is to plot cumulative percentiles
against value, which allows the percentile corresponding to a
particular value to be read from the plot. A graphical presentation of
the dose modelling in relation to the experimental data may also be
helpful in deciding which dose descriptor should be used for
lifetime risk.

84
 Table 6. Population percentiles from a two-dimensional simulation

Uncertainty
Average SD Minimum P1 P5 P10 P25 Median P75 P90 P95 P99 Maximum
Average 0.457 0.063 0.234 0.236 0.366 0.403 0.456 0.462 0.497 0.502 0.503 0.510 0.874
Minimum 0.047 0.061 0.000 0.000 0.000 0.000 0.016 0.055 0.076 0.076 0.076 0.076 0.874
P1 0.094 0.065 0.000 0.000 0.000 0.007 0.072 0.101 0.129 0.129 0.130 0.131 0.874
P5 0.146 0.068 0.000 0.000 0.000 0.069 0.144 0.148 0.178 0.179 0.180 0.180 0.874
P10 0.188 0.074 0.000 0.000 0.000 0.116 0.187 0.205 0.216 0.216 0.217 0.218 0.874
P25 0.274 0.083 0.000 0.000 0.119 0.207 0.287 0.291 0.317 0.320 0.320 0.327 0.874
Median 0.401 0.105 0.000 0.000 0.267 0.352 0.399 0.404 0.471 0.476 0.476 0.484 0.874

Variability
P75 0.586 0.064 0.388 0.394 0.519 0.531 0.561 0.568 0.651 0.657 0.657 0.667 0.874
P90 0.808 0.030 0.760 0.762 0.774 0.776 0.784 0.790 0.843 0.847 0.848 0.858 0.874
P95 0.949 0.024 0.874 0.923 0.930 0.931 0.941 0.944 0.953 0.963 1.014 1.056 1.058
P99 1.247 0.086 0.874 1.138 1.142 1.147 1.149 1.287 1.296 1.321 1.403 1.462 1.473
Maximum 2.192 0.483 0.875 1.573 1.579 1.584 1.599 2.559 2.592 2.608 2.619 2.663 2.670
Pxx = xxth percentile; SD = standard deviation.
 EHC 239: Principles for Modelling Dose–Response

Table 7. Population percentiles with confidence intervals

Percentile Average (confidence interval)

Average 0.457 (0.366, 0.503)
Minimum 0.047 (0.000, 0.076)
1st percentile 0.094 (0.000, 0.130)
5th percentile 0.146 (0.000, 0.180)
10th percentile 0.188 (0.000, 0.217)
25th percentile 0.274 (0.119, 0.320)
Median 0.401 (0.267, 0.476)
75th percentile 0.586 (0.519, 0.657)
90th percentile 0.808 (0.774, 0.848)
95th percentile 0.949 (0.930, 1.014)
99th percentile 1.247 (1.142, 1.403)
Maximum 2.192 (1.579, 2.619)

Table 8. Population percentiles with standard deviations

Percentile Average ± standard deviation

Average 0.457 ± 0.063
Minimum 0.047 ± 0.061
1st percentile 0.094 ± 0.065
5th percentile 0.146 ± 0.068
10th percentile 0.188 ± 0.074
25th percentile 0.274 ± 0.083
Median 0.401 ± 0.105
75th percentile 0.586 ± 0.064
90th percentile 0.808 ± 0.030
95th percentile 0.949 ± 0.024
99th percentile 1.247 ± 0.086
Maximum 2.192 ± 0.483

Table 9. Selected population percentiles with confidence intervals

Percentile Average (confidence interval)

Average 0.457 (0.366, 0.503)
Median 0.401 (0.267, 0.476)
90th percentile 0.808 (0.774, 0.848)
95th percentile 0.949 (0.930, 1.014)
99th percentile 1.247 (1.142, 1.403)

86
 Communicating the Results of Dose–Response Modelling

Table 10. Population mean with uncertainty estimate

Average ± standard deviation

Average 0.457 ± 0.063

“'HQVLW\












          

“&XPXODWLYH

















          

Fig. 10. Plotting frequency distributions.

Two-dimensional results are more difficult to display. Two

strategies for adding an extra dimension are illustrated in Figure 11.
The first uses three-dimensional perspective to portray the third
dimension. The second uses shading, where darker hues are used to
represent either higher density or more central values. This is
particularly of use for displaying uncertainty, as the less well
defined (more uncertain) parts of a curve appear fuzzy.

87
EHC 239: Principles for Modelling Dose–Response









 
'LVWULEXWLRQ9DOXH  



 
 


 



 &XPXODWLYH8QFHUWDLQW\
 
 
  
 

&XPXODWLYH9DULDELOLW\








&XPXODWLYHIUHTXHQF\ 



 'DWD










        

'LVWULEXWLRQYDOXH

Fig. 11. Plotting results of three-dimensional simulations.

88
 Communicating the Results of Dose–Response Modelling

7.7 Risk assessment context and questions

The output of the DRM should be directed towards addressing

specific questions about the likelihood of adverse health effects in
response to exposure to chemicals. This would build on
conventional risk assessment procedures that have been accepted
internationally as the indicator for determining acceptable levels of
exposure. These rely on the identification of a NOAEL/no-
observed-effect level (NOEL) for a critical end-point in the effect
data and incorporation of uncertainty factors to allow for
interspecies and interindividual variation.

DRM offers the potential to provide additional information for

the risk manager, specifically a more scientifically robust method
for determining the health-based guidance values (e.g. ADI) using
the BMD and better information on the likelihood of effects at low
doses that are below the levels observed in biological systems. The
mathematical models will also provide estimates of the statistical
uncertainty surrounding estimates of likely effect.

Whether traditional safety-based assessments or DRM

assessments are carried out, the risk manager will still require
information on the toxicology of the adverse health effect and the
robustness of the determination of the health-based guidance value
to help inform the management options. This may include the
following:

x a discussion of the strength and weight of evidence;

x uncertainties and gaps in the data;
x information on the nature and severity of the (critical) effect;
x limitations in the interpretation;
x assumptions made in the analysis; and
x qualitative assessment of the potential effects of exceeding the
health-based guidance value.

7.8 Synopsis of approach to modelling

DRM involves six basic steps: data selection, model selection,

statistical linkage, parameter estimation, implementation, and
evaluation (see chapter 4, Table 1). In undertaking a DRM exercise,

89
 EHC 239: Principles for Modelling Dose–Response

two factors that will impact the types of outputs and that may be of
importance to the risk manager are briefly described below.

7.8.1 Data sets

Traditional safety assessments focus primarily on a single

critical end-point, whereas DRM gives the potential for separating
out multiple end-points. Modelling outcomes may be based on data
from single or multiple experiments. In the latter situation, meta-
analysis may integrate the results of several independent studies that
are considered to be “combinable”.

The risk manager could see four types of data from the
modelling evaluations: namely, quantal, count, continuous, and
ordinal categorical data. The risk manager will need to understand
what data sets were modelled and, if quantitative information from
more than one data set is presented, will need guidance on the
rationale for forwarding the additional data set information and for
synthesizing this additional information. This guidance may include
information about the consistency (or inconsistency) of the
quantitative response across the end-points. Such information could
be used by the risk manager to strengthen (or weaken) his or her
confidence in the quantitative evaluation of the potential for health
impacts.

If DRM information is available from human epidemiological

evaluation, then an understanding of both the strengths as well as
the possible limitations (often in the quantitative exposure
information) of the data set may also temper or strengthen the
qualitative or quantitative assessment from the animal studies.

7.8.2 Uncertainty

DRM should capture the relative uncertainties in the estimates

of risk. This information will allow the generation of confidence
limits on health-based guidance values. However, such confidence
limits will still capture only one part of the uncertainty inherent in
these estimates. The risk manager will need to know what
uncertainty is accounted for in the information provided, and the
risk assessment information will need to clearly indicate what
uncertainty is not accounted for in a quantitative assessment.

90
 Communicating the Results of Dose–Response Modelling

One approach that has been used to capture variability in

population response is calculation of population percentiles.
Availability of dose–response functions when linked with
population-based exposure assessments has allowed risk managers
to calculate percentiles of populations above target exposure or
intake levels. Likewise, dose–response functions have also been
utilized to calculate percentiles of the population above target risk
levels.

One of the advantages of DRM is that the confidence limit

around the BMD can be calculated. From the conservative point of
view, the lower limit of the dose is most important. However, this is
not the same as to say that the confidence limit of the health-based
guidance values can be calculated, as the uncertainty factors will
obscure such estimates.

7.9 Explaining/interpreting the output of the dose–

response analysis

Advice to the risk manager should describe the uncertainties

inherent in such an approach to the use of dose–response data, such
as uncertainties in the slope estimate in animals, the relevance of
this slope to humans (such an approach is more appropriate if the
response is a continuous variable, rather than quantal), and the
appropriateness of the uncertainty factor applied to allow for
species differences and human variability.

7.9.1 Outputs in the observable biological range

The output of the analysis takes the form of a numerical

quantity—at present, commonly a TDI or ADI derived from a
NOAEL, which is a single point in the dose–response relationship.
The dose–response analysis uses more of the available information
by fitting a mathematical model to all the data in the observable
biological range and then determining the dose associated with a
specified response level. A statistical lower bound (e.g. the 95%
lower bound on the dose) is often used to account for statistical
uncertainties (a BMDL) and for the level of health protection
required by the risk manager. As with the NOAEL, the BMDL can
be used as the starting point for deriving a health-based guidance
value and/or MOE. However, unlike the NOAEL, the BMD

91
 EHC 239: Principles for Modelling Dose–Response

approach uses the whole range of experimental dose–response data,

and therefore it is not limited by the doses selected by the
investigators.

7.9.1.1 Health-based guidance values

On the basis of current practice, it appears that the BMD

approach leads to doses that are usually quite similar to NOAELs
for the studies in question (see section 7.3). In the same way as for
the derivation of the ADI/TDI, uncertainty factors, for example 100,
are applied to the BMDL to obtain the health-based guidance value.
However, the confidence intervals that are possible in the case of
the BMD-derived health-based guidance value provide the risk
manager with an increased understanding of the uncertainty
associated with the risk assessment. This allows a more informed
decision to be made when choosing among risk management
options.

7.9.1.2 Margin of exposure

An MOE is determined by comparing the point of departure

(the BMDL) with the actual or estimated human exposure. The
MOE is used when limited toxicological or human data exist but the
hazard identification and characterization data are insufficient to set
a health-based guidance value. Alternatively, the MOE approach is
used when it is inappropriate to derive a health-based guidance
value owing to the nature of the effect, such as for substances that
are genotoxic and carcinogenic.

The acceptability of an MOE depends on its magnitude and is

ultimately a risk management decision. To aid that decision, the risk
assessor should provide information on the nature of the toxicity
involved and nature and magnitude of the uncertainties, from both
the toxicological and exposure perspectives. Although the risk
assessor should not provide an assessment of the acceptability of the
MOE, guidance on its adequacy, taking into account the
severity/nature of the toxicity, uncertainties, and variability, should
be given—for example, in terms of high, medium, or low concern.
The use of all the data by the dose–response analysis enables the
uncertainties to be better defined. The MOE can also be used by the
risk manager for priority setting.

92
 Communicating the Results of Dose–Response Modelling

There is no internationally accepted value for an MOE for a

genotoxic and carcinogenic compound such that the exposure
would not be a significant health risk. However, several institutions
and countries have used the MOE approach, and their conclusions
provide examples of MOE values that have been considered
acceptable:

x The National Health and Medical Research Council in

Australia concluded that a guideline dose for carcinogens
present in soil could be calculated by application of uncertainty
factors up to 50 000 to the BMD (not BMDL). The factor
applied in any particular case would depend on the nature of
the effects (National Health and Medical Research Council,
1999).

x The reciprocal of the MOE, the exposure potency index (EPI),

has been used by Health Canada for genotoxic and
carcinogenic compounds in their Human Health Risk
Assessment for Priority Substances under the Canadian
Environmental Protection Act (Health Canada, 1994). MOE
values of <5000, 5000–500 000, and >500 000 indicate high,
medium, and low priority, respectively.

x The Committee on Carcinogenicity in the United Kingdom

considered derivation of the minimal risk level for a genotoxic
and carcinogen compound. One proposal was that an adequate
MOE for carcinogenicity might be 10 000 (Gaylor et al., 1999;
Gold et al., 2003). A particular carcinogenic impurity posed a
negligible carcinogenic risk if an uncertainty factor of 10 000
was applied to the estimated 5% BMD (BMD5) (Committee on
Carcinogenicity, 2003). The MOE for average intakes for
acrylamide in men in Norway has been estimated using the T25
value1 and the LED10 (the lower bound on the effective dose
for a 10% increase in risk) (approximately equivalent to
BMDL10) methods. These approaches result in MOE values of
1306 and 1225 for T25 and LED10, respectively.

1 The tumorigenic descriptor T is the chronic daily dose that will give 25% of the
25
animals tumours above background at a specific tissue site. The T25 is determined by
linear interpolation from the lowest dose giving a statistically significant increase in
tumours (Dybing et al., 1997).

93
 EHC 239: Principles for Modelling Dose–Response

x The 64th (WHO, 2006) and 67th (WHO, 2007a) meetings of

JECFA used MOE approaches for the evaluation of several
substances that were genotoxic and carcinogenic. The 64th
JECFA developed general considerations for the formulation of
advice on compounds that are both genotoxic and carcinogenic.
This meeting established MOEs for acrylamide, ethyl
carbamate, polybrominated diphenyl ethers, and polycyclic
aromatic hydrocarbons. The 67th JECFA established an MOE
for 1,3-dichloro-2-propanol.

x A joint European Food Safety Authority/WHO conference on

the risk assessment of substances that are both genotoxic and
carcinogenic (Barlow et al., 2006) compared the approaches
that are currently used. “This conference concluded that the
MOE approach was a useful and pragmatic option.…”

x O’Brien et al. (2006) presented a critical appraisal of the

approaches to the risk assessment of genotoxic carcinogens in
food and concluded that “Overall, MOE is the most appropriate
default approach because it combines information on potency
and exposure, without the generation of numerical risk
estimates of unknown reliability.” They presented case-studies
on the calculation of MOEs for acrylamide, aflatoxin B1,
benzo(a)pyrene, dimethylnitrosamine, ethyl carbamate, and 2-
amino-1-methyl-6-phenylimidazo(4,5b)pyridine.

7.9.2 Outputs outside the observable biological range

DRM evaluations can produce information in several formats,

including dose–response functions that allow, along with estimates
of exposure, the prediction of risks at specified exposure levels and
functions that allow the estimation of exposure levels resulting in
specified risks. In addition, DRM exercises can provide uncertainty
analyses. The availability of such outputs from DRM exercises can
provide both opportunities for additional assessment as well as
challenges in interpretation for the risk manager.

Three different methods have been used or proposed for

quantitative risk assessment by regulatory authorities in the United
States and Europe for non-threshold (genotoxic) carcinogens. In the
area of food safety, the United States Food and Drug
Administration has used a simple, direct method for low-dose

94
 Communicating the Results of Dose–Response Modelling

cancer risk assessment. A point on the dose–response curve is

chosen below which the data no longer appear to be reliable (e.g. 1–
10% tumour incidence), and a straight line is drawn from the upper
confidence limit on risk at that point to the origin (Gaylor et al.,
1997). The linearized multistage model was previously extensively
used by the United States Environmental Protection Agency
(USEPA, 1986). The LED10 method was later proposed by the
USEPA (1996), and the T25 (Dybing et al., 1997; Sanner et al.,
2001) method has been used in Europe (European Commission,
1999; SCCNFP, 2003). Lifetime cancer hazards may be estimated
by linear extrapolation using LED10 and T25 as starting points. The
results obtained with these extrapolation methods are in most cases
nearly indistinguishable (Sanner et al., 2001). A measure for an
assessment of concern may be arrived at by comparing the
calculated risks for some specific scenario of human exposure to
such substances, with some default policy-determined risk level.

7.9.2.1 Prediction of risks at specified exposure levels

One type of output from DRM is the prediction of risks at

specified exposure levels. This output can take the generic form of
predicting “X number of health-impacted individuals at exposure
Y”. Examples of such estimates have been used to predict the
number of excess lung cancer deaths due to smoking two packs of
cigarettes per day, the number of excess skin cancers from arsenic-
contaminated water, and the number of excess mortality cases due
to air pollution. In the optimal case, such estimates are supported by
parallel assessments that describe the uncertainty in such estimates,
by providing additional information on the range of estimates,
rather than a single value. The risk manager can then make such
statements as “Up to X individuals may be impacted by exposure
Y”. This same information can allow the risk manager to see how
low the estimates of the health impact may be; when confidence
limits are included in such estimates, many uncertain health impacts
can be shown to include the potential for no health impacts.
Assumptions inherent in such estimates that can impact
interpretation by the risk manager include choice of models, choice
of end-points, and limitations in initial data sets that were
extrapolated.

95
 EHC 239: Principles for Modelling Dose–Response

One use of such information has been to evaluate the effect of

different maximum limits for a chemical on risks. This type of
consideration was included when JECFA evaluated aflatoxin B1 and
the impact of different maximum limits on risk (WHO, 1999,
2007b). Similar assessments have also been performed for lead and
fumonisins B1 and B2 (Carrington et al., 1996; Humphreys et al.,
2001). For example, the health impacts of current particulate
standards (WHO, 2000, 2003) have been estimated. Availability of
such estimates can provide additional information for risk managers
to conduct cost–benefit analyses, risk–benefit assessments, and
evaluations of public health interventions.

7.9.2.2 Prediction of exposure levels producing specified risk levels

Another type of output from DRM is risk level estimates. In

these estimates, a specific level of risk is evaluated and the amount
of exposure that would be estimated to result in that risk is
determined. For example, a common level of risk related to
carcinogen exposures that has been evaluated in the United States
has been 10í6 over a lifetime. Estimates of exposure that would
result in that level of risk have been determined, and such estimates
have been made for approximately 100 environmental pollutants
(http://www.epa.gov/iris). For the risk manager, availability of such
estimates can allow for development of risk-based consistency in
proposed regulatory actions.

7.9.2.3 Uncertainty analyses

A third type of output from DRM is that linked with

uncertainty analysis. One example of such approaches is when the
DRM output is linked with distributions of population effects with
confidence intervals. The result from such analysis is a distribution
of potential population risks. For example, in Figure 12, the outputs
for three models and two data sets were used to generate a set of
3000 different model parameters (two data sets, three models, 500
bootstraps).

One approach that has been used to extrapolate dose–response

models beyond bioassay data has focused on the use of biomarker
data to extend the dose–response curve 1–2 orders of magnitude
closer to environmentally relevant exposures. Such approaches can
be facilitated when DRM data are available.

96
 Communicating the Results of Dose–Response Modelling

0DPPDU\7XPRXU,QFLGHQFH

(

(

(

(
)UHTXHQF\

(

(

(

(

(

(

(
( ( ( ( ( ( ( (

$FU\ODPLGH PJNJERG\ZHLJKWGD\

Fig. 12. Integrated uncertainty analysis for mammary tumours. The dark line
is the central (median) estimate, and the dotted lines are the 5% and 95%
confidence limits.

All these modelling approaches exhibit similar limitations and

difficulties. A benefit is that DRM allows for the transfer of more
quantitative toxicological data into risk manager assessment
methods such as cost–benefit and risk–benefit analyses. The
limitation is the question of whether the model outputs are accurate
and representative of public health impacts.

7.10 Issues for risk managers

7.10.1 Risk assessment issues

7.10.1.1 Population versus individual effects

The potential health effect at the population level can be

informed by DRM. However, as the behaviour, environment, or
biological characteristics may vary among individuals, a dose–
response model may need to describe or model these characteristics
to produce a prediction of adverse health effects in the population.

97
 EHC 239: Principles for Modelling Dose–Response

The output of the dose–response model should identify the degree

of any subpopulation effects.

7.10.1.2 Risk characterization

The actual risk to the population of an adverse health effect

requires consideration of both the likelihood and severity of the
effect, as determined from the dose–response model when
combined with the exposure to the chemical in the population under
consideration. The exposure may be determined from consumption
surveys, measurement of environmental media, direct contact
information, or biomarkers (e.g. IPCS, 2000; Kroes et al., 2002).

Consideration of the DRM data together with exposure data

will help identify populations at risk. This information, together
with knowledge about the severity of the adverse health effects, will
inform the risk management options.

7.10.2 Risk management issues

7.10.2.1 Risk management options

A risk assessment can be used to establish that a risk is of a

sufficient magnitude that regulation or other type of intervention
may be warranted. DRM can then be used to evaluate the
consequences of possible interventions that aim to reduce the risk.
That is, a model may be used to estimate change in the likelihood of
the adverse health effect occurring following implementation of a
particular intervention. To date, alternative risk management
options have been evaluated using DRM in a limited number of
cases. For example, at the request of the Codex Committee on Food
Additives and Contaminants, the 49th JECFA analysed the
application of two hypothetical standards for aflatoxin
contamination in food in model populations (WHO, 1999).

A range of risk management interventions are available, with

the types of interventions varying from a ban on a particular product
(e.g. carcinogenic antibiotics, DDT), establishing regulatory limits
(e.g. aflatoxins), advice on consumption or use patterns (e.g.
consumption of predatory fish that accumulate high levels of
methylmercury), and control at source of production (e.g. emissions
of dioxins).

98
 Communicating the Results of Dose–Response Modelling

7.10.2.2 Cost–benefit and risk–benefit analyses

While health risk management decisions should be based on

risk assessments, a number of other factors will influence the final
decisions. In particular, it may also be necessary to undertake a
cost–benefit analysis (e.g. health costs to the community from
exposure to aflatoxins versus the cost of implementation of a
management strategy) and/or risk–benefit analysis (e.g. risk
associated with methylmercury in fish versus nutritional benefits of
fish consumption) and to assess the feasibility of the intervention,
availability of alternatives, and loss of products of economic value.
These factors are beyond the scope of the assessment of the risks
and will need to reflect wider societal factors.

7.10.2.3 Acceptable level of risk

Different institutions and countries may make different risk

management decisions based on different perceptions of the risk
that is deemed to be acceptable to society. The ADI, which usually
incorporates a composite uncertainty factor of 100 when based on
animal studies, has been accepted by international institutions and
countries as a health-based guidance value. Although DRM can
give a prediction of the risk at various exposures, there is no
international agreement on how to interpret this new information,
the appropriate criteria for making or reversing risk decisions, or the
acceptable level of risk determined using this technique.

A predicted risk level, such as 10í6, determined from dose–

response analysis has been used by some countries and institutions
as being not appreciable or negligible (virtually safe dose).
Variations around the calculated risk by a factor of about 10 trigger
further consideration of the qualitative aspects of the risk
assessment, such as variability and uncertainty (Sanner et al., 2001;
SCCNFP, 2003). In the case of compounds in drinking-water
considered to be genotoxic carcinogens, WHO has assigned
guideline values associated with an estimated upper-bound excess
lifetime cancer risk of 10í5 determined by a mathematical model
(WHO, 2004b). The United States Occupational Safety and Health
Administration has considered a lifetime cancer risk for workers
higher than 10í3 to represent an unacceptably high risk, and its goal
is to reduce this risk to less than 10í5 (OSHA, 1983, 1984).

99
EHC 239: Principles for Modelling Dose–Response

Proposals for the application of lifetime risk estimates in

establishing tolerable risk levels have also been published in Europe
(Bos et al., 2004).

100
 8. CONCLUSIONS AND RECOMMENDATIONS

8.1 Conclusions

x Full DRM can be considered a more sophisticated or robust alternative to

the NOAEL approach in all cases where suitable dose–response data are
available (e.g. several dose groups with different response levels).

x For quantal dose–response data, the interest is often in low response

(incidence) levels. This may call for low-dose extrapolation by several
orders of magnitude (e.g. for tumour incidences). However, equally
plausible dose–response risk models may result in highly divergent low
estimates. A currently applied approach is to estimate a BMD10 and
linearly extrapolate from that point downwards, as a conservative
approach. Another option, currently under development, is to apply a
Bayesian approach that considers the various models all together.

x For continuous dose–response data, two approaches of DRM exist. One is

to transform the continuous data into quantal data. The other is to consider
continuous dose–response data as information on the severity of the effect
and therefore as a function of dose. In the latter approach, measurable
changes of effect are often close to response levels considered as adverse
(e.g. 10% inhibition of cholinesterase), and the low-dose extrapolation
problem is minor or non-existent.

x For the purpose of deriving an ADI, TDI, or RfD, DRM may be used for
deriving a BMD, to be used as a point of departure in the same way as the
NOAEL is used (i.e. the same uncertainty factors would be applied to the
BMD as to the NOAEL).

x DRM may also be used for estimating risks at a given (human) exposure
level. For risks in terms of incidences (quantal data), this may involve
low-dose extrapolation.

x DRM exercises can provide information on uncertainties associated with

the data and identify factors contributing to uncertainties in risk estimates.

x Application of DRM for all end-points can be cost prohibitive, so it is

efficient to pre-select the apparently more sensitive end-points. In some

101
 EHC 239: Principles for Modelling Dose–Response

cases, however, it is not easy to identify the most sensitive end-points by

visual inspection, so all of the end-points may need to be modelled.

x The BMD and the BMDL should always be reported, so that the quality of
the data and the model fit are clear and potencies can be compared on the
basis of the BMD.

x The output of the different models used in DRM should be presented.

8.2 Recommendations

x Toxicity testing protocols (e.g. Organisation for Economic Co-operation

and Development guidelines) should be reviewed for optimization for
BMD and other DRM approaches, including optimal designs for the
number of animals and number of doses for different dose–response
curves. Additional research is needed for the development of optimal
study designs. Guidance should be developed for combining existing
studies with a view to DRM.

x Better guidance needs to be developed for combined analysis of different

data sets for more precisely estimating BMDs.

x Better understanding of when and how to use the BMR needs to be

developed.

x Better understanding of the shape of the dose–response curve at low doses

needs to be developed. Additional research is needed to determine the
biological basis for extrapolation (e.g. by using biomarkers, tumour
precursors, genetically modified animals, and toxicokinetics for target
dose estimation).

x Improved guidance needs to be developed for risk communication based

on the results of DRM and probabilistic assessment techniques. This
should include communication of the types of uncertainty and the relation
to statistical variability, imprecision, and the use of confidence intervals.

x The use of DRM should be reviewed and additional general principles for
its use developed when more experience becomes available.

102
REFERENCES

Akaike H (1973) Information theory and an extension of the maximum likelihood principle.
In: Petrov BN & Csaki F ed. Proceedings of the second international symposium on
information theory. Budapest, Akademiai Kiado, pp 267–281.

Allen BC, Kavlock RJ, Kimmel CA, & Faustman EM (1994) Dose–response assessment
for developmental toxicity: II. Comparison of generic benchmark dose estimates with
NOAELs. Fundam Appl Toxicol, 23: 487–495.

Allen BC, Strong PL, Price CJ, Hubbard SA, & Daston GP (1996) Benchmark dose
analysis of developmental toxicity in rats exposed to boric acid. Fundam Appl Toxicol, 32:
194–204.

Barlow S, Renwick AG, Kleiner J, Bridges JW, Busk L, Dybing E, Edler L, Eisenbrand G,
Fink-Gremmels J, Knaap A, Kroes R, Liem D, Müller DJG, Page S, Schlatter J, Tritscher
A, Rolland V, Tueting W, & Würtzen, G (2006) Risk assessment of substances that are
both genotoxic and carcinogenic. Food Chem Toxicol, 44(10): 1636–1650.

Barnes DG & Dourson ML (1988) Reference dose (RfD): Description and use in health
risk assessments. Regul Toxicol Pharmacol, 8: 471–486.

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. Regul Toxicol Pharmacol, 21: 296–306.

Bos PMJ, Baars BJ, & van Raaij MTM (2004) Risk assessment of peak exposure to
genotoxic carcinogens: A pragmatic approach. Toxicol Lett, 151: 43–50.

Bozdogan H (1987) Model-selection and Akaike’s information criterion (AIC): The general
theory and its analytical extensions. Psychometrika, 52: 345–370.

Burnham KP & Anderson DR (2002) Model selection and multimodel inference, 2nd ed.
New York, Springer.

Carrington CD & Bolger PM (2000) A pooled analysis of the Iraqi and Seychelles
methylmercury studies. Hum Ecol Risk Assess, 6: 323–340.

Carrington CD, Bolger PM, & Scheuplein RJ (1996) Risk analysis of dietary lead
exposure. Food Addit Contam, 13: 61–76.

Cogliano VJ, Baan RA, Straif K, Grosse Y, Secretan MB, El Ghissassi F, & Kleihues P
(2004) The science and practice of carcinogen identification and evaluation. Environ
Health Perspect, 112(13): 1269–1274 (http://ehp.niehs.nih.gov/docs/2004/6950/
abstract.html).

103
EHC 239: Principles for Modelling Dose–Response

Committee on Carcinogenicity (2003) Minutes of United Kingdom Committee on

Carcinogenicity of Chemicals in Food, Consumer Products and the Environment meeting,
26 June 2003 (http://www.advisorybodies.doh.gov.uk/coc/meetings/coc032.htm).

Crump KS (1984) A new method for determining allowable daily intakes. Fundam Appl
Toxicol, 4: 854–871.

Crump KS (1995) Calculation of benchmark doses from continuous data. Risk Anal,
15(1): 79–89.

Crump K (2002) Critical issues in benchmark calculations from continuous data. Crit Rev
Toxicol, 32(3): 133–153.

Davidian M & Giltinan DM (1995) Nonlinear models for repeated measurement data.
London, Chapman and Hall.

Dourson ML & Stara JF (1983) Regulatory history and experimental support of uncertainty
(safety) factors. Regul Toxicol Pharmacol, 2: 224–238.

Dourson ML, Hertzberg RC, Hartung R, & Blackburn K (1985) Novel approaches for the
estimation of acceptable daily intake. Toxicol Ind Health, 1: 23–41.

Dybing E, Sanner T, Roelfzema H, Kroese D, & Tennant RW (1997) A simplified

carcinogenic potency index. Description of the system and study of correlations between
carcinogenic potency and species/site specificity and mutagenicity. Pharmacol Toxicol,
80: 272–279.

Edler LK, Dourson M, Kleiner J, Mileson B, Nordmann H, Renwick A, Slob W, Walton K, &
Würtzen G (2002) Mathematical modelling and quantitative methods. Food Chem Toxicol,
40: 283–326.

Efron B (1987) Better bootstrap confidence intervals. J Am Stat Soc, 82: 171–185.

Efron B & Tibshirani R (1993) An introduction to the bootstrap. London, Chapman and
Hall.

European Commission (1999) European Commission guidelines for setting specific

concentration limits for carcinogens in Annex 1 of Directive 67/548/EEC. Inclusion of
potency considerations. Brussels, European Commission, Working Group on the
Classification and Labelling of Dangerous Substances.

European Commission (2000) White paper on food safety. Commission of the European
Communities, 12 January 2000, Comm (1999) 719 final (http://ec.europa.eu/dgs/
health_consumer/library/pub/pub06_en.pdf).

Evans JS, Graham JD, Gray GM, & Sielken RL (1994) A distributional approach to
characterizing low-dose cancer risk. Risk Anal, 14: 25–34.

Evans M, Hastings N, & Peacock B (1993) Statistical distributions, 2nd ed. New York,
John Wiley & Sons.

104
 References

Fitzgerald DJ, Robinson NI, & Pester BA (2004) Application of benzo(a)pyrene and coal
tar tumor dose–response data to a modified benchmark dose method of guideline
development. Environ Health Perspect, 112: 1341–1346.

Fowles JR, Alexeeff GV, & Dodge D (1999) The use of benchmark dose methodology
with acute inhalation lethality data. Regul Toxicol Pharmacol, 29: 262–278.

Garvin T (2001) Analytical paradigms: The epistemological distances between scientists,

policy makers, and the public. Risk Anal, 21(3): 443–455.

Gaylor DW (1989) Quantitative risk analysis for quantal reproductive and developmental
effects. Environ Health Perspect, 79: 243–246.

Gaylor D & Aylward L (2004) An evaluation of benchmark dose methodology for non-
cancer continuous-data health effects in animals due to exposures to dioxin (TCDD).
Regul Toxicol Pharmacol, 40: 9–17.

Gaylor D & Slikker W (1994) Modelling for risk assessment of neurotoxic effects. Risk
Anal, 14(3): 333–338.

Gaylor D & Slikker W (2004) Role of the standard deviation in the estimation of
benchmark doses with continuous data. Risk Anal, 6: 1683–1687.

Gaylor DW, Axelrad JA, Brown RP, Cavagnaro JA, Cyr WH, Hulebak KI, Lorentzen RJ,
Miller MA, Mulligan LT, & Schwetz BA (1997) Health risk assessment practices in the
U.S. Food and Drug Administration. Regul Toxicol Pharmacol, 26: 307–321.

Gaylor DW, Kodell RL, Chen JJ, & Krewski D (1999) A unified approach to risk
assessment for cancer and noncancer endpoints based on benchmark doses and
uncertainty/safety factors. Regul Toxicol Pharmacol, 29: 151–157.

Gelman A, Carlin JB, Stern HS, & Rubin DB (2004) Bayesian data analysis, 2nd ed. Boca
Raton, Florida, Chapman and Hall/CRC.

Ghani AC, Ferguson NM, Donnelly CA, & Anderson RM (2000) Predicted vCJD mortality
in Great Britain. Nature, 406: 583–584.

Gold LS, Gaylor DW, & Slone TH (2003) Comparison of cancer risk estimates based on a
variety of risk assessment methodologies. Regul Toxicol Pharmacol, 237: 45–53.

Hacking I (1976) The great decision. In: The emergence of probability. A philosophical
study of early ideas about probability, induction and statistical inference. Cambridge,
Cambridge University Press, pp 63–72.

Hasselblad V & Jarabek AM (1995) Dose–response analysis of toxic chemicals. In: Berry
DA & Stangl DK ed. Bayesian biostatistics. New York, Marcel Dekker.

Health Canada (1994) Human health risk assessment for priority substances. Ottawa,
Ontario, Minister of Supply and Services Canada (http://www.hc-sc.gc.ca/ewh-
semt/pubs/contaminants/existsub/approach/index_e.html).

Hertzberg RC (1991) Quantitative extrapolation of toxicological finds. In: Krewski D &

Franklin C ed. Statistics in toxicology. New York, Gordon & Breach Science.

105
EHC 239: Principles for Modelling Dose–Response

Hertzberg RC & Miller M (1985) A statistical model for species extrapolation using
categorical response data. Toxicol Ind Health, 1(4): 43–57.

Hertzberg RC & Wymer L (1991) Modelling the severity of toxic effects. In: Proceedings of
the 84th annual meeting and exhibition of the Air and Waste Management Association,
16–21 June 1991, Vancouver, British Columbia. Pittsburgh, Pennsylvania, Air and Waste
Management Association.

Humphreys SH, Carrington CD, & Bolger PM (2001) A quantitative risk assessment for
fumonisins B1 and B2 in corn. Food Addit Contam, 18: 211–220.

IPCS (1987) Principles for the safety assessment of food additives and contaminants in
food. Geneva, World Health Organization, International Programme on Chemical Safety
(Environmental Health Criteria 70; http://www.inchem.org/documents/ehc/ehc/ehc70.htm).

IPCS (1994) Assessing human health risks of chemicals: Derivation of guidance values
for health-based exposure limits. Geneva, World Health Organization, International
Programme on Chemical Safety (Environmental Health Criteria 170;
http://www.inchem.org/documents/ehc/ehc/ehc170.htm).

IPCS (1999) Principles for the assessment of risks to human health from exposure to
chemicals. Geneva, World Health Organization, International Programme on Chemical
Safety (Environmental Health Criteria 210; http://www.inchem.org/documents/ehc/ehc/
ehc210.htm).

IPCS (2000) Human exposure assessment. Geneva, World Health Organization,

International Programme on Chemical Safety (Environmental Health Criteria 214;
http://www.inchem.org/documents/ehc/ehc/ehc214.htm).

IPCS (2004) IPCS risk assessment terminology. Geneva, World Health Organization,
International Programme on Chemical Safety (Harmonization Project Document No. 1;
http://www.who.int/ipcs/methods/harmonization/areas/ipcsterminologyparts1and2.pdf).

IPCS (2005) Chemical-specific adjustment factors for interspecies differences and human
variability: Guidance document for use of data in dose/concentration–response
assessment. Geneva, World Health Organization, International Programme on Chemical
Safety (Harmonization Project Document No. 2; http://whqlibdoc.who.int/publications/
2005/9241546786_eng.pdf).

Kim AH, Kohn MC, Portier CJ, & Walker NJ (2002) Impact of physiologically based
pharmacokinetic modeling on benchmark dose calculations for TCDD-induced
biochemical responses. Regul Toxicol Pharmacol, 36: 287–296.

Kimmel CA & Gaylor DW (1988) Issues in qualitative and quantitative risk analysis for
development toxicology. Risk Anal, 8: 15–20.

Kodell RL, Chen JJ, & Gaylor DW (1995) Neurotoxicity modeling for risk assessment.
Regul Toxicol Pharmacol, 22(1): 24–29.

Kroes R, Müller D, Lambe J, Löwik MRH, van Klavern J, Kleiner J, Massey R, Mayer S,
Urieta I, Verger P, & Visconti A (2002) Assessment of intake from the diet. Food Chem
Toxicol, 40: 327–385.

106
 References

Lehman AJ & Fitzhugh OG (1954) Ten-fold safety factor studies. Assoc Food Drug Off US
Q Bull, XVIII(1): 33–35.

Moerbeek M, Piersma AH, & Slob W (2004) A comparison of three methods for
calculating confidence intervals for the benchmark dose. Risk Anal, 24(1): 31–40.

Murrell JA, Portier CJ, & Morris RW (1998) Characterizing dose–response. I. Critical
assessment of the benchmark dose concept. Risk Anal, 18: 13–26.

National Health and Medical Research Council (1999) Toxicity assessment for
carcinogenic soil contaminants. Canberra, Commonwealth of Australia.

NRC (1983) Risk assessment in the federal government: Managing the process.
Washington, DC, National Research Council, National Academy of Science Press.

NRC (1994) Science and judgement in risk assessment. Washington, DC, National
Research Council, National Academy of Science Press.

NRC (1996) Understanding risk: Informing decisions in a democratic society. Washington,

DC, National Research Council, National Academy of Science Press.

NTP (2002) Report on carcinogens, 10th ed. Research Triangle Park, North Carolina,
United States Department of Health and Human Services, National Institutes of Health,
National Toxicology Program (http://ntp-server.niehs.nih.gov/
index.cfm?objectid=72016262-BDB7-CEBA-FA60E922B18C2540).

O’Brien J, Renwick AG, Constable A, Dybing E, Müller DJG, Schlatter J, Slob W, Tueting
W, van Benthem J, Williams GM, & Wolfreys A (2006) Approaches to the risk assessment
of genotoxic carcinogens in food: A critical appraisal. Food Chem Toxicol, 44(10): 1613–
1635.

OSHA (1983) Occupational exposure to inorganic arsenic: Supplemental statement of

reasons for the final rule, 48 F.R. 1,864. Washington, DC, United States Department of
Labor, Occupational Safety and Health Administration.

OSHA (1984) Occupational exposure to ethylene oxide: Final standard, 49 F.R. 46,936.
Washington, DC, United States Department of Labor, Occupational Safety and Health
Administration.

Piersma AH, Verhoef A, te Biesebeek JD, Pieters MN, & Slob W (2000) Developmental
toxicity of butyl benzyl phthalate in the rat using a multiple dose study design. Reprod
Toxicol, 14: 417–425.

Pieters MN, Bakker M, & Slob W (2004) Reduced intake of deoxynivalenol in the
Netherlands: A risk assessment update. Toxicol Lett, 153: 145–153.

Portier CJ (1994) Biostatistical issues in the design and analysis of animal carcinogenicity
experiments. Environ Health Perspect, 102(Suppl 1): 5–8.

Portier C & Kohn M (1996) A biologically-based model for the carcinogenic effects of
2378-TCDD in female Sprague-Dawley rats. Organohalogen Compds, 29: 222–227.

107
EHC 239: Principles for Modelling Dose–Response

Presidential Commission (1997) Risk assessment and risk management in regulatory

decision-making, Vol 2. Washington, DC, The Presidential/Congressional Commission on
Risk Assessment and Risk Management.

Raftery A, Madigan D, & Hoeting J (1997) Bayesian model averaging for linear regression
models. J Am Stat Assoc, 92: 179–191.

Renwick AG & Lazarus NR (1998) Human variability and noncancer risk assessment—an
analysis of the default uncertainty factor. Regul Toxicol Pharmacol, 27: 3–20.

Renwick AG, Barlow SM, Hertz-Picciotto I, Boobis AR, Dybing E, Edler L, Eisenbrand G,
Grieg JB, Kleiner J, Lambe J, Müller DJG, Smith MR, Tritscher A, Tuijtelaars S, van den
Brandt PA, Walker R, & Kroes R (2003) Risk characterization of chemicals in food and
diet. Food Chem Toxicol, 41: 1211–1271.

Rescher N (1969) Many-valued logic. New York, McGraw-Hill.

Sand S, Falk Filipsson A, & Victorin K (2002) Evaluation of the benchmark dose method
for dichotomous data: Model dependence and model selection. Regul Toxicol Pharmacol,
36: 184–197.

Sand S, von Rosen D, & Falk Filipsson A (2003) Benchmark calculations in risk
assessment using continuous dose response information: The influence of variance and
the determination of a cut-off value. Risk Anal, 23: 1059–1068.

Sanner T, Dybing E, Willems MI, & Kroese ED (2001) A simple method for quantitative
risk assessment of non-threshold carcinogens based on the dose descriptor T25.
Pharmacol Toxicol, 88: 331–341.

SCCNFP (2003) The SCCNFP’s notes of guidance for the testing of cosmetic ingredients
and their safety evaluation, 5th rev. Adopted by the SCCNFP during the 25th plenary
meeting of 20 October 2003. Brussels, The Scientific Committee on Cosmetic Products
and Non-Food Products Intended for Consumers.

Sherman CD & Portier CJ (1998) Multistage carcinogenesis models. In: Encyclopedia of

biostatistics. Sussex, Wiley & Sons, pp 2808–2814.

Slob W & Pieters MN (1998) A probabilistic approach for deriving acceptable human
intake limits and human health risks from toxicological studies: General framework. Risk
Anal, 18: 787–798.

Slob W, Moerbeck M, Rauniomaa E, & Piersma AH (2005) A statistical evaluation of

toxicity designs for the estimation of the benchmark dose in continuous endpoints. Toxicol
Sci, 84: 167–185.

Thompson KM (2002) Variability and uncertainty meet risk management and risk
communication. Risk Anal, 22(3): 647–654.

Timchalk C, Nolan RJ, Mendrala AL, Dittenber DA, Brzak KA, & Mattsson JL (2002) A
physiologically based pharmacokinetic and pharmacodynamic (PBPK/PD) model for the
organophosphate pesticide chlorpyrifos in rats and humans. Toxicol Sci, 66: 34–53.

108
 References

USEPA (1986) Guidelines for carcinogen risk assessment. United States Environmental
Protection Agency. Fed Regist, 51: 33992–34003.

USEPA (1996) Proposed guidelines for carcinogen risk assessment. United States
Environmental Protection Agency. Fed Regist, 61: 17960–180011.

USEPA (2005) Guidelines for carcinogen risk assessment 2005. Washington, DC, United
States Environmental Protection Agency, Risk Assessment Forum (EPA/630/P-03/001F;
http://cfpub.epa.gov/ncea/cfm/recordisplay.cfm?deid=116283).

WHO (1997) Food consumption and exposure assessment of chemicals. Report of a

FAO/WHO consultation, 10–14 February 1997 (WHO/FSF/97.5; available from
Department of Food Safety, World Health Organization, 20 Avenue Appia, CH-1211
Geneva 27, Switzerland).

WHO (1999) Evaluation of certain food additives and contaminants. Forty-ninth report of
the Joint FAO/WHO Expert Committee on Food Additives. Geneva, World Health
Organization.

WHO (2000) Air quality guidelines for Europe, 2nd ed. Copenhagen, World Health
Organization Regional Office for Europe.

WHO (2002) Evaluation of certain food additives and contaminants. Fifty-seventh report
of the Joint FAO/WHO Expert Committee on Food Additives. Geneva, World Health
Organization (Technical Report Series 909).

WHO (2003) Health aspects of air pollution with particulate matter, ozone and nitrogen
dioxide. Report of a Working Group. Copenhagen, World Health Organization Regional
Office for Europe (http://www.euro.who.int/document/e79097.pdf).

WHO (2004a) Risk assessment of Listeria monocytogenes in ready-to-eat foods. Geneva,

World Health Organization (Microbiological Risk Assessment Series, No. 5, Technical
Report; http://www.who.int/foodsafety/publications/micro/mra_listeria/en/index.html).

WHO (2004b) Guidelines for drinking-water quality, 3rd ed. Vol. 1. Recommendations.
Geneva, World Health Organization (http://www.who.int/water_sanitation_health/
dwq/gdwq3rev/en/).

WHO (2006) Evaluation of certain food contaminants. Sixty-fourth report of the Joint
FAO/WHO Expert Committee on Food Additives. Geneva, World Health Organization
(WHO Technical Report Series No. 930; http://whqlibdoc.who.int/trs/
WHO_TRS_930_eng.pdf).

WHO (2007a) Evaluation of certain food additives and contaminants. Sixty-seventh report
of the Joint FAO/WHO Expert Committee on Food Additives. Geneva, World Health
Organization (WHO Technical Report Series No. 940; http://whqlibdoc.who.int/
trs/WHO_TRS_940_eng.pdf).

WHO (2007b) Evaluation of certain food additives and contaminants. Sixty-eighth report
of the Joint FAO/WHO Expert Committee on Food Additives. Geneva, World Health
Organization (WHO Technical Report Series No. 947; http://whqlibdoc.who.int/
trs/WHO_TRS_947_eng.pdf).

109
EHC 239: Principles for Modelling Dose–Response

Williams GM & Iatropoulos MJ (2002) Alteration of liver cell function and proliferation:
differentiation between adaptation and toxicity. Toxicol Pathol, 30(1): 41–53.

Woutersen RA, Jonker D, Stevenson H, te Biesebeek JD, & Slob W (2001) The
benchmark approach applied to a 28-day toxicity study with Rhodorsil Silane in rats: The
impact of increasing the number of dose groups. Food Chem Toxicol, 39: 697–707.

110
 ANNEX 1: TERMINOLOGY

Acceptable daily intake (ADI)/tolerable daily intake (TDI)/reference

dose (RfD): Estimated maximum amount of an agent, expressed on
a body mass basis, to which an individual in a (sub)population may
be exposed daily over the individual’s lifetime without appreciable
health risk.

Acceptable risk: A risk management term. The acceptability of risk

depends on scientific data, on social, economic, and political
factors, and on the perceived benefits arising from exposure to an
agent.

Additional risk (extra risk): The additional proportion of total

animals that respond in the presence of the dose, or the probability
of response at dose d, P(d), minus the probability of response in the
absence of exposure, P(0).

Adverse effect: Change in the morphology, physiology, growth,

development, reproduction, or lifespan of an organism, system, or
(sub)population that results in an impairment of functional capacity,
an impairment of the capacity to compensate for additional stress, or
an increase in susceptibility to other influences.

Akaike information criterion: A statistical procedure that provides a

measure of the goodness of fit of a dose–response model to a set of
data.

Assessment factor: Numerical adjustment to extrapolate from

experimentally determined (dose–response) relationships to
estimate the exposure to an agent below which an adverse effect is
not likely to occur (see Safety factor and Uncertainty factor).

Benchmark concentration (BMC): The concentration of a substance

that is associated with a specified low incidence of risk of a health
effect, or the concentration associated with a specified measure or
change of a biological effect.

111
EHC 239: Principles for Modelling Dose–Response

Benchmark dose (BMD): A dose of a substance associated with a

specified low incidence of risk, generally in the range of 1–10%, of
a health effect; or the dose associated with a specified measure or
change of a biological effect.

Benchmark dose lower confidence limit (BMDL): A lower one-sided

confidence limit on the BMD.

Benchmark response (BMR): The response, generally expressed as

in excess of background, at which a benchmark dose or
concentration is desired.

Bernoulli distribution: A theoretical distribution of the number of

successes in a finite set of independent trials with a constant
probability of success. It is a discrete distribution having two
possible outcomes labelled by n = 0 and n = 1, in which n = 1
(“success”) occurs with probability p and n = 0 (“failure”) occurs
with probability q { 1 í p, where 0 < p < 1.

Binomial distribution: The statistical distribution of the probabilities

of observing 0, 1, 2, … , n events in a sample of n independent trials
each with the same individual probability that the event occurs.

Bootstrap: A statistical technique based on multiple resampling with

replacement of the sample values or resampling of estimated
distributions of the sample values that is used to calculate
confidence limits or perform statistical tests for complex situations
or where the distribution of an estimate or test statistic cannot be
assumed.

Cancer potency (cancer slope factor): A number that estimates the

cancer risk (incidence) for a lifetime exposure to a substance per
unit of dose, which is generally expressed as mg/kg body weight per
day.

Categorical data: Results obtained where observations or

measurements on individuals or samples are stratified according to
degree or severity of an effect (e.g. none, mild, moderate, or
severe).

112
 Annex 1: Terminology

Categorical default factor: A factor based on common

characteristics of a group of compounds (e.g. physical/chemical
properties or pathways of metabolism).

Chemical-specific adjustment factor (CSAF): A factor based on

quantitative chemical-specific toxicokinetic or toxicodynamic data,
which replaces some or all of the default uncertainty factor.

Chi-square test: A statistical test used to examine the deviation of

an observed number of events from an expected number of events.

Clustered data: Measurements collected on some grouping of

individuals (e.g. litters in reproductive and developmental studies).

Confidence interval (one-sided): An interval below the estimated

upper confidence limit, or an interval above the estimated lower
confidence limit, that is expected to include the true value of an
estimated parameter with a specified confidence (percentage of the
time).

Confidence interval (two-sided): An estimated interval from the

lower to upper confidence limit of an estimate of a parameter. This
interval is expected to include the true value of the parameter with a
specified confidence percentage (e.g. 95% of such intervals are
expected to include the true values of the estimated parameters).

Confidence limit: An estimated value below (or above) which the

true value of an estimated parameter is expected to lie for a
specified percentage of such estimated limits.

Constrained dose–response model: Estimates of one or more

parameters of the model restricted to a specified range (e.g. equal to
or greater than zero).

Continuous data: Effects measured on a continuum, e.g. organ

weight or enzyme concentration, as opposed to quantal or
categorical data, where effects are classified by assignment to a
class.

113
EHC 239: Principles for Modelling Dose–Response

Convergence: A parameter approach that estimates a single value

with increasing sample size or increasing number of computer
iterations.

Covariate: An independent variable other than dose that may

influence the outcome of an effect (e.g. age, body weight, or
polymorphism).

Critical effect: The adverse effect, or its known precursor, that is

relevant to human risk assessment and that occurs in the
dose/concentration scale in the most sensitive animal species.

Degrees of freedom: For dose–response model fitting, the number

of data points minus the number of model parameters estimated
from the data.

Default value: Pragmatic, fixed, or standard value used in the

absence of relevant data.

Dichotomous data: Quantal data where an effect for an individual

may be classified by one of two possibilities (e.g. dead or alive),
with or without a specific type of tumour.

Dispersion: Variation (differences) from a central (mean or median)

value.

Dose: Total amount of an agent administered to or taken up or

absorbed by an organism, system, or (sub)population.

Dose–response: Relationship between the amount of an agent

administered to, taken up by, or absorbed by an organism, system,
or (sub)population and the change developed in that organism,
system, or (sub)population in reaction to the agent.

Dose–response assessment: Analysis of the relationship between

the total amount of an agent administered to, taken up by, or
absorbed by an organism, system, or (sub)population and the
changes developed in that organism, system, or (sub)population in
reaction to that agent, and inferences derived from such an analysis
with respect to the entire population.

114
 Annex 1: Terminology

Dose–response model: A mathematical relationship (function) that

relates (predicts) a measure of an effect to a dose.

Dose–response trend: Relationship between incidence or severity of

a biological effect and a function of dose. Simply the slope for a
linear dose–response.

EDx: Effective dose associated with a biological effect in x% of the

individuals. Dose may be the external exposure often expressed in
milligrams of the substance per day per kilogram body weight
raised to a power (generally 1, 3/4, or 2/3) or area under the curve
(AUC) in blood or target tissue where the substance remains in the
body over a period of time.

Estimate: An empirical value derived from data for a parameter.

Exposure: Concentration or amount of a particular agent that

reaches a target organism, system, or (sub)population in a specific
frequency for a defined duration.

Gamma distribution: A unimodal statistical distribution (relative

proportion of responders as a function of some measure) restricted
to effects greater than or equal to zero that can describe a wide
variety of shapes (e.g. flat, peaked, asymmetrical).

Gaussian (normal) distribution: A unimodal symmetrical (bell-

shaped) distribution where the most prevalent value is the mean
(average) and the spread is measured by the standard deviation.
Mathematically, the distribution varies from minus infinity with
zero probability to plus infinity with zero probability.

Goodness of fit: A statistic that measures the dispersion of data

about a dose–response curve in order to provide a test for rejection
of a model due to lack of an adequate fit (e.g. a p-value < 0.1).

Hazard identification: The identification of the type and nature of

adverse effects that an agent has an inherent capacity to cause in an
organism, system, or (sub)population.

115
EHC 239: Principles for Modelling Dose–Response

Hill equation: A dose–response curve, frequently used for enzyme

kinetics, that monotonically approaches an asymptote (maximum
value) as a function of dose raised to a power.

Hybrid model: For continuous data, establishes abnormal values

based on the extremes in controls (unexposed individuals or
animals) and estimates the risk of abnormal levels as a function of
dose.

Incidence: Proportion or probability of individuals or animals

exhibiting an effect that varies from zero to one, sometimes
expressed as a percentage from 0% to 100%.

Independence: The result in one animal or individual does not

influence the result in another animal or individual.

Intercept term: The estimated value at zero dose or the dose

corresponding to a zero effect.

Least squares: A statistical procedure that estimates the values of

dose–response parameters such that the sum of squares of
deviations of data points from their estimated values is minimized
(i.e. minimizes the estimated variance).

Likelihood function: Relative probabilities that various values of

population parameters would arise from the sample observations.

Likelihood ratio: Ratio of the probability that the observed data

arise from a set of model parameters relative to the maximum
probability that arises from the set of maximum likelihood
estimates.

Linear dose–response model: The amount of change in a response is

proportional to the amount of change in some function of dose.

Linearized multistage model: Dose–response model based on the

multistage model of carcinogenesis that is restricted to a form that is
approximately linear at low doses.

Local maximum: Mathematical solution that maximizes a function

in a region that may not be the overall global maximum.

116
 Annex 1: Terminology

Logistic model: A sigmoidal (S-shaped) function that relates the

proportion of individuals with a specified characteristic to an
independent variable.

Lognormal distribution: A mathematical description where the

natural logarithm of a random variable has a normal distribution.

Log transformation: Logarithm of raw data.

Lowest-observed-adverse-effect level (LOAEL): The lowest

concentration or dose of a substance, found by experiment or
observation, that causes an adverse alteration of morphology,
functional capacity, growth, development, or lifespan of the target
organisms distinguishable from normal (control) organisms of the
same species and strain under the same defined conditions of
exposure.

Lowest-observed-effect level (LOEL): The lowest concentration or

dose of a substance, found by experiment or observation, that
causes any alteration of morphology, functional capacity, growth,
development, or lifespan of the target organisms distinguishable
from normal (control) organisms of the same species and strain
under the same defined conditions of exposure.

Margin of exposure (MOE): Ratio of the no-observed-adverse-

effect level (NOAEL) or benchmark dose lower confidence limit
(BMDL) for the critical effect to the theoretical, predicted, or
estimated exposure dose or concentration.

Maximum likelihood estimate: Estimate of a population parameter

most likely to have produced the sample observations.

Mechanism of action: A detailed description of the precise chain of

events from the molecular level to gross macroscopic or
histopathological toxicity.

Michaelis-Menten equation: A dose–response curve, frequently

used for enzyme kinetics, with maximum slope at zero dose that
approaches a maximum asymptote at increasing dose.

117
EHC 239: Principles for Modelling Dose–Response

Mode of action: A series of events that may lead to induction of the

relevant end-point of toxicity for which the weight of evidence
supports plausibility.

Monotonic dose–response: A dose–response that never decreases as

dose increases. A monotonic function may be flat (constant) up to a
threshold dose or may be flat at high doses if a biological limit (e.g.
saturation) is attained.

Multinomial: Animals or individuals may be classified by more than

two (binomial) categories (e.g. in a reproductive study, fetuses may
be dead, alive normal, or alive abnormal).

Negligible risk: A risk management term. In cases where a

quantitative risk estimate has been made, it is any risk less than an
upper-bound incremental lifetime risk calculated using conservative
risk assessment techniques such as the benchmark dose.

Non-linear dose–response model: Mathematical relationship that

cannot be expressed simply as the change in response being
proportional to the amount of change of some function of dose.

No-observed-adverse-effect level (NOAEL): The highest

concentration or dose of a substance, found by experiment or
observation, that causes no detectable adverse alteration of
morphology, functional capacity, growth, development, or lifespan
of the target organisms under defined conditions of exposure.

No-observed-effect level (NOEL): The highest concentration or dose

of a substance, found by experiment or observation, that causes no
detectable alteration of morphology, functional capacity, growth,
development, or lifespan of the target organisms under defined
conditions of exposure.

Normal distribution: A mathematical description where a

continuous random variable x with a mean P and a variance V2 has a
probability density function:

 —
 

ı
3 [ Hí [í
ı ʌ

118
 Annex 1: Terminology

Objective function: Choice of function that is optimized for

maximum likelihood estimation.

Ordinal data: Integers designating the rank, order, or counts.

Parameter: A value used to numerically describe a population of

values (e.g. the mean and standard deviation); or a value used to
describe a dose–response curve (e.g. the intercept and the slope of a
linear dose–response).

Point of departure: The point on a dose–response curve established

from experimental data (e.g. the benchmark dose), generally
corresponding to an estimated low effect level (e.g. 1–10%
incidence of an effect). Depending on the mode of action and
available data, some form of extrapolation below the point of
departure may be employed for low-dose risk assessment, or the
point of departure may be divided by a series of uncertainty factors
to arrive at a reference dose. Points of departure include the BMD,
BMDL, LOAEL, and carcinogenic potency estimates, such as the
T25.

Polynomial: A mathematical function of the sum of a constant,

linear term, quadratic term, cubic term, etc.

Probability: The proportion (on a scale of 0 to 1) of cases for which

a particular event occurs. Zero indicates the event never occurs, and
one indicates the event always occurs.

Probability distribution: A mathematical description of the relative

probabilities of all possible outcomes of a measurement.

Probit function: Assumes that the relative probabilities of effects as

a function of dose are described by a normal distribution. The
cumulative probability as a function of dose has a sigmoidal shape.

Profile likelihood: A plot of the likelihood function versus the

estimated value of a parameter.

P-value: In testing a hypothesis, the probability of a type I error

(false positive). The probability that the sample (experimental)
results are compatible with a specific hypothesis.

119
EHC 239: Principles for Modelling Dose–Response

Quadratic term: A quantity in a mathematical formula that is raised

to the second power (squared).

Quantal data: Dichotomous (binomial) classification where an

individual or animal is placed in one of two categories (e.g. dead or
alive, with or without a particular type of tumour, normal or
abnormal level of a hormone).

Quantile: Percentile (cumulative probability) of a distribution that

ranges from zero to the 100th percentile.

Regression analysis: A statistical process that produces a

mathematical function (regression equation) that relates a dependent
variable (biological effect) to an independent variable (e.g. dose
rate, duration of exposure, age).

Repeated measures: A biological end-point is measured for the

same individual or animal at different times (ages).

Response: Change developed in the state or dynamics of an

organism, system, or (sub)population in reaction to exposure to an
agent.

Residual variance: The variance in experimental measurements

remaining after accounting for the variance due to the independent
variables (e.g. dose rate, duration of exposure, age). Typically
referred to as the inherent unaccountable experimental variation.

Risk: The probability of an adverse effect in an organism, system, or

(sub)population caused under specified circumstances by exposure
to an agent.

Risk assessment: A process intended to calculate or estimate the risk

to a given target organism, system, or (sub)population, including the
identification of attendant uncertainties, following exposure to a
particular agent, taking into account the inherent characteristics of
the agent of concern as well as the characteristics of the specific
target system.

Risk characterization: The qualitative and, wherever possible,

quantitative determination, including attendant uncertainties, of the

120
 Annex 1: Terminology

probability of occurrence of known and potential adverse effects of

an agent in a given organism, system, or (sub)population, under
defined exposure conditions.

Safety factor: Composite (reductive) factor by which an observed or

estimated no-observed-adverse-effect level (NOAEL) is divided to
arrive at a criterion or standard that is considered safe or without
appreciable risk (see Assessment factor and Uncertainty factor).

Severity: The degree to which an effect changes and impairs the

functional capacity of an organ system.

Shape parameter: The exponent on dose in a dose–response

function that dictates the curvature of the function.

Threshold: Dose or exposure concentration of an agent below which

a stated effect is not observed or expected to occur.

Threshold of toxicological concern: An exposure threshold value

below which there is a very low probability of an appreciable risk to
human health.

Toxicodynamics: The process of interaction of chemical substances

with target sites and the subsequent reactions leading to adverse
effects.

Toxicokinetics: The process of the uptake of potentially toxic

substances by the body, the biotransformation they undergo, the
distribution of the substances and their metabolites in the tissues,
and the elimination of the substances and their metabolites from the
body. Both the amounts and the concentrations of the substances
and their metabolites are studied. The term has essentially the same
meaning as pharmacokinetics, but the latter term should be
restricted to the study of pharmaceutical substances.

Uncertainty: Imperfect knowledge concerning the present or future

state of an organism, system, or (sub)population under
consideration.

Uncertainty factor: Reductive factor by which an observed or

estimated no-observed-adverse-effect level (NOAEL) is divided to

121
EHC 239: Principles for Modelling Dose–Response

arrive at a criterion or standard that is considered safe or without

appreciable risk (see Assessment factor and Safety factor).

Unconstrained dose–response model: No restrictions imposed on

the estimates of parameters.

Upper-tail probability: Probability that a variable exceeds a

specified value.

Validation: Process by which the reliability and relevance of a

particular approach, method, process, or assessment is established
for a defined purpose.

Variability: Observable diversity in biological sensitivity or

response and in exposure parameters.

Variance: Measure of variability, standard deviation squared.

Weibull: Form of a dose–response curve characterized by a

relatively shallow slope at low doses that increases sharply as dose
increases before levelling off at high doses.

Weighted least squares estimate: Parameter estimate obtained by

minimizing the sum of squares of observed and estimated values
weighted by a function, frequently the reciprocal of the variance of
an observation.

122
 RESUME, CONCLUSIONS ET RECOMMANDATIONS

1. Résumé

La modélisation de la relation dose-réponse (DRM), destinée à

évaluer quantitativement les risques et enfin à étayer les décisions
de santé publique concernant les expositions à des produits
chimiques, peut être décrite comme un processus en six étapes. Les
quatre premières étapes – sélection des données, choix du modèle,
mise en relation statistique et estimation des paramètres –
constituent une analyse de la relation dose-réponse. Ces étapes
composent le processus permettant d’obtenir une description
mathématique des données, en vue de prédire les réponses à des
doses connues ou d’établir des estimations de doses à partir d’une
réponse donnée. La cinquième étape réalise une synthèse des
résultats de l’analyse dose-réponse et des estimations de
l’exposition afin de guider les décisions de santé publique. L’étape
finale, qui peut éventuellement intervenir plus tôt dans la DRM,
évalue la qualité de la relation dose-réponse et la sensibilité des
prédictions aux hypothèses ayant servi à l’analyse.

La caractérisation de la relation dose-réponse dans les études

chez l’homme et chez l’animal constitue une composante majeure
de la caractérisation des dangers et sert à extrapoler les incidences
des effets nocifs sur la gamme de niveaux d’exposition humaine. Au
cours des années, diverses méthodes ont été mises au point pour
traiter de telles relations, améliorer l’extrapolation pour les faibles
doses et dériver des valeurs guides reposant sur des considérations
sanitaires telles que les doses journalières admissibles (DJA), les
doses journalières tolérables (DJT) et les doses de référence (Dref).
La DRM peut s’avérer utile dans les évaluations des risques en
permettant un meilleur usage des données disponibles et en
fournissant des outils pour évaluer la qualité des données et les
incertitudes résultantes sur les estimations de la relation dose-
réponse.

D’une manière générale, les estimations obtenues par DRM

sont établies à partir de données provenant de l’ensemble de la
courbe dose-réponse pour l’effet critique. L’approche standard
reposant sur la dose sans effet nocif observé (DSENO) peut être
considérée comme un cas spécial et simplifié d’analyse de la

123
EHC 239: Principles for Modelling Dose–Response

relation dose-réponse, dans la mesure où elle identifie une dose

unique supposée ne pas avoir d’effet nocif appréciable. La modèle
DRM reflète les caractéristiques de la courbe dose-réponse, en
permettant notamment d’estimer sa pente. Dans le cas d’un cadre de
régression, il indique l’écart-type et l’intervalle de confiance pour
les paramètres de modélisation. L’un des inconvénients de
l’approche DSENO réside dans l’impossibilité de quantifier les
degrés de variabilité et d’incertitude, alors que d’autres modèles de
la relation dose-réponse peuvent faciliter l’analyse de ces grandeurs.
L’utilisation d’un modèle dose-réponse peut permettre d’optimiser
la conception de l’étude et de préciser les besoins en matière
d’études supplémentaires. L’approche DSENO intègre des
informations biologiques à travers l’application d’un jugement
d’expert, néanmoins subjectif. Une DRM complète est en mesure de
fournir une analyse plus « riche en éléments scientifiques » grâce à
l’inclusion quantitative plus formelle dans les modèles de facteurs et
de covariables, par exemple. Les estimations obtenues à partir de
cette DRM facilitent la comparaison dans un cadre commun entre
des expériences, des effets et des composés qui diffèrent sur le plan
quantitatif. La modélisation DRM peut aussi conduire à de
meilleures évaluations des risques et de l’innocuité et offre la
possibilité d’étudier les probabilités d’effets se manifestant en
dehors de la plage observable.

Le choix des modèles à utiliser dépend du type de donnée. Il

faut sélectionner un modèle dose-réponse et un modèle décrivant la
variabilité des données. Une fois qu’on dispose de modèles adaptés
à un jeu de données, on peut évaluer leur degré de représentativité
pour ces données par des mesures de la qualité de l’ajustement. On
peut en outre comparer leur capacité à décrire les données. Les
incertitudes portant sur les inférences tirées de ces modèles se
répartissent dans quatre catégories : incertitudes statistiques sur les
inférences dues à la variabilité des réponses entre les sujets des
expériences, erreurs expérimentales (randomisation imparfaite,
erreurs de dosage, localisation défavorable de la dose, par exemple),
variabilité d’une expérience à l’autre due aux différences inévitables
dans l’exécution de l’expérience et incertitude due au fait que l’on
ne connaît pas le « vrai modèle » décrivant les données. L’analyse
de la relation dose-réponse doit, dans la mesure du possible, prendre
en compte l’ensemble de ces quatre sources de variabilité et
d’incertitude.

124
 Résumé, Conclusions et Recommandations

Le calcul des doses de référence (BMD) est une application

particulièrement importante de la DRM. Les BMD sont les doses
pour lesquelles on détermine par déduction qu’il se produira un
niveau donnée de réponse. Lorsqu’on dispose de données
appropriées, les BMD offrent une alternative à l’approche DSENO
pour le calcul de valeurs guides reposant sur des considérations
sanitaires. Lorsqu’une extrapolation s’avère nécessaire, il convient
de représenter l’incertitude associée à la prédiction. Il est dans ce
cas particulièrement important d’indiquer l’incertitude liée au
modèle.

Une DRM complète peut apporter des informations

supplémentaires au gestionnaire de risques. La sortie du modèle doit
être conçue pour répondre à certaines questions concernant la
probabilité d’effets sanitaires nocifs. Elle peut être présentée
essentiellement de trois façons. Premièrement, elle peut servir à
établir des valeurs guides reposant sur des considérations sanitaires
telles que les DJA, les DJT ou les Dref, d’une manière analogue aux
procédures actuellement appliquées à partir de la DSENO ou de la
dose minimale avec effets nocifs observés (DMENO). La DRM
peut être une méthode plus sûre sur le plan scientifique pour
déterminer ces valeurs guides. Deuxièmement, la sortie de la DRM
peut être utilisée en gestion des risques pour estimer une marge
d’exposition (ME), par détermination du rapport de la dose
correspondant à une limite donnée de la réponse à un niveau
d’exposition humaine. Troisièmement, sur la base de la relation
dose-réponse modélisée, cette sortie peut être une estimation
quantitative de l’ampleur du risque ou de l’effet sanitaire pour un
niveau d’exposition humaine, moyennant l’hypothèse généralement
acceptée que les facteurs d’incertitude utilisés couvrent les
incertitudes associées aux différences de sensibilité entre individus
et espèces. La DRM peut fournir de meilleures informations sur la
probabilité des effets pour les doses faibles et inférieures aux
niveaux observés dans les systèmes biologiques, ainsi que de
meilleures estimations des incertitudes statistiques entachant les
estimations des effets probables.

La multiplicité des jeux de données et des incertitudes peut

influer sur le type de sortie des exercices de DRM et avoir de
l’importance pour les gestionnaires de risques. On peut utiliser la
DRM avec des données d’exposition pour identifier les sous-

125
EHC 239: Principles for Modelling Dose–Response

populations à risque. Elle peut aussi aider les gestionnaires dans la

détermination des priorités et dans l’évaluation des conséquences
d’interventions proposées pour réduire les risques. Pour la
communication à propos des risques, l’application des techniques de
DRM offre des opportunités, mais comporte aussi des difficultés.
Les évaluations par DRM peuvent produire des informations sous
plusieurs formats, et notamment sous forme de fonctions dose-
réponse permettant, avec les estimations de l’exposition, de prédire
les risques pour des niveaux d’exposition donnés, ainsi que de
fonctions permettant inversement d’estimer les niveaux
d’exposition à l’origine de risques donnés. On obtient ainsi
notamment des estimations du risque potentiel d’absorption plus
importante qu’une valeur guide reposant sur des considérations
sanitaires, DJA par exemple. Ces évaluations offrent aussi des
approches pour comparer les risques ou les bénéfices concurrents et
s’intéresser aux incertitudes susceptibles d’influer sur les risques
prédits. Toutefois, à moins que la situation en termes de risque ne
soit envisagée à l’échelle de la population, il existe un problème de
communication à propos du risque car lorsqu’on présente le niveau
de risque dans des situations où aucun niveau d’exposition n’est
dépourvu de risque, la modélisation prévoit qu’un certain
pourcentage de la population subira des effets jugés nocifs. Il faut
reconnaître que l’utilisation de la DRM impose aux données des
exigences en termes de qualité et de quantité et nécessite des
compétences spécifiques.

L’utilisation courante des estimations tirées der la DRM

pourrait, du point de vue de la gestion des risques, permettre une
meilleure caractérisation avant la prise de décisions en :

x apportant des informations sur les valeurs guides (ampleur et

types des impacts sanitaires) ;
x montrant les bénéfices de différentes actions réglementaires ;
x fournissant au décideur une appréciation des données « plus
que ponctuelle » ;
x favorisant la cohérence dans les décisions, moyennant des
ajustements appropriés pour tenir compte des différences entre
les effets, les niveaux d’effet, les espèces et les types d’étude ;
et en
x permettant en continu et en permanence des interactions
itératives entre l’évaluateur et le gestionnaire de risques.

126
 Résumé, Conclusions et Recommandations

L’utilisation de la DRM et des techniques d’évaluation

probabiliste pour décrire quantitativement la variabilité et
l’incertitude génère de nouvelles difficultés dans la communication
à propos des risques. Ces difficultés résident notamment dans :
x l’explication de la prévision, pour un certain pourcentage de la
population, d’un dépassement du niveau de sécurité et/ou de
l’apparition d’effets nocifs ;
x l’explication du niveau de risque dans les cas où on suppose
qu’aucun niveau d’exposition n’est dépourvu de risque ;
x la comparaison entre risques ou bénéfices concurrents ;
x la mise en lumière d’incertitudes influant sur le risque prédit ;
et
x l’explication du fait qu’en matière de risque, une estimation
indique ce qui peut se passer au niveau d’une population, plutôt
qu’à celui d’un individu, ce qui, notons le, vaut aussi pour
l’approche DJA/DJT.

2. Conclusions

x La DRM complète peut être considérée comme une alternative

plus élaborée et plus robuste à l’approche DSENO dans tous les
cas où l’on dispose de données appropriées sur la relation dose-
réponse (pour plusieurs groupes de dose et différents niveaux
d’exposition, par exemple).
x Pour les données dose-réponse ponctuelles, on s’intéresse
souvent aux faibles niveaux de réponse (d’incidence). Il est
parfois nécessaire, dans cette perspective, d’extrapoler sur
plusieurs ordres de grandeur (pour l’incidence des tumeurs, par
exemple). Cependant, des modèles également plausibles de la
relation dose-réponse peuvent fournir des estimations fortement
divergentes pour les faibles valeurs. Une approche actuellement
appliquée en tant que méthode prudente consiste à estimer la
BMD10 (dose pour un risque de 10 %) et à extrapoler
linéairement à partir de ce point vers les valeurs descendantes.
Une autre solution, actuellement en cours de développement,
applique une approche bayésienne, considérant globalement les
divers modèles.
x Pour les données dose-réponse continues, il existe deux
approches de type DRM. L’une comprend la transformation des

127
EHC 239: Principles for Modelling Dose–Response

données continues en données ponctuelles. L’autre considère

les données dose-réponse continues comme des informations
sur la gravité de l’effet et donc comme une fonction de la dose.
Dans cette dernière approche, des variations mesurables de
l’effet sont souvent proches des niveaux de réponses considérés
comme nocifs (par exemple, inhibition de 10 % de la
cholinestérase) et le problème de l’extrapolation à faible dose
est mineur ou ne se pose pas.

x Pour dériver une DJA, une DJT ou une Dref, on peut faire
appel à la DRM pour déterminer une BMD, qui sera utilisée
comme point de départ de la même façon qu’une DSENO
(c’est-à-dire qu’on appliquera les mêmes facteurs d’incertitude
à la BMD qu’à la DSENO).

x La DRM peut aussi être employée pour estimer les risques

correspondant à un niveau d’exposition (humaine) donné. Pour
évaluer les risques en termes d’incidence (données
ponctuelles), cette opération peut devoir inclure une
extrapolation aux faibles doses.

x Les exercices de DRM peuvent apporter des informations sur

les incertitudes associées aux données et identifier des facteurs
contribuant aux incertitudes sur les estimations des risques.

x L’application de la DRM à tous les points finaux peut être

extrêmement onéreuse, il est donc plus efficace de
présélectionner les points finaux apparemment les plus
sensibles. Dans certains cas cependant, il n’est pas facile
d’identifier visuellement ces points de sorte qu’il peut être
nécessaire de modéliser tous les points finaux.

x La BMD et la borne inférieure de l’intervalle de confiance de la

BMD (BMDL) doivent toujours être indiquées de manière à ce
que la qualité des données et de l’ajustement du modèle
apparaisse clairement et que l’on puisse procéder à une
comparaison de puissances à partir de la BMD.

x Il convient de présenter la sortie des différents modèles de

DRM.

128
 Résumé, Conclusions et Recommandations

3. Recommandations

x Les protocoles d’évaluation de la toxicité (par exemple les

Lignes directrices de l’Organisation de coopération et de
développement économiques) doivent être examinés pour
optimiser l’approche utilisant la BMD et d’autres approches de
type DRM, notamment pour choisir au mieux les nombres
d’animaux et de doses pour les différentes courbes dose-
réponse. Des recherches supplémentaires sont nécessaires pour
développer des types d’étude optimaux. Il faut également
élaborer des conseils pour combiner les études existantes en
vue d’une DRM.

x Il faut mettre au point des recommandations pour l’analyse

combinée de différents jeux de données en vue d’une
estimation plus précise des BMD.

x Il faut aussi parvenir à mieux comprendre quand et comment

utiliser la réponse de référence (BMR).

x La forme de la courbe dose-réponse aux faibles doses doit être

mieux interprétée. Des recherches supplémentaires sont
nécessaires pour déterminer la base biologique de
l’extrapolation (en faisant appel, par exemple, à des marqueurs
biologiques, à des précurseurs de tumeur, à des animaux
génétiquement modifiés ou à la toxico-cinétique, pour estimer
la dose cible).

x Il faut élaborer de meilleures recommandations pour la

communication à propos des risques sur la base des résultats de
la DRM et des techniques d’évaluation probabiliste. Cette
communication devra couvrir les types d’incertitude, leur
relation avec la variabilité statistique, l’imprécision et
l’utilisation des intervalles de confiance.

x L’utilisation de la DRM doit faire l’objet d’un bilan et des

principes généraux supplémentaires devront être développés à
mesure que l’on disposera de plus d’expérience.

129
 RESUMEN, CONCLUSIONES Y
RECOMENDACIONES

1. Resumen

La creación de modelos de la relación dosis-respuesta, para su

utilización en la evaluación cuantitativa del riesgo y en último
término para documentar las decisiones en materia de salud pública,
se puede describir como un proceso en seis etapas. Las cuatro
primeras etapas—selección de datos, selección del modelo,
vinculación estadística y estimación de los parámetros—
constituyen el análisis de la relación dosis-respuesta. Estas etapas
están relacionadas con el proceso mediante el cual se obtiene una
descripción matemática de los datos, a fin de evaluar respuestas
previstas para dosis conocidas u obtener estimaciones de la dosis
cuando lo que interesa es una respuesta determinada. La quinta
etapa consiste en la integración de los resultados del análisis de la
relación dosis-respuesta en las estimaciones de la exposición, con el
objetivo de orientar las decisiones relativas a la salud pública. La
última etapa, que se puede elegir aplicar antes, consiste en una
evaluación de la calidad del análisis de la relación dosis-respuesta y
de la sensibilidad de las predicciones de los modelos con respecto a
las hipótesis utilizadas en el análisis.

La caracterización de las relaciones dosis-respuesta en estudios

realizados en animales y personas ha sido un componente
importante de la caracterización del peligro y se ha utilizado en la
extrapolación de incidencias de efectos adversos en la gama de los
niveles de exposición humana. Durante años se han elaborado
diversos métodos para ajustar dichas relaciones, mejorar la
extrapolación a dosis bajas y obtener valores guía basados en la
salud, como la ingesta diaria admisible (IDA), la ingesta diaria
tolerable (IDT) y las dosis de referencia. La creación de modelos
puede ser útil en las evaluaciones del riesgo para utilizar mejor los
datos disponibles y para suministrar instrumentos de evaluación de
la calidad de los datos y las consiguientes incertidumbres en las
estimaciones de la relación dosis-respuesta.

En general, las estimaciones de los modelos de la relación

dosis-respuesta se basan en los datos obtenidos de la totalidad de la

130
 Resumen, Conclusiones y Recomendaciones

curva correspondiente a dicha relación para el efecto crítico. El

método normalizado de la concentración sin efectos adversos
observados (NOAEL) se puede considerar como un caso especial
simplificado de análisis de la relación dosis-respuesta, puesto que
identifica una dosis única que se supone que no tiene un efecto
adverso apreciable. El modelo de la relación dosis-respuesta refleja
las características de la curva de dicha relación, en particular porque
proporciona estimaciones de la pendiente. En el caso de un marco
de regresión, proporciona el error estándar y los intervalos de
confianza para los parámetros del modelo. La utilización del método
de la NOAEL tiene el inconveniente de que no es posible
cuantificar el grado de variabilidad e incertidumbre que puede
haber, mientras que otros modelos de la relación dosis-respuesta
pueden facilitar el análisis de la sensibilidad y la incertidumbre. El
examen de un modelo de dosis-respuesta puede mejorar al máximo
la formulación del estudio y aclarar la necesidad de estudios
adicionales. El método de la NOAEL incorpora información
biológica mediante la aplicación de un parecer “experto”, pero
subjetivo. La creación de modelos de la relación dosis-respuesta
completos permitiría un análisis más “científico”, por ejemplo
mediante la inclusión cuantitativa más oficial de factores y
covariantes en los modelos. Las estimaciones derivadas de los
modelos de dosis-respuesta mejoran la capacidad para comparar
experimentos, efectos y compuestos con diferencias cuantitativas en
el ámbito de un marco común. Los modelos pueden mejorar las
evaluaciones del riesgo y de la inocuidad, ofreciendo al mismo
tiempo oportunidades para examinar la probabilidad de los efectos
fuera de la gama observable.

La elección de los modelos que se van a utilizar depende del

tipo de datos. Dichos modelos deben incluir un patrón para la
relación dosis-respuesta y otro para la variabilidad de los datos. Una
vez ajustados los modelos a una serie de datos, se puede evaluar el
grado en que los describen individualmente utilizando medidas de la
precisión del ajuste. Además, se puede comparar entre ellos la
capacidad para describir los datos. Las incertidumbres sobre las
consecuencias que puedan derivarse de dichos modelos entran en
cuatro categorías principales: incertidumbre estadística de las
consecuencias debida a la variabilidad entre las respuestas de los
sujetos objeto de experimentación, errores experimentales (por
ejemplo, distribución al azar imperfecta, errores de dosificación,

131
EHC 239: Principles for Modelling Dose–Response

localización desfavorable de las dosis), variabilidad entre

experimentos debida a diferencias inevitables en su realización e
incertidumbre debida al hecho de que no se conoce el “verdadero
modelo” para los datos. Siempre que sea posible, en el análisis de la
relación dosis-respuesta hay que abordar las cuatro fuentes de
variabilidad e incertidumbre.

Una aplicación particularmente importante de los modelos de

dosis-respuesta es el cálculo de las dosis de referencia. Son las dosis
con las cuales se deduce que se producirá un determinado nivel de
respuesta. Cuando se dispone de datos apropiados, las dosis de
referencia son una alternativa al método de la NOAEL para calcular
los valores guía basados en la salud. Cuando es necesaria una
extrapolación, se debe representar la incertidumbre asociada con
una predicción. En este caso es particularmente importante incluir la
incertidumbre del modelo.

La creación de modelos de la relación dosis-respuesta

completos ofrece la posibilidad de proporcionar información
adicional a los gestores del riesgo. Los resultados de los modelos se
deben orientar hacia el examen de cuestiones específicas relativas a
la probabilidad de efectos adversos en la salud. Se pueden presentar
de tres maneras principales. En primer lugar, se pueden utilizar para
el establecimiento de un valor guía basado en la salud, por ejemplo
una IDA, una IDT o unas dosis de referencia, de manera análoga a
los procedimientos actuales basados en la NOAEL o la
concentración más baja con efectos adversos observados (LOAEL).
Los modelos de dosis-respuesta pueden ser un método más sólido
desde el punto de vista científico para determinar valores guía
basados en la salud. En segundo lugar, los resultados de dichos
modelos se pueden utilizar en la gestión del riesgo para estimar un
margen de exposición, mediante el cálculo de la relación entre la
dosis correspondiente a un límite determinado de respuesta y un
nivel de exposición humana. En tercer lugar, sobre la base de la
relación dosis-respuesta obtenida mediante el modelo, el resultado
puede ser una estimación cuantitativa de la magnitud del
riesgo/efecto en la salud para el nivel de exposición humana, con la
hipótesis generalmente aceptada de que los factores de
incertidumbre utilizados incluyen las incertidumbres relativas a las
diferencias de sensibilidad intraespecíficas e interespecíficas. Los
modelos de dosis-respuesta pueden proporcionar mejor información
sobre la probabilidad de efectos con dosis bajas, inferiores a los
132
 Resumen, Conclusiones y Recomendaciones

niveles observados en los sistemas biológicos, y pueden

proporcionar asimismo mejores estimaciones de las incertidumbres
estadísticas de los efectos probables.

Dos factores que pueden influir en el tipo de resultados

obtenidos de la aplicación de los modelos de dosis-respuesta y que
pueden ser importantes para el gestor del riesgo son las series de
datos múltiples y las incertidumbres. Los modelos se pueden utilizar
con datos de exposición para identificar las subpoblaciones en
situación de riesgo. También se pueden emplear para ayudar a los
gestores del riesgo a establecer prioridades y evaluar las
consecuencias de las intervenciones propuestas encaminadas a
reducir el riesgo. Para la comunicación del riesgo, la utilización de
técnicas con modelos de dosis-respuesta ofrece oportunidades y
retos. Las evaluaciones con estos modelos pueden generar
información de varios tipos, como funciones de la relación dosis-
respuesta que permiten, junto con las estimaciones de la exposición,
la predicción de los riesgos con niveles específicos de exposición y
funciones que permiten la estimación de los niveles de exposición
que dan lugar a riesgos determinados. Esto incluye las estimaciones
del posible riesgo de ingestas por encima de un valor guía basado en
la salud, por ejemplo la IDA. Las evaluaciones con los modelos de
dosis-respuesta también ofrecen métodos para comparar riesgos o
beneficios competitivos y permiten concentrar la atención en las
incertidumbres que pueden influir en el riesgo pronosticado. Sin
embargo, salvo que la situación del riesgo se examine en la
población, su comunicación presenta el problema de que, aun
explicando el nivel de riesgo en esas circunstancias en las que no
hay un nivel inocuo de exposición, cabe predecir que cierto
porcentaje de la población va a registrar algunos efectos
considerados adversos. Hay que reconocer que la utilización de los
modelos de la relación dosis-respuesta requiere cierta cantidad y
calidad de datos, así como conocimientos técnicos específicos.

El uso potencial “continuo” de las estimaciones derivadas de

los modelos de dosis-respuesta puede, desde una perspectiva de
gestión del riesgo, mejorar la caracterización para la adopción de
decisiones, porque:

133
 EHC 239: Principles for Modelling Dose–Response

x facilita información sobre lo que ocurre por encima del valor

guía basado en la salud (magnitud y tipos de efectos en la
salud);
x demuestra los beneficios de distintas medidas normativas;
x ofrece a los encargados de la adopción de decisiones una
apreciación de los datos desde más de un punto de vista;
x promueve la coherencia en las decisiones, si se hacen ajustes
apropiados para las diferencias en los efectos, el nivel de los
efectos, las especies y la formulación del estudio; y
x facilita una interacción iterativa entre el asesor del riesgo y el
gestor del riesgo de manera continua e ininterrumpida.

La utilización de modelos de la relación dosis respuesta y

de técnicas de evaluación probabilística para describir de
manera cuantitativa la variabilidad y la incertidumbre incorpora
nuevos retos a la comunicación del riesgo. Algunos de ellos son
los siguientes:

x explicar que se prevé que un cierto porcentaje de la población

superará el nivel de inocuidad y/o sufrirá un efecto adverso;
x explicar el nivel de riesgo en esas circunstancias en las que se
supone que no hay un nivel inocuo de exposición;
x comparar los riesgos o los beneficios en pugna;
x prestar una atención especial a las incertidumbres que influyen
en el riesgo pronosticado; y
x explicar que una estimación del riesgo se refiere a lo que puede
ocurrir a la población, más que a nivel individual, y señalar que
esto es lo que ocurre también con el enfoque de la IDA/IDT.

2. Conclusiones

x La creación de modelos de la relación dosis-respuesta

completos se puede considerar un método alternativo más
complejo o válido que el de la NOAEL en todos los casos en
que se disponga de datos apropiados de la relación dosis-
respuesta (por ejemplo, para varios grupos de dosis con
distintos niveles de respuesta).

x Para los datos cuantales de la relación dosis-respuesta, el

interés radica con frecuencia en los niveles bajos de respuesta
(incidencia). Esto puede exigir una extrapolación a dosis más

134
 Resumen, Conclusiones y Recomendaciones

bajas en varios órdenes de magnitud (por ejemplo, para las

incidencias de tumores). Sin embargo, los modelos del riesgo
de la relación dosis-respuesta que sean igualmente admisibles
pueden dar lugar a estimaciones bajas muy divergentes. Un
método aplicado actualmente, considerado prudente, consiste
en estimar una dosis de referencia10 (dosis con un riesgo del
10%) y hacer una extrapolación de manera lineal descendente
desde ese punto. Otra opción, todavía en preparación, consiste
en aplicar un método bayesiano, que examina los distintos
modelos en conjunto.

x Para la obtención de datos continuos de la relación dosis-

respuesta hay dos sistemas de utilización de los modelos. Uno
consiste en transformar los datos continuos en datos cuantales.
El otro en considerar los datos continuos de la relación dosis-
respuesta como información de la gravedad del efecto y, por
consiguiente, como una función de la dosis. En el segundo
sistema, los cambios mensurables de los efectos suelen estar
cerca de los niveles de respuesta considerados adversos (por
ejemplo, la inhibición del 10% de la colinesterasa) y el
problema de la extrapolación a dosis bajas es insignificante o
inexistente.

x Con el fin de obtener un valor de la IDA, la IDT o las dosis de

referencia, se pueden utilizar los modelos de dosis-respuesta
para derivar una dosis de referencia, que se utilizará como
punto de partida de la misma manera que se utiliza la NOAEL
(es decir, se aplicarían a la dosis de referencia los mismos
factores de incertidumbre que a la NOAEL).

x También se pueden utilizar modelos de la relación dosis-

respuesta para estimar los riesgos en un determinado nivel de
exposición (humana). Para los riesgos expresados como
incidencias (datos cuantales) puede ser necesaria la
extrapolación a dosis bajas.

x El uso de modelos de dosis-respuesta puede proporcionar

información sobre las incertidumbres asociadas con los datos e
identificar los factores que contribuyen a ellas en las
estimaciones del riesgo.

135
 EHC 239: Principles for Modelling Dose–Response

x La aplicación de modelos de la relación dosis-respuesta a todos

los efectos finales puede tener un costo prohibitivo, de manera
que sería útil realizar una selección previa de los efectos finales
aparentemente más sensibles. Sin embargo, en algunos casos no
es fácil identificar los más sensibles mediante una inspección
visual, de manera que hay que aplicar el modelo a todos ellos.

x Se debería notificar siempre la dosis de referencia y su límite

inferior de confianza, de manera que la calidad de los datos y el
ajuste del modelo sean claros y se puedan comparar sus
potencias basándose en la dosis de referencia.

x Se deben presentar los resultados de los distintos métodos

utilizados en los modelos de dosis-respuesta.

3. Recomendaciones

x Se deben examinar los protocolos de las pruebas de toxicidad

(por ejemplo, las directrices de la Organización de Cooperación
y Desarrollo Económicos) para conseguir unos resultados
óptimos de las dosis de referencia y demás métodos basados en
modelos de la relación dosis-respuesta, por ejemplo las
formulaciones óptimas correspondientes al número de animales
y el número de dosis para diferentes curvas de la relación dosis-
respuesta.

x Hay que elaborar mejores orientaciones para el análisis

combinado de distintas series de datos, a fin de estimar las
dosis de referencia con mayor precisión.

x Es necesario fomentar un mayor conocimiento de cuándo y

cómo se ha de utilizar la respuesta de referencia.

x Hay que tratar de conocer mejor la forma de la curva de la

relación dosis-respuesta a dosis bajas. Se requieren nuevas
investigaciones para determinar la base biológica de la
extrapolación (por ejemplo, utilizando biomarcadores,
precursores de tumores, animales modificados genéticamente y
la tóxicocinética para la estimación de dosis específicas).

136
 Resumen, Conclusiones y Recomendaciones

x Es necesario elaborar orientaciones mejores para la

comunicación del riesgo basada en los resultados de los
modelos dosis-respuesta y de las técnicas de evaluación
probabilística. Deben incluir la comunicación de los tipos de
incertidumbre y la relación con la variabilidad estadística, la
imprecisión y la utilización de intervalos de confianza.

x Se debe examinar la utilización de modelos de la relación

dosis-respuesta y se han de elaborar principios generales
adicionales para su uso cuando se disponga de más experiencia.

137
 THE ENVIRONMENTAL HEALTH CRITERIA SERIES (continued)

Acetaldehyde (No. 167, 1995) Chlorinated paraffins (No. 181, 1996)

Acetone (No. 207, 1998) Chlorine and hydrogen chloride
Acetonitrile (No. 154, 1993) (No. 21, 1982)
Acrolein (No. 127, 1991) Chloroalkyl ethers, selected
Acrylamide (No. 49, 1985) (No. 201, 1998)
Acrylic acid (No. 191, 1997) Chlorobenzenes other than
Acrylonitrile (No. 28, 1983) hexachlorobenzene (No. 128, 1991)
Aged population, principles for evaluating Chlorofluorocarbons, fully halogenated
the effects of chemicals (No. 144, 1992) (No. 113, 1990)
Aldicarb (No. 121, 1991) Chlorofluorocarbons, partially
Aldrin and dieldrin (No. 91, 1989) halogenated (ethane derivatives)
Allergic hypersensitization associated with (No. 139, 1992)
exposure to chemicals, principles and (methane derivatives) (No. 126, 1991)
methods for assessing (No. 212, 1999)
Allethrins (No. 87, 1989) Chloroform (No. 163, 1994)
Aluminium (No. 194, 1997) Chlorophenols (No. 93, 1989)
Amitrole (No. 158, 1994) Chlorothalonil (No. 183, 1996)
Ammonia (No. 54, 1986) Chromium (No. 61, 1988)
Anticoagulant rodenticides Chrysotile asbestos (No. 203, 1998)
(No. 175, 1995) Copper (No. 200, 1998)
st
Arsenic (No. 18, 1981, 1 edition) Cresols (No. 168, 1995)
Arsenic and arsenic compounds Cyhalothrin (No. 99, 1990)
nd
(No. 224, 2001, 2 edition) Cypermethrin (No. 82, 1989)
Asbestos and other natural mineral fibres Cypermethrin, alpha- (No. 142, 1992)
(No. 53, 1986) DDT and its derivatives (No. 9, 1979)
Assessment of risks to human health DDT and its derivatives –
from exposure to chemicals, principles environmental aspects (No. 83, 1989)
for the (No. 210, 1999) Deltamethrin (No. 97, 1990)
Autoimmunity associated with exposure Demeton-S-methyl (No. 197, 1997)
to chemicals, principles and methods for Dermal absorption (No. 235, 2006)
assessing (No. 236, 2006) Diaminotoluenes (No. 74, 1987)
Bacillus thuringiensis (No. 217, 1999) Diazinon (No. 198, 1997)
Barium (No. 107, 1990) 1,2-Dibromoethane (No. 177, 1996)
Benomyl (No. 148, 1993) Di-n-butyl phthalate (No. 189, 1997)
Benzene (No. 150, 1993) 1,2-Dichloroethane
Beryllium (No. 106, 1990) (No. 62, 1987, 1st edition)
Biomarkers and risk assessment: (No. 176, 1995, 2nd edition)
concepts and principles (No. 155, 1993) 2,4-Dichlorophenoxyacetic acid
Biomarkers in risk assessment: validity (2,4-D) (No. 29, 1984)
and 2,4-Dichlorophenoxyacetic acid –
validation (No. 222, 2001) environmental aspects (No. 84, 1989)
Biotoxins, aquatic (marine and 1,3-Dichloropropene, 1,2-dichloropropane
freshwater) (No. 37, 1984) and mixtures (No. 146, 1993)
Boron (No. 204, 1998) Dichlorvos (No. 79, 1988)
Brominated diphenylethers Dinitro-ortho-cresol (No. 220, 2000)
(No. 162, 1994) Diesel fuel and exhaust emissions
Butanols – four isomers (No. 65, 1987) (No. 171, 1996)
Cadmium (No. 134, 1992) Diethylhexyl phthalate (No. 131, 1992)
Cadmium – environmental aspects Diflubenzuron (No. 184, 1996)
(No. 135, 1992) Dimethoate (No. 90, 1989)
Camphechlor (No. 45, 1984) Dimethylformamide (No. 114, 1991)
Carbamate pesticides: a general Dimethyl sulfate (No. 48, 1985)
introduction (No. 64, 1986) Diseases of suspected chemical
Carbaryl (No. 153, 1994) etiology and their prevention,
Carbendazim (No. 149, 1993) principles of studies on (No. 72, 1987)
Carbon disulfide (No. 10, 1979) Disinfectants and disinfectant by-products
Carbon monoxide (No. 216, 1999)
(No. 13, 1979, 1st edition) Dithiocarbamate pesticides,
(No. 213, 1999, 2nd edition) ethylenethiourea, and propylenethiourea:
Carbon tetrachloride (No. 208, 1999) a general introduction (No. 78, 1988)
Carcinogens, summary report on the Electromagnetic fields (No. 137, 1992)
evaluation of short-term in vitro tests Elemental speciation in human health
(No. 47, 1985) risk assessment (No. 234, 2006)
Carcinogens, summary report on the Endosulfan (No. 40, 1984)
evaluation of short-term in vivo tests Endrin (No. 130, 1992)
(No. 109, 1990) Environmental epidemiology, guidelines
Chlordane (No. 34, 1984) on studies in (No. 27, 1983)
Chlordimeform (No. 199, 1997) Epichlorohydrin (No. 33, 1984)
Chlordecone (No. 43, 1984) Essential trace elements: Principles and
Chlorendic acid and anhydride
(No. 185, 1996)
 THE ENVIRONMENTAL HEALTH CRITERIA SERIES (continued)

methods for the assessment of risk Methyl parathion (No. 145, 1992)
(No. 228, 2001) Methyl tertiary-butyl ether (No. 206, 1998)
Ethylbenzene (No. 186, 1996) Mirex (No. 44, 1984)
Ethylene oxide (No. 55, 1985) Modelling dose–response for the risk
Extremely low frequency (ELF) fields assessment of chemicals, Principles for
(No. 36, 1984) (No. 239, 2009)
(No. 238, 2007) Morpholine (No. 179, 1996)
Fenitrothion (No. 133, 1992) Mutagenic and carcinogenic chemicals,
Fenvalerate (No. 95, 1990) guide to short-term tests for detecting
Flame retardants: a general introduction (No. 51, 1985)
(No. 192, 1997) Mycotoxins (No. 11, 1979)
Flame retardants: tris(chloropropyl) Mycotoxins, selected: ochratoxins,
phosphate and tris(2-chloroethyl) trichothecenes, ergot (No. 105, 1990)
phosphate (No. 209, 1998) Nephrotoxicity associated with exposure
Flame retardants: tris(2-butoxyethyl) to chemicals, principles and methods for
phosphate, tris(2-ethylhexyl) phosphate the assessment of (No. 119, 1991)
and tetrakis(hydroxymethyl) Neurotoxicity associated with exposure to
phosphonium salts (No. 218, 2000) chemicals, principles and methods for the
Fluorides (No. 227, 2001) assessment of (No. 60, 1986)
Fluorine and fluorides (No. 36, 1984) Neurotoxicity risk assessment for human
Food additives and contaminants in food, health, principles and approaches
principles for the safety assessment of (No. 223, 2001)
(No. 70, 1987) Nickel (No. 108, 1991)
Formaldehyde (No. 89, 1989) Nitrates, nitrites, and N-nitroso
Fumonisin B1 (No. 219, 2000) compounds (No. 5, 1978)a
Genetic effects in human populations, Nitrobenzene (No. 230, 2003)
guidelines for the study of (No. 46, 1985) Nitrogen oxides
Glyphosate (No. 159, 1994) (No. 4, 1977, 1st edition)a
Guidance values for human (No. 188, 1997, 2nd edition)
exposure limits (No. 170, 1994) 2-Nitropropane (No. 138, 1992)
Health risks in children associated with Nitro-and nitro-oxypolycyclic aromatic
exposure to chemicals, principles for hydrocarbons, selected (No. 229, 2003)
evaluating (No. 237, 2006) Noise (No. 12, 1980)a
Heptachlor (No. 38, 1984) Organophosphorus insecticides:
Hexachlorobenzene (No. 195, 1997) a general introduction (No. 63, 1986)
Hexachlorobutadiene (No. 156, 1994) Palladium (No. 226, 2001)
Alpha- and beta-hexachlorocyclohexanes Paraquat and diquat (No. 39, 1984)
(No. 123, 1992) Pentachlorophenol (No. 71, 1987)
Hexachlorocyclopentadiene Permethrin (No. 94, 1990)
(No. 120, 1991) Pesticide residues in food, principles for
n-Hexane (No. 122, 1991) the toxicological assessment of
Human exposure assessment (No. 104, 1990)
(No. 214, 2000) Petroleum products, selected
Hydrazine (No. 68, 1987) (No. 20, 1982)
Hydrogen sulfide (No. 19, 1981) Phenol (No. 161, 1994)
Hydroquinone (No. 157, 1994) d-Phenothrin (No. 96, 1990)
Immunotoxicity associated with exposure Phosgene (No. 193, 1997)
to chemicals, principles and methods for Phosphine and selected metal
assessment (No. 180, 1996) phosphides (No. 73, 1988)
Infancy and early childhood, principles for Photochemical oxidants (No. 7, 1978)
evaluating health risks from chemicals Platinum (No. 125, 1991)
during (No. 59, 1986) Polybrominated biphenyls (No. 152, 1994)
Isobenzan (No. 129, 1991) Polybrominated dibenzo-p-dioxins and
Isophorone (No. 174, 1995) dibenzofurans (No. 205, 1998)
Kelevan (No. 66, 1986) Polychlorinated biphenylsa and terphenyls
Lasers and optical aradiation (No. 23, 1982) (No. 2, 1976, 1st edition)
Lead (No. 3, 1977) (No. 140, 1992, 2nd edition)
Lead, inorganic (No. 165, 1995) Polychlorinated dibenzo-p-dioxins and
Lead – environmental aspects dibenzofurans (No. 88, 1989)
(No. 85, 1989) Polycyclic aromatic hydrocarbons,
Lindane (No. 124, 1991) selected non-heterocyclic (No. 202, 1998)
Linear alkylbenzene sulfonates Progeny, principles for evaluating health
and related compounds (No. 169, 1996) risks associated with exposure to
Magnetic fields (No. 69, 1987) chemicals during pregnancy
Man-made mineral fibres (No. 77, 1988) (No. 30, 1984)
Manganese (No. 17, 1981) 1-Propanol (No. 102, 1990)
Mercury (No. 1, 1976)a 2-Propanol (No. 103, 1990)
Mercury – environmental aspects Propachlor (No. 147, 1993)
(No. 86, 1989) Propylene oxide (No. 56, 1985)
Mercury, inorganic (No. 118, 1991) Pyrrolizidine alkaloids (No. 80, 1988)
Methanol (No. 196, 1997) Quintozene (No. 41, 1984)
Methomyl (No. 178, 1996) Quality management for chemical
2-Methoxyethanol, 2-ethoxyethanol, and safety testing (No. 141, 1992)
their acetates (No. 115, 1990) Radiofrequency and microwaves
Methyl bromide (No. 166, 1995) (No. 16, 1981)
Methylene chloride Radionuclides, selected (No. 25, 1983)
(No. 32, 1984, 1st edition) Reproduction, principles for evaluating
(No. 164, 1996, 2nd edition) health risks associated with exposure
Methyl ethyl ketone (No. 143, 1992) to chemicals (No. 225, 2001)
Methyl isobutyl ketone (No. 117, 1990) Resmethrins (No. 92, 1989)
Methylmercury (No. 101, 1990)
_______
a
Out of print
 THE ENVIRONMENTAL HEALTH CRITERIA SERIES (continued)

Selenium (No. 58, 1986) Toxicity of chemicals (Part 1), principles

Static fields (No. 232, 2006) and methods for evaluating the
Synthetic organic fibres, selected (No. 6, 1978)
(No. 151, 1993) Toxicokinetic studies, principles of
Styrene (No. 26, 1983) (No. 57, 1986)
Sulfur oxides and suspended particulate Transgenic animal mutagenicity assays
matter (No. 8, 1979) (No. 233, 2006)
Tecnazene (No. 42, 1984) Tributyl phosphate (No. 112, 1991)
Tetrabromobisphenol A and derivatives Tributyltin compounds (No. 116, 1990)
(No. 172, 1995) Trichlorfon (No. 132, 1992)
Tetrachloroethylene (No. 31, 1984) 1,1,1-Trichloroethane (No. 136, 1992)
Tetradifon (No. 67, 1986) Trichloroethylene (No. 50, 1985)
Tetramethrin (No. 98, 1990) Tricresyl phosphate (No. 110, 1990)
Thallium (No. 182, 1996) Triphenyl phosphate (No. 111, 1991)
Thiocarbamate pesticides: a general Tris- and bis(2,3-dibromopropyl)
introduction (No. 76, 1988) phosphate (No. 173, 1995)
Tin and organotin compounds Ultrasound (No. 22, 1982)
(No. 15, 1980) Ultraviolet radiation
Titanium (No. 24, 1982) (No. 14, 1979, 1st edition)
Tobacco use and exposure to other (No. 160, 1994, 2nd edition)
agents (No. 211, 1999) Vanadium (No. 81, 1988)
Toluene (No. 52, 1986) Vinyl chloride (No. 215, 1999)
Toluene diisocyanates (No. 75, 1987) Vinylidene chloride (No. 100, 1990)
White spirit (No. 187, 1996)
Xylenes (No. 190, 1997)
Zinc (No. 221, 2001)

CONCISE INTERNATIONAL CHEMICAL ASSESSMENT DOCUMENT SERIES

CICADs are IPCS risk assessment documents that provide concise but critical summaries of the
relevant scientific information concerning the potential effects of chemicals upon human health
and/or the environment

Acrolein (No. 43, 2002) Cobalt and inorganic cobalt compounds

Acrylonitrile (No. 39, 2002) (No. 69, 2006)
Arsine: human health aspects Crystalline silica, quartz (No. 24, 2000)
(No. 47, 2002) Cumene (No. 18, 1999)
Asphalt (bitumen) (No. 59, 2004) 1,2-Diaminoethane (No. 15, 1999)
Azodicarbonamide (No. 16, 1999) 3,3'-Dichlorobenzidine (No. 2, 1998)
Barium and barium compounds 1,2-Dichloroethane (No. 1, 1998)
(No. 33, 2001) 1,1-Dichloroethene (Vinylidene chloride)
Benzoic acid and sodium benzoate (No. 51, 2003)
(No. 26, 2000) 2,2-Dichloro-1,1,1-trifluoroethane
Benzyl butyl phthalate (No. 17, 1999) (HCFC-123) (No. 23, 2000)
Beryllium and beryllium compounds Diethyl phthalate (No. 52, 2003)
(No. 32, 2001) Diethylene glycol dimethyl
Biphenyl (No. 6, 1999) ether (No. 41, 2002)
Bromoethane (No. 42, 2002) Dimethylformamide, N,N- (No. 31, 2001)
1,3-Butadiene: human health aspects Diphenylmethane diisocyanate (MDI)
(No. 30, 2001) (No. 27, 2001)
2-Butenal (No. 74, 2008) Elemental mercury and inorganic mercury
2-Butoxyethanol (No. 10, 1998) compounds; human health aspects
2-Butoxyethanol (update) (No. 67, 2005) (No. 50, 2004)
Butyl acetates (No. 64, 2005) Ethylenediamine (No. 15, 1999)
Carbon disulfide (No. 46, 2002) Ethylene glycol: environmental aspects
Chloral hydrate (No. 25, 2000) (No. 22, 2000)
Chlorinated naphthalenes (No. 34, 2001) Ethylene glycol: human health aspects
Chlorine dioxide (No. 37, 2002) (No. 45, 2002))
4-Chloroaniline (No. 48, 2003) Ethylene oxide (No. 54, 2003)
Chlorobenzenes other than Formaldehyde (No. 40, 2002)
hexachlorobenzene: environmental 2-Furaldehyde (No. 21, 2000)
aspects (No. 60, 2004) Glyoxal (No. 57, 2004)
Chloroform (No. 58, 2004) Heptachlor (No. 70, 2006)
Coal tar creosote (No. 62, 2004)
CONCISE INTERNATIONAL CHEMICAL ASSESSMENT DOCUMENT SERIES
(continued)

HCFC-123 (No. 23, 2000) Phenylhydrazine (No. 19, 2000)

Hydrogen cyanide and cyanides: human N-Phenyl-1-naphthylamine (No. 9, 1998)
health aspects (No. 61, 2004) Polychlorinated biphenyls: human health
Hydrogen sulfide: human health aspects aspects (No. 55, 2003)
(No. 53, 2003) Resorcinol (No. 71, 2006)
Limonene (No. 5, 1998) Silver and silver compounds:
Manganese and its compounds environmental aspects (No. 44, 2002)
(No. 12, 1999) 1,1,2,2-Tetrachloroethane (No. 3, 1998)
Manganese and its compounds: Tetrachloroethene (No. 68, 2006)
environmental aspects (No. 63, 2004) 1,1,2,2-Tetrafluoroethane (No. 11, 1998)
Mercury, elemental, and inorganic Thiourea (No. 49, 2003)
mercury compounds: human health Tin and inorganic tin compounds
aspects (No. 50, 2003) (No. 65, 2005)
Methyl and ethyl cyanoacrylates (No. 36, o-Toluidine (No. 7, 1998)
2001) 2,4,6-Tribromophenol and other simple
Methyl chloride (No. 28, 2001) brominated phenols (No. 66, 2005)
Methyl methacrylate (No. 4, 1998) Tributylin oxide (No. 14, 1999)
N-Methyl-2-pyrrolidone (No. 35, 2001) 1,2,3-Trichloropropane (No. 56, 2003)
Methyltin, butyltin, and octyltin Triglycidyl isocyanurate (No. 8, 1998)
compounds, Mono- and disubstituted Triphenyltin compounds (No. 13, 1999)
(No. 73, 2006) Vanadium pentoxide and other inorganic
Mononitrophenols (No. 20, 2000) vanadium compounds (No. 29, 2001)
N-Nitrosodimethylamine (No. 38, 2002)

To order further copies of monographs in these series, please contact WHO Press,
World Health Organization, 1211 Geneva 27, Switzerland (Fax: +41-22-
791 4857; E-mail: [email protected])