Modelling cumulative risk
Hilko van der Voet
Biometris, DLO, Wageningen University and Research Centre
Third ACROPOLIS consortium meeting
31 March 2011, Milano
Contents
• State of the art for modelling of single pesticides
– Exposure assessment
– Risk assessment
• ACROPOLIS: modelling for multiple pesticides in
a Common Assessment Group
– Cumulative exposure assessment
– Cumulative risk assessment
Risk assessment
integrated risk
assessment
A.G. Renwick et al. (2003)
Exposure Assessment
MCRA
Individual Margin of Exposure
1
– MCRA 7 calculates exposure
distributions for single compounds
– Percentiles and number of people
exceeding a limit value (e.g. ARfD)
– Acute or chronic risks
– Processing factors, variability factors,
modelling of non-detects, covariates, ...
– Drill-downs
– Uncertainty analysis
10
100
1000
10000
Risk Assessment: the IPRA
model
van der Voet & Slob (2007), Risk Analysis 27: 351-371
Individual Margin of Exposure
1
10
100
1000
10000
Individual Margin of Exposure
 Exposure assessment and hazard characterisation combined into an
integrated probabilistic model (IPRA)
 Margin of Exposure replaced by Individual Margin of Exposure (IMoE)
 Analysis of variability and uncertainty kept separate
 Proposed instruments for risk managers:
IMoE safety bar, IMoEp1 and/or IMoEL
Individual Margin of Exposure
1
10
IMoEL
100
1000
10000
IMoEp1
Van der Voet et al. (2009)
Example: Comparison of risks
• Decisions of fungicide use are an example of risk-benefit analysis
– Fungicides may have toxic effects (hazard)
– Fungicides may reduce risk of mycotoxin production (benefit)
Individual Margin of Exposure
1
10
100
1000
10000
100000
1000000
10000000
5% effect on BW from
mycotoxin
5% effect on
erythrocyte count from
fungicide
50% cases of
hepatocytomegaly from
fungicide
Muri et al. (2009)
Cumulative assessments
• Common Assessment Groups refer to
multiple compounds with for the purpose
of the assessment will be assumed to
have the same health effect
• Potency differences are captured in
Relative Potency Factors (RPFs)
– Estimated from data
– Therefore RPF estimates will be not exactly
known but uncertain
•
Estimating RPFs from doseExample Organophosphates
et al. 2009)
response(Bosgra
data
– Dose-response data EPA
– Parallel curves fitted by PROAST
Probabilistic models cumulative
exposure
• It is important to describe the variation
between persons (Person Oriented
Models) in the relevant population
• Which population is used?
– Models with predefined populations or
subpopulations thereof: e.g. US models
DEEM/Calendex, LifeLine, CARES, SHEDS
– Model applicable to user-defined populations:
Acropolis model based on MCRA
Data for cumulative exposure
• Consumption data: national survey data
• Residue data: need to collect at the level of
individual samples so that correlations between
pesticides are represented
– use of pesticides A and B may be exclusive
– or they may be used always together
– or anything in between ...
• Problem: residue data matrix contain many
missing values (MVs) and non-detects (NDs)
Cumulative exposure: residue
data
non-detect
(< 0.05)
missing value
(nonmeasurement)
positive
value
Cumulative exposure
assessment
In the EFSA triazole project (van Klaveren et al.
2009) two approaches for cumulative exposure
assessment using single-residue modelling
methods were compared:
1. First add, then analyse


Calculate RPF-weighted sum of concentrations per sample
then exposure assessment for ‘single’ compound
2. First analyse, then add


Parallel exposure assessment runs for the compounds
then RPF-weighted summing of intakes using same sequence
of simulated consumers
Approach 1: First add, then
analyse
Assumes that the total set of samples is representative for a
food
Advantage:
– incorporates correlations between compounds
• negative correlation: lower exposure
• positive correlation: higher exposure
• Disadvantage:
– requires data for all compounds in all samples
• for non-measured compounds effectively a concentration 0 is assumed
• estimated exposure may be too low
Approach 2: First analyse, then
add
Assumes that per compound the set of samples
with measurements is representative for a food
• Advantage:
– each compound may have its own set of samples
• Disadvantage:
– does not incorporates correlations between
compounds
Example triazoles
van Klaveren et al. (2009)
• Netherlands: not much difference
– most samples were analysed for most triazoles
• France: Approach 2 more conservative
– many samples analysed for only part of triazoles
ACROPOLIS approach
• Combine advantages of Approaches 1 and 2 by
– Fitting a multivariate model to the combined residue data
– Allow for patterns of missing information
– Allow for measurements below a Limit of reporting (non-detects)
• Detailed models are under investigation
– Correlation between pesticides may exist
• Regarding the use frequencies
• Regarding the resulting concentrations
– We know fairly certain that each pesticide is only used in a
fraction of cases, so there must be many ‘true zeroes’
– Some models may allow the use of additional data from
Pesticide Usage Surveys
FERA PUS data. Example : Wheat
(GB, 2008)
1400
•
800
•
600
•
200
400
•
•
0
Number of fields
1000
•
1200
Wheat
1
2
3
4
Number of triazoles per field
5
6
Proportion of wheat fields treated
with a triazole is 0.95
12 different triazoles are used for
wheat in GB
111 different combinations of up to
6 triazoles used
Most fields use a combination of
2 or 3 triazoles
Only 25 fields were treated with 6
different triazoles
Conclusion: many of the nondetects and missing values must
be true zeroes
Example : Wheat GB, 2008
(FERA)
Wheat Overall Usage
250
•
150
•
100
Triadimenol
Triadimenol
Tetraconazole
Tetraconazole
Tebuconazole
Tebuconazole
Prothioconazole
Prothioconazole
Propiconazole
Propiconazole
Metconazole
Metconazole
Flutriafol
Flutriafol
Flusilazole
Flusilazole
Fluquinconazole
Fluquinconazole
Epoxiconazole
Epoxiconazole
Cyproconazole
Cyproconazole
0
Prothioconazole is applied
most (in total 272.88 kg/ha)
either individually (2 fields)
or in combination (1465
fields)
Prothioconazole used in GB
but not in The Netherlands
•
Suggests GB data for
wheat may not be
appropriate to make
assumptions for some
countries in Europe
•
Data available for other
countries?
50
Bromuconazole
Bromuconazole
Total triazole applied (kg/ha)
200
Example modelling of
correlation
simulated from
bivariate normal
distribution,
means 3 and 7
sds 2 and 3
correlation 0.8
Which distributions are
appropriate?
• We need a statistical model for cumulative
exposure
• Options:
– multivariate lognormal (convenient)
– Mixture of true zeroes and lognormal
– other parametric or non-parametric
multivariate distributions
Uncertainty approaches
• Uncertainty about inputs and model form
 uncertainty about quantities of interest
– e.g. fraction of population exceeding a limit value
• Sources of information on uncertainty
– Data, e.g. implicit in small sample or s.e. from literature
– Expert judgment (needs ‘elicitation’)
• Main approaches to address uncertainty:
– modelling based on available data or expert judgment
– qualitative assessment of uncertainties by experts, summarized
in uncertainty tables
Quantitative and qualitative
approaches
Uncertainty PoCE
35
% contribution
30
25
20
15
10
5
0
MC
cons
conc
proc
anim al
Uncertainty source
inter
intra
Updated view on data
needed for cumulative
assessments
• Consumption survey data
• Residue monitoring data or field trial data (preregistration)
• Food conversion (linking food as eaten to food
as measured)
• Data on processing, unit variability
• Pesticide usage data
• Dose response data for critical health effect to
estimate RPFs (or for direct use)
Cumulative Risk Assessment
• For integrating exposure assessment and hazard
characterisation two approaches are possible:
– Two-step approach:
• First, perform cumulative exposure assessment using RPFweighted sum
• Secondly, calculate MoE or IMoE distribution using
toxicology data for the index compound
• Examples in Bosgra et al. (2009), Müller et al. (2009)
– One-step approach:
• single-pesticide IMoE distributions from a cumulative IPRA
analysis can be directly combined into a cumulative IMoE
distribution (for details see van der Voet et al. 2009)
• This would circumvent the explicit calculation of RPFs
Conclusions
• Modelling cumulative exposure and risk already
possible, further developed in ACROPOLIS
• Patterns and amount of missing values and nondetects may be a problem
• Pesticide usage survey data may be useful
• Future: Integrated models may replace separate
estimation of RPF and use of RPF models
• ACROPOLIS system: bring many data together
in one platform, accessible to all stakeholders