Probabilistic Methods for Assessing
Dietary Exposure to Pesticides
Andy Hart (Fera, UK)
Technical Meeting with Stakeholders on Cumulative Risk Assessment
EFSA, 11 February 2014
Outline of presentation
p
• EFSA Guidance on probabilistic assessment
• Deterministic and probabilistic
p
• Variability and uncertainty
• Needs and challenges in probabilistic assessment
• How these are addressed by the EFSA Guidance
• Communicating results with Risk Managers
2
EFSA Guidance
Main sections of Guidance:
• Introduction
pp
• Tiered approach
• Problem definition
• Acute exposure
p
• Chronic exposure
• Cumulative exposure
• Outputs
• Evaluating uncertainties
• Key issues for reporting & peer review
• Interpretation of results & options for risk managers
• Validation
• Software quality requirements
• Conclusions
C
l i
• Appendix: Case studies
3
Deterministic vs. Probabilistic
• Real
R l exposures are variable
i bl and
d uncertain
t i
• Deterministic exposure assessment
– Uses p
point estimates for inputs
p
((consumption,
p
concentration, processing effects, etc.)
– Generates point estimate of exposure (a single value)
– Intended to be ‘conservative’
• Probabilistic exposure assessment
– Uses distributions to take account of variability and
uncertainty
t i t off inputs
i
t
– Generates a distribution of exposures
– Intended
I t d d to
t representt the
th reall distribution
di t ib ti
4
Cum
mulattive % of po
opulattion
Variability
y and uncertainty
y
100
Variability
of actual
exposures
80
60
40
20
0
0
20
40
60
Exposure mg/kg bw/day
80
100
5
Cum
mulattive % of po
opulattion
Variability
y and uncertainty
y
100
variability
of actual
exposures
80
60
40
Uncertainty in our knowledge of exposure
20
0
0
20
40
60
Exposure mg/kg bw/day
80
100
6
Cum
mulattive % of po
opulattion
Variability
y and uncertainty
y
100
actual
exposures
are unknown
80
60
40
uncertainty in our knowledge of exposure
20
0
0
20
40
60
Exposure mg/kg bw/day
80
100
7
Cum
mulattive % of po
opulattion
Variability
y and uncertainty
y
100
?
80
60
40
uncertainty in our knowledge of exposure
20
Deterministic point estimate
Based on conservative assumptions
0
0
20
40
60
Exposure mg/kg bw/day
80
100
8
A common view of probabilistic
assessment
• Option for refinement when deterministic estimate
raises concern
• More realistic estimates of exposure
• Avoids over
over-conservative
conservative first tier assumptions
• Load my data into available software & press ‘go’
go
• Compare
p
95th p
percentile exposure
p
to reference
dose
9
Cum
mulattive % of po
opulattion
100
5% of EU population = 25 million
80
Pesticides: ‘shall not have any harmful
effects on human health’ Reg. 1107/2009
60
Need to characterise the frequency
g
of upper tail exposures
& magnitude
40
20
95th percentile exposure
compare to reference dose
?
0
0
20
40
60
Exposure mg/kg bw/day
80
100
10
Cum
mulattive % of po
opulattion
100
Pesticides: ‘shall not have any harmful
effects on human health’ Reg. 1107/2009
80
60
A high level of certainty is implied...
need to show how much certainty there is
40
Need confidence intervals for the
distribution of exposures
20
0
0
20
40
60
Exposure mg/kg bw/day
80
100
11
Cum
mulattive % of po
opulattion
100
80
Many uncertainties are difficult to quantify
probabilistically,
b bili ti ll e.g. ffor pesticides:
ti id
60
•
•
•
•
40
limited samples of data
distribution shapes uncertain (esp.
(esp tails)
many non-detects
dependencies between parameters
Use alternative assumptions to explore
the impact of difficult uncertainties
20
0
0
20
40
60
Exposure mg/kg bw/day
80
100
12
Compare to Reference Dose:
stop
Further refinement required
stop
Cum
mulattive % of po
opulattion
Optimistic Pessimistic
100
80
Many uncertainties are difficult to quantify
probabilistically,
b bili ti ll e.g. ffor pesticides:
ti id
60
•
•
•
•
40
limited samples of data
distribution shapes uncertain (esp.
(esp tails)
many non-detects
dependencies between parameters
Use alternative assumptions to explore
the impact of difficult uncertainties
20
0
0
20
40
60
Exposure mg/kg bw/day
80
100
13
Cum
mulattive % of po
opulattion
Optimistic Pessimistic
100
80
?
?
But... it is never possible to quantify
all uncertainties
60
40
Need to identify
y and evaluate
unquantified uncertainties
20
0
0
20
40
60
Exposure mg/kg bw/day
80
100
14
Cum
mulattive % of po
opulattion
Optimistic Pessimistic
100
Parametric modelling
g may
yg
generate
unrealistically high values
80
Empirical bootstrap will underestimate upper tail
60
Statistical methods may
y over- or
underestimate upper tails
40
Need ‘drill-down’ examination of
estimates in upper tail
20
0
0
20
40
60
Exposure mg/kg bw/day
80
100
15
Key needs in probabilistic
assessment
• Focus on characterising upper tail exposures,
yp
percentile of the distribution
not on an arbitrary
• Give confidence intervals for quantified
uncertainties
• Use alternative assumptions to assess
uncertainties that are difficult to quantify
probabilistically
b bili ti ll
• Evaluate the impact of unquantified uncertainties
• ‘Drill-down’ to check for over- and
underestimation in upper tail
16
PPR Panel Guidance
• How does it address the needs identified above?
17
PPR Panel Guidance
• Use alternative assumptions to assess uncertainties
that are difficult to quantify probabilistically
• Start with a Basic Assessment using
g Optimistic
p
and
Pessimistic assumptions for:
–
–
–
–
–
–
–
–
Residue distribution (bootstrap vs. lognormal)
S
Sampling
li uncertainty
t i t (b
(bootstrap
t t
vs. parametric)
ti )
Unit-to-unit variability (none vs. conservative est.)
Non detects (zero vs.
Non-detects
vs Limit of Reporting)
Processing effects (zero vs. deterministic estimate)
% crop treated (approx. estimate vs. 100%)
Residues in water (zero, legal limit)
etc...
• If important, quantify further in Refined Assessment
18
PPR Panel Guidance
• Focus on characterising upper tail, not an arbitrary percentile
• Give confidence intervals for q
quantified uncertainties
Exposure
levels
expressed
as % of
reference
dose and/or
Margin of
Exposure
Frequency of
exceedance
per million
Optimistic &
P
Pessimistic
i i ti
assumptions
Estimated
frequencies
& confidence
intervals
‘<‘ = below
resolution of
model
19
PPR Panel Guidance
• Evaluate the impact of unquantified uncertainties
• ‘Drill-down’ to check for over- and underestimation in upper
pp tail
Evaluation of
unquantified
uncertainties
Findings
f
from
drilld ill
down check
of tail results
20
PPR Panel Guidance
• Evaluate the impact of unquantified uncertainties
21
Cumulative exposure
p
• Everything so far applies equally to both single
substance and cumulative assessments
• In addition, for cumulative assessments:
– Use Relative Potency Factors to combine exposure
contributions of different substances in CAG
– For acute exposure
exposure, take account of correlations
between concentrations of different substances in
same food type
• Use statistical models to fill gaps in residue database
Section 6 of Guidance
22
An additional challenge...
g
• Communication and interpretation of results
• Exposure assessment characterises the upper tail:
– Person(-day)s
( y) p
per million exceeding
g effect level
– Confidence intervals for quantified uncertainties
– Potential impact of unquantified uncertainties
• Question for toxicologists:
– What is the likelihood & severity of effects for these
upper tail exposures?
• Question for risk managers:
– Are
A results
lt consistent
i t t with
ith ‘not
‘ t any harmful
h
f l effects’
ff t ’ ?
23
Summary
y
• Focus
F
on characterising
h
t i i upper ttailil exposures, nott on an
y percentile of the distribution
arbitrary
• Give confidence intervals for quantified uncertainties
• Use alternative assumptions to assess uncertainties that
are difficult to quantify probabilistically
• Evaluate
E l t the
th impact
i
t off unquantified
tifi d uncertainties
t i ti
• ‘Drill
Drill-down
down’ to check for over
over- & underestimation in tail
• Communicate and interpret results
24