probabilistic methods for assessing dietary exposure to ... · probabilistic methods for assessing...

Probabilistic Methods for AssessingDietary Exposure to Pesticides

Andy Hart (Fera, UK)

Technical Meeting with Stakeholders on Cumulative Risk AssessmentEFSA, 11 February 2014

Outline of presentationp

• EFSA Guidance on probabilistic assessment• EFSA Guidance on probabilistic assessment

• Deterministic and probabilisticp

• 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:Main sections of Guidance:• Introduction• Tiered approachpp• Problem definition• Acute exposurep• Chronic exposure• Cumulative exposure• Outputs• Evaluating uncertainties• Key issues for reporting & peer review• Interpretation of results & options for risk managers• Validation• Software quality requirements

C l i• Conclusions• Appendix: Case studies 3

Deterministic vs. Probabilistic

R l i bl d t i• Real exposures are variable and uncertain

• Deterministic exposure assessment– Uses point estimates for inputs (consumption, p p ( 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

t i t f i tuncertainty of inputs– Generates a distribution of exposures

I t d d t t th l di t ib ti– Intended to represent the real distribution

4

Variability and uncertaintyy y

100

tion

80

opul

at Variabilityof actualexposures

60

% o

f po exposures

40

tive

%

20

mul

at

00 20 40 60 80 100

Cum

5

0 20 40 60 80 100Exposure mg/kg bw/day


100

tion

80

opul

at variabilityof actual exposures

60

% o

f po exposures

40

tive

%

Uncertainty in our knowledge of exposure

20

mul

at

00 20 40 60 80 100

Cum

6



100

tion

actual 80

opul

at exposures are unknown

60

% o

f po

40

tive

%

uncertainty in our knowledge of exposure

20

mul

at

00 20 40 60 80 100

Cum

7



100

tion ?

80

opul

at

60

% o

f po

40

tive

%

uncertainty in our knowledge of exposure

20

mul

at

Deterministic point estimate

00 20 40 60 80 100

Cum Deterministic point estimate

Based on conservative assumptions

8


A common view of probabilistic assessmentassessment

• Option for refinement when deterministic estimate raises concernraises concern

• More realistic estimates of exposureMore realistic estimates of exposure

• Avoids over-conservative first tier assumptionsAvoids over conservative first tier assumptions

• Load my data into available software & press ‘go’Load my data into available software & press go

• Compare 95th percentile exposure to reference p p pdose

9

100

tion 5% of EU population = 25 million

80

opul

at Pesticides: ‘shall not have any harmful effects on human health’ Reg. 1107/2009

60

% o

f po

40

tive

% Need to characterise the frequency & magnitude of upper tail exposures

20

mul

at

95th percentile exposure

g

?0

0 20 40 60 80 100

Cum compare to reference dose ?

10


100

tion

80

opul

at Pesticides: ‘shall not have any harmful effects on human health’ Reg. 1107/2009

60

% o

f po

40

tive

% A high level of certainty is implied...need to show how much certainty there is

20

mul

at

Need confidence intervals for the distribution of exposures

00 20 40 60 80 100

Cum distribution of exposures

11


100

tion

80

opul

at

Many uncertainties are difficult to quantify b bili ti ll f ti id

60

% o

f po probabilistically, e.g. for pesticides:

• limited samples of data• distribution shapes uncertain (esp tails)

40

tive

% • distribution shapes uncertain (esp. tails)• many non-detects• dependencies between parameters

20

mul

at

Use alternative assumptions to explore the impact of difficult uncertainties

00 20 40 60 80 100

Cum the impact of difficult uncertainties

12


Compare to Reference Dose:

Optimistic Pessimistic

Further refinement required stopstop

100

tion


80

opul

at

Many uncertainties are difficult to quantify b bili ti ll f ti id

60

% o

f po probabilistically, e.g. for pesticides:

• limited samples of data• distribution shapes uncertain (esp tails)

40

tive

% • distribution shapes uncertain (esp. tails)• many non-detects• dependencies between parameters

20

mul

at

Use alternative assumptions to explore the impact of difficult uncertainties

00 20 40 60 80 100

Cum the impact of difficult uncertainties

13


Optimistic Pessimistic100

tion


80

opul

at ? ?

60

% o

f po

But... it is never possible to quantify all uncertainties

40

tive

%

Need to identify and evaluate20

mul

at

yunquantified uncertainties

00 20 40 60 80 100

Cum

14


Optimistic Pessimistic100

tion


Parametric modelling may generate

80

opul

at

Empirical bootstrap will underestimate upper tail

g y gunrealistically high values

60

% o

f po

Statistical methods may over- or 40

tive

%

yunderestimate upper tails

20

mul

at Need ‘drill-down’ examination of estimates in upper tail

00 20 40 60 80 100

Cum

15


Key needs in probabilistic assessmentassessment

• Focus on characterising upper tail exposures, not on an arbitrary percentile of the distributiony p

• Give confidence intervals for quantifieduncertainties

• Use alternative assumptions to assess• Use alternative assumptions to assess uncertainties that are difficult to quantify

b bili ti llprobabilistically

• Evaluate the impact of unquantified uncertainties• 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 Optimistic and

that are difficult to quantify probabilistically

g pPessimistic assumptions for:– Residue distribution (bootstrap vs. lognormal)

S li t i t (b t t t i )– Sampling uncertainty (bootstrap vs. parametric)– Unit-to-unit variability (none vs. conservative est.)

Non detects (zero vs Limit of Reporting)– Non-detects (zero vs. Limit of Reporting)– Processing effects (zero vs. deterministic estimate)– % crop treated (approx. estimate vs. 100%)% crop treated (approx. estimate vs. 100%)– Residues in water (zero, legal limit)– etc...

• If important, quantify further in Refined Assessment18

PPR Panel Guidance

• Focus on characterising upper tail, not an arbitrary percentile• Give confidence intervals for quantified uncertaintiesq

ExposureFrequency of exceedanceExposure

levels expressed

as % of

per million

Optimistic & P i i tireference

dose and/or Margin of Exposure

Pessimistic assumptions

ExposureEstimated

frequencies & confidence

intervalsintervals

‘<‘ = below resolution of

19

resolution of model

PPR Panel Guidance

• Evaluate the impact of unquantified uncertainties• ‘Drill-down’ to check for over- and underestimation in upper tailpp

Evaluation ofEvaluation of unquantifieduncertainties

Findings f d illfrom drill-

down check of tail results

2020

PPR Panel Guidance

• Evaluate the impact of unquantified uncertainties

21

Cumulative exposurep

• Everything so far applies equally to both single substance and cumulative assessmentssubstance and cumulative assessments

• In addition, for cumulative assessments:– Use Relative Potency Factors to combine exposureUse Relative Potency Factors to combine exposure

contributions of different substances in CAG– For acute exposure take account of correlations– For acute exposure, take account of correlations

between concentrations of different substances in same food typesame food type

• Use statistical models to fill gaps in residue database

22Section 6 of Guidance

An additional challenge...g

• Communication and interpretation of results• Exposure assessment characterises the upper tail:

– Person(-day)s per million exceeding effect level( y) p g– Confidence intervals for quantified uncertainties– Potential impact of unquantified uncertaintiesPotential impact of unquantified uncertainties

• Question for toxicologists:– What is the likelihood & severity of effects for these

upper tail exposures?

• Question for risk managers:A lt i t t ith ‘ t h f l ff t ’ ?– Are results consistent with ‘not any harmful effects’ ?

23

Summaryy

F h t i i t il t• Focus on characterising upper tail exposures, not on an arbitrary percentile of the distributiony

• Give confidence intervals for quantified uncertainties

• Use alternative assumptions to assess uncertainties that are difficult to quantify probabilistically

E l t th i t f tifi d t i ti• Evaluate the impact of unquantified uncertainties

• ‘Drill down’ to check for over & underestimation in tail• Drill-down to check for over- & underestimation in tail

• Communicate and interpret resultsCommunicate and interpret results

24

probabilistic methods for assessing dietary exposure to ... · probabilistic methods for assessing...

Documents