No Slide Title

Download Report

Transcript No Slide Title

Categorical Regression as a
Predictive Tool for Determining Risks at
Doses above the Reference Dose (RfD)
Richard Hertzberg
Biomathematics Consulting
Atlanta, GA
Beyond Science and Decisions
From Issue Identification to Dose-Response Assessment
Austin, TX
March 16, 2010
1
Overview
• Goal: Estimate risk for dose > RfD
• Why RfD and BMD will not work
• How categorical regression works
• Pros and Cons
• Future
2
The RfD is Limited
(pun intended)
“If it's zero degrees outside today and it's supposed to
be twice as cold tomorrow, how cold is it going to be?”
(ba-da-bum)
• For lower doses, the RfD does not inform us of risk:
Safe is still safe.
3
Why Severity Categories?
• Reference Dose Limitations
–Bounding value: minimal to no risk for lower doses
–Benchmark dose based on modeling the critical effect
–Other effects not included in calculation of RfD
• Doses higher than the RfD
–Cannot estimate risk except of the critical effect
–Need dose-response information on all “toxic effects”
• Multiple responses and measures (need lots of data)
–Such information is rarely in any single study
4
Desirable Effects Information
• Critical Effect—effect observed at the lowest dose
–Do NOT want: P(critical effect | dose>RfD)
• Secondary Effects—observed at higher doses, also includes:
–Effects mediated by chemical metabolites
–Effects that are not adverse (e.g., enzyme induction)
• Variations in Effects (e.g., from chemical mixture exposures)
–Co-occurring effects might be worse than any by itself
• Usually require toxicologist’s judgment on severity
–Mixture Issues
• Joint toxic action may occur: dose- or response-additivity;
toxicological interactions (e.g., synergism, antagonism)
5
What Are We Modeling?
RfD: An estimate (with uncertainty spanning perhaps an order of
magnitude) of a daily exposure to the human population (including
sensitive subgroups) that is likely to be without an appreciable
risk of deleterious effects during a lifetime.
sensitive subgroups
be conservative
without an appreciable risk
r < 0.01 (?)
likely
P( ... ) > 0.95 (?)
deleterious effects
“adverse effects”
P( "risk of adverse effect" < 0.01 | dose<RfD) > 0.95
6
What Are We Modeling?
P( “adverse effect” | dose=D)
for this presentation,
D is a dose > RfD
7
How Categorical Regression Works
• Meta-analysis of exposure-response data
– A method for combining health effects data across studies,
endpoints, exposure durations and species
• Basic approach to modeling (e.g., Dourson et al., 1997)
– Using toxicological judgment, each dose group (or individual animal)
of every study is assigned to a severity category
– Using link function, e.g. logistic regression, severity categories are
regressed on dose (and possibly duration)
– Models predict the probability that an effect severity will be
observed, given dose
• Example also of five pesticides (Teuschler et al., 1999)
– Toxicological effects data from multiple bioassays
– Results compared with RfD
8
Categorical Regression Model
• Logistic model for the i th severity category
• The Probability that severity is less than or equal to level i,
given dose is expressed as:
P s  i  
expi   * log dose
1  expi   * log dose
where
s = severity
i = an effect level (1=NOEL/NOAEL, 2=AEL, 3=FEL)
αi = an intercept term for level i
β = a slope factor related to dose
9
Example 1. Frequency of Effect Categories for
Aldicarb Exposure in Humans*
Study
Haines,
1971
Wyld et al.,
1992
Dose
(mg/kg/d)
Group
Size
Frequency of Responders within Categories of**
No
Effects
Nonadverse
Effects
Adverse
Effects
Frank
Effects
0.025
4
0
0
4
0
0.050
4
0
0
4
0
0.10
4
0
0
2
2
0.0
22
22
0
0
0
0.010
8
8
0
0
0
0.025
12
2
9
1
0
0.050
12
0
9
3
0
0.075
4
0
0
4
0
*Source: Dourson et al. 1997. Categorical Regression of Toxicity Data, A Case
Study Using Aldicarb. Reg. Toxicology and Pharmacology, 25:121-129
**Numbers reflect a judgment that whole blood (Haines, 1971) or red blood cell
(Wyld et al., 1992) cholinesterase inhibition of 20% or greater is considered an
adverse effect. This percentage can be debated and is a source of uncertainty.
10
Example Severity Assignments for Human
Health Effects from Aldicarb Exposures
Category
4 - Frank
Effects
Site
Effect
Cholinergic effects
Severe abdominal pain, nausea and/or
vomiting, diarrhea, seizures, severe
disorientation or confusion, excitation
Whole Body
Mortality
Brain, whole blood or red blood
cell acetylcholinesterase
Inhibition (e.g., of 20% or greater)
Cholinergic effects
Mild: Muscular weakness or twitching,
blurred vision and/or watery eyes,
pinpoint pupils, excess salivation,
sweating or clamminess
Nervous system
Hyperactivity or altered patterns of
locomotion
2 - NonAdverse
Effects
Plasma, whole blood or red
blood cell acetylcholinesterase
Inhibition (e.g., observed, but less than
20%)
1 – No
Effects
All
No effect
3Adverse
Effects
11
(Adapted from Dourson et al. ,1997)
Aldicarb: Multiple Severities
Information for the risk manager!
12
Example 2. Health Effects of Concern
for 5 Pesticides*
Chemical
Primary Effect
Secondary Effects
Diazinon
Cholinesterase inhibition Nervous system disturbance
Disulfoton
Cholinesterase inhibition Myopia
Optic nerve degeneration
Impaired reproductive performance
Impaired fetal development
EPTC
Degenerative
cardiomyopathy
Muscle and nerve degeneration
Fenamiphos Cholinesterase inhibition Maternal and developmental toxicity
Lindane
13
Liver, kidney toxicity
Immunosuppressive effects
Decreased adrenocortical function
Reproductive effects
Fetal viability
*Source: Teuschler et al. 1999. Health Risk Above the Reference Dose for Multiple
Chemicals. Reg. Tox. And Pharm. 30:S19-S26.
Animal Study Data Records Modeled
Using a 3 Category Model
Chemical
No. of Data
Records
Modeled
Percentage of Dose Groups in
Each Severity Category*
1- NOAEL or
NOEL
2 - AEL
3 - FEL
Diazinon
77
25
62
13
Disulfoton
47
17
70
13
EPTC
58
50
45
5
Fenamiphos
44
32
50
18
Lindane
99
47
32
21
*NOAEL = No Observed Adverse Effect Level,
AEL = Adverse Effect Level, FEL = Frank Effect Level
14
Pros and Cons
• Advantages:
–Provides a consistent basis for calculating risk above the RfD
–Can use available data, even marginal studies and dose
group level information
–Accounts for severity of toxic effect by combining studies on
multiple effects
• Limitations:
–Animal to human extrapolation is still needed
–Data are transformed into categories, losing information
• Cannot track toxic MOA progression with increasing dose
–More tox judgment needed than merely NOAEL vs AEL
15
Closer to Goal of
P(toxicity | dose>RfD)
RfD & BMD
clear, off-target
16
Categorical Regression
more on target, but fuzzy
DDT: oral, many species
17
More Pros and Cons
• Is this human RISK estimation?
–If includes human incidence data, then YES
–If human data are on dose groups, then Not Exactly
–If animal data, then No
• Not Exactly? No?
–Dose group data: risk=P(dose group shows toxicity)
–Animal data: unknown relevance to human probability
(tolerance) distribution (same with BMD)
–But, in either case: “risk” can inform regulatory decisions
18
Categorical Regression
Modeling Results
Chemical
Chi2
Test*
Diazinon
Disulfoton
1
0.6
EPTC
Fenamiphos
Lindane
0.3
0.1
0.2
Parameter Results
α1
α2
β
-8.1
5.1
-4.7
-9.9
-0.9
-5.6
3.8
9.4
-5.7
-7.2
-0.6
-5.1
-0.1
1.8
-1.7
P(Adverse or Frank Effect) is equal to 1 – P(s < 1), i.e., one minus the probability of
observing a nonadverse or no effect.
*Tests proportional odds assumption of common slope parameter across categories
Source: Teuschler and Hertzberg, 2008, presented at SRA.
19
Reference Doses (RfDs) and All
Effects Toxicity Doses (AETDs)
Chemical
RfD*
mg/kg/d
Animal Study
RfD**
NOAEL
UF
mg/kg/d
Diazinon
0.00009
100
Disulfoton
0.00004
EPTC
Human
AETD***
mg/kg/d
ED05
UF
Human
Equivalent
ED05
mg/kg/d
0.009
0.0005
10
0.005
1000
0.8
0.0005
10
0.005
0.025
100
2.5
0.145
10
1.45
Fenamiphos
0.00025
100
0.025
0.001
10
0.01
Lindane
0.0003
1000
0.33
0.0017
10
0.017
Meta-analysis
used multiple
durations
andwhich
species
*RfDs from IRIS (accessed
2008), except
for Diazinon
was derived in
Teuschler
et al.,
1999those UFs
so
did not
need
**Uncertainty Factors (UF) = 10 for Interspecies, 10 for Intraspecies, and 10
for subchronic to chronic (Lindane), 10 for LOAEL to NOAEL (Disulfoton)
***AETD = ED05 from categorical regression, UF of 10 for intraspecies
20
Future?
• Multiple effect ED’s appear to provide a less conservative
screen for risk than RfD-based ED’s
• More data = better modeling
–Duration influence
–‘omics data for precursors and adverse effects
–Improved species conversions
• Uncertainties need to be considered and discussed
–Expert judgment used in identifying severity categories
–Interpretation of using dose group data
21
Acknowledgements
• Linda Teuschler, US EPA
• Mike Dourson and Lynne Haber, Toxicology Excellence for
Risk Assessment
• Bill Stiteler, Syracuse Research Corporation
22