an introduction to bayesisan decision analysis
TRANSCRIPT
Enterprise EHS Software Solutions
Monica Melkonian, MS, CIHProduct Manager- Industrial
Hygiene
- THE BASICS OF BDA METHODOLOGY- BDA DECISION CHARTS- BDA AND THE AIHA EXPOSURE MODEL- QUANTIFYING PROFESSIONAL JUDGMENT
Paul Hewett, owner, Exposure Assessment Solutions Inc
Enterprise EHS Software Solutions
ADVANTAGES OF BDA• Can set a plausible parameter space
• Output is a set of Decision Charts
• Can incorporate Professional Judgment
• Best applied to small datasets
• Provides feedback
• Consistent with …AIHA Exposure Banding Model
EU Control Banding Model
pharmaceutical Control Banding or PB-OEL Models
• Can be applied to datasets that contain non-detects
3
Enterprise EHS Software Solutions
AIHA EXPOSURE ASSESSMENTStart
Basic Characterization
Exposure Assessment
Unacceptable Exposure
UncertainAcceptable Exposure
Reassessment
ControlFurther Information
Gathering
The exposure assessment and management program of most companies is based on the AIHA approach.
Category 0, 1, 2 or 3
Category 4 or 5
Enterprise EHS Software Solutions
AIHA EXPOSURE CONTROL BANDING MODEL
Exposure Control Category Cutoff (%OEL) Confidence level
0 X0.95 < 1%
High
Medium
Low
1 1% < X0.95 < 10%
2 10% < X0.95 < 50%
3 50% < X0.95 < 100%
4 X0.95 > 100%
5
Enterprise EHS Software Solutions
TYPICAL ACTIONS
Exposure Control Category Recommended Actions
0 (<1% of OEL) No action
1 (<10% of OEL) general HazCom
2 (10-50% of OEL) + chemical specific HazCom
3 (50-100% of OEL) + exposure surveillance, medical surveillance, work practice analysis
4 (>100% of OEL) + respirators & engineering controls, work practice controls, validate respirator selection
Multiples of OEL (e.g., based on respirator APFs)
+ immediate engineering controls or process shutdown, validate respirator selection
6
Enterprise EHS Software Solutions
• The goal of an exposure assessment is to select the exposure category that most likely contains the true 95th percentile.
• The resulting “actions” will then be risk-based …
that is, proportional to the expected degree of risk,
as indicated by the exposure category.
• How do we pick the correct exposure category and certainty level?
7
EXPOSURE ASSESSMENT GOAL
Enterprise EHS Software Solutions
THE “DATA DRIVEN” APPROACH
• Collect a large n dataset.
• Calculate the “sample” 95th percentile (X0.95)
plus the 95%LCL and 95%UCL.
• Compare sample X0.95 to the category cutoffs
Which category does it fall in?
• Determine a certainty level.
Compare LCL and UCL to category cutoffs.
• With large n little to no judgment is necessary. Nearly all IHs will reach the same decision.
8
Enterprise EHS Software Solutions
EXAMPLE
• An IH collected 17 representative measurements and calculated standard statistics (OEL=1):
gm = 0.12 and gsd = 2.0
x0.95 = 0.38
38% of the OEL; suggests a Category 2 exposure profile
95%LCL = 0.26 and 95%UCL = 0.67
Category 3 is also a possibility. Categories 1 and 4 can be eliminated from consideration.
Final Rating : Category 2, Medium Certainty
9
Enterprise EHS Software Solutions
How do we pick a category and certainty level… when we have only a few measurements?
• n=1; x = {0.05} (OEL=1)
• n=2; x = {0.05, 0.2}
X0.95 = 0.50 ( 95%LCL=0.16, 95%UCL=1.5x1010 )
• n=3; x = {0.05, 0.2, 0.1}
X0.95 = 0.31 ( 95%LCL=0.16, 95%UCL=20 )
• Which category: 0, 1, 2, 3, or 4 ?
10
PICKING A CATEGORY
Enterprise EHS Software Solutions
PICKING A CATEGORY
• x = {<0.1, 0.2, <0.1} (OEL=1)
X0.95 = ?
• x = {<0.05, 0.2}
X0.95 = ?
• x = {<0.05} or {<0.25} or {<0.75}
X0.95 = ?
• Which category: 0, 1, 2, 3, or 4 ?11
…when some or most of the measurements are non-detects?
Enterprise EHS Software Solutions
BAYESIAN DECISION ANALYSIS (BDA)
• An adjunct to the calculation and interpretation of traditional statistics.
• The goal of BDA is to estimate the probability that the true 95th percentile exposure falls within a particular category, or Exposure Rating.
• BDA can explicitly incorporate professional judgment.
12
Enterprise EHS Software Solutions 13
ADVANTAGES
• Best suited for small datasets
• Can handle non-detects
• Professional judgment can be incorporated
• Encourages the improvement of professional judgment
• The output is a set of easy to interpret “decision charts”
13x = {0.05, 0.2, 0.1}
Enterprise EHS Software Solutions
BDA DECISION CHARTS
• Prior decision chart Represents our professional judgment
regarding the probability of each of the five Exposure Ratings.
• Likelihood decision chart The set of probabilities of each exposure
rating calculated using only the collected data.
• Posterior decision chart The set of probabilities of each exposure
rating, updated using the information from the prior decision chart.
x = {0.05, 0.2, 0.1}
Enterprise EHS Software Solutions
• In order to apply Bayesian analysis to industrial hygiene “decision making” we need the following:
A model for classifying occupational exposure profiles into exposure categories.e.g., the AIHA Exposure Control Banding Model
A distributional modele.g., the lognormal model
A decision statistice.g., the 95th percentile
15
BDA DECISION MAKING
Enterprise EHS Software Solutions
When n is small, confidence intervals are often extremely broad.
• X = {0.20, 0.05, 0.10} (OEL=1)
• n = 3
• gm = 0.1 95%LCL = 0.03, 95%UCL = 0.32
• gsd = 2.0 95%LCL = 1.5, 95%UCL = 21
• x0.95 = 0.31 95%LCL = 0.16, 95%UCL = 20
16
CONFIDENCE INTERVALS
Enterprise EHS Software Solutions
Exposure Rating Cutoff (%OEL)
0 X0.95 < 1%
1 1%< X0.95 <10%
2 10%< X0.95 <50%
3 50%< X0.95 <100%
4 X0.95 > 100%
An “Exposure Rating” represents a population of exposure profiles.
17
EXPOSURE RATINGS
Enterprise EHS Software Solutions
GM
0.0001 0.001 0.01 0.1 1
GS
D
20
10
PARAMETER SPACE (OEL =1)
0 1 2 3 4
18
Enterprise EHS Software Solutions
GM
0.001 0.01 0.1 1
GS
D
5
4
3
2
1
0 1 2 3 4
19
KEY CONCEPT: Parameter Space with a plausible upper limit for the true GSD (OEL=1)
KEY CONCEPTS
Enterprise EHS Software Solutions
GM
0.001 0.01 0.1 1
GS
D
5
4
3
2
1
0 1 2 3 4
20
KEY CONCEPT: Decision Charts
Enterprise EHS Software Solutions
Bayes’ Theorem –The foundation of Bayesian statistics
Posterior Likelihood Prior
Correction Factor21
BAYES’ THEOREM
Enterprise EHS Software Solutions
Applied to all lognormal distributions within parameter space
22
BAYES’ THEOREM
KEY CONCEPT: Populationi = AIHA exposure category(i.e., all combinations of geometric mean (G) and geometric standard deviation (D) within the ith exposure category)
Enterprise EHS Software Solutions
DECISION CHARTS
• Example Prior Decision Distributions
Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.05
0.2
0.5
0.2
0.05
Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.2 0.2 0.2 0.2 0.2
Non-informative priorInformative prior:Category 2, Medium Certainty
23
Enterprise EHS Software Solutions
GENERIC “PRIOR DECISION CHARTS”
• Professional Judgment prior
When the user picks an Initial Rating and Certainty Level a recommended Prior Decision Chart is loaded.
The default category probabilities represent a “best guess” as to what a generic prior should look like.
24
Enterprise EHS Software Solutions
Prior
Exposure Rating
0 1 2 3 4
De
cis
ion
Pro
ba
bilit
y
1
0.8
0.6
0.4
0.2
0
0.6
0.2 0.150.04 0.01
Prior
Exposure Rating
0 1 2 3 4
De
cis
ion
Pro
ba
bilit
y
1
0.8
0.6
0.4
0.2
0
0.5
0.250.18
0.05 0.02
Prior
Exposure Rating
0 1 2 3 4
De
cis
ion
Pro
ba
bilit
y
1
0.8
0.6
0.4
0.2
0
0.40.3
0.21
0.06 0.03
Non-informative prior “decisionchart”
IR=Initial RatingCL=Certainty Level
IR=Category 0CL=low
IR=Category 0CL=medium
IR=Category 0CL=high
Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.2 0.2 0.2 0.2 0.2
25
Enterprise EHS Software Solutions 26
Non-informative prior “decisionchart”
IR=Category 1CL=low
IR=Category 1CL=medium
IR=Category 1CL=high
Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.2 0.2 0.2 0.2 0.2
Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.30.4
0.180.07 0.05
Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.25
0.5
0.160.06 0.03
Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.2
0.6
0.140.04 0.02
IR=Initial RatingCL=Certainty Level
Enterprise EHS Software Solutions 27
Non-informative prior “decisionchart”
IR=Category 2CL=low
IR=Category 2CL=medium
IR=Category 2CL=high
Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.2 0.2 0.2 0.2 0.2
IR=Initial RatingCL=Certainty Level
Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.07
0.23
0.4
0.23
0.07
Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.05
0.2
0.5
0.2
0.05
Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.03
0.17
0.6
0.17
0.03
Enterprise EHS Software Solutions 28
Non-informative prior “decisionchart”
IR=Category 3CL=low
IR=Category 3CL=medium
IR=Category 3CL=high
Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.2 0.2 0.2 0.2 0.2
IR=Initial RatingCL=Certainty Level
Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.05 0.070.18
0.40.3
Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.03 0.060.16
0.5
0.25
Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.02 0.040.14
0.6
0.2
Enterprise EHS Software Solutions 29
Non-informative prior “decisionchart”
IR=Category 4CL=low
IR=Category 4CL=medium
IR=Category 4CL=high
Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.2 0.2 0.2 0.2 0.2
IR=Initial RatingCL=Certainty Level Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.03 0.06
0.210.3
0.4
Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.02 0.050.18
0.25
0.5
Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.01 0.040.15
0.2
0.6
Enterprise EHS Software Solutions
X={ 0.20, 0.05, 0.10 }
Likelihood
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0 0.002
0.66
0.2290.109
30
EXAMPLE LIKELIHOOD DECISION CHART
Enterprise EHS Software Solutions
EXAMPLE POSTERIOR DECISION CHARTS
Posterior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0 0.001
0.865
0.120.014
Posterior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0 0.002
0.66
0.2290.109
Using the non-informative prior
Using the informative prior
31
Enterprise EHS Software Solutions 32
PRIOR
LIKELIHOOD
POSTERIOR
Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.2 0.2 0.2 0.2 0.2
Prior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0.05
0.2
0.5
0.2
0.05
Likelihood
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0 0.002
0.66
0.2290.109
Posterior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0 0.001
0.865
0.120.014
Posterior
Exposure Rating
0 1 2 3 4
Decis
ion P
robability 1
0.8
0.6
0.4
0.2
0
0 0.002
0.66
0.2290.109
NON-INFORMATIVE INFORMATIVE
Enterprise EHS Software Solutions
EXAMPLE APPLICATIONS OF BDA
• General analysis of occupational exposure data
• Reach a decision when n is small
• Leverage professional judgment
• Provide feedback
• Assist in respirator selection
• Analyze censored datasets
33
Enterprise EHS Software Solutions
GENERAL ANALYSIS OF OCCUPATIONAL EXPOSURE DATA
• OEL=0.2 mg/m3
• n = 4
• x = {0.015, 0.008, 0.006, 0.016}
mg/m3
• In principle, BDA can be applied to any sample size (but is limited here to n<250).
34
Enterprise EHS Software Solutions
REACH A DECISION WHEN N IS SMALL
• OEL=1 ppm
• n = 1
• x = 0.05 ppm
• BDA can be applied to sample sizes as low as n=1.
35
Enterprise EHS Software Solutions
REACH A DECISION WHEN N IS SMALL
• OEL=1 ppm
• n = 1
• x = 0.99 ppm
• “Yes, the measurement is <OEL. But I strongly suspect that the exposure profile is not acceptable.”
• BDA leads to the same conclusion.
36
Enterprise EHS Software Solutions
LEVERAGE PROFESSIONAL JUDGEMENT
• OEL=1 ppm
• n = 1
• x = 0.05 ppm
• Professional judgment can sharpen the decision.
37
Enterprise EHS Software Solutions
PROVIDE FEEDBACK
• OEL=1 ppm
• n = 3
• x1 = 0.25 ppm
• x2 = 0.50 ppm
• x3 = 1.00 ppm
• The Prior is inconsistent with the Likelihood.
• BDA can be used to help improve professional judgment.
38
Enterprise EHS Software Solutions
ASSIST IN RESPIRATOR SELECTION
• OEL=1 ppm
• n = 3
• x1 = 0.99 ppm
• x2 = 0.50 ppm
• x3 = 2.0 ppm
• Decision = Category 4
• BDA can be used to guide PPE selection.
39
Enterprise EHS Software Solutions
ANALYZE CENSORED DATASETS
• OEL=1 ppm
• n = 1
• x < 0.05
• LOD = 0.05
• BDA can be applied to censored datasets, even datasets that are 100% censored and have different detection limits.
40
Enterprise EHS Software Solutions
Likelihood
Exposure Rating
0 1 2 3 4
De
cis
ion
Pro
ba
bilit
y
1
0.8
0.6
0.4
0.2
0
0 0.003
0.393
0.253
0.352
Likelihood
Exposure Rating
0 1 2 3 4
De
cis
ion
Pro
ba
bilit
y
1
0.8
0.6
0.4
0.2
0
0.05
0.804
0.13
0.012
0.005
ONE SAMPLE EXAMPLES (OEL=1)
• X = {0.005} ppm
• X = {0.01} ppm
• X = {0.05} ppm
• X = {0.25} ppm
Likelihood
Exposure Rating
0 1 2 3 4
De
cis
ion
Pro
ba
bilit
y
1
0.8
0.6
0.4
0.2
0
0.394
0.542
0.059
0.004
0.001
41
Likelihood
Exposure Rating
0 1 2 3 4
De
cis
ion
Pro
ba
bilit
y
1
0.8
0.6
0.4
0.2
0
0.000
0.394
0.459
0.082
0.064
Enterprise EHS Software Solutions 4242
OEL = 1 PPM
• X = {0.49} ppm
• X = {0.75} ppm
• X = {0.99} ppm
• X = {1.5} ppm
Likelihood
Exposure Rating
0 1 2 3 4
De
cis
ion
Pro
ba
bilit
y
1
0.8
0.6
0.4
0.2
0
0 0.000
0.056
0.351
0.593
Likelihood
Exposure Rating
0 1 2 3 4
De
cis
ion
Pro
ba
bilit
y
1
0.8
0.6
0.4
0.2
0
0 0.000
0.013
0.189
0.798
Likelihood
Exposure Rating
0 1 2 3 4
De
cis
ion
Pro
ba
bilit
y1
0.8
0.6
0.4
0.2
0
0 0 0.002
0.012
0.986
Likelihood
Exposure Rating
0 1 2 3 4
De
cis
ion
Pro
ba
bilit
y
1
0.8
0.6
0.4
0.2
0
0 0 0.006
0.048
0.946
42
Enterprise EHS Software Solutions
LARGE SAMPLE EXAMPLE
• An IH collected 17 representative measurements and calculated standard statistics (OEL=1):
X0.95 = 0.38 95%LCL = 0.26 and 95%UCL = 0.67
Final Rating : Category 2, Medium Certainty
• When n is large the BDA
results and standard
statistical analysis generally
lead to the same decision.
Likelihood
Exposure Rating
0 1 2 3 4
De
cis
ion
Pro
ba
bilit
y
1
0.8
0.6
0.4
0.2
0
0 0
0.79
0.201
0.009
43
Enterprise EHS Software Solutions
ASSUMPTIONS
• The lognormal distribution is a reasonable approximation of the true exposure profile.
• The true exposure profile falls somewhere within Parameter Space.
If the true GSD approaches or exceeds 4 the decision probabilities for the upper categories may be underestimates (and can potentially mislead).
44
Enterprise EHS Software Solutions
CAUTIONS
• BDA does not give us a “data driven” decision from a small n dataset.
• BDA simply gives us a different way to look at the data, in terms of …
exposure categories and
decision probabilities that the true 95th percentile falls within a category.
45
Enterprise EHS Software Solutions
CAUTIONS
• Training is necessary to properly use the BDA method and interpret the decision charts.
• Management should agree on …
the company risk management goals
the decision statistic
the use of the AIHA exposure rating method
how BDA fits into the overall program.
46
Enterprise EHS Software Solutions
• BDA is new “decision making” tool for IHs.
• Use it in conjunction with your other tools: graphs and goodness-of-fit tests
descriptive statistics
compliance statistics.
• Do the BDA results suggest a different interpretation of your dataset?
• Which interpretation is most likely correct? The BDA interpretation or that reached using your current data analysis tools?
SUMMARY
47
Enterprise EHS Software Solutions
WATCH THE WEBINAR
48
If you want to listen to the presentation that goes alongside this deck, you can do so for free here:http://www2.medgate.com/bayesian-webinar