Ferson
Ferson
Ferson
Scott Ferson
Applied Biomathematics
scott@ramas.com
Background
• Most risk assessments are deterministic
and deliberately conservative
• However ...
– degree of conservatism is opaque,
unquantified, and can be inconsistent
• Distribution of values
(exceedance risk)
X
Continuous distributions
• Normal 1
Cumulative probability
• Uniform 0.8
• Exponential 0.6
• Lognormal 0.4
• Triangular 0.2
• Weibull 0
0 2 4 6 8
• Beta
• Laplace
• Many, many more
Discrete distributions
1
Cumulative probability
• Binominal 0.8
0.6
• Poisson
0.4
• Discrete uniform
0.2
• Reciprocal 0
0 2 4 6 8
• Well, almost
• Ensembles
• Distribution shapes and tails
• Dependencies
• Backcalculation
Computational methods
• Analytical approaches
– Laplace and Mellin transforms
– Delta method (just the mean and variance)
convolution
+ =
X Y X+Y
Generating random deviates
• Can make realizations from an arbitrary
distribution using uniform deviates
Cumulative probability
uniform deviate
0.8
0.6
0.4
0.2
0
0 50 100 150
Value of random variable
How to specify distributions
• Default distributions
comes right out of the book
• Fitted or empirical distributions
usually not enough data available
• Extrapolations and surrogate data
requires professional judgment
• Elicitation from experts
expensive, controversial when experts disagree
• Maximum entropy criterion
inconsistent through changes of scale
1.000
.750
.500
.250
.000
-5.00 0.00 5.00 10.00 15.00
.750
100 Trials:
close-up of 0.01
.500
0.00
.000
0.02
1.000
0.00
.000
0.00 125.00 250.00 375.00 500.00
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 9
II
I
1
I or II
0
Backcalculation
• Cleanup and remediation planning requires
backcalculation
conc × intake
dose =
body mass
0.04
0.2
0.02
0.00 0.0
0 2 4 6 8 0 2 4 6 8
Planned dose Body mass
0.06 0.03
0.04 0.02
0.02 0.01
0.00 0.00
0 2 4 6 8 0 2 4 6 8
Intake Concentration
0.06
Large doses in the realized distribution (arising
0.05 from the allowed distribution of concentrations)
are more common than were originally planned.
0.04
planned
0.03
0.02
0.01 realized
0
0 2 4 6 8
Dose
Normal approximation
• If A+B=C, compute B as C−A under the
assumption that the correlation
between A and C is r = sA/sC
A=lognormal(5,1.5) C=triangular(0,12,15.5)
1 .0 0
.7 5 realized
.5 0 planned
.2 5
.0
-5.0 0 -0.63 3 .7 5 8 .1 3 12.50 0 5 10 15 20
C−A, with r =sA/sC C
Iterative (trial & error) approach
• Initialize B with C−A
• This distribution is too wide
• Transform density p(x) to p(x)m
• Rescale so that area remains one
• Whenever m > 1 dispersion decreases
• Repeat until you get an acceptable fit
0.06
By trial and error, you may be able to find a distribution
0.05 of concentrations that yields something close to the
planned distribution of doses. This is not always
0.04 possible however.
conc× intake
= dose
body mass
0.03
intake
ln + ln(conc) = ln(dose)
0.02 body mass
1442443 1 424 3 1 424 3
A B C
0.01
0
0 2 4 6 8
Dose
Monte Carlo methods
• How?
– replace each point estimate with a distribution
– repeatedly sample from each distribution
– tally answers in histogram
• Why?
– simple to implement
– fairly simple to explain
– summarizes entire distribution of risk
• Why not?
– requires a lot of empirical information (or guesses)
– routine assumptions may be “non-protective”
Steps
• Clarify the questions and gather data
• Formulate the model (identify variables
and their interrelationships)
• Specify distributions for each input
• Specify dependencies and correlations
• Run simulation
• Conduct sensitivity studies
• Present results
• Discuss limitations of the assessment
Limitations of Monte Carlo
• Needs precise, stationary distributions
• Nonlinear dependencies often ignored
• Unknown dependencies often ignored
• Assumes model is well understood
• Can be computationally expensive
• Backcalculations difficult or impossible
• Treats incertitude like variability
• Thought by some to be too difficult
Pitfalls of Monte Carlo
• Muddled model
• Infeasible correlation matrices
• Inconsistent independence assumptions
• Multiply instantiated variables
• Too few replications
• No sensitivity studies
• Confounding of different populations
Kinds of uncertainty
• Variability
– Stochasticity
– Heterogeneity
– Spatial variation
– Individuality and genetics
• Incertitude
– Measurement error
– Ignorance
– Model uncertainty
• Vagueness, confusion, etc.
Why distinguish them?
• Empirical effort can reduce incertitude
0 0 0
0 20 0 40 0 80