hobbs: better ignorant than misled: including uncertainty in forecasts supporting management and...
TRANSCRIPT
Motivation Bayes Evaluating actions Example Conclusions
Better ignorant than misled: Includinguncertainty in forecasts supporting management
and policy
N. Thompson Hobbs
Natural Resource Ecology Laboratory, Department of Ecosystem Science andSustainability, and Graduate Degree Program in Ecology,
Colorado State University
Uncertainty Analysis: A Critical Step in Ecological SynthesisAnnual Meeting of the Ecological Society of America, 8/5/2013
1 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Honest synthesis
Published
research
Monitoring
Experiments Physical
laws
Theory
Management and Policy
Process
studies
2 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
The problem of management:
What actions will allow us to meetgoals for the future?
3 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Synthesis with deterministic models
4 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Orrin H. Pilkey and Linda Pilkey-Jarvis 2007
5 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
A broadly applicable approach to ecological modeling insupport of management
[E,θprocess,θdata |data] ∝
[data|θdata ,Et ] [Et |θprocess ,Et−1] [θprocess ,θdata ,E0]
data
EtEt�1
✓data✓process
Process
Parameters
6 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Posterior predictive distributions[E ′|data] =∫
θθθ
∫E [E ′|E ,θdata ,θprocess ] [E ,θdata ,θprocess |data]dEdθdataθprocess
Time
Stat
e of
Pop
ulat
ion
(E o
r E’
)St
ate
of E
colo
gica
l Sys
tem
(E o
r E')
7 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Posterior predictive distribution of future states, E ′
Future state of system, E'
Prob
abilit
y de
nsity
Time
Stat
e of
Pop
ulat
ion
(E o
r E’
)St
ate
of E
colo
gica
l Sys
tem
(E o
r E')
8 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
How to evaluate actions?
Objective: reduce state below a target
Future state of system, E'
Pro
babi
lity
dens
ity
Objective
9 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Objective: maintain state within acceptable range
Future state of system, E'
Pro
babi
lity
dens
ity
Range
10 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Objective: increase state above a target
Future state of system, E'
Pro
babi
lity
dens
ity
Objective
11 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Action: do nothing
Future state of system, E'
Pro
babi
lity
dens
ity
Objective
12 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Action: implement managment
Future state of system, E'
Pro
babi
lity
dens
ity
Objective
Management
13 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Net effect of management
Future state of system, E'
Pro
babi
lity
dens
ity
Objective
14 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Net effect of management
Future state of system, E'
Pro
babi
lity
dens
ity
Objective
15 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Example: Managing brucellosis in Yellowstone Bison
16 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Goal: Reduce probability of infectionby half in five years.
Action: Annually vaccinate 200sero-positive females.
17 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Bayesian matrix model with multiple sources of data
Serology Ysero[t]
Mark recaptureyϕ[t]
Removalsyr[t]
Censusyn.obs[t]yn.sd[t]
Aerial classificationsyratio.calf[t]
yσ[t]
A[t]n[t�1] � r[t]
n[t]asd, bsd
psight
Ground classificationsyp.µ[t]yp.σ[t]
�p
Removal age, sexyr.age[t]
Removal serologyyr.sero[t]
rcomp[t]
rsero[t]
�, fc, fn, fp, , s, v
rinfect[t]
18 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Effect of vaccination: Treat 200 sero-positive / year
0 1 2 3 4 5
0.00
0.05
0.10
0.15
0.20
0.25
Years into the future
Pro
babi
lity
of tr
ansm
issi
on
Do nothingVaccinate
19 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Effect of vaccination
0 1 2 3 4 5
0.00
0.10
0.20
0.30
Years into the future
Pro
babi
lity
of tr
ansm
issi
on
20 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Effect of vaccination
Objective: reduce transmission probability by half
0.0 0.1 0.2 0.3 0.4
02
46
810
Five years in the future
Probablity of tranmission
Density
Do nothingVaccinate
21 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Comparison of alternatives
Management action Probability of meeting goal
Do nothing .05
Vaccinate 200 sero-positives .33
Cull 200 sero-positives .29
Cull 200 females .15
Boundary hunting .03
22 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Multiple objectives, multiple actions
Objectives
I Reduce P(infection) byhalf
I Sero prevalence < 40%
I Population size between3000 - 3500
I Appropriate demographiccomposition
Actions
I Vaccination
I Remove sero-positives
I Remove sero-negatives
I Boundary hunting (or removal)
23 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Closing
I Value
I Provides honest forecasts relevant to actions and goals.I Informs the conversation
I Limitations
I Demonic intrusions aren’t included.I Forecasting horizons are short.
24 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
Develop Bayesian model(s) of system
Forecast system behavior under different policy
options
Implement policy (or better, policies)
Observe system behavior
Update model(s)
Walters, C. J. 1986. Adaptive Management of Renewable Resources. Macmillan, New York.
25 / 26
institution-logo
Motivation Bayes Evaluating actions Example Conclusions
May 21-30, 2013 Google “NREL Bayes”
Building capacity in Bayesian modeling
for practicing ecologists
Award DEB 000347455
A workshop for faculty and agency researchers. $1000 stipend. See www.nrel.colostate.edu/projects/bayesworkshop/
26 / 26