‫הטיות בהערכת סיכונים‬
‫‪Bias in Risk Assessment‬‬
‫יניב מרדכי‬
‫אוניברסיטת תל‪-‬אביב‬
‫הפקולטה להנדסה‬
‫רקע‬
‫• ניהול סיכונים בפרויקטים מצריך הערכת סיכון‬
‫איכותית‪ ,‬לצורך קבלת החלטות מבוססות נתוני סיכון‪.‬‬
‫• הערכת הסיכון היא בד"כ סובייקטיבית‪ ,‬כלומר‬
‫מבוססת על הערכותיהם של אנשי מקצוע ע"ס ניסיון‬
‫ואינטואיציה‪ ,‬ולא אובייקטיבית או סטטיסטית‪.‬‬
‫• ישנם גורמים שונים המביאים להטיות בהערכת סיכון‬
‫ע"י אנשי מקצוע‪.‬‬
‫‪2‬‬
‫סיכון = התשובה לשלוש שאלות‬
)‫(ועוד אחת‬
• What can go wrong?
• What is the likelihood?
• What are the consequences?
• Kaplan and Garrick (1981)
• What is the time domain?
• Haimes (1998)
3
‫הודאות‬-‫רצוי של אי‬-‫סיכון = החלק הלא‬
Probability Density Function
h(t)
Opportunity
T ~ h(t)
4
Risk is about
this part of
uncertainty
T
c
Target Value
t
:‫הערכת סיכון‬
‫סובייקטיבית‬/‫אובייקטיבית‬
“(Objective) Probability does not exist!”, De-Finetti (1937)
Subjective
Objective
Art
Science
Intuition
Evidence
Relative
Subjectivists
5
Vs.
Absolute
Frequentists
‫סיכון רב‪-‬מימדי עם חלופות להמרת‬
‫סיכון בין יעדי הפרויקט‬
‫‪6‬‬
Bias in
Risk Assessment
7
Sources of Bias
•
•
•
•
8
Information and knowledge gaps
Modeling Bias
Cognitive Bias
Strategic Bias
Information and Knowledge gaps
• One assessor knows more than the other about a
certain topic.
• Information gaps are themselves risks, and
information acquisition is a means to reduce risk
• Information does not necessarily reduce
uncertainty, but does reduce Uncertainty About
Uncertainty
9
Methodological Bias
• The Assessment Method leads to
assessment bias.
• Method components:
– Set space partitioning.
– Resolution.
– Diversity (over results, risks, objectives)
– Predetermined aspects.
– Focus.
– Utility/Likelihood mixup
10
Risk Modeling
• Move from target domain to risk domain.
• Goal: Minimize overall Project Risk, by
selecting Risk handling strategies.
• Measures of Risk:
– Estimated Risk (NxN)
– Expected Risk
– Expected Disutility of Risk
11
Estimating Risk with 5x5 Risk
Impact Grids (is wrong…)
Likelihood
12
5
5
10
15
20
25
4
4
8
12
16
20
3
3
6
9
12
15
2
2
4
6
8
10
1
1
2
3
4
5
1
2
3
4
Severity
5
• The most popular risk
assessment technique
• Assess Likelihood Score.
• Assess Severity Score.
• Impact = Product of Scores
• Color provides additional
classification
(low,medium,high,extreme)
• Other variation: 3x3, 10x10
• Scores are sometimes
replaced by real values
Estimating Risk with 5x5 Risk
Impact Grids (is wrong…)
• Why this method is flawed:
Likelihood
13
5
5
10
15
20
25
4
4
8
12
16
20
3
3
6
9
12
15
2
2
4
6
8
10
1
1
2
3
4
5
1
2
3
4
Severity
5
– “Bernoulli” Fallacy
– How much is
“1”,”2”,”3”,“4”,”5”?
– What are we assessing
(mean, mode, median)?
– Score product is algebraically
invalid.
– Loss of information.
• [Pennock & Haimes (2002),
Williams (1996)]
Partitioned Multi-Objective
Risk Analysis
14
The Fallacy of Averages:
What Would Life Be Like If…
• Highways were constructed to accommodate the
average traffic load of vehicles or average weight.
• Communication networks were sufficient to handle
only the average data transfer rate.
• Airports were designed to handle the average
airliner traffic.
• Emergency services were staffed to attend to the
daily average number of emergency calls.
15
Eliciting Subjective Probabilities:
The Fractile Method - Raiffa (1968)
Continue identifying the median for each half-range until
Construct
Identify x25
the
Identify
and
assessment
Identify
x75Interpolate
-xmedians
and
as amedian
xof
CDF,
each
not
half-range
as a PDF
50 x-0 the
100
sufficient precision is achieved
1.00
0.75
0.50
0.25
0.00
1
16
2
3
4
5
6
7
8
9
10
Risk Disutility
• A Disutility Function quantifies
the adverse effect of each
possible deviation (from target
value).
• Disutility is similar to Utility
– Complies with all Utility Theory
axioms
– Increases as deviation
increases
• Expected Disutility
– The expectation of the
disutilityfunction, the smaller the
better.
• The objective is to minimize
Expected Disutility of Risk
17
5
4
3
2
1
100000
120000
140000
160000
180000
200000
Cognitive Bias
• Stems from (mostly unaware) perceptional
deviations, fallacies and deceptions.
• Research led by Tversky & Kahneman
(1970’s-80’s).
• Some Cognitive Biases may be observed
and fixed.
18
Cognitive Bias - examples
• Availability
– Human tendency to base assessment on vivid memory.
– Memorable experiences and impressions lead to
exaggerated assessments.
– Examples: car accidents, airplane crashes.
• Size
– We have difficulty grasping absolutely or relatively large
numbers, and think more clearly in smaller numbers.
– We tend to ignore the effects of the Law of Large
Numbers, and refer to estimates from smaller samples as
equally representative as estimates from larger ones, or
even as better ones
19
Cognitive Bias - examples
• Anchoring
– Taking an initial assessment based on historical
information, memory or cues, and adjusting it according
to circumstances of the new assessment
– Frequently insufficient to neutralize the initial anchor
• Representativeness
– Person relies on one estimate to make an assessment of
some other correlated estimate.
– Bias due to the intuitive attempt to rely on the inverse
conditional probability as an anchor, and ignore the
absolute probabilities of the events. (Bayes' Theorem)
20
Strategic Behavior
• Stems from personal cost/benefit considerations.
• Appears in collaborative but non-cooperative
settings.
• Applications in Social Choice: Voting Scheme,
Point Estimates and Single Peaked Preferences.
• Assessment combination methods like Averaging
and Bayesian Inference are vulnerable to Strategic
Behavior.
21
Strategy-Proof Assessment
• The Median of a set of scores is a strategy-proof
assessment (Moulin 1980).
• Game Theoretically, once the decision maker
chooses the Median, no one gains benefit from
being untruthful.
• For a continuous/discrete distribution, the median
assessment of the cumulative probability
distribution per each point/quantile is strategy-proof
• This is the MEDAS setting: Median Distribution
Assessment Scheme.
22
Illustration:
Multiple Assessments (CDF)
group_id 265 type (All)
function_id
264
1014
1764
2514
3264
F(X)
1.1
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
23
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
x
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Illustration:
Multiple Assessments +MEDAS (CDF)
group_id 265 type (All)
function_id
0
264
1014
1764
2514
3264
F(X)
1.1
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
24
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
x
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Illustration:
MEDAS
group_id 265 type (All)
function_id
0
F(X)
1.1
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
25
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
x
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Illustration:
PDF of MEDAS
function_id (All) group_id 265
Total
Total
0.3
f(x)_
0.25
0.2
0.15
0.1
0.05
0
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
x
26
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
MEDAS - Illustration (1/3)
G1(x)
0.9
0.8
F3(x)
0.7
F2(x)
0.6
F1(0.4) higher
than strategy:
F2(0.4) is
Assessor
best
Assessor
1 is1’s
confronted
with the
median,
F2(0.4)
On x=0.4,
but>F1(0.4)
causes
G1(x)greater
is
> below
G1(0.4)
disutility
the
is
following
situation
and
has
to
provide
F
*
(0.4)
≤
F
(0.4)=M(0.4)
1 candidate
2 will1:be
pointless.
F2(0.4)
to assessor
median
range
median.
F1(x)
2
Weakly2>[F2(0.4)-G1(0.4)]
Sincere.
[M(0.4)-G1(0.4)]
0.5
0.4
0.3
0.4
0.5
G1(x)
F2(x)
F3(x)
0.6
F1=G1(x)
Median
27
MEDAS - Illustration (2/3)
G1(x)
0.9
0.8
F3(x)
0.7
F2(x)
0.6
Assessor
best
strategy:
Assessor
F1(0.6)<
1F2(0.6)
is1’s
confronted
is median,
withbut
the
F3(0.6)
On x=0.6,
>F1(0.6)
G1(x) is
> above
G1(0.6)
the
is
following
causes
greater
situation
and has
to assessor
to provide
1:
F1*(0.6)
≥disutility
Fmedian
3(0.6)=M(0.6)
pointless.
candidate
F3(0.6)
will
be
range
median.
2>[F2(0.6)-G1(0.6)]2
[M(0.6)-G1(0.6)]
F1(x)
Weakly
Sincere.
0.5
0.4
0.3
0.4
0.5
G1(x)
F2(x)
F3(x)
0.6
F1=G1(x)
Median
28
MEDAS - Illustration (3/3)
Assessor
best
strategy:
F1(0.5)≠G1(0.5)
is
still
median,
On x=0.5, 1’s
G1(x)
is within
thebut
causes
greater=disutility
to
assessor 1:
candidate
range.
F1*(0.5)
Gmedian
1(0.5)=M(0.5)
F1(0.5)=G1(0.5)
is the 2median!
[F1(0.5)-G1(0.5)]
>0
Strongly
Sincere.
0.9
0.8
G1(x)
F3(x)
0.7
F2(x)
0.6
0.5
0.4
0.3
0.4
0.5
G1(x)
F2(x)
F3(x)
0.6
F1=G1(x)
Median
29
MEDAS - characteristics
• Works on CDF form (not PDF)
• Relaxes “Single-Peakedness”.
• Works for single point, quantile and continuous
cases.
• Always induces sincere behavior as dominant
strategy
• “Winning” (having one’s assessment elected to
represent) is not the goal.
• Does not require payoff considerations.
• Does not repair cognitive bias.
30
Challenges
• Definition
– Sources, targets, objectives,
utility functions.
• Assessment
– Let each assessor assess the
effect of each source on each
target under each strategy
– Conduct a strategy-proof
assessment process
• Decision making
– Evaluate trade-off according to
predefined preference relations.
– Select optimal risk handling
strategies
31
• Implementation
– Translate decisions to practical
risk mitigation steps.
– Monitor progress and
effectiveness.
• Iteration
– Reassess and reevaluate
– Repeat Risk Management
Process
LEAN Risk Management
• The challenge:
To found an effective, successful and easy risk management
process.
• The response:
– Conduct Risk Management workshops, based on the LEAN
philosophy and methodology.
– Bring all stakeholders together, extract all knowledge and information
about uncertainty.
– Conduct methodological analysis and assessment of risks.
– Produce and execute a risk handling plan, with immediate effects.
32
Summary
• Risk Assessment - not what you thought!
• High Fidelity Risk Assessment requires an initial
effort, and continuous maintenance, but has clearly
significant ROI.
• Risk Assessment requires appropriate and valid
methods and sometimes even professional risk
analysis escort.
• Risk management workshops are a constructive
and effective way to produce quality risk
assessments.
33
‫תודה!‬
‫‪34‬‬