Engineering Risk Benefit Analysis

1.155, 2.943, 3.577, 6.938, 10.816, 13.621, 16.862, 22.82, ESD.72

CBA 4.

Including Uncertainty

George E. Apostolakis Massachusetts Institute of Technology

Spring 2007
CBA 4. Including Uncertainty 1

Uncertainty

Practically any CBA requires consideration of uncertainty. Most methodologies in use are ad hoc, due to the intrinsic difficulty of the generalized problem.

CBA 4. Including Uncertainty

Methods 1. Scenario analysis 2. Adjustments of interest rates 3. Decision Theory 4. Simplified probabilistic models
CBA 4. Including Uncertainty 3

Scenario Analysis
Preparation and analysis of scenarios: -- Optimistic or most favorable estimate -- Most likely or best estimate or fair estimate -- Pessimistic or least favorable estimate Interpretation is difficult without assignment of probabilities to scenarios. Benefit: Brings additional information into the process.
CBA 4. Including Uncertainty 4

Example
A new machine is to be purchased for producing units in a new manner. Pessim. Fair Optim. Annual number of units: 900 1,000 1,100 Savings per unit: $50 55 60 Operating costs: $2,000 1,600 1,200 etc. PW: -$49,000 22,000 120,000

What are we to do with such information?

CBA 4. Including Uncertainty 5

Developing Scenarios
For each element in the problem, e.g., interest rate and costs, define the three values. We really dont know how conservative (pessimistic) the final answer is. People are bad processors of information. Point estimates tend to cluster around the median value. Possibility of displacement bias. Extremes greater than the 75th or smaller than the 25th percentile are difficult to imagine. Overconfidence.
CBA 4. Including Uncertainty 6

Forecasting Oil Prices

D. Draper, Assessment and propagation of model uncertainty, Journal of the Royal Statist. Soc., B (1995) 45-97.

In 1980, 43 economists and energy experts forecast the price of oil from 1981 to 2020 to aid in policy planning. They used 10 leading econometric models under each of 12 scenarios embodying a variety of assumptions about inputs, such as supply, demand, and growth rates.
CBA 4. Including Uncertainty 7

The Plausible Scenario

One scenario was termed as the plausible median case. It represented the general trends to be expected. The 10 models were applied to the plausible scenario. Results for 1986: Actual price: $13 Range of predictions: $27 to $ 51.
CBA 4. Including Uncertainty 8

High Interest Rates

Justify choice of alternatives using a high interest rate, e.g., 30%. Example Total annual income: $55,000 Capital cost: $80,000 Annual capital recovery with return: $80,000 (A/P, 30%, 6yrs) = $30,272 Annual operating cost $28,600 Net annual profit: 55,000 (30,272 + 28,600) = = -$3,872
CBA 4. Including Uncertainty 9

Example (contd)
The high rate of 30% is intended to cover uncertainty. If the annual income were $60,000, then the net annual profit would be 60,000 - (30,272 + 28,600) = = $1,128 and the venture would be accepted. A high interest rate does not guarantee that all uncertainties are accounted for. Its choice is arbitrary.

CBA 4. Including Uncertainty

10

Decision Theory: Manufacturing Example Decision: To continue producing old product (O) or convert to a new product (N). The payoffs depend on the market conditions: s: strong market for the new product w: weak market for the new product
CBA 4. Including Uncertainty 11

Manufacturing Example Payoffs

Earnings (payoffs): L1: $15,000/yr, old product, L2: $30,000/yr, new product and the market is strong, L3: -$10,000/yr, new product and the market is weak

Demand and Probabilities: Period 2 (5 yrs) Period 1 (5 yrs) s1 s2 s1 w2 w1 w2

Probabilities 0.4 = P(s1s2) 0.4 = P(s1w2) 0.2 = P(w1w2)

P(s1) = P(s1s2) + P(s1w2) = 0.8; P(w1) = 0.2; P(s2/s1) = 0.5; P(w2/w1) = 1.0
CBA 4. Including Uncertainty 12

Decision Tree
Decision Options States of Nature P1 P2 s2 s1 w2 w2 PW of Payoffs

$243.26K $96.94K

w1

-$81.09K O $121.61K
CBA 4. Including Uncertainty 13

Calculation of the Payoffs

(1 + 0.04 ) 5 1 PWs1 = 30 x ( P / A ,0.04 ,5 ) = 30 x = 133 .52 K 5 0.04 x (1 + 0.04 )

PWs 2 = 30 x ( P / A , 0 .04 , 5 ) x ( P / F , 0 .04 , 5 ) = (1 + 0.04 ) 5 1 1 = 30 x x = 109 .74 K 5 5 0 .04 x (1 + 0.04 ) (1 + 0 .04 )

PWs1 s 2 = 133 .52 + 109 .74 = 243 .26 K

10 PWw1 = 133.52x = 44.51K 30
PWs1w 2 = 133.52 36.58 = 96.94K PWw1w 2 = 44.51 36.58 = 81.09K
CBA 4. Including Uncertainty 14

PW w 2 = 109 .74 x

10 = 36 .58 K 30

Calculation of the EMV

Old Product
(1 + 0.04 )10 1 = 121 .61K PWO = 15 x ( P / A ,0.04 ,10 ) = 15 x 10 0.04 x (1 + 0.04 )

New Product EMVN = 243.26x0.4 + 96.94x0.4 -81.09x0.2 = 119.86K Decision Stay with the old product?

CBA 4. Including Uncertainty

15

Calculation using Utilities

Let the utility of payoffs be U(x) = 1.18 ln(x+5) - 1.29 -2 x 2 (x in $M) [U(2) = 1, U(-2)= 0] U(243.26) = 0.665; U(121.61) = 0.637; U(96.94) = 0.632; U(-81.09) = 0.590 Old Product EU(O) = 0.637 New Product EU(N) = 0.665x0.4 + 0.632x0.4 + 0.590x0.2 = 0.6368 Decision Stay with the old product?
CBA 4. Including Uncertainty 16

Probabilistic Models We have:

XT X1 X2 PW[ X(T)] = X 0 + + + .... + 2 (1 + i ) (1 + i ) (1 + i )T

(1)

where X j B j C j are the net benefits in year j. All Xj are r.v.s, PW[X(T)] is a r.v. Note that (1) is of the form:
Y = a0 X 0 + a1X 1 + a2 X 2 + ... + aT X T
CBA 4. Including Uncertainty 17

Analysis
Computing the probability density function (pdf) of PW is usually difficult in practice. Try to compute the quantities E[PW], the expected value of PW, and
2 PW

i.e., the variance of PW.

CBA 4. Including Uncertainty

18

Fundamental Relationships from Probability Theory (1) Let Z = aW+b (Z and W are r.v.s, a and b constants) E[Z] = a E[W] + b
2 Z 2 2 = a W

CBA 4. Including Uncertainty

19

Fundamental Relationships from Probability Theory (2)

Let

Z = W1 + W2

(W1 , W2 : independent r.v.s)

E [ Z ] = E [W1 ] + E [W2 ]
2 2 2 Z = W + W2 1

Note: Extends to any number of mutually independent r.v.s.

CBA 4. Including Uncertainty 20

Fundamental Relationships from Probability Theory (3)

Let

Z = aW1 + bW2 + c

E[Z] = aE[ W1 ] + bE[ W2 ] + c

2 2 2 2 2 = a + b W2 Z W1

If W1 and W2 are normal, then Z is also normal. We often assume that Z is normal even if W1 and W2 are not.
CBA 4. Including Uncertainty 21

Example: Reliability Physics

(RPRA 3, slide 30) A capacitor is placed across a power source. Assume that surge voltages occur on the line at a rate of one per month and they are normally distributed with a mean value of 100 volts and a standard deviation of 15 volts. The breakdown voltage of the capacitor is 130 volts. Suppose that the breakdown voltage is also normally distributed with standard deviation of 15 volts.
CBA 4. Including Uncertainty 22

Example (2)
The capacitor fails when the surge voltage, S, is greater than the capacity, C. S: rv with E[S] = 100, S = 15 volts C: rv with E[C] = 130, C = 15 volts Define a new rv D C S = aC + bS Then, E[D] = 130 100 = 30 volts and

D = C + S = 21.21
CBA 4. Including Uncertainty

volts

23

Example (3)
D is also normally distributed, therefore Pd/sv(D < 0) = P( Z < -(30/21.21)) = P(Z < - 1.41) = = P(Z > 1.41) = 0.5 0.42 = 0.08 RPRA 3, page 31, shows that Pd/sv = conditional

probability of damage given a surge voltage = P(surge voltage>130 volts/surge voltage)

= P(Z > 130 100 ) = P(Z > 2) = 15 = 1 P ( Z < 2 ) = 1 0 . 9772 = 0 . 0228 < 0 . 08

The uncertainty in the breakdown voltage increased the failure probability.

CBA 4. Including Uncertainty 24

Assuming Independence of Xj
X1 X2 XT PWInd [ X(T)] = X 0 + + + .... + 2 (1 + i ) (1 + i ) (1 + i )T

Xj ( j = 0, 1,........,T) are mutually independent 2 r.v.s with known E [ X j ] and X (= 2 j)

E[PWInd { X(T)}] = E[YInd ] =
2 2 = PW ,Ind Y ,Ind = T

(1 + i ) j
j= 0

E[ X j ]

(1 + i )2 j
j= 0
25

2 j

Note: Gaussian approximation for pdf of Y may work well in this case.
CBA 4. Including Uncertainty

Example: Project A
T = 3 yrs.; i = 8%; Initial Cost = $10K
Probability, p 0.10 0.25 0.30 0.25 0.10 1.00 Net Benefits t=1 t=2 t=3 $3K $4K $5K $6K $7K $3K $4K $5K $6K $7K $3K $4K $5K $6K $7K
26

CBA 4. Including Uncertainty

Example (2)
Denote with X1, X2, and X3, the net benefits of A in years 1, 2 and 3, respectively. Note: Net benefits, X1, X2, and X3, do not have to be identically distributed or symmetric or discrete; these choices are made just to keep the example simple. Denote with Y the present worth of A, PW(A). Then: PWInd(A) = YInd = -10K + X1/(1.08) + X2/(1.08)2 +X3/(1.08)3
CBA 4. Including Uncertainty 27

Observations
Y is a random variable (takes more than one value with different probabilities for any given implementation of project A) Value of Y will be determined by the values of the combination of X1, X2, and X3 that will actually materialize Corresponding a priori probability of any value of Y is equal to probability of that particular combination of X1, X2 and X3.
CBA 4. Including Uncertainty 28

Expectation and Variance of Annual Net Benefits

It is easy to determine the expected value and variance of each of X1, X2, and X3, separately: E[X1] = 0.1x3+0.25x4+0.3x5 +0.25x6+0.1x7 = $5,000

(Similarly, we have E[X2] = $5K and E[X3] = $5K. )

2 = 0.1x(3-5)2 + 0.25x(4-5)2 + 0.3x(5-5)2 + 0.25x(6-5)2 + X 1

+ 0.1x(7-5)2 = 1,300,000 (1,140)2

or,

X 1 $1,140 = X 2 = X 3

CBA 4. Including Uncertainty

29

Independence: Calculations
Assume that the net benefits obtained from Project A in years 1, 2 and 3 are determined independently of one another. This means the probability of the combination {X1 = 3, X2 = 6, X3 = 4} is equal to P(X1 = 3, X2 = 6, X3 = 4) = P(X1 =3)P(X2 = 6)P(X3 =4) = = (0.1)(0.25)(0.25) = 0.00625 that is, with probability 0.00625, the r.v. Y, i.e., the PW of Project A, will take on the value PWInd = YInd = -10K + 3K/(1.08) + 6K/(1.08)2 + 4K/(1.08)3 $1,097.14

CBA 4. Including Uncertainty

30

Independence: Calculations (2)

Note that Y can take on a total of 125 (= 555) different values of independent outcomes in years 1, 2 and 3. Using the expressions on slide 21 we get E[YInd]= -10K+E[X1]/(1.08)+E[X2]/(1.08)2+E[X3]/(1.08)3 = = -10K + (5K)[1/1.08 + 1/(1.08)2 + 1/(1.08)3] = -10K+ (5K)(P/A, 0.08, 3) -10K + (5K)(2.5771) $2,885
2 2 2 2+ 2 2]2+ [1/(1.08) [1/(1.08)3]2 = [1/1.08] Y X1 X2 X3 ,Ind = (1,140)2[1/(1.08)2 + 1/(1.08)4 + + 1/(1.08)6] = = (1,140)2(2.2225), or, Y ,Ind = (1,140)(2.2225)1/2 $1,700
CBA 4. Including Uncertainty 31

Conclusion

Project A will, "on average," have a net present value equal to about $2,885 and a standard deviation of approximately $1,700 around that average.

CBA 4. Including Uncertainty

32

What can we do with this information?

Alternative 1 2 3 4 E[Y] = E[PW] 20 5 15 17
Y

15 7 12 16

CBA 4. Including Uncertainty

33

Possible Decision Criteria

Choose the alternative with the highest mean value (A1). Note how close A4 is and how large the standard deviations are. Depending on the uncertainties, A1 and A4 may be indistinguishable. Minimize the probability of loss. Assume normal distributions and find P(PW < 0).
CBA 4. Including Uncertainty 34

Probability of Loss
A1: P(PW < 0) = P(Z < -(20/15)) = P(Z < -1.33) = = 0.09 best alternative A2: P(PW < 0) = P(Z < -(5/7)) = P(Z < -0.71) = = 0.24 A3: P(PW < 0) = P(Z < -(15/12)) = P(Z < -1.25) = = 0.10 A4: P(PW < 0) = P(Z < -(17/16)) = P(Z < -1.06) = = 0.14
CBA 4. Including Uncertainty 35

Probability of Loss for Project A

PInd(PWInd < 0) = P(Z < -(2885/1700)) = P(Z < -1.7) = 0.04 The fundamental assumption is that of independence of the annual benefits.

CBA 4. Including Uncertainty

36

Complete Dependence of Xj
Once the net benefits, X1, for year 1 are known, we shall also know exactly the net benefits for years 2 and 3.
Probability, p 0.10 0.25 0.30 0.25 0.10 1.00 Net Benefits t=1 t=2 t=3 $3K $4K $5K $6K $7K $3K $4K $5K $6K $7K $3K $4K $5K $6K $7K
37

CBA 4. Including Uncertainty

Mean and Variance (Dependence)

1 1 1 + + .... + PWDep [ X (T )] = X 0 + X 2 T + ( i ) 1 ( i ) ( i ) + + 1 1

PWDep (A) = YDep = -10K + X[1/(1.08) + 1/(1.08)2 +1/(1.08)3] = -10K + X(P/A, 8, 3)

From slide 15 we get

E[YDep] = -10K + E[X](P/A, 8, 3) $2,885, as before

or
2 2 = Y ,Dep X [(P/A, 8, 3)]2 =(1,140)2[2.5771]2

Y ,Dep= (1,140)(2.5771)

$2,938
38

CBA 4. Including Uncertainty

Comparison
X1 X2 XT PWInd [ X(T)] = X 0 + + + .... + 2 (1 + i ) (1 + i ) (1 + i )T

1 1 1 PW + + .... + Dep[ X(T)] = X0 + X 2 T ( 1 i ) + (1 + i) (1 + i)

In both the independent and dependent cases the mean values are the same ($2,885). The standard deviation in the dependent case ($2,938) is 73% larger than that of the independent case ($1,700). PDep(PW < 0) = P(Z < - (2885/2938)) = P(Z < -0.98) = 0.16 Compare with PInd(PW < 0) = 0.04 (slide 32) These two cases are considered as bounding the problem.
CBA 4. Including Uncertainty 39