Value of Information in Decision Trees

19

19.1 VALUE OF INFORMATION

Useful concept for Evaluating potential information-gathering activities Comparing importance of multiple uncertainties

19.2 EXPECTED VALUE OF PERFECT INFORMATION

Several computational methods Flipping tree, moving an event set of branches, appropriate for any decision tree Payoff table, most appropriate only for single-stage tree (one set of uncertain outcomes with no subsequent decisions) Expected improvement All three methods start by determining Expected Value Under Uncertainty, EVUU, which is the expected value of the optimal strategy without any additional information. To use these methods, you need (a) a model of your decision problem under uncertainty with payoffs and probabilities and (b) a willingness to summarize a payoff distribution (payoffs with associated probabilities) using expected value. The methods can be modified to use certain equivalents for a decision maker who is not risk neutral.

224

Chapter 19 Value of Information in Decision Trees

Expected Value of Perfect Information, Reordered Tree

Figure 19.1 Structure, Cash Flows, Endpoint Values, and Probabilities
0.5 High Sales $400,000 $700,000 0.3 Medium Sales $100,000 -$300,000 $400,000 0.2 Low Sales 1 $100,000 -$200,000

Introduce Product

Don't Introduce $0 $0

Figure 19.2 Rollback Expected Values

0.5 High Sales $400,000

Introduce Product $190,000

0.3 Medium Sales $100,000

0.2 Low Sales 1 $190,000 -$200,000

Don't Introduce $0

The two figures above show what is called the prior problem, i.e., the decision problem under uncertainty before obtaining any additional information.

19.2 Expected Value of Perfect Information

225

Figure 19.3 Structure Using Perfect Prediction

High Sales

Introduce Product

Medium Sales

"High Sales"

Low Sales

Don't Introduce

High Sales

Introduce Product

Medium Sales

Perfect Prediction

"Medium Sales"

Low Sales

Don't Introduce

High Sales

Introduce Product

Medium Sales

"Low Sales"

Low Sales

Don't Introduce

Before you get a perfect prediction, you are uncertain about what that prediction will be. If you originally think the probability of High Sales is 0.5, then you should also think the probability is 0.5 that a perfect prediction will tell you that sales will be high. After you get a prediction of "High Sales," the probability of actually having high sales is 1.0.

226

Chapter 19 Value of Information in Decision Trees

Figure 19.4 Rollback Using Free Perfect Prediction

1.0 High Sales $400,000 0.0 Medium Sales $100,000 $400,000 0.5 "High Sales" 1 $400,000 Don't Introduce $0 0.0 High Sales $400,000 1.0 Medium Sales $100,000 $100,000 Perfect Prediction $230,000 0.3 "Medium Sales" 1 $100,000 Don't Introduce $0 0.0 High Sales $400,000 0.0 Medium Sales $100,000 -$200,000 0.2 "Low Sales" 2 $0 Don't Introduce $0 1.0 Low Sales -$200,000 0.0 Low Sales -$200,000 0.0 Low Sales -$200,000

Introduce Product

Introduce Product

Introduce Product

EVUU: Expected Value Under Uncertainty the expected value of the best strategy without any additional information EVPP EVPI Expected Value using a (free) Perfect Prediction Expected Value of Perfect Information

EVPI = EVPP EVUU In this example, EVPI = $230,000 $190,000 = $40,000 For a perfect prediction, the information message "Low Sales" is the same as the event Low Sales, so the detailed structure shown above is not needed. A shortcut approach is to "flip" the original decision tree, shown in Figure 19.2, rearranging the order of the decision node and event node, to obtain the tree shown below.

19.2 Expected Value of Perfect Information

227

Figure 19.5 Shortcut EVPP

Introduce Product 0.5 High Sales 1 $400,000 Don't Introduce $0 $400,000

Introduce Product Perfect Prediction $230,000 0.3 Medium Sales 1 $100,000 Don't Introduce $0 $100,000

Introduce Product 0.2 Low Sales 2 $0 Don't Introduce $0 -$200,000

Expected Value of Perfect Information, Payoff Table

This method is most appropriate only for a single-stage decision tree (one set of uncertain outcomes with no subsequent decisions). Figure 19.6 Payoff Table for Prior Problem with Expected Values
Probability 0.5 0.3 0.2 Event High Sales Medium Sales Low Sales Expected Value Alternatives Introduce Don't $400,000 $0 $100,000 $0 -$200,000 $0 $190,000 $0

For each row in the body of the payoff table, if you receive a perfect prediction that the event in that row will occur, which alternative would you choose and what would your payoff be? Before you receive the prediction, you don't know which of the payoffs you will receive (either $400,000 or $100,000 or $0), so you summarize the payoff distribution using expected value, EVPP.

228

Chapter 19 Value of Information in Decision Trees

Figure 19.7 Payoff Table with EVPP

Probability 0.5 0.3 0.2 Event High Sales Medium Sales Low Sales Expected Value Alternatives Introduce Don't $400,000 $0 $100,000 $0 -$200,000 $0 $190,000 EVUU $0 Payoff Using Perfect Prediction $400,000 $100,000 $0 $230,000 EVPP

EVPI = $230,000 $190,000 = $40,000

Expected Value of Perfect Information, Expected Improvement

Like the payoff table method, this method is most appropriate only for a single-stage decision tree. (1) Use the prior decision tree or prior payoff table to find EVUU (the expected value of the best strategy without any additional information). (2) If you are committed to the best strategy, consider each outcome of the uncertain event and whether you would change your choice if you received a perfect prediction that the event was going to occur. In the example, you would not change your choice if you are told that sales will be high or medium. However, if you are told that sales will be low, you would change your choice from Introduce to Don't. (3) Determine how much your payoff will improve in each of the cases. In the example, your payoff will not improve if you are told that sales will be high or medium, but your payoff will improve by $200,000 (from $200,000 to $0) if you are told that sales will be low. (4) Compute expected improvement associated with having the perfect prediction by weighting each improvement by its associated probability. In the example, the improvements associated with a perfect prediction of high, medium, and low are $0, $0, and $200,000, respectively, with probabilities 0.5, 0.3, 0.2. EVPI = Expected Improvement = 0.5*0 + 0.3*0 + 0.2*200,000 = $40,000

19.2 Expected Value of Perfect Information

229

Expected Value of Perfect Information, Single-Season Product

Figure 19.8 Prior Problem, Four Alternatives and Three Outcomes
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 A B Single-Season Product Data Price $3.00 None $0 $0.00 0 Equip. Size Small Medium $1,000 $2,000 $0.90 $0.70 4500 5500 Large $3,000 $0.50 6500 C D E F

Fixed Cost Var. Cost Capacity Payoff Table

Prob. 0.3 0.4 0.3

Demand 3000 4000 5000 Exp.Val.

None $0 $0 $0 $0

Equip. Size Small Medium $5,300 $4,900 $7,400 $7,200 $8,450 $9,500 $7,085 $7,200

Large $4,500 $7,000 $9,500 $7,000

C16 formula: copied to C16:F18 C20 formula: copied to C20:F20

=($B$5-C$9)*MIN(C$10,$B16)-C$8

=SUMPRODUCT($A16:$A18,C16:C18)

Figure 19.9 EVPP

A B C D E 14 Equip. Size 15 Prob. Demand None Small Medium 0.3 3000 $0 $5,300 $4,900 16 0.4 4000 $0 $7,400 $7,200 17 0.3 5000 $0 $8,450 $9,500 18 19 Exp.Val. $0 $7,085 $7,200 20 21 22 H16 formula =MAX(C16:F16) copied to H16:H18 23 C20 formula copied to H20 F Large $4,500 $7,000 $9,500 $7,000 G H Payoff Using Perfect Prediction $5,300 $7,400 $9,500 $7,400 I

EVPI = EVPP EVUU = $7,400 $7,200 = $200

230

Chapter 19 Value of Information in Decision Trees

Figure 19.10 Basic Probability Decision Tree

High Sales Introduce Product

Low Sales Success Prediction

Don't Introduce

High Sales Introduce Product

Low Sales Market Survey Inconclusive

Don't Introduce

High Sales Introduce Product

Low Sales Failure Prediction

Don't Introduce

High Sales Introduce Product

Low Sales Don't Survey

Don't Introduce

19.3 DriveTek Post-Contract-Award Problem

231

Figure 19.11 DriveTek EVPI Magnetic Success/Failure

Use mechanical method +$80,000 +$80,000 0.5 Electronic success +$150,000 0.5 Awarded contract 2 +$90,000 Try electronic method +$90,000 +$150,000 0.5 Electronic failure +$30,000 +$30,000

0.7
Magnetic success Prepare proposal Try magnetic method +$20,000 +$84,000 +$120,000 +$120,000

0.3
Magnetic failure $0 $0

No Additional Information 1 +$20,000 +$20,000 1 0.5 Not awarded contract -$50,000 -$50,000

Don't prepare proposal $0 $0

Use mechanical method +$80,000 +$80,000 0.5 Electronic success +$150,000 0.5 Awarded contract 3 +$120,000 Try electronic method +$90,000 +$150,000 0.5 Electronic failure +$30,000 +$30,000

1.0
2 +$30,500 Prepare proposal Try magnetic method +$35,000 +$120,000 +$120,000 Magnetic success +$120,000

0.0
Magnetic failure $0 $0

0.7
"Magnetic Success" 1 +$35,000 0.5 Not awarded contract

-$50,000 -$50,000

Don't prepare proposal $0 $0

Use mechanical method +$80,000 +$80,000 0.5 Electronic success +$150,000 +$30,500 0.5 Awarded contract 2 +$90,000 Try electronic method +$90,000 +$150,000 0.5 Electronic failure +$30,000 +$30,000

Perfect Prediction

0.0
Magnetic success Prepare proposal Try magnetic method +$20,000 $0 +$120,000 +$120,000

1.0
Magnetic failure $0 $0

0.3
"Magnetic Failure" 1 +$20,000 0.5 Not awarded contract

-$50,000 -$50,000

Don't prepare proposal $0 $0

19.3 DRIVETEK POST-CONTRACT-AWARD PROBLEM

DriveTek decided to prepare the proposal, and it turned out that they were awarded the contract. The $50,000 cost and $250,000 up-front payment are in the past. The current decision is to determine which method to use to satisfy the contract at minimum expected cost.

232

Chapter 19 Value of Information in Decision Trees

The following decision trees show costs for cash flows, terminal values, and rollback values. The rollback method uses TreePlans option to minimize cost of immediate successors. Figure 19.12 Costs for Cash Flows and Terminal Values
Use mechanical $120,000 $120,000 0.5 Electronic success $50,000 Try electronic $50,000 $0 0.5 Electronic failure $170,000 $120,000 0.7 Magnetic success $80,000 Try magnetic $80,000 $0 0.3 Magnetic failure $200,000 $120,000

Figure 19.13 Expected Cost Under Uncertainty

Use mechanical $120,000 $120,000 0.5 Electronic success $50,000 Try electronic 2 $110,000 $110,000 $50,000 0.5 Electronic failure $170,000 $170,000 0.7 Magnetic success $80,000 Try magnetic $116,000 0.3 Magnetic failure $200,000 $200,000 $80,000

ECUU = Expected Cost Under Uncertainty = $110,000

19.3 DriveTek Post-Contract-Award Problem

233

Figure 19.14 Expected Cost with Perfect Prediction for Electronic Uncertainty
Use mechanical $120,000 $120,000 1.0 Electronic success $50,000 0.5 "Electronic success" 2 $50,000 Try electronic $50,000 $50,000 0.0 Electronic failure $170,000 $170,000 0.7 Magnetic success $80,000 Try magnetic $116,000 0.3 Magnetic failure $200,000 $83,000 $200,000 $80,000

Use mechanical $120,000 $120,000 0.0 Electronic success $50,000 0.5 "Electronic failure" 3 $116,000 Try electronic $170,000 $50,000 1.0 Electronic failure $170,000 $170,000 0.7 Magnetic success $80,000 Try magnetic $116,000 0.3 Magnetic failure $200,000 $200,000 $80,000

ECPP(Electronic) = Expected Cost with Perfect Prediction for Electronic Uncertainty = $83,000 EVPI(Electronic) = ECUU ECPP(Electronic) = $110,000 $83,000 = $27,000

234

Chapter 19 Value of Information in Decision Trees

Figure 19.15 Expected Cost with Perfect Prediction for Magnetic Uncertainty
Use mechanical $120,000 $120,000 0.5 Electronic success $50,000 0.7 "Magnetic success" 3 $80,000 Try electronic $110,000 $50,000 0.5 Electronic failure $170,000 $170,000 1.0 Magnetic success $80,000 Try magnetic $80,000 0.0 Magnetic failure $200,000 $89,000 $200,000 $80,000

Use mechanical $120,000 $120,000 0.5 Electronic success $50,000 0.3 "Magnetic failure" 2 $110,000 Try electronic $110,000 $50,000 0.5 Electronic failure $170,000 $170,000 0.0 Magnetic success $80,000 Try magnetic $200,000 1.0 Magnetic failure $200,000 $200,000 $80,000

ECPP(Magnetic) = Expected Cost with Perfect Prediction for Magnetic Uncertainty = $89,000 EVPI(Magnetic) = ECUU ECPP(Magnetic) = $110,000 $89,000 = $21,000

19.3 DriveTek Post-Contract-Award Problem

235

Figure 19.16 Expected Cost with Perfect Prediction for Both Uncertainties
Use mechanical $120,000 $120,000 1.0 Electronic success $50,000 0.7 "Magnetic success" 2 0 $50,000 Try electronic $50,000 $50,000 0.0 Electronic failure $170,000 $170,000 1.0 Magnetic success $80,000 Try magnetic 0.5 "Electronic success" $50,000 $80,000 0.0 Magnetic failure $200,000 $200,000 $80,000

Use mechanical $120,000 $120,000 1.0 Electronic success $50,000 0.3 "Magnetic failure" 2 0 $50,000 Try electronic $50,000 $50,000 0.0 Electronic failure $170,000 $170,000 0.0 Magnetic success $80,000 Try magnetic $200,000 1.0 Magnetic failure $200,000 $71,000 $200,000 $80,000

Use mechanical $120,000 $120,000 0.0 Electronic success $50,000 0.7 "Magnetic success" 3 0 $80,000 Try electronic $170,000 $50,000 1.0 Electronic failure $170,000 $170,000 1.0 Magnetic success $80,000 Try magnetic 0.5 "Electronic failure" $92,000 $80,000 0.0 Magnetic failure $200,000 $200,000 $80,000

Use mechanical $120,000 $120,000 0.0 Electronic success $50,000 0.3 "Magnetic failure" 1 0 $120,000 Try electronic $170,000 $50,000 1.0 Electronic failure $170,000 $170,000 0.0 Magnetic success $80,000 Try magnetic $200,000 1.0 Magnetic failure $200,000 $200,000 $80,000

236

Chapter 19 Value of Information in Decision Trees

ECPP(Both) = Expected Cost with Perfect Prediction for Both Uncertainties = $71,000 EVPI(Both) = ECUU ECPP(Both) = $110,000 $71,000 = $39,000 EVPI(Electronic) + EVPI(Magnetic) = $27,000 + $21,000 = $48,000 Here, EVPI(Both) EVPI(Electronic) + EVPI(Magnetic) And, in general, as here, EVPIs are not additive. In some special cases, EVPI(Two Events) = EVPI(First Event) + EVPI(Second Event)

19.4 SENSITIVITY ANALYSIS VS EVPI

Working Paper Title: Do Sensitivity Analyses Really Capture Problem Sensitivity? An Empirical Analysis Based on Information Value Authors: James C. Felli, Naval Postgraduate School and Gordon B. Hazen, Northwestern University Date: March 1998 The most common methods of sensitivity analysis (SA) in decision-analytic modeling are based either on proximity in parameter-space to decision thresholds or on the range of payoffs that accompany parameter variation. As an alternative, we propose the use of the expected value of perfect information (EVPI) as a sensitivity measure and argue from first principles that it is the proper measure of decision sensitivity. EVPI has significant advantages over conventional SA, especially in the multiparametric case, where graphical SA breaks down. In realistically sized problems, simple one- and two-way SAs may not fully capture parameter interactions, raising the disturbing possibility that many published decision analyses might be overconfident in their policy recommendations. To investigate the extent of this potential problem, we re-examined 25 decision analyses drawn from the published literature and calculated EVPI values for parameters on which sensitivity analyses had been performed, as well as the entire set of problem parameters. While we expected EVPI values to indicate greater problem sensitivity than conventional SA due to revealed parameter interaction, we in fact found the opposite: compared to EVPI, the one- and twoparameter SAs accompanying these problems dramatically overestimated problem sensitivity to input parameters. This phenomenon can be explained by invoking the flat maxima principle enunciated by von Winterfeldt and Edwards. http://www.mccombs.utexas.edu/faculty/jim.dyer/DA_WP/WP980019.pdf