Decision Analysis under Risk and Uncertainty

Introduction

Managers routinely make decisions before outcomes are known. A firm may choose a production capacity before demand is observed, enter a market before competitors’ reactions are clear, or approve an investment before regulatory conditions are settled. Decision analysis provides a disciplined way to represent these choices, incorporate evidence, compare alternatives, and recognize the limits of what is known.

The central distinction is between risk and uncertainty. Under risk, relevant states of nature are known and probabilities can be assigned to them. Under uncertainty, the states may be known but credible probabilities are unavailable, disputed, or highly subjective. The distinction is practical rather than absolute: the same strategic situation may be treated as risk by one organization with extensive data and as uncertainty by another with limited evidence.

Decision analysis does not eliminate uncertainty. It makes assumptions visible, clarifies trade-offs, and shows how conclusions change when assumptions change. It is therefore especially useful for strategic decisions involving irreversible commitments, asymmetric consequences, and incomplete information.

Learning Objectives

By the end of this chapter, the reader should be able to:

1. Distinguish decisions under certainty, risk, and uncertainty.
2. Construct and interpret payoff tables.
3. Apply expected monetary value, expected opportunity loss, and utility-based criteria.
4. Build and solve decision trees, including sequential decisions.
5. Conduct sensitivity and break-even analysis.
6. Explain the value of perfect and sample information.
7. Use Bayes’ rule to update beliefs when new evidence arrives.
8. Apply introductory game-theoretic reasoning to competitive decisions.
9. Evaluate ethical, managerial, and governance implications of analytical decisions.

1. The Architecture of a Decision Problem

A well-defined decision problem contains five elements:

• Decision maker: the individual or organization choosing an action.
• Alternatives: feasible actions available to the decision maker.
• States of nature: external conditions not controlled by the decision maker.
• Outcomes or payoffs: consequences produced by each action–state combination.
• Information: prior knowledge, observations, forecasts, and signals relevant to the states.

A decision alternative should be described at the level at which management can actually act. “Grow” is often too vague; “build a 50,000-unit facility” is more operational. States of nature should be mutually exclusive and collectively exhaustive for the decision being analyzed. For example, demand states might be “low,” “medium,” and “high,” provided that every relevant demand outcome belongs to one and only one category.

A payoff may be profit, contribution margin, net present value, market share, service level, risk-adjusted return, or another measurable objective. Payoffs must be comparable. Mixing annual profit in one cell with lifetime revenue in another produces a formally neat but substantively invalid table.

Certainty, risk, and uncertainty

• Certainty: one relevant state is known to occur.
• Risk: several states are possible and probabilities are available.
• Uncertainty: several states are possible but probabilities are unavailable, unreliable, or contested.

The classification affects the decision rule. Under risk, expected value and expected utility are often appropriate. Under uncertainty, managers may use conservative, optimistic, regret-based, or robustness-oriented criteria, while recognizing that these rules express preferences toward ambiguity rather than discovering a single objectively correct answer.

2. Payoff Tables and Decision Criteria

A payoff table lists alternatives in rows and states of nature in columns. Each cell gives the payoff resulting from that combination.

Consider a firm choosing among a small, medium, or large facility. Estimated annual profit (in $000) is as follows:

Suppose management estimates that low demand has probability 0.60 and high demand has probability 0.40. The expected monetary value (EMV) of alternative a is:

Here, s indexes states of nature, p(s) is the probability of state s, and V(a,s) is the monetary payoff from action a when state s occurs. Probabilities must be non-negative and sum to 1; all payoffs must use the same monetary units and time basis.

The EMVs are:

• Small: 0.60(120) + 0.40(150) = $132,000.
• Medium: 0.60(80) + 0.40(260) = $152,000.
• Large: 0.60(−40) + 0.40(400) = $136,000.

On an expected-profit basis, the medium facility is preferred. This conclusion does not mean that the medium facility produces the highest profit in every state. It means that, given the stated probabilities and a risk-neutral objective measured in money, it has the highest average payoff.

Dominance

An alternative is dominated if another alternative produces at least as high a payoff in every state and a strictly higher payoff in at least one state. A dominated alternative can be removed when the payoffs and states fully capture the decision. Dominance is powerful because it does not require probabilities.

In the example, no facility is dominated: the small facility performs best under low demand, while the large facility performs best under high demand.

Expected opportunity loss

Opportunity loss, or regret, measures the difference between the best payoff available after the state is known and the payoff from the selected action. For each state:

The expected opportunity loss (EOL) is:

For the facility example, the best payoff is $120,000 under low demand and $400,000 under high demand.

The medium facility has the lowest EOL. Under the same probability model and payoff structure, the minimum EOL equals the value of perfect information, as explained later.

Decision criteria under uncertainty

When probabilities cannot be justified, several criteria are available. They should be presented as policy choices, not as objective laws.

For the facility table, maximax selects the large facility because its best outcome is 400; maximin selects the small facility because its worst outcome is 120. These different recommendations demonstrate that uncertainty cannot be handled without expressing an attitude toward adverse outcomes and ambiguity.

3. Risk Preferences and Utility

EMV treats each monetary unit as equally valuable regardless of the decision maker’s wealth or risk tolerance. This is the risk-neutral assumption. Many managers are risk-averse: they prefer a certain moderate outcome to a gamble with the same expected monetary value. Others may accept greater variability in pursuit of high upside.

Utility theory represents preferences through a utility function U(x), where x is a payoff and utility reflects the decision maker’s satisfaction or preference. The expected utility of alternative a is:

The preferred action maximizes expected utility, not necessarily expected money. Utility is ordinal: a difference of 10 utility units is not automatically twice as desirable as a difference of 5 units. Utility values should be elicited consistently from the decision maker or estimated from observed preferences.

A common risk-averse utility function is the exponential form:

where r > 0 is the coefficient of absolute risk aversion and e is the base of the natural logarithm. This form is useful because it is increasing and concave in x: more money increases utility, but at a diminishing rate. Its use assumes that the exponential representation is an acceptable approximation and that the units of x are handled consistently; changing dollars to thousands of dollars requires adjusting r.

A simpler illustrative utility scale can be constructed by assigning utility 0 to the worst outcome and 1 to the best, then eliciting intermediate values through preference questions. Such a scale is meaningful only if the decision maker’s trade-offs are coherent and the consequences included are sufficiently complete.

Risk premium and certainty equivalent

The certainty equivalent (CE) is the certain monetary amount that gives the same utility as a risky prospect:

The risk premium (RP) is:

For a risk-averse decision maker, the risk premium is normally positive. Utility analysis is particularly important when outcomes include bankruptcy risk, reputational harm, safety consequences, career effects, or irreversible strategic commitments that are not adequately represented by accounting profit.

4. Decision Trees and Sequential Decisions

A payoff table is appropriate when the decision is made once and all relevant information is summarized in advance. A decision tree is more suitable when decisions occur sequentially and later choices depend on observed events.

Standard notation uses:

• Square nodes for decision points.
• Circular nodes for chance events.
• Branches for alternatives or states.
• Terminal nodes for final payoffs.

The tree is solved by folding back from right to left. At a chance node, calculate the probability-weighted value of branches. At a decision node, select the branch with the highest expected value or expected utility, subject to the decision maker’s objective.

Figure: A sequential decision tree in which research evidence informs a later launch decision.

Worked decision-tree example

A company can launch a product immediately or pay $20,000 for a market test. Without testing, the probability of strong demand is 0.40. Immediate launch yields $180,000 if demand is strong and −$100,000 if demand is weak. Not launching yields $0.

The EMV of immediate launch is:

Therefore, without testing, launch is preferred to not launching.

Suppose the test produces a favourable or unfavourable signal. Historical validation indicates:

• P(favourable signal | strong demand) = 0.80
• P(favourable signal | weak demand) = 0.20

The probability of a favourable signal is:

Thus, P(F) = 0.80(0.40) + 0.20(0.60) = 0.44, and P(U) = 0.56. Bayes’ rule gives:

So, P(S|F) = 0.32 ÷ 0.44 ≈ 0.727. After a favourable signal, the EMV of launching is approximately 0.727(180,000) + 0.273(−100,000) = $103,640, so launch.

For an unfavourable signal, P(S|U) = 0.08 ÷ 0.56 ≈ 0.143. The EMV of launching is approximately 0.143(180,000) + 0.857(−100,000) = −$59,260, so stop.

The expected value after testing, before paying for it, is:

Subtracting the test cost gives approximately $25,600. Since this exceeds the $12,000 value of immediate launch, the test is preferred under risk neutrality. The test changes the action in some information states and creates value through selective commitment.

5. Sensitivity and Break-Even Analysis

A recommendation is only as credible as its robustness. Sensitivity analysis varies one or more inputs and observes whether the preferred alternative changes. Important inputs include probabilities, payoffs, costs, demand forecasts, discount rates, and utility parameters.

For the facility example, compare medium and large facilities. Let p be the probability of high demand. The medium facility is preferred when:

Expanding gives 80 + 180p ≥ −40 + 440p. Therefore, 120 ≥ 260p, or:

If the probability of high demand is below approximately 46.15%, medium is preferred; above that threshold, large is preferred. This is a break-even probability. It tells management exactly which belief must change to reverse the decision.

Useful sensitivity displays include:

• One-way analysis: vary one input while holding others constant.
• Two-way analysis: vary two inputs to identify regions of preferred action.
• Scenario analysis: assess coherent combinations such as recession, base case, and expansion.
• Tornado diagrams: rank variables by their influence on the outcome.
• Monte Carlo simulation: assign distributions to uncertain inputs and repeatedly sample outcomes.

Sensitivity analysis is not a substitute for better data. It identifies which data are worth improving and which assumptions are immaterial to the decision.

6. The Value of Information

Information has value only if it can change an action or improve the outcome conditional on what is learned. Information that merely confirms a decision without affecting behavior may have little or no decision value, although it may have compliance or learning value.

Expected value of perfect information

Perfect information reveals the true state before the decision is made. Its expected value is:

For the facility example, the expected payoff with perfect information is:

0.60(120) + 0.40(400) = $232,000.

The best EMV without perfect information is $152,000. Therefore:

This equals the minimum EOL in the table. EVPI is an upper bound: no imperfect research should cost more than the value of perfect information when the objective and model are otherwise unchanged.

Expected value of sample information

Sample information is imperfect evidence, such as a survey, pilot, forecast, or market test. Its expected value before acquisition is:

The net value is:

In the product-test example, the value with testing before cost is $45,600, while the best no-test value is $12,000. EVSI is $33,600. After a $20,000 test cost, the net value is $13,600, which is positive.

The analysis assumes that probabilities and test characteristics are correctly specified, that the organization will follow the optimal contingent policy, and that the test cost includes all relevant direct and indirect costs.

7. Bayesian Updating

Bayesian analysis formalizes learning from evidence. A prior probability expresses beliefs before observing evidence; a likelihood describes how probable the evidence is under each state; the posterior probability expresses revised beliefs.

For state s and evidence e:

where:

The denominator ensures that posterior probabilities sum to 1. Bayesian updating is not restricted to subjective beliefs: priors may come from historical frequencies, while likelihoods may come from validated experiments or predictive models.

Good practice requires separating:

• the state being predicted,
• the evidence actually observed,
• the reliability of the evidence,
• the prior information source, and
• the decision consequence of the posterior belief.

A common error is to confuse P(evidence | state) with P(state | evidence). A highly accurate alarm does not automatically imply that a detected condition is highly probable; the base rate matters.

In organizations, Bayesian methods support credit scoring, fraud detection, demand forecasting, medical and safety screening, and adaptive strategy. Governance is essential because biased historical data can produce biased priors or likelihood estimates, and because a mathematically correct posterior can still be ethically unacceptable if the underlying data collection is discriminatory.

8. Introductory Game-Theoretic Reasoning

Decision analysis often treats the environment as passive. In competitive markets, however, another decision maker may respond strategically. Game theory studies situations in which each player’s payoff depends partly on the actions of others.

A normal-form game can be represented by a payoff matrix. Consider two competing firms choosing High price or Low price. Each cell lists payoffs as (Firm A, Firm B):

A dominant strategy yields at least as good a payoff as every alternative regardless of the opponent’s action. Here, Low is dominant for each firm: 10 exceeds 8 when the rival chooses High, and 5 exceeds 3 when the rival chooses Low. The resulting Nash equilibrium is (Low, Low), because neither firm can improve its payoff by changing strategy unilaterally.

A Nash equilibrium is stable given the other players’ strategies; it is not necessarily socially optimal. Both firms would prefer (High, High), but each has an incentive to undercut the other. This tension resembles a prisoner’s dilemma.

Important cautions include:

• Payoffs must include relevant future effects, not merely one-period profit.
• Repeated interaction can support cooperation through credible retaliation or reward.
• A threat matters only if it is credible and feasible.
• Mixed strategies may be relevant when no pure strategy is stable.
• Regulation and competition law constrain collusion and exclusionary conduct.

Game-theoretic reasoning should not be used to claim certainty about competitor behavior. It is a structured way to explore incentives, reactions, and strategic vulnerabilities.

9. Illustrative Case: Northstar’s Capacity Decision

Background

Northstar Components, a fictional industrial supplier, must choose whether to expand production before a major customer finalizes its procurement plan. The expansion requires a non-recoverable investment. Demand may be high, medium, or low. Management has historical forecasts but recognizes that a competitor may enter the market.

Problem

The alternatives are to maintain current capacity, expand moderately, or build a large facility. Estimated three-year incremental net present values, in millions of dollars, are:

The forecast probabilities are 0.35, 0.40, and 0.25 respectively.

Analysis

The EMVs are:

• Maintain = 0.35(1.0) + 0.40(1.2) + 0.25(1.3) = 1.145 million.
• Moderate = 0.35(0.2) + 0.40(2.0) + 0.25(3.0) = 1.670 million.
• Large = 0.35(−2.0) + 0.40(1.0) + 0.25(5.5) = 1.025 million.

A risk-neutral analysis therefore selects moderate expansion. However, the large expansion has substantial upside and a −$2 million downside. If Northstar faces severe liquidity constraints or values survival more than expected NPV, expected utility may select maintain capacity or moderate expansion even more strongly.

Management then considers a customer survey costing $100,000. Historical validation suggests that a positive survey is more likely when demand is high, while a negative survey is more likely when demand is low. The team should calculate posterior probabilities for each survey outcome, solve the decision tree, and compare EVSI with the $100,000 cost. It should also conduct sensitivity analysis around the high-demand probability and the probability that a competitor enters.

Approach and solution

The recommended governance process is:

1. Document the three alternatives and the definition of each demand state.
2. Validate that all NPVs use the same discount rate, horizon, and treatment of taxes and working capital.
3. Elicit probabilities from historical data and expert judgment separately, then reconcile differences.
4. Model the survey’s sensitivity and specificity rather than assuming it is informative.
5. Calculate EMV and expected utility under liquidity and risk constraints.
6. Identify the break-even probability at which large expansion overtakes moderate expansion.
7. Establish a staged investment option if capacity can be expanded in modules.
8. Specify what evidence would trigger acceleration, delay, or cancellation.

Outcome

The numerical evidence supports moderate expansion, but the most strategically robust solution may be a staged commitment: secure permits and long-lead equipment while postponing the final construction decision until customer evidence improves. This option should be modeled explicitly because real options—such as the ability to defer, abandon, or expand—can make a flexible strategy more valuable than a static payoff table suggests.

Learning points

• The highest possible payoff is not automatically the best decision.
• Forecast probabilities and payoff estimates should be stress-tested.
• Information has value only when it affects action.
• Flexibility can be a strategic asset.
• A recommendation should state both its preferred action and the conditions that would reverse it.

10. Limitations, Ethics, and Best Practice

Decision models can create false precision. A result such as $1.67 million may appear exact even when probabilities, demand categories, and cash-flow forecasts are uncertain. Analysts should report ranges, scenarios, and key assumptions rather than only a single number.

Important limitations include model misspecification, omitted outcomes, dependence among uncertainties, behavioral bias, poor probability calibration, and data instability. Correlated risks should not be treated as independent merely for convenience. For example, recession may simultaneously reduce demand, increase financing costs, and increase competitor distress.

Ethically responsible decision analysis requires:

• transparency about assumptions and data sources;
• sensitivity analysis for vulnerable stakeholders;
• avoidance of discriminatory proxy variables;
• protection of confidential and personal information;
• documentation of who selected the objective function;
• human review where automated decisions affect rights or livelihoods;
• mechanisms for appeal, audit, and correction.

AI and predictive analytics can improve scenario generation and Bayesian forecasting, but they do not determine organizational values. Generative systems may produce unsupported probabilities, while machine-learning systems may perform poorly under distributional change. Human experts should validate model structure, monitor drift, and distinguish statistical prediction from causal explanation.

Best practice is to combine quantitative analysis with managerial judgment through a documented cycle: define the decision, model alternatives and states, estimate probabilities and payoffs, calculate the preferred policy, test sensitivity, assess information value, examine ethical consequences, and monitor outcomes after implementation. Post-decision review is essential for learning, but it should distinguish a poor decision made with reasonable information from a good decision that happened to encounter an adverse outcome.

Chapter Summary

Decision analysis provides a structured approach to choices involving uncertain outcomes. Payoff tables organize alternatives, states, probabilities, and consequences. Under risk, expected monetary value compares probability-weighted payoffs, while expected opportunity loss measures the cost of not choosing the best action after the state is known. Under uncertainty, maximax, maximin, minimax regret, Hurwicz, and Laplace criteria express different attitudes toward ambiguity.

Expected monetary value assumes risk neutrality. Expected utility incorporates risk preferences, and the certainty equivalent translates a risky prospect into an equivalent certain amount. Decision trees extend payoff tables to sequential decisions by combining chance probabilities with later choices and solving through backward induction.

Sensitivity and break-even analysis reveal whether recommendations are robust. The expected value of perfect information is an upper bound on what information could be worth; the expected value of sample information accounts for imperfect evidence and should be compared with acquisition cost. Bayes’ rule updates prior beliefs using new evidence and supports adaptive decisions. Game-theoretic reasoning adds strategic interdependence by examining incentives, dominant strategies, Nash equilibrium, credibility, and competitor responses.

Effective decision analysis is not merely calculation. It requires valid definitions, consistent units, credible data, explicit assumptions, ethical governance, and recognition of model limitations. The strongest managerial recommendation identifies a preferred action, quantifies uncertainty, states the conditions that would reverse the choice, and preserves flexibility where possible.

Key Terms

Decision alternative: An action available to the decision maker.

State of nature: An external condition not controlled by the decision maker.

Payoff: The measurable consequence of an action under a particular state.

Risk: A situation in which probabilities can be assigned to possible states.

Uncertainty: A situation in which probabilities are unavailable, unreliable, or disputed.

Expected monetary value: The probability-weighted average monetary payoff of an alternative.

Opportunity loss: The difference between the best payoff in a state and the payoff from a selected action.

Expected opportunity loss: The probability-weighted average opportunity loss.

Dominance: A relationship in which one alternative is at least as good in every state and better in at least one.

Utility: A representation of preference or satisfaction associated with an outcome.

Risk aversion: Preference for less risky outcomes when expected monetary value is otherwise comparable.

Certainty equivalent: A certain amount producing the same utility as a risky prospect.

Decision tree: A graphical representation of sequential decisions and chance events.

Sensitivity analysis: Examination of how results change when inputs vary.

Value of information: The improvement in expected decision outcome attributable to information.

Perfect information: Information that reveals the true state before action is selected.

Bayes’ rule: A method for updating probabilities using observed evidence.

Nash equilibrium: A set of strategies from which no player benefits by deviating alone.

Real option: The value of managerial flexibility to defer, expand, contract, or abandon an action.

Model risk: The possibility that a model’s structure, assumptions, or data are unsuitable.

Review Questions

5 Short Answer Questions

1. Distinguish decision-making under risk from decision-making under uncertainty.
2. What is the purpose of a payoff table?
3. Why can an alternative with the highest possible payoff fail to have the highest EMV?
4. Define expected value of perfect information.
5. What is the difference between a prior and a posterior probability?

5 Long Answer Questions

1. Explain how EMV, expected utility, and certainty equivalent differ, and identify situations in which each is useful.
2. Describe how a decision tree is constructed and solved using backward induction.
3. Compare maximax, maximin, minimax regret, Hurwicz, and Laplace criteria under uncertainty.
4. Explain how sensitivity analysis improves the quality and credibility of strategic recommendations.
5. Discuss the ethical and governance risks associated with using AI-generated probabilities in high-impact decisions.

3 Analytical / Application Questions

1. A firm must choose between two projects. Project A pays $100,000 in a favourable market and $40,000 in an unfavourable market. Project B pays $180,000 in a favourable market and loses $20,000 in an unfavourable market. If the probability of a favourable market is 0.55, calculate each project’s EMV and identify the preferred project under risk neutrality. What additional information would be needed to assess a risk-averse decision maker?
2. Using Bayes’ rule, analyze a screening test with P(condition) = 0.10, P(positive | condition) = 0.90, and P(positive | no condition) = 0.05. Calculate P(condition | positive) and explain why the result differs from the test’s sensitivity.
3. Construct a two-player payoff matrix for a competitive pricing decision in which each firm has High and Low price strategies. Identify whether either player has a dominant strategy and determine any Nash equilibrium. Explain how repeated interaction or regulation might change the strategic outcome.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes risk from uncertainty?

A. Risk has no possible outcomes, whereas uncertainty has many outcomes.

B. Risk permits credible probability assignments, whereas uncertainty does not.

C. Risk concerns only financial outcomes, whereas uncertainty concerns only nonfinancial outcomes.

D. Risk is always controllable, whereas uncertainty is always uncontrollable.

Correct Answer: B. Risk permits credible probability assignments, whereas uncertainty does not.

Question 2

Using the facility example, which alternative has the highest EMV when low-demand probability is 0.60 and high-demand probability is 0.40?

A. Small facility, $132,000

B. Medium facility, $152,000

C. Large facility, $136,000

D. All alternatives have the same EMV

Correct Answer: B. Medium facility, $152,000

Question 3

What does a dominated alternative have in common with another alternative?

A. It has a higher payoff in every state.

B. It has identical payoffs in every state.

C. Another alternative is at least as good in every state and better in at least one.

D. It has the highest possible payoff but the lowest expected payoff.

Correct Answer: C. Another alternative is at least as good in every state and better in at least one.

Question 4

For a risk-averse decision maker, the certainty equivalent of a risky prospect is generally:

A. Greater than its EMV

B. Equal to its maximum possible payoff

C. Less than its EMV

D. Equal to its worst possible payoff

Correct Answer: C. Less than its EMV

Question 5

In a decision tree, what operation is performed at a chance node during fold-back analysis?

A. Select the branch with the largest payoff without using probabilities.

B. Calculate the probability-weighted value of its outgoing branches.

C. Discard all branches with negative payoffs.

D. Replace every payoff with its utility value automatically.

Correct Answer: B. Calculate the probability-weighted value of its outgoing branches.

Question 6

For the facility example, what is the break-even probability of high demand at which medium and large facilities have equal EMV?

A. 0.2500

B. 0.4000

C. 0.4615 approximately

D. 0.6000

Correct Answer: C. 0.4615 approximately

Question 7

Which expression represents expected value of perfect information?

A. Best EMV minus expected payoff with perfect information

B. Expected payoff with perfect information minus best EMV

C. Maximum payoff minus minimum payoff

D. Expected opportunity loss plus research cost

Correct Answer: B. Expected payoff with perfect information minus best EMV

Question 8

Bayes’ rule is primarily used to:

A. Convert monetary payoffs into utilities

B. Identify a dominant strategy in a game

C. Update probabilities after observing evidence

D. Calculate the maximum possible payoff

Correct Answer: C. Update probabilities after observing evidence

Question 9

In the pricing game presented, why is Low price a dominant strategy for each firm?

A. Low price gives the same payoff regardless of the rival’s action.

B. Low price gives at least as high a payoff against either rival action and a higher payoff in each comparison.

C. High price is illegal in the market.

D. Both firms agreed in advance to choose Low price.

Correct Answer: B. Low price gives at least as high a payoff against either rival action and a higher payoff in each comparison.

Question 10

What is the principal purpose of sensitivity analysis?

A. To guarantee that the selected decision will succeed

B. To remove all uncertainty from the model

C. To determine how changes in assumptions affect the recommendation

D. To replace managerial judgment with a single forecast

Correct Answer: C. To determine how changes in assumptions affect the recommendation