Decision Analysis Under Risk and Uncertainty

Introduction

Operations managers routinely make decisions before the future is known. A firm may need to choose plant capacity before demand is observed, approve a product launch before market acceptance is clear, or select a supplier before prices and reliability are fully established. Decision analysis provides a disciplined way to represent such choices, incorporate information about uncertain conditions, and compare alternatives transparently.

The central distinction is between risk and uncertainty. Under risk, the relevant future states are known and their probabilities can be estimated with reasonable confidence. Under uncertainty, the states may be known but reliable probabilities are unavailable, disputed, or too sensitive to assumptions. In both settings, the decision maker must separate three elements:

1. Actions or alternatives that management can control.
2. States of nature that influence outcomes but cannot be controlled.
3. Consequences or payoffs resulting from each action–state combination.

Decision analysis does not eliminate uncertainty. Its value lies in making assumptions explicit, identifying trade-offs, testing sensitivity, and clarifying what additional information would be worth obtaining.

Learning Objectives

After studying this chapter, you should be able to:

• construct and interpret payoff tables;
• distinguish decisions under certainty, risk and uncertainty;
• calculate expected monetary value and expected opportunity loss;
• construct and solve decision trees for sequential decisions;
• apply utility theory when monetary outcomes do not adequately represent preferences;
• update probabilities using Bayes’ theorem;
• evaluate the expected value of perfect and sample information;
• conduct sensitivity, robust and scenario analysis;
• recognise behavioural, ethical and governance limitations in quantitative decision models.

1. The Structure of a Decision Problem

A well-formed decision problem defines the decision timing, feasible actions, uncertain states, payoffs, probabilities where available, and the objective or criterion used to select an action.

Let action a_i be one of i = 1, 2, …, m, and let state of nature s_j be one of j = 1, 2, …, n. The payoff from action a_i when state s_j occurs is denoted by v_ij. Payoffs may be profits, costs, service levels, contribution margins, patient outcomes, or a multidimensional performance measure converted into a common scale.

A payoff table has actions in rows and states in columns. The entry at their intersection is the consequence of that combination. The table must use a consistent perspective: for example, all entries may be net present value in dollars, or all may be annual operating costs. Mixing revenues, costs and profits without adjustment leads to invalid comparisons.

Decision environments

Probabilities should represent uncertainty about states, not preferences for actions. A probability distribution is therefore conceptually different from a manager’s utility function.

2. Payoff Tables and Basic Criteria

Expected monetary value

When probabilities p_j are assigned to states, with p_j ≥ 0 and ∑_j=1ⁿ p_j = 1, the expected monetary value (EMV) of action a_i is:

Here, v_ij is measured in the same monetary unit for every state. EMV is appropriate when repeated decisions are made under similar conditions, or when the decision maker is approximately risk neutral. For a one-time decision, EMV alone may be inappropriate if large losses threaten organisational survival.

Worked example: capacity selection

An operations manager must choose a small, medium or large facility. Demand may be low, medium or high. Estimated annual profits are as follows, in thousands of currency units.

For the medium facility:

For the large facility:

The small facility has the highest EMV, 420 thousand. However, the medium facility produces a higher payoff than the small facility in medium and high demand. The recommendation therefore depends on the organisation’s tolerance for downside exposure, growth strategy and confidence in the probabilities.

Expected opportunity loss

Regret, or opportunity loss, measures the cost of not selecting the best action after the state becomes known. For each state, identify the maximum payoff and subtract each action’s payoff from that maximum:

The expected opportunity loss (EOL) is:

The action with minimum EOL is equivalent to the action with maximum EMV when the same payoff table and probabilities are used. In the example, the statewise maxima are 420, 560 and 820. EOL values are:

• Small: 0.30(0)+0.45(140)+0.25(400)=163 thousand.
• Medium: 0.30(120)+0.45(0)+0.25(260)=101 thousand.
• Large: 0.30(320)+0.45(160)+0.25(0)=168 thousand.

The medium facility has the lowest EOL, confirming that it has the highest EMV under these probabilities. EOL is particularly useful for explaining the cost of an imperfect decision.

Decision criteria under uncertainty

When probabilities are not defensible, several criteria may be used. Each expresses a different attitude toward uncertainty.

• Maximax: choose the action with the highest possible payoff. This is optimistic and favours opportunities with high upside.
• Maximin: choose the action whose worst payoff is highest. This is conservative and protects against the most adverse state.
• Minimax regret: calculate the regret table and choose the action with the smallest maximum regret.
• Laplace criterion: assume states are equally likely and maximise the average payoff. This is a formal convention, not evidence that states truly have equal probabilities.
• Hurwicz criterion: combine the best and worst outcomes using optimism coefficient α, where 0 ≤ α ≤ 1:

The choice among these criteria should be justified, not hidden. A robust analysis often reports several criteria rather than pretending that one is universally correct.

3. Decision Trees and Sequential Decisions

A payoff table is suitable for a single decision made before the state is observed. A decision tree is more appropriate when decisions occur sequentially and later choices depend on new information.

A decision node is usually represented conceptually as a choice point; a chance node represents an uncertain event. Branches from a chance node must have probabilities that sum to one. The tree is solved by rollback: calculate expected values at the final chance nodes, select the best action at each decision node, and move backward to the initial decision.

Figure: A generic sequential decision tree. Squares indicate decision points and circles indicate chance events.

Worked sequential example

A company can launch immediately, conduct a market test, or abandon the project. A favourable test occurs with probability 0.40 and costs 30 thousand. Conditional on a favourable test, the probability of high demand is 0.75; conditional on an unfavourable test, it is 0.20. A launch yields 500 thousand in high demand and 100 thousand in low demand. Deferral after testing yields zero operating payoff but avoids launch expenditure; abandoning yields zero. Assume the launch payoffs already include launch cost.

After a favourable test, the expected launch payoff is:

After an unfavourable test, it is:

Thus launch is preferred after either test result if deferral yields zero. The expected value of testing is:

If no test is conducted, the prior probability of high demand is 0.40(0.75)+0.60(0.20)=0.42. Immediate launch has expected payoff:

The no-test launch is therefore preferable to testing in this illustrative case: 268 thousand versus 238 thousand. The test is informative, but its cost exceeds the benefit of tailoring the decision to its result. This conclusion changes if the test is cheaper, more accurate, or if launching under low demand has a much larger loss.

4. Utility Theory and Risk Attitudes

Expected monetary value assumes that a monetary gain of 100 has the same decision significance wherever it occurs. Managers and organisations are often risk averse: they prefer a certain moderate return to a gamble with the same expected return, particularly when losses threaten liquidity, employment or regulatory compliance.

A utility function U(x) maps an outcome x into preference units. The expected utility of an action is:

An action with the greatest expected utility is preferred under the specified utility model. A risk-averse utility function is concave; for example, for positive wealth, U(W)=√W is concave. A linear function represents risk neutrality, while a convex function represents risk seeking over the relevant range.

Utility functions must be elicited carefully. One common method identifies a best outcome and a worst outcome, then asks the decision maker to compare intermediate outcomes with lotteries between these anchors. The certainty equivalent is the sure monetary amount regarded as equally desirable to the risky prospect. The difference between expected value and certainty equivalent is the risk premium:

Utility analysis is not a licence to conceal distributional consequences. A project can have high expected utility for shareholders while imposing unacceptable risks on workers, communities or customers. Multiple stakeholders and constraints may need to be represented explicitly.

5. Bayesian Updating

Probabilities should be revised when credible evidence arrives. Let [[MATH(H)]] denote a hypothesis or state, and [[MATH(E)]] new evidence. Bayes’ theorem is:

where P(H) is the prior probability, P(E|H) is the likelihood of observing the evidence if the hypothesis is true, and P(H|E) is the posterior probability. For mutually exclusive and exhaustive hypotheses H_k:

Example: demand forecast evidence

Suppose high demand has prior probability 0.40. A forecasting signal is favourable with probability 0.80 when demand is high and 0.30 when demand is low. Then:

A favourable signal raises the probability of high demand from 0.40 to 0.64. The posterior should then be used in the subsequent decision tree, not the original prior. A common error is to use the signal’s accuracy as the posterior probability without applying the base rate.

Bayesian analysis is particularly valuable in predictive maintenance, quality control, demand sensing, fraud detection and clinical operations. Its quality depends on the prior, the likelihood model and the independence assumptions used to combine evidence.

6. The Value of Information

Information has economic value only when it can change an action or improve outcomes. The expected value of perfect information (EVPI) is the most the decision maker should pay to learn the true state before acting.

Let EMV^* be the maximum EMV before information and let the best payoff in each state be selected after perfect information. Then:

EVPI is also the minimum expected opportunity loss across actions. It is an upper bound: no imperfect forecast can be worth more than perfect information.

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

The best EMV without information is 420, so:

A sample or market test has an expected value of sample information (EVSI) equal to the expected value after using the test, before paying for it, minus the best EMV without the test. If the test costs C, its net value is EVSI − C. The test should be purchased only if this net value is positive, subject to strategic, ethical and operational considerations.

Information may also have nonmonetary value, such as improving safety or compliance. Conversely, a technically accurate forecast can have low value if decisions cannot be changed in time.

7. Sensitivity, Robustness and Scenario Planning

Sensitivity analysis

Sensitivity analysis varies one or more assumptions to determine whether the recommended action changes. Important variables include state probabilities, demand volume, prices, costs, capacity, lead time and payoff estimates. A break-even probability can make the analysis actionable.

For two actions A and B with two states, A is preferred when:

Solving this inequality gives the probability threshold at which the recommendation changes. Sensitivity analysis should also test correlated changes, such as high demand accompanied by higher input prices, rather than varying every parameter independently.

Robust decision analysis

A recommendation is robust when it performs acceptably across a wide range of plausible assumptions, models or states. Robustness is not identical to choosing the highest average payoff. A robust alternative may sacrifice some upside to reduce severe downside or dependence on uncertain forecasts.

Useful practices include:

• reporting a range of outcomes rather than a single estimate;
• identifying assumptions that drive the decision;
• stress-testing severe but plausible conditions;
• comparing regret, downside loss and service performance;
• using constraints such as minimum cash flow or maximum tolerated disruption;
• selecting flexible actions that preserve future options.

A robust optimisation formulation may represent uncertain parameters θ in an uncertainty set Θ and seek an action that performs well under the worst plausible parameter value:

This is a modelling framework, not a universal prescription. An excessively broad uncertainty set can produce an unnecessarily conservative decision.

Scenario planning

Scenario planning develops internally coherent narratives about alternative futures, such as rapid growth, regulated transition, supply disruption or technological substitution. Scenarios are not necessarily probability forecasts. Their purpose is to challenge assumptions, identify leading indicators and design contingent actions.

A high-quality scenario process distinguishes:

• predetermined trends, such as contractual commitments;
• critical uncertainties, such as technology adoption;
• early warning indicators;
• actions that are robust across scenarios;
• trigger points for switching strategy.

Scenario planning complements, rather than replaces, quantitative analysis. A scenario can be translated into a payoff table, but forcing every narrative into a precise probability may create false precision.

8. Illustrative Case: Regional Distribution Centre

Background

A fictional consumer-goods company, Northstar Distribution, serves a rapidly growing region. It must decide whether to expand an existing warehouse, build a large automated centre, or lease flexible third-party capacity for two years. Demand is uncertain because a major competitor may enter the market.

Problem

The operations team estimates three demand states over the planning horizon: contraction, stable growth and rapid growth. It also recognises that automation has a high fixed commitment but lower unit handling cost. Finance requests a recommendation based only on expected profit; operations argues that service reliability and reversibility should also be considered.

Analysis

The team first constructs a payoff table in net present value. It then performs four analyses:

1. EMV analysis: uses a base probability distribution supplied by the forecasting group.
2. Stress analysis: doubles the probability of contraction and examines cash-flow consequences.
3. Scenario analysis: develops competitor-entry, regional-boom and supply-disruption narratives.
4. Option analysis: values the ability to expand leased capacity after observing demand.

The lease option has a lower expected profit than the automated centre in the regional-boom scenario, but it avoids irreversible capital expenditure in the competitor-entry scenario. The team therefore models the lease as a sequential decision: lease an initial capacity, observe demand after six months, then expand, maintain or exit.

Approach and solution

Rollback analysis shows that the value of the lease’s flexibility is material. However, the result depends on expansion lead time. If additional leased capacity can be secured within two weeks, the flexible option is preferred across most plausible probability distributions. If expansion requires six months, the service penalty makes the existing-warehouse expansion preferable.

The executive committee adopts a contingent policy rather than a single forecast-based choice:

• lease an initial block of capacity;
• monitor order growth, competitor entry and utilisation weekly;
• expand when utilisation exceeds a specified threshold for four consecutive weeks;
• retain a minimum service-level buffer;
• review the decision if supplier prices exceed a predefined trigger.

Outcome and learning points

The case illustrates that decision quality is not simply the selection of the largest EMV. Timing, reversibility, information arrival, service constraints and downside exposure materially affect the appropriate strategy. It also demonstrates why operational flexibility is an economic asset: it allows the organisation to defer commitment until uncertainty is reduced.

9. Implementation, Governance and Ethical Issues

Decision models should be auditable. A professional model should record data sources, probability elicitation procedures, payoff definitions, time horizons, discount rates, excluded consequences and sensitivity results. Assumptions should be version-controlled, especially when models influence capital approval, staffing or customer access.

Important governance practices include:

• separating model development from approval where feasible;
• documenting who supplied probabilities and who bears consequences;
• validating calculations independently;
• monitoring forecast accuracy and calibration over time;
• defining escalation thresholds for severe downside outcomes;
• preserving human review for high-impact decisions.

Ethically, expected value can conceal unequal effects. A policy that maximises average service level may disadvantage remote customers; an algorithmic staffing decision may reduce opportunities for a protected group; a risk model may use historical data that reflect discriminatory practices. Decision makers should examine distributional outcomes, privacy, explainability and the right to challenge automated recommendations.

AI and machine-learning systems can improve probability estimates, detect patterns and generate scenarios. They do not remove the need to define actions, consequences and objectives. Predictive accuracy is not the same as decision value: a highly accurate prediction may not change an operational choice, while a modestly accurate early-warning signal may be valuable if it enables timely mitigation. AI outputs should therefore enter a governed decision process with uncertainty intervals, drift monitoring, bias assessment and human accountability.

10. Common Errors and Best Practices

Common errors include treating uncertain outcomes as probabilities without evidence, confusing conditional and marginal probabilities, ignoring the cost of information, mixing nominal and real monetary values, double-counting risk, and selecting a criterion without explaining the underlying preference.

Best practice is to:

1. define the decision and decision deadline;
2. identify controllable actions and mutually exclusive states;
3. construct a consistent payoff table;
4. distinguish prior, conditional and posterior probabilities;
5. select EMV, utility or an uncertainty criterion deliberately;
6. represent sequential choices with a decision tree;
7. test sensitivity and stress conditions;
8. evaluate information and flexibility;
9. communicate assumptions, ranges and limitations;
10. monitor outcomes and update the model after implementation.

Chapter Summary

Decision analysis represents managerial choices through actions, uncertain states and payoffs. Payoff tables support one-stage decisions, while decision trees capture sequential choices and enable rollback analysis. Under risk, expected monetary value combines payoffs with probabilities, and expected opportunity loss expresses the cost of choosing imperfectly. Under uncertainty, maximin, maximax, minimax regret, Laplace and Hurwicz criteria reflect different attitudes toward incomplete knowledge.

Utility theory incorporates risk preferences when monetary expectation is insufficient. Bayes’ theorem updates beliefs when new evidence arrives. Perfect information provides an upper bound on information value, while sample information is worthwhile only when its expected decision benefit exceeds its cost. Sensitivity analysis reveals decision thresholds; robust analysis tests performance across plausible assumptions; and scenario planning develops coherent alternative futures and contingent actions.

Effective decision analysis is not mechanical. It requires defensible data, transparent assumptions, recognition of model uncertainty, attention to flexibility and timing, and ethical consideration of who gains and who bears risk. The strongest managerial recommendation is often a monitored, reversible policy with explicit trigger points rather than a one-time commitment based on a single forecast.

Key Terms

Decision analysis: A structured approach to selecting actions under certainty, risk or uncertainty.

Action: A controllable alternative available to the decision maker.

State of nature: An uncertain condition that affects outcomes but is not controlled by the decision maker.

Payoff: The consequence associated with an action–state combination.

Payoff table: A matrix listing payoffs for alternatives and states.

Risk: A decision environment in which probabilities can be assigned to states.

Uncertainty: A decision environment in which reliable probabilities are unavailable or disputed.

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

Expected opportunity loss: The probability-weighted regret associated with an action.

Regret: The loss from failing to select the best action after the state is known.

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

Rollback analysis: Solving a decision tree from its terminal outcomes back to the initial decision.

Utility: A preference measure assigned to outcomes.

Risk premium: The difference between expected monetary value and certainty equivalent.

Prior probability: Belief about a state before new evidence is observed.

Posterior probability: Updated belief after incorporating evidence.

Expected value of perfect information: Maximum expected benefit from knowing the true state before acting.

Expected value of sample information: Expected benefit of imperfect information before its acquisition cost.

Robustness: Acceptable performance across a range of plausible assumptions or futures.

Scenario planning: Exploration of coherent alternative future environments to support contingent strategy.

Review Questions

Short Answer Questions

1. Distinguish risk from uncertainty in managerial decision analysis.
2. What are the three basic components of a payoff table?
3. Why may expected monetary value be unsuitable for a highly risk-averse organisation?
4. What is the purpose of rollback analysis in a decision tree?
5. Distinguish a prior probability from a posterior probability.

Long Answer Questions

1. Explain the construction and use of payoff tables, including EMV and expected opportunity loss.
2. Compare maximin, maximax, minimax regret, Laplace and Hurwicz criteria.
3. Explain how utility theory changes the analysis of risky monetary outcomes.
4. Derive and interpret the expected value of perfect information and expected value of sample information.
5. Discuss how sensitivity analysis, robust analysis and scenario planning can be integrated in an operations decision.

Analytical / Application Questions

1. A firm has three capacity alternatives and four demand states but cannot defend probability estimates. Design an analysis using at least three uncertainty criteria and explain how management should interpret conflicting recommendations.
2. A diagnostic test has different false-positive and false-negative rates for two demand states. Construct a Bayesian updating procedure and explain how the posterior probabilities would enter a sequential capacity decision.
3. Design a decision tree for a distribution-centre lease followed by an expansion decision after six months. Identify the information needed, the trigger conditions, the relevant payoffs and the operational risks that a purely financial EMV may omit.

Multiple Choice Questions with Answers

Question 1

Which criterion selects the action with the highest worst-case payoff?

A. Maximax

B. Maximin

C. Laplace

D. Expected opportunity loss

Correct Answer: B. Maximin

Question 2

For an action with payoffs 100 and 300 under probabilities 0.4 and 0.6, what is its EMV?

A. 180

B. 200

C. 220

D. 240

Correct Answer: C. 220

Question 3

In a decision tree, a chance node represents:

A. A controllable managerial choice

B. An uncertain event

C. A final accounting adjustment

D. A utility function

Correct Answer: B. An uncertain event

Question 4

If P(H)=0.4, P(E|H)=0.8 and P(E|not H)=0.3, what is P(H|E)?

A. 0.40

B. 0.50

C. 0.64

D. 0.80

Correct Answer: C. 0.64

Question 5

Expected opportunity loss is calculated using:

A. The minimum payoff in each state

B. The difference between the statewise maximum payoff and the action payoff

C. Only the largest overall payoff

D. The average of all probabilities

Correct Answer: B. The difference between the statewise maximum payoff and the action payoff

Question 6

A concave utility function generally represents:

A. Risk aversion

B. Risk neutrality

C. Risk seeking

D. Probability neutrality

Correct Answer: A. Risk aversion

Question 7

The expected value of perfect information is:

A. Always negative

B. The value with perfect information minus the best value without it

C. The cost of collecting any forecast

D. The maximum regret of an action

Correct Answer: B. The value with perfect information minus the best value without it

Question 8

Which statement about sample information is correct?

A. It is always worth buying if it is accurate

B. Its value cannot exceed the value of perfect information

C. It eliminates all uncertainty

D. It uses no probabilities

Correct Answer: B. Its value cannot exceed the value of perfect information

Question 9

Scenario planning primarily aims to:

A. Prove one forecast is correct

B. Assign equal probabilities to all futures

C. Explore coherent alternative futures and contingent actions

D. Replace operational monitoring

Correct Answer: C. Explore coherent alternative futures and contingent actions

Question 10

A robust decision is one that:

A. Has the highest payoff in one optimistic scenario

B. Performs acceptably across plausible assumptions and futures

C. Ignores downside outcomes

D. Requires perfect information

Correct Answer: B. Performs acceptably across plausible assumptions and futures