Decision Problem Structuring and Model Formulation

Introduction

Prescriptive analytics converts information into recommended action. Its central question is not merely “What is likely to happen?” but “What should a decision-maker do, given objectives, uncertainty, resources, and competing interests?” The quality of a prescriptive model depends heavily on how the decision problem is structured before an optimization model, simulation, or decision tree is built.

Poorly structured problems produce technically elegant but managerially irrelevant solutions. A model may optimize the wrong objective, omit an important stakeholder, treat a controllable choice as uncertain, or rely on data that cannot support the proposed decision. Problem structuring therefore provides the bridge between managerial judgment and analytical formulation.

This chapter develops a systematic approach to framing structured, semi-structured, and unstructured decisions. It explains influence diagrams, objective hierarchies, alternatives, states of nature, constraints, assumptions, stakeholder analysis, data requirements, model boundaries, and baseline models. A detailed illustrative case demonstrates how these elements are integrated.

Learning Objectives

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

1. Distinguish structured, semi-structured, and unstructured decisions.
2. Identify decision variables, uncertain events, outcomes, objectives, and constraints.
3. Represent causal and informational relationships using influence-diagram logic.
4. Construct objective hierarchies and recognize conflicting stakeholder objectives.
5. Define model boundaries, assumptions, data requirements, and suitable baselines.
6. Formulate a decision problem suitable for decision trees, optimization, simulation, or other prescriptive methods.
7. Critically evaluate limitations, ethical risks, and governance requirements.

1. The Nature of Decision Problems

A decision problem consists of at least five elements: a decision-maker, one or more available actions, relevant uncertainties, consequences of actions, and criteria for judging those consequences. In formal terms, an action may be represented by a decision variable x, an uncertain condition by a state of nature s, and the resulting consequence by y = f(x,s). The function f need not be a simple equation; it may be a simulation, business rule, or expert assessment.

Structured decisions

A structured decision has clearly defined alternatives, predictable procedures, and an established objective. Examples include replenishing inventory when stock falls below a reorder point, assigning employees to shifts subject to coverage rules, and approving a transaction when documented conditions are satisfied. Such decisions are often suitable for mathematical programming, rule-based automation, or database workflows.

Semi-structured decisions

Semi-structured decisions combine formal analysis with managerial judgment. Examples include selecting a supplier, setting a promotional budget, deciding whether to expand capacity, or prioritizing loan applications. Some relationships and constraints can be modeled, while preferences, exceptional circumstances, and strategic considerations require human interpretation.

Unstructured decisions

Unstructured decisions involve ambiguous objectives, incomplete information, novel situations, and contested interpretations. Examples include entering an unfamiliar international market, responding to a major reputational crisis, or designing a long-term sustainability strategy. Analytics can clarify scenarios and consequences, but it cannot eliminate judgment. In such contexts, the principal analytical contribution may be problem exploration rather than automated recommendation.

These categories are not permanent properties of decisions. A previously unstructured problem can become more structured as the organization develops data, policies, and experience. Conversely, a structured process may become semi-structured when conditions change.

2. From Business Situation to Decision Model

A business situation is not yet a model. Structuring requires progressively narrowing the situation into a decision statement.

A useful decision statement specifies:

• Decision owner: Who has authority to act?
• Decision timing: When must the decision be made, and over what horizon?
• Alternatives: What actions are genuinely available?
• Uncertainty: What events are outside the decision-maker’s control?
• Consequences: What outcomes follow from each action and state?
• Objectives: What does “better” mean?
• Constraints: What limits feasible action?
• Information: What is known now, and what may become known later?
• Scope: Which effects and stakeholders are included?

A concise formulation is:

Here, x is the decision, X is the set of feasible alternatives, s is a state of nature, U is utility or value, g_j is the jth resource, policy, or performance constraint, and b_j is its permitted limit. This is a general representation, not necessarily a complete solvable model. If uncertainty is represented probabilistically, expected utility is often written as:

where S is the set of possible states and P(s) is the probability of state s, with probabilities nonnegative and summing to one. Expected utility is appropriate only when the states are sufficiently defined, probabilities are defensible, and the decision-maker accepts expectation as an appropriate evaluation principle. Risk-averse decisions may require a different objective, such as a downside-risk measure.

3. Alternatives, Decisions, and States of Nature

Alternatives and decision variables

An alternative is a complete or partial course of action. It may be discrete—select supplier A, B, or C—or continuous, such as deciding an order quantity. In a multi-period model, the decision may be a policy that specifies what to do after future information arrives.

Decision variables must be controllable, measurable, and linked to an action. “Improve customer satisfaction” is an objective, not a decision variable. “Increase service staffing by two employees during the evening shift” is a decision.

Alternatives should be mutually distinguishable and sufficiently complete. A list that includes “launch a product” and “launch a product with a pilot” may mix different levels of detail. Analysts should define the decision granularity explicitly.

States of nature

A state of nature is a condition outside the decision-maker’s control that affects consequences. Examples include high or low demand, a supplier disruption, an interest-rate scenario, or customer adoption. States should be collectively exhaustive for the chosen model boundary, although they need not describe every imaginable future.

A key distinction is timing. If demand is observed before the order decision, it is information. If demand is realized after ordering, it is a state of nature. The same variable can therefore be a forecast input in one model and an outcome uncertainty in another.

Payoffs and consequences

A payoff table records the consequence of each alternative under each state. Consider a retailer choosing between a large and small seasonal order:

If high demand has probability 0.6 and low demand probability 0.4, the expected monetary values are:

Under the assumptions that the probabilities and monetary payoffs are valid and that expected monetary value is the decision criterion, the small order is preferred. This conclusion does not mean it is universally best: a risk-averse manager may prefer the lower downside exposure of the small order, while a growth-oriented manager may accept more risk.

4. Influence-Diagram Thinking

An influence diagram is a compact graphical representation of decisions, uncertainties, and outcomes. Its common elements are:

• Decision node: a controllable choice, often drawn as a rectangle.
• Chance node: an uncertain event or variable, often drawn as an oval.
• Value node: an objective or payoff, often drawn as a diamond.
• Directed arrow: a relevant informational, causal, or functional dependency.

For example, a retailer’s capacity decision may be represented conceptually as:

Market research → Demand forecast → Capacity decision → Sales and cost → Profit
        ↓                              ↓                 ↑
 Competitor actions ───────────────→ Demand ──────────────

The diagram should be read carefully. An arrow from research to a forecast indicates that research informs the forecast; it does not prove that research causes demand. An arrow from capacity to cost may represent a modeled relationship. Influence diagrams are therefore useful for clarifying information and dependency, but they are not automatically causal identification models.

A well-constructed diagram helps answer four questions:

1. What can the organization control?
2. What is uncertain?
3. What information is available before each decision?
4. How are outcomes evaluated?

Influence diagrams also expose missing variables. If a model recommends staffing without representing service demand, labor availability, or service-level consequences, its structure is suspect.

5. Objective Hierarchies and Multiple Criteria

Organizations rarely pursue one uncontested objective. A hospital may seek clinical quality, access, financial sustainability, and equity. A logistics firm may balance cost, delivery speed, resilience, emissions, and employee safety.

An objective hierarchy decomposes a broad mission into measurable subordinate objectives:

Organizational mission
├── Financial sustainability
│   ├── Operating margin
│   └── Cash requirement
├── Customer value
│   ├── Reliability
│   └── Waiting time
└── Responsible operations
    ├── Emissions
    └── Equity of access

The hierarchy supports measurement and discussion, but it does not by itself resolve trade-offs. A multi-objective value function may be written as:

where v_k(x) is the normalized value of objective k, w_k is its weight, and ∑w_k = 1 when weights are normalized. This additive form assumes that the objectives can be meaningfully normalized and that the trade-offs represented by the weights are acceptable. It may be inappropriate when objectives interact strongly or when minimum thresholds must be met. In such cases, constraints, lexicographic priorities, goal programming, or Pareto analysis may be preferable.

Normalization must be transparent. For a benefit criterion where higher values are better, one possible score is:

where z_k(x) is the observed performance and the minimum and maximum are agreed reference points. The result is dimensionless; the reference range must be meaningful and should not be selected merely to make a favored alternative appear attractive.

Stakeholder analysis is essential because objective weights are often political or ethical judgments, not technical facts. Analysts should identify who gains, who bears costs, who is represented in the data, and who has authority to approve the decision. Stakeholders may include customers, employees, owners, regulators, suppliers, communities, and future generations.

6. Constraints and Feasibility

Constraints define what cannot or should not be done. They may be:

• Resource constraints: budget, staff, vehicles, materials, or computing capacity.
• Technical constraints: capacity, compatibility, or process requirements.
• Policy constraints: internal rules and approval thresholds.
• Legal constraints: privacy, employment, competition, safety, and nondiscrimination requirements.
• Service constraints: minimum fill rates, response times, or quality levels.
• Ethical constraints: unacceptable discrimination, unsafe practice, or exploitation.

A budget constraint might be represented as:

where x_i is the quantity of activity i, c_i is its unit cost, and B is the available budget. Units must be consistent: if costs are dollars per unit, the left-hand side is dollars.

Constraints should be classified as hard or soft. A hard constraint may not be violated. A soft constraint can be violated at a stated penalty, such as a small overtime cost for exceeding a preferred staffing limit. Treating every preference as a hard constraint can make a model infeasible; treating legal or safety requirements as soft can be unacceptable.

Feasibility should be tested before optimization. If no alternative satisfies all hard constraints, the analyst must identify the conflict rather than quietly relax a rule. Infeasibility analysis can reveal that the budget, service target, and capacity assumptions cannot coexist.

7. Assumptions, Model Boundaries, and Abstraction

Every model is an abstraction. Model boundaries specify what is included, excluded, aggregated, and treated as fixed. A boundary statement might read: “The model selects weekly distribution quantities for the next twelve weeks across three warehouses. It includes transport and inventory costs, assumes approved suppliers remain available, and excludes long-term facility investment.”

Important assumptions concern:

• time horizon and time intervals;
• demand and price behavior;
• availability and quality of data;
• independence or dependence among uncertain events;
• capacity and operating rules;
• treatment of taxes, inflation, and discounting;
• behavior of customers, employees, and competitors;
• whether relationships are causal, predictive, or merely correlational.

Assumptions should be recorded in an assumption register with an owner, rationale, evidence, and sensitivity test. A model is more credible when users can see which conclusions depend on uncertain assumptions.

Overly narrow boundaries can create suboptimization. For example, minimizing transport cost while excluding late-delivery penalties may increase total cost. Overly broad boundaries can make the model opaque, data-intensive, and difficult to validate. The appropriate boundary is the smallest one that captures material consequences of the decision.

8. Data Requirements and Baselines

Data requirements follow from the decision structure, not the other way around. For each variable, specify its definition, unit, time grain, source, owner, quality, and intended use.

Data quality includes accuracy, completeness, timeliness, consistency, provenance, and representativeness. A large dataset does not compensate for a poorly defined variable. In particular, historical decisions may reflect prior bias, so using them as labels can reproduce rather than correct inequity.

A baseline is the reference against which a proposed policy is evaluated. Common baselines include:

• current practice or “do nothing”;
• a simple rule, such as reorder-point control;
• a historical average;
• expert judgment;
• a random or unconstrained benchmark, where appropriate;
• a best-known operational policy.

A sophisticated model should not be accepted merely because it outperforms an arbitrary baseline. Baselines should be feasible, transparent, and relevant to the actual decision. Evaluation should compare both performance and implementation burden.

9. Choosing a Formulation Method

The structure of the problem guides method selection.

A decision tree is particularly useful when the number of alternatives and states is manageable and the sequence of decisions and information is central. It becomes unwieldy when there are many periods, correlated uncertainties, or continuous decisions. Optimization is powerful for allocation and scheduling but may hide assumptions in a compact algebraic form. Simulation can represent operational detail but usually requires a separate policy-search or comparison procedure.

10. Worked Case: Regional Vaccine-Clinic Expansion

The following is a fictional illustrative case.

Background

A regional public-health agency must decide whether to expand weekend vaccination clinics for the next flu season. It has a fixed operating budget and wants to increase access without creating unsafe staffing conditions.

Problem definition

The decision owner is the regional operations director. The decision is made before the season, but weekly attendance becomes observable after launch. Alternatives are:

• maintain the current weekday schedule;
• add clinics in high-demand districts;
• add clinics across all districts.

States of nature describe low, medium, and high attendance. Attendance is uncertain because of disease prevalence, public communication, and competing services. Relevant outcomes include vaccinations delivered, cost, waiting time, staff overtime, and access across low-income districts.

Objective hierarchy and stakeholders

The primary mission is improved public health. Subobjectives are access, safe service quality, financial stewardship, and equity. Stakeholders include residents, public-health staff, local governments, finance officers, and vulnerable populations. The agency decides that safety is a hard constraint, while access and equity are valued objectives subject to the budget.

Model formulation

Let x_j indicate whether clinic configuration j is selected, where x_j ∈ {0,1}. A simplified budget constraint is:

where C_j is the cost of configuration j and B is the approved budget. Exactly one configuration is selected:

For each state s, the agency estimates vaccinations V_js, waiting time W_js, and equity score E_js. A normalized value model could be:

The signs indicate that more vaccinations and equity are desirable, while waiting time and cost reduce value. The weights must be approved by the agency, not inferred solely from historical spending. A safety rule excludes any configuration for which projected staffing exceeds the safe threshold in any relevant state.

Data and validation

The agency requires historical appointment demand, no-show rates, district population, staffing rosters, clinic capacity, travel distance, and overtime records. Analysts must distinguish observed attendance from unmet demand: a low historical attendance rate may reflect poor access rather than low need. The current weekday schedule is the operational baseline.

The team validates the model through expert review, historical back-testing, extreme-case testing, and sensitivity analysis. It tests whether the recommended configuration changes when high-demand probability, staffing capacity, or the equity weight changes.

Outcome

Suppose the analysis shows that all-district expansion produces the greatest expected attendance but violates the safety constraint under high demand. Targeted expansion satisfies safety and improves access in underserved districts, while requiring a modest budget increase relative to the baseline. The agency selects targeted expansion and establishes a monitoring rule: if observed demand exceeds a threshold for two consecutive weeks, temporary capacity is added where safe.

Learning points

The case illustrates that the “best” alternative is not defined by attendance alone. The decision requires a hierarchy of objectives, explicit constraints, stakeholder participation, a baseline, and an adaptive implementation policy. It also shows why a model should represent information arriving after implementation rather than treating the entire season as known in advance.

11. Sensitivity, Robustness, and Model Risk

A recommendation is fragile if small changes in assumptions reverse the decision. Sensitivity analysis varies one or more inputs—such as probabilities, costs, demand, or objective weights—and observes the effect on the preferred alternative. Scenario analysis changes a coherent bundle of assumptions, such as an economic downturn or supply disruption.

Robustness asks whether an alternative performs acceptably across a range of plausible conditions, not merely whether it maximizes one estimated value. A robust policy may be preferred when probabilities are disputed or consequences of failure are severe.

Model risk arises from incorrect structure, data error, implementation failure, and misuse. Sources include omitted variables, double counting benefits, biased training data, unrealistic behavioral assumptions, and optimization around a poorly chosen metric. Validation should therefore examine conceptual validity, data validity, computational correctness, and use validity.

12. Ethics, Governance, and Responsible Use

Prescriptive analytics can allocate scarce resources and therefore distribute benefits and burdens. Ethical review should ask:

• Are affected groups represented in data and consultation?
• Does the model use legally or ethically sensitive attributes?
• Could proxy variables reproduce discrimination?
• Are explanations available to decision-makers and affected persons?
• Can a human override the recommendation, and is the override documented?
• Are privacy, security, retention, and access controls adequate?
• Is there an appeal or correction mechanism?

Fairness is not a single universal metric. Equal selection rates, equal error rates, equal outcomes, and priority for need can conflict. The relevant standard depends on law, institutional mission, and ethical reasoning. Analysts must not present a technical metric as a complete ethical justification.

AI-enabled systems increase the importance of governance. Automated recommendations should have version control, audit logs, monitoring for drift, documented data lineage, and clearly assigned accountability. Generative AI may assist with drafting assumptions or exploring scenarios, but it should not be treated as an authoritative source of probabilities, legal interpretation, or stakeholder values without verification.

13. Best-Practice Workflow

A disciplined workflow is:

1. Frame the decision in plain language.
2. Identify the decision owner, timing, and affected stakeholders.
3. Separate controllable decisions from uncertain states.
4. Map dependencies with an influence-diagram approach.
5. Develop an objective hierarchy and define trade-offs.
6. Specify alternatives and hard and soft constraints.
7. Set model boundaries and document assumptions.
8. Define data requirements and assess data quality.
9. Select a transparent and operationally relevant baseline.
10. Choose a suitable analytical method.
11. Validate structure, data, calculations, and usability.
12. Conduct sensitivity, scenario, and robustness analysis.
13. Design implementation, monitoring, and governance.
14. Communicate limitations and retain human accountability.

The analyst’s role is not to replace managerial judgment but to make assumptions, consequences, and trade-offs more explicit and contestable.

Chapter Summary

Decision problem structuring is the foundation of prescriptive analytics. Structured decisions may be automated, while semi-structured and unstructured decisions require analysis combined with judgment. A sound formulation identifies the decision-maker, timing, alternatives, states of nature, outcomes, objectives, constraints, information, and model boundary.

Influence diagrams clarify decisions, uncertainties, outcomes, and dependencies. Objective hierarchies expose competing goals, while multi-criteria formulations make trade-offs explicit. Constraints define feasibility and must distinguish legal, safety, and operational requirements from preferences. Assumptions and data requirements should be documented and tested rather than hidden inside software.

Baselines provide meaningful reference points for evaluating proposed policies. Method selection should reflect the number and type of decisions, uncertainty, timing, operational complexity, and stakeholder needs. Sensitivity and robustness analysis reveal fragile recommendations. Finally, ethical governance is essential because prescriptive models influence the distribution of opportunities, costs, risks, and public services.

Key Terms

Decision problem: A situation requiring choice among actions under objectives and constraints.

Structured decision: A decision with defined alternatives, procedures, and evaluation criteria.

Semi-structured decision: A decision combining formal analysis with managerial judgment.

Unstructured decision: A novel or ambiguous decision with uncertain objectives and procedures.

Decision variable: A controllable quantity or choice selected by the decision-maker.

State of nature: An uncertain condition outside the decision-maker’s control.

Payoff: The consequence or value associated with an action and state.

Influence diagram: A graphical representation of decisions, uncertainties, dependencies, and value.

Objective hierarchy: A decomposition of broad goals into measurable subordinate objectives.

Constraint: A condition limiting feasible actions or outcomes.

Hard constraint: A requirement that cannot be violated.

Soft constraint: A preferred condition that may be violated at a stated penalty.

Model boundary: The declared scope of variables, time, processes, and stakeholders included in a model.

Assumption register: A record of modeling assumptions, evidence, ownership, and sensitivity.

Baseline: A reference policy used to evaluate an alternative.

Expected monetary value: The probability-weighted average monetary payoff across states.

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

Robustness: The extent to which a decision performs acceptably across plausible conditions.

Model risk: The possibility that model structure, data, implementation, or use produces misleading decisions.

Stakeholder analysis: Identification of parties affected by, influencing, or accountable for a decision.

Review Questions

Short Answer Questions

1. What distinguishes a state of nature from a decision variable?
2. Why is a baseline necessary in prescriptive analytics?
3. Define a hard constraint and give one example.
4. What information does an influence diagram communicate?
5. Why should model boundaries be explicitly documented?

Long Answer Questions

1. Explain how an unstructured decision can become more structured over time. Illustrate your answer with a business example.
2. Describe how an objective hierarchy can support decisions involving cost, service quality, and equity.
3. Compare decision trees, optimization, and simulation as methods for prescriptive analytics.
4. Explain the role of stakeholder analysis in defining objectives, constraints, and ethical safeguards.
5. Discuss why data quality and model assumptions can be more important than algorithmic sophistication.

Analytical / Application Questions

1. A retailer must choose among three order quantities before uncertain seasonal demand is realized. Construct the decision elements, states of nature, payoff measures, constraints, and baseline that should be included in a decision model.
2. A hospital wants to allocate operating-room capacity among emergency, cancer, and elective procedures. Develop an objective hierarchy and identify at least five relevant constraints. Explain which should be hard and which could be soft.
3. An organization proposes an AI system to prioritize job applicants. Define the model boundary, data requirements, stakeholders, baseline, validation tests, and governance safeguards needed before deployment.

Multiple Choice Questions with Answers

Question 1

Which element is controllable by the decision-maker?

A. A state of nature

B. A decision variable

C. A random shock

D. A realized demand level

Correct Answer: B. A decision variable

Question 2

What is the primary purpose of an influence diagram?

A. To replace all stakeholder consultation

B. To calculate only historical averages

C. To show decisions, uncertainties, dependencies, and value

D. To guarantee causal identification

Correct Answer: C. To show decisions, uncertainties, dependencies, and value

Question 3

In the worked ordering example, what is the expected monetary value of the large order?

A. $92,000

B. $94,000

C. $110,000

D. $180,000

Correct Answer: A. $92,000

Question 4

Which is normally a hard constraint?

A. A preferred management style

B. A legally required safety limit

C. A negotiable service preference

D. An analyst’s suggested target

Correct Answer: B. A legally required safety limit

Question 5

What does a baseline provide?

A. A reference policy for evaluating a proposed policy

B. A guarantee that the model is unbiased

C. A replacement for model validation

D. A probability distribution for every variable

Correct Answer: A. A reference policy for evaluating a proposed policy

Question 6

Which situation is most clearly semi-structured?

A. Applying a fixed password rule

B. Choosing a supplier using cost, resilience, and managerial judgment

C. Recording a transaction timestamp

D. Checking whether a form is complete

Correct Answer: B. Choosing a supplier using cost, resilience, and managerial judgment

Question 7

In an objective hierarchy, why are weights used in an additive value model?

A. To convert every constraint into a law

B. To represent the relative importance of objectives

C. To remove uncertainty from the data

D. To prove that objectives are independent

Correct Answer: B. To represent the relative importance of objectives

Question 8

What is a major risk of using historical decisions as prediction labels?

A. They always contain too few observations

B. They cannot be stored digitally

C. They may reproduce past organizational bias

D. They eliminate the need for a baseline

Correct Answer: C. They may reproduce past organizational bias

Question 9

Which method is especially suitable when a small number of sequential decisions and uncertain events must be represented?

A. Decision tree analysis

B. Data deletion

C. Simple descriptive reporting

D. Unconstrained averaging

Correct Answer: A. Decision tree analysis

Question 10

What does sensitivity analysis examine?

A. Whether a model has any variables

B. How results change when inputs or assumptions vary

C. Whether stakeholders agree on every value

D. Whether all alternatives have identical payoffs

Correct Answer: B. How results change when inputs or assumptions vary