Two options can have the same average payoff and still be very different decisions. One may be steady. The other may combine a large gain with a serious downside. A spreadsheet that reports only the average has not finished the job.

Decision theory studies how choices should be structured when actions lead to uncertain outcomes and when the decision-maker cares about some outcomes more than others. It separates states of the world, available actions, consequences, beliefs about uncertainty and preferences over those consequences. Expected utility is one influential way to combine those ingredients into a coherent choice rule.

This article owns the formal decision problem. Strategic Decision-Making owns the broader organisational craft of framing consequential choices. Value of Information owns the question of whether further evidence is worth acquiring. Bayesian Inference and Updating owns the statistical mechanics of updating beliefs. Here we build the common decision grammar that connects them.

All numerical examples below are hypothetical teaching constructions. They are not forecasts, financial advice, medical recommendations or measured eduKate outcomes.

Reading route: start with the anatomy of a decision, move through expected value and expected utility, test dominance and risk, then examine sequential decisions, limits and the final decision-theory checklist.

1. A decision problem has five different objects

A clean decision model distinguishes five things that ordinary language often mixes together.

• Actions are the choices available to the decision-maker.
• States are uncertain conditions not controlled by that immediate choice.
• Outcomes are what happens when an action meets a state.
• Beliefs describe how plausible the states are.
• Preferences describe how desirable the outcomes are.

Confusing these objects creates subtle errors. A high probability is not the same thing as a desirable outcome. A desirable outcome is not the same thing as an available action. A prediction does not choose for us. It informs the part of the decision model concerned with uncertain states.

2. A small decision table makes the structure visible

Imagine a fictional library choosing between two arrangements for an evening event. Arrangement S is steady. Arrangement X is more ambitious but depends strongly on attendance.

Constructed utility-point outcomes for teaching only

Action	High attendance	Low attendance
Steady arrangement S	60	60
Ambitious arrangement X	100	20

The table contains no probability yet. It says only what each action would produce under each state, according to the chosen point scale. Decision theory insists that we keep this consequence structure separate from our beliefs about which state will occur.

3. Expected value is a probability-weighted average

Suppose high and low attendance are each assigned probability 0.5. The expected value of S is 60. The expected value of X is 0.5 × 100 + 0.5 × 20 = 60. On this linear point scale, the two actions tie in expectation.

Expected value is not a forecast that the outcome will literally equal 60. Neither branch of X produces 60. It is an ex ante average over possible outcomes using the stated probabilities.

That distinction matters. If a decision-maker treats an expected value as the most likely realised outcome, a probability-weighted calculation quietly turns into a false prediction.

4. The probability model belongs in the evidence layer

Where did the fifty-fifty probabilities come from? Historical attendance? A forecasting model? Expert judgement? A deliberately neutral teaching assumption? These origins are not equivalent.

Strong decision analysis labels the source of each probability and tests whether the choice is sensitive to it. If high attendance has probability p, X has expected value 100p + 20(1 − p) = 20 + 80p. It exceeds S when 20 + 80p > 60, or p > 0.5.

The threshold is useful because it turns a vague dispute into a specific evidence question: is the chance of high attendance credibly above one half?

5. Expected monetary value and expected utility are not the same idea

If consequences are measured directly in money, an expected monetary value calculation assumes that one additional dollar contributes the same decision weight regardless of the decision-maker’s existing position and regardless of scale. That can be a poor representation of preference.

Expected utility theory instead applies a utility function to consequences and averages those utility values using probabilities. Utility is not a hidden physical substance. It is a representation of preference under a specified theory.

The current Stanford Encyclopedia of Philosophy entry on Decision Theory explains the orthodox expected-utility framework and its representation theorems. Its important lesson for practical modelling is that probability and desirability enter the choice rule as different ingredients.

6. Utility can represent diminishing marginal value

Suppose a fictional decision uses monetary outcomes but the decision-maker’s utility is represented for illustration by u(w) = √w for positive wealth w. Increasing wealth from 100 to 400 changes utility from 10 to 20. Increasing wealth by the same absolute 300 again, from 400 to 700, changes utility by less than 10.

This curvature captures diminishing marginal utility in the model: extra resources still matter, but their represented contribution changes with the starting point.

The formula is not a universal psychology law. Different decision-makers and purposes can imply different preference structures. The point is to make the representation explicit rather than assume that money, marks, time and human welfare can all be treated as linear interchangeable units.

7. Risk attitude appears in the shape of utility

A concave utility function can represent risk aversion over the modelled consequence. A linear utility function corresponds to risk neutrality on that scale. A convex utility function can represent risk-seeking behaviour over the relevant range.

Consider a fifty-fifty prospect between wealth 100 and 900. Its expected wealth is 500. Under u(w)=√w, expected utility is 0.5 × 10 + 0.5 × 30 = 20. The certain wealth level with utility 20 is 400. So this utility model ranks a certain 400 as equivalent to the risky prospect whose expected wealth is 500.

The difference between 500 and 400 is a certainty-equivalent illustration of the model’s risk aversion. It is not a market price or a recommendation.

8. Utility comparisons must preserve their scale assumptions

Utility numbers are not automatically meaningful across people. Saying one person has utility 80 and another has utility 40 does not normally establish that the first person is twice as well off. Many expected-utility representations preserve preference under positive affine transformations of the utility scale.

This matters whenever a decision model aggregates multiple people. Adding individual utility scores can embed ethical assumptions that are far stronger than the mathematics reveals. Public decisions require explicit attention to distribution, rights and whose preferences enter the objective.

9. Dominance can settle some decisions without probabilities

If Action A produces an outcome at least as good as Action B in every relevant state and produces a strictly better outcome in at least one state, A dominates B under the stated consequence ordering.

A dominated option does not become sensible merely because its story is attractive. Dominance is powerful because it can remove options before we argue about uncertain probabilities.

But dominance depends on the outcomes included. If one table ignores cost, accessibility or a safety constraint, an apparent dominance result may disappear when the decision is modelled properly.

10. Stochastic dominance compares whole distributions

Some decisions have many possible outcomes rather than two states. Stochastic dominance provides ways to compare distributions without fixing one exact utility function, under stated classes of preferences.

First-order stochastic dominance roughly captures the idea that one distribution delivers at least as much probability of high outcomes across thresholds. Second-order dominance can support comparisons for broad classes of risk-averse preferences.

These ideas are useful precisely because they expose what a ranking does and does not require. They should not be treated as magic rules divorced from the outcome scale and preference assumptions.

11. A decision tree represents sequence as well as uncertainty

A one-stage table is inadequate when decisions unfold over time. A decision tree distinguishes decision nodes from chance nodes and traces the order in which choices and observations occur.

CHOOSE PILOT?
├─ NO → choose current service
└─ YES
   ├─ favourable signal → choose expanded design
   └─ unfavourable signal → keep current service

The timing matters. A signal received after an irreversible commitment cannot have the same decision value as a signal received before it.

12. Roll back a tree from the future

Decision trees are commonly evaluated by working backward. At a future decision node, choose the best action given the information available there. At a chance node, average the downstream values using the conditional probabilities. Continue backward until the initial choice has a value.

This method prevents an important error: evaluating the value of information while refusing to allow future actions to respond to that information.

13. Beliefs can change before preferences do

Suppose the consequences and preferences remain fixed but new evidence raises the probability of high attendance. The preferred action may change because the belief layer changed.

That is different from changing preferences. If the institution becomes less willing to tolerate crowding risk, the utility attached to certain outcomes may change even with the same probability forecasts.

Separating belief revision from preference revision makes disagreements easier to diagnose. Two people may agree on the facts but value outcomes differently, or value outcomes similarly while holding different beliefs about what will happen.

14. Bayesian updating connects evidence to the state probabilities

When evidence arrives, Bayesian reasoning can update probabilities for uncertain states. But Bayesian updating does not by itself determine the action. The action depends on both the updated beliefs and the consequence-preference structure.

This is why Bayesian Inference and Updating and decision theory are connected but not identical. One updates beliefs; the other evaluates choices using those beliefs.

15. Value of information asks whether learning can change action

Information has decision value when it allows the decision-maker to condition a future action on what was learned and thereby improve expected consequences.

The complete treatment belongs to How Value of Information Works. The important connection here is architectural: information is valuable not because uncertainty is intellectually uncomfortable, but because resolving some uncertainty can alter the choice.

16. An accurate signal can still have little decision value

If one action remains best after every plausible signal, information can improve beliefs without changing action. The signal may be scientifically interesting while adding little immediate value to that particular decision.

Conversely, a modestly accurate signal can be valuable when the current decision sits close to a threshold and the consequences of choosing the wrong side are large.

17. Expected utility is normative, not automatically descriptive

A normative model asks how choices should be ordered under stated rationality principles. A descriptive model asks how people actually choose. These are different jobs.

The Stanford Encyclopedia of Philosophy entry on Descriptive Decision Theory discusses models of observed human choice and the ways real decisions can depart from simple expected-utility descriptions.

We should therefore not treat a normative model as a psychological prediction merely because its mathematics is elegant.

18. Framing can change observed choice without changing the underlying outcomes

People can respond differently when equivalent outcomes are described as gains rather than losses, percentages rather than counts, or one-time events rather than repeated risks.

Decision analysis should therefore present outcomes in forms that preserve important quantities and make framing effects visible. Where possible, show both absolute and relative measures rather than allowing one presentation to dominate interpretation.

19. The Allais problem shows that plausible preferences can challenge simple axioms

Classic decision-theory puzzles such as the Allais paradox reveal systematic tensions between observed preferences and some expected-utility axioms. The lesson is not that every use of expected utility fails. It is that the axioms deserve to be understood rather than treated as invisible facts of nature.

The Stanford entry on Expected Utility provides a deeper route into the representation, arguments and objections surrounding the theory.

20. Expected regret answers a different question

Regret compares the payoff from the chosen action with the payoff that would have been obtained by the best action for the realised state. A minimax-regret rule focuses on limiting the worst shortfall relative to hindsight’s best choice.

This can be useful under deep uncertainty when probabilities are weak. But regret is not a probability-free version of expected utility. It uses a different criterion and can rank options differently.

21. Maximin protects the worst case but can be too conservative

A maximin rule chooses the action whose worst outcome is best. This can be appropriate when catastrophic downside dominates the decision and probabilities are untrustworthy.

It can also reject enormous upside to improve a relatively mild worst case. The correct lesson is not to prefer maximin universally, but to make the decision criterion explicit and match it to the consequence structure.

22. Robust decisions are not always expected-value winners

A high expected value can depend on a narrow probability model. Another option may perform reasonably across a wider set of plausible models. Robust decision-making asks whether the preferred action survives uncertainty about the model itself.

This connects directly to Sensitivity Analysis and Robustness Checks. The goal is to learn whether uncertainty about inputs, structure or preferences changes the action.

23. Deep uncertainty requires a different posture

Sometimes assigning precise probabilities is itself unjustified. The future may contain structural change, strategic opponents or unprecedented conditions.

Under deep uncertainty, useful decision rules may emphasise robustness, reversibility, staged commitment, option preservation and monitoring. The formal choice problem still matters, but the belief layer becomes a set of scenarios or model families rather than one trusted probability distribution.

24. Multi-criteria decisions need more than one outcome dimension

Real decisions often involve cost, time, reliability, accessibility, learning value and environmental consequence simultaneously. Compressing all of them into one number can be useful only if the aggregation rule is defensible.

Some criteria are constraints rather than tradeable benefits. A safety requirement may define the feasible set before any utility comparison. A rights boundary should not be assigned a small negative weight simply so a large commercial gain can overpower it.

25. Group decisions add a social-choice layer

When several people are affected, the decision-maker must decide how their interests enter the model. Averaging preference scores, using majority choice, prioritising the worst-off group or applying explicit rights constraints produce different outcomes.

Decision theory for one agent therefore does not automatically solve collective decision-making. Aggregation is a substantive governance and ethical choice.

26. Decision quality is not the same as outcome quality

A coherent decision can produce a bad realised outcome because uncertainty remains. A poor decision can succeed through luck.

Evaluate decision quality using the information, preferences and feasible actions available at the time. Preserve the model and assumptions before the outcome is known. Otherwise hindsight can rewrite a risky success as genius and a sensible failure as obvious incompetence.

27. Sequential decisions create an option to learn

A large commitment can sometimes be broken into stages: small pilot, observation, revised choice, larger commitment. This structure can create value even when the pilot itself has modest direct benefit.

The option to wait for information is not automatically beneficial. Delay can consume time, money or opportunity. A sequential model should therefore include the cost of learning and the possibility that options disappear while we wait.

28. Real options connect reversibility to uncertainty

A real option is the ability, but not the obligation, to take a future action: expand a pilot, abandon a project, switch suppliers, defer an investment or preserve a route.

Optionality matters most when future information can change what we want to do. It has little value if the future choice is already fixed or if the option expires before meaningful learning can occur.

29. Decision models can become performative

When people act on a decision model, they can change the system the model describes. A staffing model changes queues. A ranking changes applicant behaviour. A price recommendation changes demand.

The decision should therefore return to observation. Revalidate probabilities, outcomes and preference assumptions after implementation rather than treating the model as permanently correct.

30. A precise decision model can still be wrong

Most errors do not come from arithmetic. They come from a missing option, a wrong state model, a misleading outcome measure, a probability that lacks evidence, or an objective that leaves out the people who bear the cost.

Formalisation is valuable because it makes these assumptions inspectable. It becomes dangerous when the formula is used to hide them.

31. The best model is sometimes “no numerical model yet”

If consequences cannot be measured responsibly or probabilities cannot be defended, forcing a full expected-utility calculation can create false precision. The right next move may be qualitative scenario comparison, a bounding analysis or a targeted experiment.

Decision theory does not require fake numbers. It requires a clear account of what is known, what is valued and what remains unresolved.

32. A worked decision-theory audit

Return to the fictional event. S returns 60 in either attendance state. X returns 100 if attendance is high and 20 if low. Let the probability of high attendance be p.

• Expected value of S: 60.
• Expected value of X: 20 + 80p.
• Risk-neutral threshold: choose X when p > 0.5.
• Risk-averse decision: depends on utility attached to the two X outcomes, not expected points alone.
• Perfect information: would allow X in high attendance and S in low attendance.
• New evidence: matters only through how it changes the state beliefs and, potentially, the preferred action.

That compact audit exposes the whole architecture: action, state, outcome, probability, utility and information timing.

33. A decision-theory checklist

1. What exactly is being chosen?
2. What actions are genuinely available?
3. What uncertain states matter?
4. What outcome follows from each action-state combination?
5. Which outcomes are inadmissible regardless of payoff?
6. What evidence supports the state probabilities?
7. How sensitive is the choice to those probabilities?
8. Are outcome values linear, or is a utility representation needed?
9. Whose preferences are represented?
10. Are any options dominated?
11. What happens under the worst plausible state?
12. Would information arriving before commitment change action?
13. What does waiting cost?
14. Can the decision be staged or reversed?
15. Which assumptions should trigger review after implementation?

34. The educational habit beneath the mathematics

The deepest habit is not “always maximise expected utility”. It is to stop collapsing evidence, value and choice into one vague intuition.

Ask what can happen, how likely it is, what each consequence means, who bears it, what can still be changed later and what information would alter the decision. Formal decision theory is valuable because it turns those questions into a structure that can be challenged rather than a feeling that cannot.

Sources and further reading

This is an explanatory synthesis with original hypothetical calculations. It is not a substitute for domain-specific risk, legal, financial, clinical or public-policy analysis. Sources were checked for this edition on 5 September 2026.

• Stanford Encyclopedia of Philosophy — Decision Theory.
• Stanford Encyclopedia of Philosophy — Normative Theories of Rational Choice: Expected Utility.
• Stanford Encyclopedia of Philosophy — Descriptive Decision Theory.

Continue through eduKate: use Value of Information for research-before-action decisions, Sensitivity Analysis and Robustness Checks for fragile assumptions, Bayesian Inference and Updating for changing beliefs, and Strategic Decision-Making for the wider organisational context. Return to the Research Collections Directory for the broader evidence-methods route.