eduKateSG · Judgement · The Gold Standard of Education
Use multi-attribute utility models only when their independence assumptions match the decision
Choose the explanation you need, or follow all twelve chapters in order.
Full chapter index · Independent practice · Judgement subject guide
Multi-attribute utility theory helps structure choices whose outcomes are uncertain and matter on several attributes. It combines single-attribute utilities under conditions such as utility independence. Additive and multiplicative forms are earned by those conditions; they are not interchangeable spreadsheet styles.
This guide separates value under certainty from utility under risk, builds a fictional assessment, tests additive and multiplicative representations, and shows when interactions or hard constraints require a different model.
Continue from The Gold Standard of Judgement within The Gold Standard of Education: learning that transfers into life and the How X Works learning library.
Full chapter index
See the mechanism · Chapters 1–2
Build the worked model · Chapters 3–5
Test what changes · Chapters 6–8
Check evidence and limits · Chapters 9–10
Practise and continue · Chapters 11–12
Name the decision maker, alternatives and uncertain events. A consequence is a vector of attribute levels, such as cost, reliability and recovery time. A lottery assigns probabilities to possible consequence vectors.
Attributes should be understandable, measurable and sufficient for the choice without avoidable double counting. Operational measures belong beside their units and time horizons. A legal minimum or physical feasibility condition may need to screen options before utility aggregation.
Multi-attribute utility theory addresses preferences over uncertain consequences. If every outcome is treated as certain, a multi-attribute value model may be the more accurate term. Do not borrow the language of risk without eliciting risk preferences.
Evidence check: for Define attributes, consequences and lotteries, preserve the observation before the interpretation. Record decision maker, lotteries, attributes and constraints. Then state the inference, one credible alternative, and the additional observation that would distinguish them. This compact audit keeps a fluent explanation tied to evidence that another reader can inspect.
Changed-context transfer: repeat this chapter’s decision after changing one consequential condition while holding the others fixed. Calculate or predict what should change, name the threshold or mechanism, and compare with the original case. If the conclusion survives, state the range tested; if it reverses, explain why. Transfer is demonstrated by the changed case, not asserted from the first example.
Practical decision record: write one sentence for what is known, one for what remains uncertain, and one for the next safe action. Include the evidence owner, the latest useful decision time, and the condition that would stop or reverse the action. This turns a technical result into a reviewable choice without pretending uncertainty has disappeared.
For each attribute, define a least and most preferred reference level and scale utility between them, often 0 and 1. Under expected utility, intermediate points can be elicited through choices between a certain level and a lottery over the endpoints.
Suppose a decision maker is indifferent between a guaranteed reliability of 96% and a lottery with probability 0.7 of 99% and 0.3 of 90%. With endpoint utilities 1 and 0, the elicited utility of 96% is 0.7 under the stated comparison.
Repeat and check consistency. Risk attitude can differ by attribute and range. A smooth fitted curve is a model of elicited preferences, not a physical law.
Avoid certainty equivalents that respondents do not understand. Use probability wheels, plain-language consequences and practice questions. Reverse the framing and repeat selected comparisons to check stability. Large reversals may indicate learning, fatigue or that the attribute is not being interpreted consistently. Record intervals when a person can identify only a range of acceptable probabilities.
Evidence check: for Build single-attribute utility functions, preserve the observation before the interpretation. Record value under certainty and utility under risk. Then state the inference, one credible alternative, and the additional observation that would distinguish them. This compact audit keeps a fluent explanation tied to evidence that another reader can inspect.
Changed-context transfer: repeat this chapter’s decision after changing one consequential condition while holding the others fixed. Calculate or predict what should change, name the threshold or mechanism, and compare with the original case. If the conclusion survives, state the range tested; if it reverses, explain why. Transfer is demonstrated by the changed case, not asserted from the first example.
Practical decision record: write one sentence for what is known, one for what remains uncertain, and one for the next safe action. Include the evidence owner, the latest useful decision time, and the condition that would stop or reverse the action. This turns a technical result into a reviewable choice without pretending uncertainty has disappeared.
Attribute X is utility independent of Y when preferences over lotteries that vary X do not depend on the fixed level of Y, apart from permitted scaling. This is stronger than statistical independence and different from saying the attributes are uncorrelated.
For example, cost and reliability can be statistically correlated in the evidence while preferences remain utility independent after consequences are stated. Conversely, they may be statistically uncorrelated while the decision maker refuses to accept cost risk whenever reliability is below a threshold. Data relationships and preference relationships answer different questions.
Ask conditional questions at several fixed levels of other attributes. If risk attitude toward cost changes sharply when safety is low, utility independence may fail.
Independence is a judgement about preferences. Data cannot establish it alone. Record elicitation questions, respondents and disagreements so reviewers can see why a functional form was accepted.
Evidence check: for Understand utility independence, preserve the observation before the interpretation. Record single-attribute curves and elicitation. Then state the inference, one credible alternative, and the additional observation that would distinguish them. This compact audit keeps a fluent explanation tied to evidence that another reader can inspect.
Changed-context transfer: repeat this chapter’s decision after changing one consequential condition while holding the others fixed. Calculate or predict what should change, name the threshold or mechanism, and compare with the original case. If the conclusion survives, state the range tested; if it reverses, explain why. Transfer is demonstrated by the changed case, not asserted from the first example.
Practical decision record: write one sentence for what is known, one for what remains uncertain, and one for the next safe action. Include the evidence owner, the latest useful decision time, and the condition that would stop or reverse the action. This turns a technical result into a reviewable choice without pretending uncertainty has disappeared.
CHAPTER 4 OF 12 · Build the worked model
4. Distinguish mutual and additive independence
Mutual utility independence requires every subset of attributes to be utility independent of its complement. Under suitable conditions it supports multilinear forms and, with standard scaling, common additive or multiplicative representations.
Additive independence concerns preferences over probability distributions with the same marginals on each attribute. It supports an additive expected-utility form and is a demanding condition because it rules out preference for particular combinations beyond their marginal contributions.
Do not infer additive utility merely because criteria look separate on an objectives tree. Test whether combinations such as low cost with low reliability carry special concern.
Evidence check: for Distinguish mutual and additive independence, preserve the observation before the interpretation. Record every probability, utility and calculation. Then state the inference, one credible alternative, and the additional observation that would distinguish them. This compact audit keeps a fluent explanation tied to evidence that another reader can inspect.
Changed-context transfer: repeat this chapter’s decision after changing one consequential condition while holding the others fixed. Calculate or predict what should change, name the threshold or mechanism, and compare with the original case. If the conclusion survives, state the range tested; if it reverses, explain why. Transfer is demonstrated by the changed case, not asserted from the first example.
Practical decision record: write one sentence for what is known, one for what remains uncertain, and one for the next safe action. Include the evidence owner, the latest useful decision time, and the condition that would stop or reverse the action. This turns a technical result into a reviewable choice without pretending uncertainty has disappeared.
An additive representation takes u(x) = sum of k_i u_i(x_i), with non-negative scaling constants k_i summing to one. It allows utility losses on one attribute to be offset by gains on another, subject to any constraints applied first.
Suppose reliability, recovery time and cost have k values 0.50, 0.30 and 0.20. If Option A has single-attribute utilities 0.8, 0.6 and 0.5, its overall utility is 0.40 + 0.18 + 0.10 = 0.68.
The k values are scaling constants tied to attribute ranges and utility functions. They are not abstract importance percentages that can be moved to another decision unchanged.
| Attribute | Scaling constant | Single utility | Contribution |
|---|---|---|---|
| Reliability | 0.50 | 0.80 | 0.40 |
| Recovery time | 0.30 | 0.60 | 0.18 |
| Cost | 0.20 | 0.50 | 0.10 |
| Overall | 0.68 |
Evidence check: for Use the additive model as a special case, preserve the observation before the interpretation. Record utility versus statistical independence. Then state the inference, one credible alternative, and the additional observation that would distinguish them. This compact audit keeps a fluent explanation tied to evidence that another reader can inspect.
Changed-context transfer: repeat this chapter’s decision after changing one consequential condition while holding the others fixed. Calculate or predict what should change, name the threshold or mechanism, and compare with the original case. If the conclusion survives, state the range tested; if it reverses, explain why. Transfer is demonstrated by the changed case, not asserted from the first example.
Practical decision record: write one sentence for what is known, one for what remains uncertain, and one for the next safe action. Include the evidence owner, the latest useful decision time, and the condition that would stop or reverse the action. This turns a technical result into a reviewable choice without pretending uncertainty has disappeared.
When mutual utility independence holds but scaling constants do not sum to one, a multiplicative form may be used. One common normalisation is 1 + K u(x) = the product over attributes of [1 + K k_i u_i(x_i)], where K is chosen to satisfy the endpoint normalisation.
K captures interaction created by the scaling structure. As K approaches zero, the form approaches the additive case. The sign and magnitude should be derived from elicited constants, not chosen to reward a preferred option.
Solve the normalisation equation carefully and verify endpoint behaviour: the all-worst consequence should have utility zero and the all-best consequence utility one. Depending on the constants, the equation has the zero additive limit and may have a non-zero solution for the multiplicative form. State which solution is used and why.
The formula does not license arbitrary synergy. If preferences contain a specific interaction such as reliability mattering only above a security threshold, model that condition explicitly and retest independence.
Evidence check: for Understand the multiplicative representation, preserve the observation before the interpretation. Record additive and multiplicative conditions. Then state the inference, one credible alternative, and the additional observation that would distinguish them. This compact audit keeps a fluent explanation tied to evidence that another reader can inspect.
Changed-context transfer: repeat this chapter’s decision after changing one consequential condition while holding the others fixed. Calculate or predict what should change, name the threshold or mechanism, and compare with the original case. If the conclusion survives, state the range tested; if it reverses, explain why. Transfer is demonstrated by the changed case, not asserted from the first example.
Practical decision record: write one sentence for what is known, one for what remains uncertain, and one for the next safe action. Include the evidence owner, the latest useful decision time, and the condition that would stop or reverse the action. This turns a technical result into a reviewable choice without pretending uncertainty has disappeared.
CHAPTER 7 OF 12 · Test what changes
7. Compare fictional additive and multiplicative results
Consider two attributes with k1 = 0.7 and k2 = 0.6. Their sum exceeds one, so the normalised multiplicative constant is non-zero. Suppose Option P has u1 = 0.8 and u2 = 0.4, while Option Q has 0.6 and 0.7.
Calculate K from 1 + K = (1 + 0.7K)(1 + 0.6K). The non-zero solution is approximately -0.714. Substituting produces comparable overall utilities under this exact scaling contract. Show each factor and retain enough precision to reproduce the ranking.
Do not compare these outputs with an additive model that uses the same unnormalised k values as if nothing changed. The models encode different trade-off structures. The point is to test representation, not to manufacture a numerical winner.
Evidence check: for Compare fictional additive and multiplicative results, preserve the observation before the interpretation. Record K normalisation and interaction meaning. Then state the inference, one credible alternative, and the additional observation that would distinguish them. This compact audit keeps a fluent explanation tied to evidence that another reader can inspect.
Changed-context transfer: repeat this chapter’s decision after changing one consequential condition while holding the others fixed. Calculate or predict what should change, name the threshold or mechanism, and compare with the original case. If the conclusion survives, state the range tested; if it reverses, explain why. Transfer is demonstrated by the changed case, not asserted from the first example.
Practical decision record: write one sentence for what is known, one for what remains uncertain, and one for the next safe action. Include the evidence owner, the latest useful decision time, and the condition that would stop or reverse the action. This turns a technical result into a reviewable choice without pretending uncertainty has disappeared.
CHAPTER 8 OF 12 · Test what changes
8. Detect interactions and conditional preferences
Ask whether the desirability of improving one attribute depends on the level of another. Reliability gains may matter more when demand is high; response time may matter only when backup capacity is low. These are substantive interactions.
Use two-dimensional indifference assessments, scenario comparisons and residual checks against observed choices. A model that systematically mispredicts particular combinations is warning that separability fails.
Options include redefining attributes, adding an interaction term, using a multilinear model or keeping scenarios separate. More complex models require more elicitation and validation; complexity should buy explanatory accuracy.
Evidence check: for Detect interactions and conditional preferences, preserve the observation before the interpretation. Record conditional preferences and residual checks. Then state the inference, one credible alternative, and the additional observation that would distinguish them. This compact audit keeps a fluent explanation tied to evidence that another reader can inspect.
Changed-context transfer: repeat this chapter’s decision after changing one consequential condition while holding the others fixed. Calculate or predict what should change, name the threshold or mechanism, and compare with the original case. If the conclusion survives, state the range tested; if it reverses, explain why. Transfer is demonstrated by the changed case, not asserted from the first example.
Practical decision record: write one sentence for what is known, one for what remains uncertain, and one for the next safe action. Include the evidence owner, the latest useful decision time, and the condition that would stop or reverse the action. This turns a technical result into a reviewable choice without pretending uncertainty has disappeared.
CHAPTER 9 OF 12 · Check evidence and limits
9. Propagate probability and preference uncertainty
Outcome probabilities may be estimated with error, and utility assessments may be imprecise or disputed. Use ranges or distributions for both rather than treating elicited decimals as exact.
Check dependence among uncertain outcomes as well. Multiplying marginal probabilities or sampling each attribute independently can create impossible consequence combinations. Use a joint event model or scenario tree that preserves physical and operational relationships before applying utilities.
Document correlations and scenario probabilities so another analyst can reproduce the expected utility.
Run sensitivity on probabilities, single-attribute curves, scaling constants, K and constraints. Locate conditions where the preferred option reverses. Monte Carlo analysis can show ranking frequencies, but its input distributions must be justified.
Separate aleatory uncertainty from model and judgement uncertainty. A precise expected utility from uncertain inputs is not a precise decision fact.
| Layer | Variation | Question | Failure |
|---|---|---|---|
| Probability | Scenario or interval | Does expected utility reverse? | False certainty |
| Utility curve | Elicited range | Does risk attitude matter? | Linear convenience |
| Scaling constants | Plausible set | Which trade-off is decisive? | Copied weights |
| Model form | Additive vs interaction | Does independence hold? | Hidden conditionality |
Evidence check: for Propagate probability and preference uncertainty, preserve the observation before the interpretation. Record probability, utility and model sensitivity. Then state the inference, one credible alternative, and the additional observation that would distinguish them. This compact audit keeps a fluent explanation tied to evidence that another reader can inspect.
Changed-context transfer: repeat this chapter’s decision after changing one consequential condition while holding the others fixed. Calculate or predict what should change, name the threshold or mechanism, and compare with the original case. If the conclusion survives, state the range tested; if it reverses, explain why. Transfer is demonstrated by the changed case, not asserted from the first example.
Practical decision record: write one sentence for what is known, one for what remains uncertain, and one for the next safe action. Include the evidence owner, the latest useful decision time, and the condition that would stop or reverse the action. This turns a technical result into a reviewable choice without pretending uncertainty has disappeared.
CHAPTER 10 OF 12 · Check evidence and limits
10. Protect participation, constraints and auditability
Different stakeholders may face different consequences and risk attitudes. Averaging utilities can hide distributional conflict. Show group-specific models and identify who has decision authority.
Rights, safety floors and feasibility limits should not be traded away through a high aggregate score. Apply non-compensable constraints transparently.
Keep the elicitation understandable. If participants cannot explain what a lottery or scaling question means, pause and redesign it. A sophisticated formula cannot rescue invalid preference data.
Evidence check: for Protect participation, constraints and auditability, preserve the observation before the interpretation. Record participation and non-compensable limits. Then state the inference, one credible alternative, and the additional observation that would distinguish them. This compact audit keeps a fluent explanation tied to evidence that another reader can inspect.
Changed-context transfer: repeat this chapter’s decision after changing one consequential condition while holding the others fixed. Calculate or predict what should change, name the threshold or mechanism, and compare with the original case. If the conclusion survives, state the range tested; if it reverses, explain why. Transfer is demonstrated by the changed case, not asserted from the first example.
Practical decision record: write one sentence for what is known, one for what remains uncertain, and one for the next safe action. Include the evidence owner, the latest useful decision time, and the condition that would stop or reverse the action. This turns a technical result into a reviewable choice without pretending uncertainty has disappeared.
In Decision A, cost ranges from $1 million to $5 million. In Decision B, feasible costs are $4.8 million to $5 million. The cost utility curve and scaling constant must be reconsidered because the relevant swing shrank. Reusing the old k value overstates cost’s decision leverage.
Now reliability utility becomes much steeper when recovery time exceeds eight hours. The conditional pattern violates the earlier separability story. Add an interaction or split scenarios, then rerun sensitivity instead of adjusting constants until the original ranking returns.
Evidence check: for Run a changed-context transfer test, preserve the observation before the interpretation. Record the changed-range interaction cases. Then state the inference, one credible alternative, and the additional observation that would distinguish them. This compact audit keeps a fluent explanation tied to evidence that another reader can inspect.
Changed-context transfer: repeat this chapter’s decision after changing one consequential condition while holding the others fixed. Calculate or predict what should change, name the threshold or mechanism, and compare with the original case. If the conclusion survives, state the range tested; if it reverses, explain why. Transfer is demonstrated by the changed case, not asserted from the first example.
Practical decision record: write one sentence for what is known, one for what remains uncertain, and one for the next safe action. Include the evidence owner, the latest useful decision time, and the condition that would stop or reverse the action. This turns a technical result into a reviewable choice without pretending uncertainty has disappeared.
Archive alternatives, uncertain events, probabilities, attributes, ranges, single-attribute elicitation, lotteries, utility curves, independence tests, scaling constants, K solution, constraints, calculations, stakeholder roles and sensitivity runs. Common questions: Is utility independence statistical independence? No. Does a sum of weights greater than one prove error? Not in a multiplicative scaling. Is additive always simpler? Yes, but only valid under its conditions. Can utilities be copied? Not safely across changed ranges or decision makers.
Continue with swing weighting and reference levels, MACBETH cardinal value scales, and UTA preference disaggregation. For foundations, use Keeney’s open survey of multi-attribute utility analysis and Purdue Engineering’s multi-attribute utility theory module.
Evidence check: for Keep the utility route and continue, preserve the observation before the interpretation. Record the narrowest valid utility claim supported. Then state the inference, one credible alternative, and the additional observation that would distinguish them. This compact audit keeps a fluent explanation tied to evidence that another reader can inspect.
Changed-context transfer: repeat this chapter’s decision after changing one consequential condition while holding the others fixed. Calculate or predict what should change, name the threshold or mechanism, and compare with the original case. If the conclusion survives, state the range tested; if it reverses, explain why. Transfer is demonstrated by the changed case, not asserted from the first example.
Practical decision record: write one sentence for what is known, one for what remains uncertain, and one for the next safe action. Include the evidence owner, the latest useful decision time, and the condition that would stop or reverse the action. This turns a technical result into a reviewable choice without pretending uncertainty has disappeared.