Can utility values vary while independence holds?

Yes; positive rescaling and offsets can preserve conditional lottery preferences.

Does independence work in both directions automatically?

No. Each direction needs justification.

Is it a claim about causality?

No. It concerns preferences, not whether changing one attribute physically affects another.

AI Glossary / Graphs and symbolic AI / Advanced

Utility independence

Definition

Utility independence is a preference condition in multiattribute decision theory. A set of attributes X is utility independent of the remaining attributes Y when preferences between lotteries over X do not change with the fixed value of Y. It concerns the decision-maker's preferences under uncertainty, rather than statistical independence between measured variables.

What it means

Suppose outcomes describe both delivery performance and customer-support arrangements. To examine utility independence, compare uncertain delivery alternatives while holding the support arrangement fixed and identical across the alternatives. Repeat the comparison at a different fixed support arrangement. The question is whether the ordering of those same delivery lotteries changes.
The condition covers all relevant lottery comparisons, including indifference, across the specified attribute ranges. Observing agreement on one example provides evidence, not a general proof about someone's preferences. The direction also matters: independence of X from Y does not automatically establish independence of Y from X.

How it works

Within an expected-utility representation, utility independence allows conditional utility functions over X to share the same shape up to a positive change of scale and an offset. One useful form is U(x,y)=a(y)+b(y)u(x), where b(y)>0. At fixed y, the offset cancels in comparisons and the positive multiplier preserves the ordering of expected u(x).
The actual utility values can therefore change with y without changing conditional lottery preferences. This does not by itself justify an additive utility function with separate, fixed contributions from each attribute.
Preference independence is a related, weaker concept involving comparisons of certain attribute combinations. Utility independence also tests uncertain alternatives. Statistical independence instead describes a probability distribution and cannot be established or rejected merely by inspecting a person's utility function.

Three practical examples

A procurement team compares two lotteries over delivery times under a fixed warranty arrangement, then repeats under another warranty. A reversal suggests that delivery-time utility should not be elicited once and reused unchanged without examining the interaction.
A scheduling assistant asks whether a user prefers a certain amount of free time to a gamble over more or less free time, holding travel burden fixed. Repeating this under a different burden tests whether the same uncertainty preference remains appropriate.
A production planner models output quantity and service support. Even if lottery preferences over quantity remain stable at each fixed support level, the model can still give quantity a larger utility scale when support is better. Stable ordering does not require identical numerical contributions.

Think of it this way

Imagine repeating the same taste test in rooms with different lighting. Utility independence asks whether the ranking of the offered uncertain taste experiences stays the same in every fixed room. The overall experience can improve with the room; what must remain stable is the conditional comparison being tested.

Worked example

Consider a toy model with quantity attribute x in {0,1,2} and support attribute y in {0,1}. Define U(x,y)=(1+y)x+2y. At y=0, utilities across x are 0,1,2. At y=1, they are 2,4,6.
For any lottery over x, expected utility is E[x] when y=0 and 2E[x]+2 when y=1. Every ordering and tie between such lotteries is preserved, so x is utility independent of y in this model. A certain x=1 ties a half-and-half lottery over x=0 and x=2 at either support level.
However, this particular U is not a sum of a function of x and a function of y: increasing x from 0 to 1 adds 1 at y=0 but 2 at y=1. Its interaction term xy prevents that additive decomposition.
Now replace the table with utilities 0,2,3 at y=0 and 10,11,14 at y=1. Larger certain x remains preferable at both support levels. Yet the certain middle outcome beats the endpoint lottery at y=0, since 2 exceeds 1.5. At y=1 it loses, since 11 is below 12. This second model violates utility independence even though the ordering of certain x outcomes never reverses.

Illustrative case study

In a fictional equipment distributor, an analyst proposes eliciting delivery preferences separately from after-sales support and combining the resulting scores. Interviews initially suggest that faster delivery is always preferred when everything else is fixed.
A further exercise compares a dependable intermediate delivery time with a gamble between very early and very late delivery. Managers prefer the dependable option without strong support but accept the gamble when support provides a useful contingency. The direction of the certain-time ranking has not changed; the uncertain comparison has.
The team therefore records that the proposed utility-independence assumption fails for these scenarios. It models the interaction explicitly and checks further combinations before using the result in procurement decisions. The exercise prevents a convenient decomposition from being mistaken for an established preference property.

Common misconceptions and mistakes

“Uncorrelated attributes are utility independent.” Correlation concerns data, not preference comparisons.
“Utility independence means the other attribute has no value.” It may change overall utility and its scale while leaving the tested lottery order unchanged.
“Certain-outcome rankings are enough.” Lottery preferences can reverse despite identical rankings of certain outcomes.
“Utility independence automatically means additivity.” More general forms can satisfy the condition.
“One consistent answer proves independence.” The property concerns all relevant comparisons, while practical elicitation usually examines a limited selection.

Business use

The condition can reduce the burden of assessing a complex utility model, but only when its assumptions fit the decisions. Record the attribute ranges, fixed background conditions, and comparisons used to support the simplification.
Include examples near meaningful operational thresholds and unfavorable outcomes. If decision-makers disagree or reverse a lottery preference, investigate the reason instead of forcing the answers into an additive scorecard. An interaction may reveal a missing attribute or a contingency that changes the value of uncertainty. Test the resulting recommendations against complete business scenarios, since an elegant decomposition is useful only when it preserves the choices the organization intends to make.

Check your understanding

Why does the first model preserve lottery rankings over x?

Answer: Each conditional expectation is a positive affine transformation of E[x].

Why is its formula nonadditive?

Answer: The contribution of x changes with y through xy.

What disproves independence in the second table?

Answer: The middle outcome versus endpoint-lottery preference reverses.

Related terms

Utility function
Preference independence
Expected utility
Preference elicitation
multiattribute utility
additive independence
attribute interaction
conditional utility

Learn next

Distinguish utility independence from additive independence. Then test candidate decompositions with lottery comparisons that expose interactions between attributes.

Sources

Keeney, Utility Independence and Preferences for Multiattributed Consequences, author publication abstract: https://scholars.duke.edu/publication/1126388 Ormala, Supporting Technology Development Evaluation with Multiple Attribute Utility Analysis and Fuzzy Decision Analysis, sections 4.4–4.5: https://pure.iiasa.ac.at/2253/1/WP-83-059.pdf

Previous: Utility function Next: Utility nodes