Can several rules be admissible?

Yes. Their risk functions can cross, leaving no single rule uniformly better than another.

Does an admissible rule always have the lowest worst-case risk?

No. That is a minimax question, which can select among rules using a different criterion.

Can changing the loss function change admissibility?

Yes. Dominance is evaluated using the risk induced by the chosen loss, so changing what counts as a costly error can change the conclusion.

AI Glossary / Statistics and probability / Intermediate

Admissible decision rule

Definition

An admissible decision rule is a statistical rule that is not dominated by another allowed rule under a specified model and loss function. No competing rule has expected loss at least as low for every possible parameter value and strictly lower for at least one. Admissibility is a property of the full decision problem.

What it means

A decision rule maps observed data to an action, such as an estimate, classification or operational choice. Its risk is the expected loss from using it, evaluated separately at each possible state or parameter value and averaging over the data the model could generate.

Admissibility rules out a particular kind of avoidable inferiority. It does not say that one rule is best at every state, that its losses are small, or that its actions meet practical requirements omitted from the loss function. Rules can be admissible because improving one state sacrifices another.

How it works

Specify the statistical model, parameter space, possible actions, loss function and class of allowed decision rules. For each rule, calculate or characterize its risk function R(theta, delta), the expected loss when the true parameter is theta.

A competing rule dominates a candidate if its risk is no greater at every parameter value and smaller at at least one. If such a competitor exists, the candidate is inadmissible. Equal risk everywhere does not establish dominance under this definition because strict improvement somewhere is required.

Showing that a rule beats a few familiar alternatives does not prove admissibility among all allowed rules. Proofs may require theoretical arguments or a complete analysis of a finite decision problem. Randomized rules can matter if they are allowed. Conclusions can change when the parameter space, loss function or permitted rule class changes, so these ingredients must accompany the result.

Three practical examples

1. An estimation researcher checks whether a proposed estimator has uniformly larger mean squared error than a competing estimator under the same model.

2. A decision analyst eliminates operational policies that incur no smaller expected cost in any modeled demand state and higher cost in at least one.

3. A model evaluator distinguishes an undominated set of statistical procedures from a final choice based on worst-case risk or an explicitly weighted average risk.

Think of it this way

Think of comparing travel plans across possible weather conditions. If one plan costs no more in any condition and less in at least one, a cost-only decision can discard the other. If each plan is better in different conditions, neither dominates, and another preference is needed to choose between them.

Worked example

Consider an original finite decision problem with two possible states and three allowed rules. Their expected losses in states 1 and 2 are A: (1, 5), B: (3, 3) and C: (4, 6). These are expected losses.

A dominates C because 1 is below 4 and 5 is below 6. B also dominates C because 3 is below both 4 and 6. C is therefore inadmissible. A and B do not dominate each other: A is better in state 1, while B is better in state 2. Both are admissible within this stated three-rule class.

If random mixtures of these rules are also allowed, mixtures of A and B trace a tradeoff between their two risks and do not dominate either endpoint. C remains dominated. A worst-case criterion selects B, whose maximum risk is 3 rather than A's 5. If state 1 instead receives probability 0.9 and state 2 probability 0.1, A's average risk is 0.9 times 1 + 0.1 times 5 = 1.4, below B's 3. Admissibility alone does not select between them.

Illustrative case study

Fictional case study: Stockhaven compares inventory policies using expected shortage and holding costs under several modeled demand conditions. One policy is no cheaper in any condition and is more expensive in some. The team removes it from consideration under the stated cost model.

Several policies remain because their risk profiles cross. Managers then state whether they prioritize worst-case cost or a probability-weighted average, and check whether operational constraints require changes to the decision problem. They also investigate whether the modeled demand conditions adequately cover plausible situations.

This fictional case illustrates dominance screening. It does not prove that an undominated policy is robust to model errors or that all relevant business consequences have been included.

Common misconceptions and mistakes

Admissible does not mean uniquely optimal, unbiased or minimax. Minimax minimizes the largest risk, while a Bayes rule minimizes risk averaged over a specified prior. Connections between these concepts require conditions; they are not interchangeable labels.

Risk dominance concerns expected loss across all modeled parameter values, not superiority on every individual dataset or realized outcome. A simulation on a limited grid may suggest dominance but usually cannot establish it over a continuous parameter space. Similarly, failure to find a dominating rule is not itself a proof that none exists.

Business use

Admissibility provides a disciplined way to remove rules with avoidable expected losses before addressing genuine tradeoffs. It encourages teams to specify costs and uncertainty clearly instead of ranking methods by a single favorable scenario.

For practical work, verify that the loss function captures relevant consequences and that the allowed rule class reflects real constraints. A rule that is statistically attractive may be too expensive to implement or may rely on unavailable information. Include such considerations in the decision problem rather than assuming admissibility accounts for them automatically. Use theoretical claims with their assumptions and keep empirical comparisons distinct from proofs about an entire model.

Check your understanding

Rule X has risks (2, 4) and rule Y has risks (2, 3) across the only two possible states. Is X admissible when Y is allowed?

Answer: No. Y is equally good in the first state and strictly better in the second, so Y dominates X.

Related terms

Loss function
Minimax
decision rule
risk function
dominance
Bayes rule
randomized decision rule and complete class

Learn next

construct a small risk table and identify its undominated rules. Then compare the minimax choice with choices obtained under several prior probabilities, keeping the same loss function and allowed actions throughout.

Sources

Primary sources: Fithian, Statistical decision theory, University of California, Berkeley: https://www.stat.berkeley.edu/~wfithian/courses/stat210a/models.html

Kasy, Lecture notes on statistical decision theory: https://scholar.harvard.edu/files/kasy/files/decisiontheory_0.pdf

Previous: Administrative data collection Next: Admissible heuristic