Decisions & optimisation
Value of information: when is a test worthwhile?
Should a team pay for a market test or decide immediately? A worked decision tree shows when additional information changes the best action and how much that information is worth.
The short answer
Information has value when it can improve a decision. EVPI values perfect information, while EVSI values a particular imperfect test; subtracting test costs gives its expected net benefit.
Explore the interactive exampleFree · in your browser · no sign-up
At a glance
- More knowledge creates economic value when it enables a better action.
- Before costs, the value of an imperfect test lies between zero and the value of perfect information.
- Test accuracy, prior probability and the consequences of actions belong in the same decision tree.
A team can launch a new offer or decline to proceed. It could first run a market test. The useful question is not simply how accurate that test is: What is the most the team should pay for it? Answering requires connecting test results to subsequent actions. A more detailed dashboard alone does not create an economic benefit.
Start with the decision without testing
Our entirely fictional example has two demand states. Launching earns €80,000 under high demand and loses €40,000 under low demand. Not launching has value zero in either state. All amounts are comparable net values; time and other costs would otherwise need to be added. High demand initially has probability 40%.
In thousands of euros, launching is therefore the best action, with expected value 8. This does not mean the team will earn exactly €8,000 or that a loss is impossible. It evaluates the decision before the demand state is known. Decision theory explains how risk preferences can change the objective.
Perfect information as an upper bound
If demand were known with certainty beforehand, the team would launch only under high demand. Its expected value would be . The Expected Value of Perfect Information, or EVPI, is therefore thousand euros. It bounds the gross value of information about this particular state. Howard’s decision-focused treatment links information value to uncertainty, action and evaluation (Howard 1966).
This bound depends on the available actions and objectives. Adding a smaller launch changes the decision problem. EVPI is also not a recommended research budget: real tests rarely provide perfect information and may be worth substantially less.
An imperfect test and its decision tree
The test returns a positive signal with probability 80% under high demand and a negative signal with probability 80% under low demand. Sensitivity and specificity are deliberately equal here. A positive signal occurs with probability .
After a positive result, the probability of high demand becomes . After a negative result, it falls to . This is a Bayesian update. The team launches after a positive signal and declines after a negative one. The expected value of this strategy specified in advance is €20,800.
If the test costs €5,000, its expected net benefit is €7,800. Information analysis therefore distinguishes the expected value of sample information from the benefit after research costs (Rothery 2020).
Interactive: Value of information: when is a test worthwhile?
Without JavaScript, the starting example gives EVPI €24,000, EVSI €12,800 and net benefit €7,800 after a €5,000 test cost.
Three potentially surprising observations
A test with 50% sensitivity and specificity is uninformative in this model. Neither signal changes the prior probability, so its EVSI is zero. A perfect test reaches EVPI. Between these extremes, improving accuracy may initially change nothing if no possible result crosses the action threshold.
Substantial uncertainty does not guarantee substantial information value either. When the consequences of the available actions are nearly identical, little can be gained. Conversely, a moderately accurate test can be valuable when a mistaken action is expensive. Begin with a payoff table, not the size of a dataset.
Assumptions and limits
The model assumes calibrated probabilities, known test accuracy and demand that stays unchanged after testing. Waiting for the result may reduce the opportunity. Participants in the test may differ from the eventual market. These are additional assumptions, not details that a more precise calculation automatically resolves.
For your own question, record alternative actions, relevant states, consequences and test costs. Then ask explicitly: what would we do differently after each possible signal? If the answer is always “nothing”, reconsider the test design. A test should be linked to a choice before its budget is approved.
Frequently asked questions
How do EVPI and EVSI differ?
EVPI is the expected gain from resolving uncertainty completely. EVSI values the possible outcomes of a particular test before it is performed. Both compare the best informed decision with the best decision under current knowledge (Rothery 2020).
Can an accurate test be worthless?
Yes. If every possible result leads to the same optimal action, the test does not change the decision. Its gross value is then zero, even if the probability estimate improves (Howard 1966).
Sources and further reading
- PaperHoward, R. A. (1966): Information Value Theory.
- PaperRothery, C. et al. (2020): Value of Information Analytical Methods: Report 2 of the ISPOR Value of Information Analysis Emerging Good Practices Task Force.
Content last reviewed on 14 November 2026. Please let us know about any errors: impressum@nyxai.com.