Key takeaway
Information has value when it changes the action worth taking. Compare the expected improvement with every study cost, using explicit assumptions rather than treating uncertainty as a price.
Start with an action the research could change
A proposed data project can justify more diligence without justifying every possible study. The useful question is whether an additional result could improve the decision enough to cover its cost. A report that sounds reassuring but leaves the same action optimal in every case may have little decision value in the model.
Jackson and colleagues distinguish expected value of perfect information, expected value of sample information and expected net benefit of sampling. Their paper concerns health-impact modelling; the accompanying voi package documents the calculations. The method transfers as a decision calculation. Neither source supplies a success probability, buyer demand or commercial value for a data license.
Perfect information is a hypothetical benchmark: compare the expected benefit of choosing with full knowledge against choosing with current information. An imperfect study normally gives less improvement. Subtract its cost separately. The resulting expectation is conditional on the model, its available actions and the chosen objective.
A complete invented decision before buying a study
Hypothetical example: a company may start one project or do nothing. Starting costs USD 2,000: ten staff hours at USD 100 plus USD 1,000 for an outside review. Those costs are incurred if it starts, whether or not a deal results. Success produces an assumed gross receipt of USD 8,000; failure produces zero.
Use an invented 0.30 probability of success, not an observed conversion rate. The net project payoffs are therefore USD 6,000 on success and minus USD 2,000 on failure. Doing nothing gives zero. Assume no other external fees, recurring costs, tax, financing or discounting, and a decision maker comparing expected dollars.
The expected net benefit of starting now is 0.30 × 6,000 − 0.70 × 2,000 = USD 400. Under these assumptions, starting beats doing nothing. Free perfect knowledge would permit starting only on success, giving an expected USD 1,800. The improvement, USD 1,400, is the perfect-information ceiling within this model.
An imperfect study with an explicit false-positive rate
Now suppose an optional study costs USD 500 in every outcome. Assume it reports positive for 80% of eventual successes and 20% of eventual failures. These are invented study-performance inputs; they are not measured accuracy claims.
The joint probability of success and a positive result is 0.30 × 0.80 = 0.24. Failure and a positive result has probability 0.70 × 0.20 = 0.14. Starting only after a positive result gives expected project benefit 0.24 × 6,000 − 0.14 × 2,000 = USD 1,160 before the study fee. The negative-result branch makes starting unattractive under the stipulated payoffs.
Expected sample information is worth USD 1,160 − USD 400 = USD 760 here. After paying USD 500, the study policy has expected net benefit USD 660, improving on the current decision by USD 260. This is an ex-ante comparison made before seeing the study result.
| Illustrative policy | Expected net benefit | Meaning |
|---|---|---|
| Start using current information | USD 400 | Current best action among the two assumed choices |
| Free perfect knowledge; start only on success | USD 1,800 | Hypothetical ceiling, not a purchasable guarantee |
| Study; start only after positive, before fee | USD 1,160 | Imperfect information policy |
| Same study policy after USD 500 fee | USD 660 | USD 260 expected improvement over starting now |
Preserve the losing branches and the model boundary
The positive expected improvement does not make the study profitable in every outcome. A negative result with no project still loses the USD 500 study fee. A false positive followed by project failure loses USD 2,500. A positive result followed by success gives USD 5,500 after both costs.
Before commissioning a real study, replace the invented probabilities, payoff branches and study-performance assumptions with defensible inputs and sensitivity checks. Add delay, implementation, further-review and no-deal costs where relevant. Different available actions, risk preferences or constraints can change the choice.
Use earnings to keep gross receipts separate from costs and preserve the decision tree beside the assumptions. The perfect-information ceiling is not a dataset valuation or a referral-fee quote. VOID has not measured the probability of a paid outcome, and this example does not promise one.