Probability of error isn't the cost of error.

What a wrong decision costs.

Start with what it even means to "be wrong." You make the funding decision the way Module 7 said to: on expected net benefit. If the mean net benefit of adopting is positive, you adopt. That's the right call given what you know — but it's a bet, because the true net benefit could turn out negative.

So imagine you could see one particular version of the truth — one draw from the PSA cloud. Two things can happen:

• Your decision was right for this truth. You adopted, and net benefit really is positive. You lose nothing — you'd have made the same choice with perfect knowledge.
• Your decision was wrong for this truth. You adopted, but in this world net benefit is actually negative — rejecting would have been better. The gap between the net benefit you'd have got by choosing correctly and the net benefit your actual decision delivered is the opportunity loss: the health-and-money you forfeited by betting wrong.

Crucially, you only incur an opportunity loss in the runs where your decision turns out wrong. In every run where you happened to bet right, the loss is zero — perfect information wouldn't have changed a thing.

Average the loss: EVPI per patient.

Now do that across the whole cloud. Each of the thousands of PSA runs is one possible truth; in each, you compute the opportunity loss of your fixed, here-and-now decision. In most runs — the ones on the favourable side — the loss is zero. In the runs where the truth turns against your bet, there's a real loss. Average all those losses together.

That average is the expected value of perfect information (EVPI) per patient. Put formally, it's the expected net benefit you'd achieve if you could always choose correctly (perfect information), minus the net benefit of the single best decision you can make under present uncertainty:

EVPI = E[net benefit with perfect information] − net benefit of the best decision now

Read it as: the average amount of value you're leaving on the table because you have to decide without knowing the truth. If you almost always bet right, that average is tiny. If the truth swings your decision often, it's large. And notice what it's built from — the losses, weighted by how often they occur: probability of being wrong, multiplied by how much being wrong costs. That product is the whole idea.

Scale it up: population EVPI.

So far EVPI is per patient — often a modest-looking number, maybe a few hundred pounds. But a funding decision isn't made for one patient. It's made once and then applied to every patient who will receive this technology for as long as the decision stands.

So multiply:

Population EVPI = per-patient EVPI × number of future patients over the decision's lifetime

If per-patient EVPI is £560 and 20,000 patients will be treated under this decision over its relevant horizon, the population EVPI is £560 × 20,000 ≈ £11 million. That transformation is what makes EVPI a decision tool rather than a curiosity: it converts an abstract per-patient uncertainty into a concrete pound figure — the total value at stake across everyone the decision touches. And because the patient population often runs to tens or hundreds of thousands, the scale-up frequently dominates: a trivial per-patient loss becomes an enormous population one.

That population figure has a precise meaning, which the next screens unpack: it's the most it could ever be worth to eliminate this uncertainty — the ceiling on the value of any research you might commission before deciding.

Probability × consequence.

Below is the incremental net benefit of a new technology — positive on average, so the base decision is to adopt, but uncertain. Three sliders. Uncertainty widens or narrows the spread. Stakes shifts how far the average sits from the break-even line (how consequential the call is). Population sets how many future patients the decision covers. Watch per-patient and population EVPI respond.

The two big lessons are the two extremes. A huge, sprawling cloud can be worth nothing to resolve if both sides of it lead to the same decision. And a narrow, confident cloud can be worth a fortune to resolve if it straddles the decision line and applies to a vast population. EVPI is never uncertainty alone — it's uncertainty times what's riding on it. (For clarity we sample net benefit directly here; a real analysis derives it from the full PSA — same logic, one layer deeper.)

Now you.

Two anchors of intuition.

Two facts about EVPI overturn the instinct that "more uncertain" means "more valuable to study." Hold onto both.

EVPI is zero when uncertainty doesn't change the decision. If every plausible truth — the whole cloud — leads you to the same choice, then perfect information would never make you act differently, so it's worth nothing. You can be wildly uncertain about the exact net benefit and still have an EVPI of zero, as long as the uncertainty all sits on one side of the decision line. Uncertainty with no power to change your mind is free. This is the single most counter-intuitive and most important thing about value of information.

EVPI scales with consequence and population. The flip side: a small uncertainty can carry a huge EVPI when the stakes are large. A decision balanced right on the threshold, applied to hundreds of thousands of patients, can make even a narrow cloud worth tens of millions to resolve. The magnitude of the decision — how much net benefit rides on getting it right, times how many patients — is doing as much work as the uncertainty itself.

Put together: "very uncertain" is not the same as "worth researching," and "nearly certain" is not the same as "not worth researching." Only the product — probability of being wrong, times the cost of being wrong, times the number of patients — tells you which.

From EVPI to the research decision.

This is what EVPI is for. It reframes the perennial complaint "there's too much uncertainty to decide" into an answerable question: is it worth buying more information before we commit?

The logic is a straight comparison. Population EVPI is the ceiling on the value of research — the benefit if a study removed all the uncertainty. Set it against what a study would cost:

• Population EVPI < cost of the study → the uncertainty, however uncomfortable, is cheaper to live with than to resolve. Decide now on expected value; don't research.
• Population EVPI > cost of the study → the uncertainty is expensive enough that reducing it could pay for itself. Research may be worthwhile — a green light to look closer.

This is exactly the machinery behind conditional reimbursement decisions. Rather than a blunt approve/reject, an agency can choose approve with research (fund it, but gather evidence to cut the uncertainty) or only in research (fund it only within a study), and EVPI is what tells them whether the uncertainty is worth that effort at all. Two refinements live just past this lesson: you can compute the value of resolving a single parameter rather than all of them (the expected value of partial perfect information, EVPPI — useful for designing which study to run), and you can subtract the study's cost to get its expected net benefit directly. But the core move is this one: EVPI is the bridge from measuring uncertainty to deciding what to do about it. One caveat carries over from the whole module — standard EVPI is built on a PSA, so it prices parameter uncertainty, not the structural uncertainty of the last lesson.

Which has the higher EVPI?

Why this matters for HTA

EVPI is where uncertainty analysis stops being descriptive and starts driving a decision of its own — whether to decide now or to buy more evidence first.

• Use it to justify — or refuse — more research. When someone argues "we need more data before funding this," EVPI is the discipline that tests the claim. If population EVPI is below the cost of the trial being proposed, more research is a poor use of money and the decision should be made now. Uncertainty is not, by itself, a reason to delay.
• Read it as a ceiling, not a price. Population EVPI is the value of removing all uncertainty — an ideal no real study achieves. A trial that resolves only part of it is worth less. So EVPI exceeding a study's cost is a necessary condition for that study to be worthwhile, not a guarantee. It sets the upper bound on the conversation.
• Remember what it prices. Standard EVPI, built on a PSA, values parameter uncertainty. It doesn't put a number on the structural doubt from the last lesson — a model resting on a shaky extrapolation can show a low EVPI while being deeply untrustworthy. Value of information answers "is it worth learning the true values?", not "is this the right model?"

A CEAC tells you how likely you are to be wrong. EVPI tells you what that risk is worth losing sleep over — and, precisely, how much you should be willing to pay to sleep better. It's the moment uncertainty analysis turns into a decision about uncertainty itself.

The value of information, in one breath.

• A wrong decision carries an opportunity loss — the net benefit forgone by betting wrong. Averaged over the whole PSA cloud, that loss is the EVPI per patient: the value left on the table because we must decide without knowing the truth.
• EVPI = probability of being wrong × the cost of being wrong. So it's zero when uncertainty doesn't change the decision (however wide the cloud), and large when a close call meets a big population (however narrow the cloud).
• Population EVPI = per-patient EVPI × future patients over the decision's lifetime — the ceiling on what any research to resolve the uncertainty could be worth.
• Compare it to a study's cost: below cost → decide now; above cost → research may pay. It's the bridge from measuring uncertainty to conditional decisions (approve / reject / approve-with-research / only-in-research) — and, like the whole PSA, it prices parameter, not structural, uncertainty.

Every earlier tool asked how uncertain we are. EVPI asks the only question a decision-maker can act on: given that uncertainty, is it worth paying to learn more — or is it time to decide? That's where the analysis of uncertainty finally becomes a decision.

That closes Module 9. You can now build a model, run it, and interrogate every layer of its uncertainty — from a single wiggled input to the value of resolving what you don't know. But every question so far has been the same question: is this worth it? — value for money. There's a different question a health system must also answer, and a technology can pass the first while failing the second: not "is it worth it?" but "can we afford it?" That's budget impact analysis, and it opens Module 10.