Expected Value of Perfect Information

Expected Value of Perfect Information in Health Technology Assessment

EVPI puts a number on what a health system expects to lose by having to choose a treatment while the evidence is still uncertain. It is calculated from a probabilistic cost-effectiveness model and equals the expected cost of uncertainty, measured in QALYs or in money. The 2026 report on uncertainty from the Health Economics Methods Advisory (HEMA) recommends that EVPI be reported alongside every probabilistic analysis, with population EVPI as the minimum measure of what is at stake if the decision turns out to be wrong. A low EVPI tells a committee that further research cannot be worth much, while a high EVPI shows only that research might be worth funding, not that any particular study is.

Why a funding decision under uncertainty has an expected cost

An HTA body that aims to maximise population health chooses the option with the highest expected net health benefit on current evidence. The ISPOR value of information task force describes this as the choice of a risk-neutral decision maker. Because the evidence is imperfect, that choice will sometimes be the wrong one, and the HEMA report stresses that errors run in both directions. Funding a technology that later proves not cost-effective displaces health elsewhere, while delaying or denying one that later proves cost-effective also loses health, and for some patients that loss cannot be recovered.

Decision uncertainty is the chance of making such an error. EVPI goes one step further and weights each possible error by how much health it would cost. The HEMA glossary defines the cost of uncertainty in these terms, as the health expected to be lost by deciding on current evidence, combining how likely a wrong choice is with how much would be lost. The task force notes that a technology can have highly uncertain incremental costs and effects yet no decision uncertainty at the threshold used, in which case EVPI is zero.

The EVPI formula and the order of averaging

The ISPOR methods report defines EVPI as the difference between the expected value of a decision made with perfect information and the expected value of the decision made with current knowledge. In symbols it reads as follows.

$$EVPI = E_{\theta}\left[\max_{d} NB(d,\theta)\right] - \max_{d} E_{\theta}\left[NB(d,\theta)\right]$$

where $d$ indexes the options being compared, $\theta$ is the full set of uncertain model inputs, $NB(d,\theta)$ is the net benefit of option $d$ when the inputs take the values $\theta$, and $E_{\theta}$ is the expectation over the joint distribution of $\theta$.

The first term lets the decision maker pick the best option for each possible state of the world and then averages the result. The second term averages first and then picks one option, which is all that can be done before the truth is known. An average of maxima can never be smaller than the maximum of averages, so EVPI is never negative.

The same number can be reached through opportunity loss, which the task force uses in its worked example.

$$L(\theta) = \max_{d} NB(d,\theta) - NB(d^{*},\theta), \qquad EVPI = E_{\theta}\left[L(\theta)\right]$$

where $d^{*}$ is the option with the highest expected net benefit on current evidence and $L(\theta)$ is the net benefit forgone in state $\theta$ by having chosen $d^{*}$. The loss is zero wherever $d^{*}$ is also the best option.

Net benefit is usually measured as net health benefit (QALYs gained minus costs divided by the threshold) or net monetary benefit. Both depend on the cost-effectiveness threshold, so the methods report advises that EVPI be reported for explicit threshold values or plotted as a function of the threshold. The NICE manual (PMG36, section 4.10.8) asks for expected net health benefits, when appropriate, using values of GBP 25,000 and GBP 35,000 per QALY gained, the two values used in the example below.

Worked example with five equally likely draws (illustrative)

In this hypothetical case a new drug is compared with standard care. Five equally likely draws from a probabilistic sensitivity analysis give the incremental QALYs below, and the incremental cost is GBP 10,000 in every draw to keep the arithmetic simple. At GBP 25,000 per QALY the cost is worth 10,000 / 25,000 = 0.40 QALYs, so the incremental net health benefit (INHB) of the drug in each draw is its incremental QALYs minus 0.40.

1. Current information. The five INHB values sum to 0.35 QALYs, so the mean is 0.35 / 5 = 0.07 QALYs. It is positive, so the drug is adopted.
2. Perfect information. With the truth known, the better option would be chosen in each draw, giving net benefits of 0.10, 0, 0.30, 0 and 0.25 relative to standard care. Their mean is 0.65 / 5 = 0.13 QALYs.
3. EVPI. The difference is 0.13 minus 0.07, or 0.06 QALYs per patient. The loss column gives the same answer, 0.30 / 5 = 0.06.
4. Probability of error. The drug is the wrong choice in draws 2 and 4, two draws out of five, so the error probability is 40%.

At GBP 35,000 per QALY the cost is worth 10,000 / 35,000 = 0.2857 QALYs. The mean INHB rises to 0.47 minus 0.2857, or 0.1843 QALYs, and only draw 2 now favours standard care, with a loss of 0.0857 QALYs. EVPI falls to 0.0857 / 5 = 0.017 QALYs per patient and the error probability to 20%. Moving the expected net benefit further from zero shrinks both the chance and the expected size of a wrong decision.

Suppose 2,000 new patients a year face the decision for 10 years, discounted at 3.5% a year, the rate in the NICE reference case (PMG36, section 4.5.1). Discounting each year's 2,000 patients over years 0 to 9 gives 2,000 x 8.6077 = 17,215 patients. Population EVPI is 0.06 x 17,215 = 1,033 QALYs, or about GBP 25.8 million at GBP 25,000 per QALY. Any programme of research expected to cost more than that cannot be worthwhile for this decision. The population INHB of adopting now is 0.07 x 17,215 = 1,205 QALYs, so the value at stake from uncertainty is of the same order as the value of immediate access, the comparison the HEMA report proposes.

Computing EVPI from a probabilistic sensitivity analysis

The ISPOR methods report describes EVPI as a single-loop Monte Carlo calculation that needs no sampling beyond the probabilistic analysis itself. Each saved draw already holds the net benefit of every option, so EVPI follows directly.

A deterministic run at mean input values cannot stand in for this, because the report notes that in non-linear models it gives an incorrect estimate of expected net benefit. The HEMA report adds that one-way sensitivity analysis cannot give the probability of a different decision or the expected loss from decision uncertainty. It also observes that computing EVPI alongside a probabilistic analysis is no longer a substantial undertaking, since free tools such as the Sheffield Accelerated Value of Information (SAVI) application make it a routine step. It recommends that every scenario analysis be accompanied by a probabilistic analysis of that scenario and, at a minimum, its EVPI.

Population EVPI and the time horizon

EVPI is calculated per patient, but information, once gathered, informs every later choice between the same options. The methods report therefore scales the per-person value by the discounted size of the beneficiary population.

$$EVPI_{pop} = EVPI \times \sum_{t=0}^{T} \frac{I_t}{(1+r)^{t}}$$

where $EVPI_{pop}$ is population EVPI, $I_t$ is the number of patients facing the decision in period $t$, $T$ is the last period in which the information stays useful ($T = 9$ for the 10-year example above) and $r$ is the discount rate per period.

The beneficiary population may be the prevalent cohort, future incident cohorts or both, and the horizon is harder to fix. The task force lists factors such as time to patent expiry, the launch of an in-class substitute, new diagnostic or screening tests and expected price changes, and it warns that VOI results can be highly sensitive to the horizon. Its good practice recommendation is to justify the horizon chosen and to test alternatives in scenario analyses. The HEMA report argues for population-level results because decisions have their impact on populations, letting the health gain from immediate access be set against the population value of further research.

How EVPI informs HTA, pricing and research decisions

The task force treats population EVPI as a first screen. If it is less than the expected cost of research, further research is not worthwhile and the value of information process should stop. If it is greater, research is potentially worthwhile, but this is a necessary rather than a sufficient condition, because perfect information cannot be bought with a study of finite size.

The HEMA report sets out several further uses in HTA.

• Evidence before submission. EVPI gives the maximum value of the research a manufacturer could have done before its HTA submission, an upper bound on the consequences of its research and development choices.
• Access against research. Where the population health impact of uncertainty is high compared with the impact of immediate access, decision makers may consider linking funding to further research. The report then asks whether research can reduce the key uncertainties, how long it would take, whether it would cost less than EVPI and whether access would incur costs that could not be recovered.
• Price. When the doubt is about the size of the benefit rather than whether there is one, a lower price raises the expected INHB of access and lowers both the cost of uncertainty and the value of feasible research. The report notes that the underlying uncertainty in the evidence remains.
• No demand for evidence. A low value of information despite high uncertainty is, in the report's words, a reason not to demand more evidence.

PMG36 does not ask for EVPI in its economic evaluation chapter, but before recommending a technology for use while more research is done (section 6.4.14) the committee weighs the potential value of additional evidence and the net benefits to current patients against those to current and future patients. The HEMA report notes that in England value of information has an established, if intermittent, role in research prioritisation and only in research guidance, and that other settings have no established route for a VOI result to reach coverage decisions.

EVPI next to error probability, EVPPI, EVSI and ENBS

EVPI and the probability of cost-effectiveness answer different questions. The probability shows how often an option is the best one, so for the chosen option its complement is how often that choice is wrong, while EVPI adds how much each error costs, so a small chance of a costly error can give a larger EVPI than a large chance of a trivial one. The HEMA report recommends that reports combine outputs on the probability of an incorrect decision and its consequences, presenting cost-effectiveness acceptability curves alongside EVPI.

EVPI is the widest member of the value of information family. EVPPI values perfect information on a chosen group of parameters and so shows where research should focus, EVSI values a specific study design that reduces rather than removes uncertainty, and the expected net benefit of sampling subtracts the study's cost from population EVSI. The methods report notes that the EVPPI for a group of parameters is an upper limit on the EVSI of any study informative about them.

Limitations and common misreadings

EVPI values only the uncertainty that the model represents. The methods report warns that leaving structural uncertainty unquantified is equivalent to assuming the model is perfect, and that a parameter given no distribution is treated as known with certainty. The HEMA report adds that the distributions and correlations in a probabilistic analysis, and so any VOI metric computed from them, rest on the same evidence and modelling assumptions, so EVPI is only as credible as the parameter uncertainty analysis behind it.

Common misreadings include the following.

• EVPI as the value of a trial. No study yields perfect information, so EVPI is a ceiling, and the value of an actual design needs EVSI.
• Large input uncertainty as high EVPI. Uncertainty that never changes the preferred option gives an EVPI of zero.
• A fixed population figure. Population EVPI depends heavily on the assumed beneficiary population and horizon, which should be justified and varied.
• A complete answer. The HEMA report names two judgements the methods do not capture, whether the research would in fact be undertaken and reported, and whether the system has the capacity to reassess the decision when it reports.

EVPI also assumes a risk-neutral decision maker. The HEMA report reviews the long debate over whether HTA bodies should be risk averse, a stance that would lower the acceptable price of an uncertain technology.

Sources

• Fenwick E, Steuten L, Knies S, Ghabri S, Basu A, Murray JF, Koffijberg HE, Strong M, Sanders Schmidler GD, Rothery C. Value of information analysis for research decisions: an introduction. Report 1 of the ISPOR Value of Information Analysis Emerging Good Practices Task Force. Value in Health. 2020;23(2):139-150. https://doi.org/10.1016/j.jval.2020.01.001
• National Institute for Health and Care Excellence. NICE technology appraisal and highly specialised technologies guidance: the manual (PMG36). London: NICE; 2022, last updated 13 October 2026. Sections 4.5.1, 4.10.8 and 6.4.14. https://www.nice.org.uk/process/pmg36
• Rothery C, Strong M, Koffijberg HE, Basu A, Ghabri S, Knies S, Murray JF, Sanders Schmidler GD, Steuten L, Fenwick E. Value of information analytical methods: Report 2 of the ISPOR Value of Information Analysis Emerging Good Practices Task Force. Value in Health. 2020;23(3):277-286. https://doi.org/10.1016/j.jval.2020.01.004
• Steuten L, Thavorn K, Cipriano LE, Guertin JR, McQueen B, Schmidler G, Sullivan S, Wailoo A, Sculpher M. Methods to analyze and present uncertainty in economic evaluation to support HTA decisions. HEMA Report 2. Health Economics Methods Advisory; 2026. https://hemamethods.org/wp-content/uploads/2026/10/HEMA-Report-2_For-Publication_100726.pdf