Expected Value of Perfect Information (EVPI) in Health-Economic Decision Analysis

What EVPI Measures

Every health-economic decision analysis is built on parameters nobody knows exactly: the probability of a treatment-related adverse event, the utility decrement of a health state, the cost of managing a complication, the transition rate between disease stages in a Markov model. A cost-effectiveness analysis has to make a decision — adopt the new technology or not — using the best current estimate of each parameter, even though that estimate is uncertain. Expected Value of Perfect Information (EVPI) asks a specific, decision-relevant question about that uncertainty: if every parameter in the model could be known with certainty, right now, how much better would the resulting decision be, on average, than the decision made under current uncertainty?

That “how much better” is expressed in the same currency as the decision itself — typically net monetary benefit, using the analysis’s own cost-effectiveness threshold to convert QALY gains and costs into a single monetary value. EVPI is therefore not a statistical property of the model — it is not a p-value, a confidence interval, or a measure of how “spread out” the parameter distributions are. It is an economic quantity: the maximum amount a decision-maker should, in principle, be willing to pay to eliminate all uncertainty before deciding. If EVPI is zero, no amount of further research into any parameter could change which option is chosen on average, because the current decision is already correct in every plausible state of the world the model considers. If EVPI is large relative to the cost of research, that is the headline signal that funding further study is potentially worthwhile — before a single trial design has been proposed.

This reframes “we need more data” from a vague methodological instinct into a quantified, decision-relevant threshold, which is precisely why EVPI has become a standard companion analysis to a cost-effectiveness evaluation submitted to a reimbursement body, and why NICE’s technology appraisal process and comparable health technology assessment (HTA) agencies increasingly expect it, or something like it, alongside the base-case ICER.

How EVPI Is Calculated From a Probabilistic Sensitivity Analysis

EVPI is computed directly from the probabilistic sensitivity analysis (PSA) that should already exist for a properly conducted cost-utility analysis. In a PSA, every uncertain parameter is assigned a distribution reflecting the analyst’s current uncertainty about its true value (a beta distribution for a probability, a gamma distribution for a cost, a normal distribution for a log-relative-risk, and so on), and the decision model is run many thousands of times, each time drawing a random value for every parameter from its distribution — the same Monte Carlo simulation logic used more broadly in Monte Carlo methods. Each run produces a net monetary benefit for every option under evaluation at that particular draw of parameter values.

From those simulation results, EVPI follows from two numbers:

• The expected value under current information (EV under uncertainty). Average the net monetary benefit of the single best option (chosen once, in advance, based on expected values across all parameters) across every simulation run.
• The expected value with perfect information (EV with PI). For each individual simulation run, take the net monetary benefit of whichever option is best in that specific run — i.e., assume the decision-maker could see that run’s true parameter values and choose optimally for them — then average that best-in-each-run value across all runs.

EVPI is the difference between the two: EV with PI minus EV under uncertainty. Because “choose the best option per run” can never do worse than “choose one option for all runs,” EVPI is mathematically guaranteed to be non-negative. It is exactly zero only when the same option is optimal in every single simulation run — that is, when no combination of parameter values the model considers plausible would ever change the decision. In practice, most PSAs show at least some sensitivity of the optimal choice to parameter uncertainty, which is what gives EVPI a non-trivial value per decision.

This is a per-patient (or per-decision) quantity, expressed in the same units as the model’s net benefit — usually a monetary figure. On its own it answers “how much is uncertainty costing us, per patient treated under this decision,” which is informative but not yet the number that determines whether commissioning a trial makes financial sense across a health system.

From EVPI to Population-Level EVPI

A per-patient EVPI of even a modest amount becomes a very different figure once it is scaled to everyone the decision will actually affect. Population-level EVPI multiplies the per-patient value by the number of patients expected to be affected by the decision over its decision-relevant time horizon — the period over which the current technology or guidance is expected to remain in force before it is next reviewed, superseded, or the underlying uncertainty is otherwise resolved by other means. Both components require real judgment, not just a plug-in figure:

• Population size is typically drawn from incidence/prevalence data for the eligible population, adjusted for realistic uptake rather than the full theoretical eligible pool.
• Time horizon is bounded by how long the decision is likely to stand before new evidence, a scheduled guidance review, patent expiry, or a competing technology would render the current uncertainty moot regardless of whether new research is commissioned. A common simplifying assumption applies a discount rate to future years of the horizon, consistent with how the underlying cost-effectiveness model discounts future costs and QALYs.

Population EVPI is what actually answers the research-prioritisation question, because it is directly comparable to the cost of running a trial or study. A per-patient EVPI that looks trivial can represent a very large population figure once multiplied across, say, a five-year horizon and an incidence measured in the tens of thousands annually — which is exactly the point: individual-patient uncertainty that looks unremarkable in isolation can still justify a substantial, system-level research investment once its true reach is accounted for.

Using Population EVPI to Judge Whether Research Is Worth Funding

The decision rule that follows from population EVPI is straightforward in principle: further research into the model’s parameters is potentially worth funding only if population EVPI exceeds the expected cost of the research needed to (approximately) resolve the relevant uncertainty. If population EVPI is smaller than any plausible research budget, no feasible study could pay for itself in decision-value terms, and the honest conclusion is to proceed with the technology decision as currently informed rather than commission further study. If population EVPI comfortably exceeds realistic research costs, that is the quantitative case for funding a trial, registry, or further analysis — independent of whether a specific study has even been designed yet.

Two caveats matter in practice. First, EVPI represents the value of resolving all parameter uncertainty simultaneously, which no real study ever achieves — it is a ceiling, not an estimate of what a specific proposed trial would deliver. That is the gap EVPPI closes (below), and ultimately what Expected Value of Sample Information (EVSI) addresses by attaching a value to a specific study’s finite sample size rather than to perfect knowledge. Second, EVPI compares against research cost, not against whether a positive finding is likely — a trial that confirms the current best estimate still has scientific value even if it doesn’t change the reimbursement decision, and EVPI speaks only to the decision-analytic case for further study, not that broader value.

EVPPI: The More Actionable Refinement

A large population EVPI tells a funder that uncertainty is expensive, but not which uncertainty. A decision model built around a new intervention typically has dozens of uncertain parameters — efficacy, adverse-event rates, utility values, resource-use costs, discontinuation rates — and resolving all of them with a single trial is rarely feasible or necessary. Expected Value of Partial Perfect Information (EVPPI) answers the more useful question: what would it be worth to learn the true value of one parameter, or one subset of parameters, while every other parameter remains as uncertain as it currently is?

Mechanically, EVPPI is calculated the same way as EVPI, but “perfect information” is applied selectively: only the parameter subset of interest is fixed at its “known” value in each inner evaluation, while every other parameter still varies according to its PSA distribution. Full nested Monte Carlo simulation for EVPPI is computationally expensive, so most modern applications instead use regression-based or Gaussian-process approximation methods (popularised through the NICE Decision Support Unit’s technical support documents) to estimate EVPPI across many parameter subsets from a single PSA run.

The practical output is a ranked list: which single parameter, or which small group of parameters (e.g., “the relative treatment effect on progression-free survival” or “the utility decrement associated with the primary adverse event”), carries the largest share of total EVPI. That ranking is what turns an abstract “more research is justified” conclusion into an actionable research-design brief — it tells a funder or investigator which specific endpoints and parameters a proposed trial actually needs to measure precisely to capture most of the available decision value, rather than commissioning a broad, unfocused study that re-estimates parameters the model was never particularly sensitive to in the first place.

Data and Reporting Requirements

EVPI and EVPPI analysis is only as credible as the PSA it is built on: distributions for every parameter should be justified from the data source they are drawn from (not assigned arbitrarily), correlation between parameters should be preserved where a common data source or causal link implies it, and the number of Monte Carlo iterations should be large enough that EVPI estimates are stable rather than an artefact of simulation noise. When submitting an EVPI/EVPPI analysis alongside a cost-effectiveness model — whether to a funder deciding on a trial or to an HTA body under frameworks like the CHEERS 2022 reporting checklist — report the per-patient and population-level figures separately, and state the population size, time-horizon, and valuation-threshold assumptions explicitly, since population EVPI is directly sensitive to all three.

Common Pitfalls

• Treating EVPI as a standalone endpoint rather than a threshold comparison. A population EVPI of any given absolute size is meaningless without comparing it to a realistic research-cost estimate; report both sides of the comparison, not EVPI alone.
• Using an unrealistic or unstated time horizon. Extending the horizon indefinitely inflates population EVPI arbitrarily; the horizon should be defensible against how long the specific decision is actually expected to stand.
• Skipping EVPPI and going straight from a large EVPI to a broad, unfocused trial design. EVPI justifies research in general; EVPPI is what should shape what that research actually measures.
• Confusing EVPI with EVSI. EVPI is the ceiling value of eliminating all uncertainty; Expected Value of Sample Information (EVSI) is the realistic value of one specific, finite-sample study design, which is always less than or equal to EVPI for the same parameters.
• Running too few PSA iterations. Because EVPI is a difference between two averages taken over the same simulation set, it is more sensitive to simulation noise than the base-case ICER is — a PSA that looks adequate for reporting a cost-effectiveness plane can still be too small for a stable EVPI/EVPPI estimate.

Frequently Asked Questions

What is the difference between EVPI and EVPPI?

EVPI values eliminating uncertainty in every parameter in the model at once; EVPPI values eliminating uncertainty in one parameter, or a specific subset of parameters, while every other parameter stays as uncertain as it currently is. EVPPI is what tells a funder which parameters are actually worth researching.

How is EVPI actually calculated?

From the same probabilistic sensitivity analysis used to report a cost-effectiveness result: EVPI is the average, across all PSA simulation runs, of the best net-benefit outcome achievable in each individual run, minus the net benefit of committing in advance to whichever single option has the highest expected value across all runs.

What is population-level EVPI, and why does per-patient EVPI alone not answer the funding question?

Population EVPI multiplies the per-patient figure by the number of patients affected over the decision’s relevant time horizon. A trial-funding decision has to compare research cost against the total value uncertainty is costing the health system, not the value per individual patient, so the population-scaled figure is the one that belongs in a funding case.

Does a positive EVPI automatically mean a trial should be funded?

No. It means further research is potentially worth its cost. The comparison that actually matters is population EVPI against the realistic cost of a study capable of resolving the relevant uncertainty — and because EVPI values perfect information, the more realistic comparison for a specific proposed trial design is against EVSI (expected value of sample information), not EVPI directly.

What software is used to run EVPI/EVPPI analysis?

Most PSA-based decision models are built in R (packages such as BCEA and voi are widely used specifically for EVPI/EVPPI calculation) or in a spreadsheet/Excel model with a Monte Carlo add-in; regression-based EVPPI approximation methods are typically implemented in R given the volume of simulation output involved.

How does EVPI relate to the cost-effectiveness threshold and the ICER?

The threshold used to judge whether an ICER represents good value is the same willingness-to-pay figure used to convert QALY gains into monetary net benefit inside the EVPI calculation. A different threshold assumption changes which option is “optimal” in each PSA run and therefore changes the EVPI estimate itself, not just the pass/fail read on the base-case ICER.