Before commissioning another survey, ask one question: what could the survey reveal that would make us choose differently?
The answer may be important. A modest piece of evidence can prevent an expensive mistake. But the answer may also be nothing: the same action would remain preferable across every result the proposed study could reasonably produce. The study might still be interesting. Its immediate value for that particular decision would be limited.
Value of information is the expected improvement in a specified decision that becomes possible when additional information is available. It depends on the available actions, uncertain states, consequences, current knowledge and the information a proposed investigation could actually deliver. It is not a score for how impressive a dataset looks or how many pages a report contains.
This guide develops the calculation from an original fictional library-planning example. All probabilities and payoff points below are invented for explanation. They are not eduKate operating data, service forecasts, prices or financial advice. The example uses an explicitly simplified decision model so that every step can be checked.
Reading route: Start with the decision before the research, calculate perfect information and imperfect information, then examine a useful-looking test with zero decision value, complementary questions and the stopping rule.
The decision comes before the information budget
Suppose a fictional library must choose between two service designs. The reliable design, R, offers a modest but stable benefit. The flexible design, F, performs very well if demand is high but poorly if demand is low. The library must choose before the demand state is known.
To make the arithmetic simple, represent consequences in agreed utility points. These points are a constructed scoring scale, not money. Assume the decision-maker is willing to compare expected values on that scale and that the options already satisfy relevant non-negotiable constraints.
| Action | High-demand state | Low-demand state |
|---|---|---|
| Reliable design R | 60 points | 60 points |
| Flexible design F | 100 points | 0 points |
Assume the current probability of each demand state is one half. R has an expected value of 60. F has an expected value of 0.5 × 100 + 0.5 × 0 = 50. Without more information, the model therefore recommends R.
This is the baseline against which additional evidence must be evaluated. The correct comparison is not between ignorance and omniscience. It is between the best action available using current information and the best contingent choices made possible by new information.
The ISPOR introductory report on value of information describes this connection between uncertain decisions and the potential benefit of further evidence. Its original application is healthcare resource allocation; the library example here is a separate mathematical illustration of the general decision logic.
Perfect information gives an upper benchmark, not a promise
Now imagine an oracle tells the library the true demand state before it chooses. In the high-demand state, it selects F and obtains 100 points. In the low-demand state, it selects R and obtains 60. Before hearing the oracle, the expected value of this informed policy is 0.5 × 100 + 0.5 × 60 = 80.
The expected value of perfect information, EVPI, is therefore 80 − 60 = 20 points. The first 60 points were already achievable without the oracle. Only the additional 20 are attributable to the possibility of choosing differently after learning the state.
Do not compare 80 with the uninformed value of F, which is 50. That would overstate the value of information by comparing informed choices with an inferior baseline action. The baseline must use the best admissible decision under current knowledge.
The oracle also needs a precise meaning. Here perfect information means learning the high-or-low state before this decision. In another model, it might mean learning an uncertain parameter while future individual outcomes remain random. Those are different information questions and can produce different values.
Real research rarely supplies perfect information. EVPI is useful because it establishes an upper benchmark under the stated decision model. A fallible study about the same uncertainty cannot outperform a correctly used revelation of the whole relevant state when actions and consequences are otherwise unchanged.
Why the order of averaging and choosing matters
Without new information, choose one action after averaging over the uncertainty. With perfect information, choose the best action separately in each state, then average the value of those informed choices.
Current decision value = best, over actions, of the expected payoff.
Perfect-information value before cost = expected payoff of the best action in each revealed state, minus current decision value.
That exchange in the order of choosing and averaging is the heart of the calculation. Information is valuable because it can make the action depend on what is learned.
The same reasoning explains why uncertainty alone is insufficient. If R had the highest payoff in both demand states, learning which state would occur would not change the preferred action. A question could be uncertain and interesting while having no positive decision value in that particular payoff model.
Conversely, a small uncertainty can have considerable value when the available actions have sharply different consequences on either side of a decision boundary. The amount of uncertainty and the importance of uncertainty are not the same quantity.
An imperfect survey can still improve the decision
Replace the oracle with a proposed survey that returns either a positive or negative signal. Assume a positive signal occurs with probability 0.8 when demand is high, and a negative signal occurs with probability 0.8 when demand is low. These are conditional performance assumptions for the fictional survey, not empirical claims about surveys generally.
With the initial fifty-fifty demand probabilities, a positive signal has probability 0.5. The probability of high demand given a positive signal is (0.8 × 0.5) ÷ 0.5 = 0.8. Following a negative signal, the probability of high demand is 0.2.
After a positive signal, F has expected value 80 and R still has value 60, so choose F. After a negative signal, F has expected value 20, so choose R. Before the survey is run, the expected value of this signal-dependent policy is 0.5 × 80 + 0.5 × 60 = 70.
The expected value of sample information, EVSI, is 70 − 60 = 10 points. The study is imperfect but useful: it sometimes supplies enough evidence to change the decision appropriately.
Assume its full cost is four points on the same additive utility scale. The expected net gain is then 10 − 4 = 6 points. This subtraction is valid because the example explicitly expresses both benefit and cost in compatible units. A real analysis cannot subtract dollars from assessment points or staff hours from a probability without a justified decision framework.
The distinction between EVPI, partial perfect information, EVSI and expected net benefit of sampling is set out in ISPOR’s analytical-methods report. Our numbers demonstrate the concepts; they do not reproduce an empirical case from that report.
The survey must be evaluated before its result is known
It is easy to overvalue research after a striking result. Once a positive signal has arrived, the expected value of F is 80. But the survey’s ex ante value was not 80 − 60 = 20. Before commissioning it, the library had to account for both possible signals and the actions each would prompt.
Value of information is therefore a calculation about a research policy: collect a particular kind of evidence, update the relevant beliefs, choose according to a stated rule and accept the resulting consequences. It includes results that would be disappointing, inconclusive or supportive of the current plan.
In a more complex study, there may be many possible datasets rather than two signals. The analysis must represent the data the proposed design could generate. Heath and colleagues’ tutorial on simulating study data for EVSI explains why this data-generation step is a substantive part of the calculation, including design issues such as missingness and correlation.
A precise calculation based on an unrealistic study-performance model remains unrealistic. The expected value belongs to the study that can actually be conducted, not an idealised study borrowed from a proposal.
Positive expected value does not guarantee a good realised outcome
In the fictional example, a positive signal can occur when demand is low. The policy then selects F and receives zero before the survey cost. That particular outcome is worse than the sixty points R would have delivered.
This does not contradict an EVSI of ten. Expected value averages over possible outcomes using the stipulated probabilities. It is a property of the decision policy before the uncertainty resolves, not a guarantee that every realised use of information improves the result.
Nor does one lucky outcome prove that commissioning a study was a good decision. A poor decision policy can occasionally succeed. Evaluation should distinguish the quality of the information model, the decision rule and the realised outcome.
For an institution, preserve what was believed and planned before the result arrived. Otherwise an after-the-fact story can make every success look inevitable and every failure look unforeseeable. That destroys the ability to learn whether research was commissioned for the right reasons.
The same test can have zero value for a different prior decision
Keep the payoff table and survey performance unchanged, but suppose the current probability of high demand is 0.9. R remains worth 60. F is now worth 90, so the uninformed decision is F.
A positive signal has probability 0.8 × 0.9 + 0.2 × 0.1 = 0.74. The posterior probability of high demand is 0.72 ÷ 0.74, approximately 0.973. After a negative signal, it is 0.18 ÷ 0.26, approximately 0.692.
F is preferred whenever the probability of high demand exceeds 0.6, because its expected value is 100 times that probability. Both survey results leave the probability above 0.6. The library therefore chooses F after either result.
The survey changes beliefs but not action. Its gross EVSI is zero in this model. After a four-point cost, its net value for the immediate decision is negative four.
Perfect information still has value: choose F when demand is high and R when it is low, for an expected 0.9 × 100 + 0.1 × 60 = 96. EVPI is therefore six points. The distinction is important: better information could matter, but this particular study is not strong enough to change the decision.
This constructed example also separates information value from general accuracy. The survey has exactly the same conditional performance in both prior scenarios. Its decision value changes because the current decision and the threshold have changed.
Gross information value and net research value are different
Under a coherent decision model, with unchanged feasible actions and the ability to ignore a signal, additional information cannot reduce the optimised expected payoff before its costs. The decision-maker can always follow the old policy. This is a mathematical option argument, not a claim that every real information system is harmless.
Costs can make the net value negative. So can delay, distraction, implementation burden or disclosure risks that the simple payoff table omitted. Information may also change who can act or what options remain available. Those effects belong in the decision problem rather than being dismissed by quoting a non-negativity result whose assumptions no longer hold.
Suppose the fictional survey takes so long that the library loses access to F. It no longer has the same decision after the information arrives. The earlier EVSI calculation is not applicable because it assumed both designs remained feasible.
The useful statement is conditional: information expands the set of contingent decisions under the stated model; whether acquiring it is worthwhile depends on the full consequences of acquisition and use.
Perfect information helps rule out overpriced investigations
In the initial fifty-fifty example, perfect information is worth twenty points. A proposed study costing twenty-five points cannot be justified solely by improving this one decision under this model, even before asking how accurate the study is. It costs more than the maximum possible informational benefit.
This does not mean the study has no other purpose. It might support other decisions, create a reusable dataset or satisfy a legitimate accountability requirement. Those benefits should be named and evaluated, not quietly added after a weak calculation.
EVPI can therefore be a useful screening calculation. It identifies when a detailed EVSI exercise may be unnecessary. Where even perfect information has little decision value, an elaborate research design needs a different justification.
Be careful about the unit of analysis. Twenty points for one decision is not automatically twenty points for an entire institution or twenty points per person. Specify whether the payoff concerns a single case, a programme, a cohort or a period of operations before multiplying anything.
Partial perfect information asks which uncertainty is worth resolving
A realistic decision may depend on demand, implementation cost, user response and the reliability of an external partner. Learning everything perfectly is neither possible nor necessary. The expected value of partial perfect information asks how useful it would be to resolve a selected subset while leaving the remaining uncertainty in place.
This can help organise research priorities. A large uncertainty in a parameter does not guarantee high decision value. A parameter may barely affect which action is best. Another may be less uncertain but sit near a decision threshold.
Jackson and colleagues’ review of value-of-information methods connects these calculations to evidence synthesis and research prioritisation. The decision perspective differs from simply ranking inputs by the amount of output variance they explain.
Partial values should not be added casually. Learning about one factor can change how useful it is to learn about another. A research portfolio is a joint decision problem, not always a sum of independent topic scores.
Two individually unhelpful answers can be valuable together
A small mathematical example makes complementarity visible. Two independent fair binary variables, X and Z, each take values zero or one. You must predict whether they are equal. A correct prediction earns one point; an incorrect prediction earns zero.
Without information, equality and inequality are equally likely, so the best expected payoff is one half. Learning X alone does not help: Z is still equally likely to match or differ. Learning Z alone does not help either. Each isolated piece of perfect information has zero value for this decision.
Learning both reveals the answer with certainty and raises the payoff to one. Their joint information value is one half, even though the sum of their separate values is zero.
This is an original illustrative calculation, not a model of a real library. It shows why research questions can be complementary. In an applied problem, knowing demand may be useful only if delivery capacity is also understood. Knowing an intervention’s effect may be useful only if its local implementation cost is known.
It also warns against assuming a universal rule of diminishing information value. Additional evidence can become less useful as uncertainty shrinks, but complementary information can create exceptions. The ordering and combination of investigations belong in the analysis.
Sample size is a design choice, not a synonym for value
A larger study can yield more precise information, but its additional decision value must be compared with its additional cost and delay. If a modest study already makes the relevant choice clear under the model, much greater precision may add little to that decision.
Conversely, a small convenience sample may be cheap but uninformative about the target. Its nominal size is not the relevant property. The expected data distribution depends on whom the study reaches, what it measures, nonresponse, measurement error and the analysis planned.
The EVSI study-simulation tutorial provides a methodological route for representing proposed data collection rather than equating a larger sample with a better investment. Our two-signal example is deliberately simpler, but it makes the same design requirement visible: specify what evidence the study can produce.
For a fictional library survey, fifty carefully selected interviews about the decisive access barrier might be more relevant than thousands of responses to a general satisfaction question. That is a proposal to investigate relevance, not a universal claim that small qualitative studies outperform large quantitative ones.
Evidence must concern the decision’s target
Research can be precise about the wrong population. A survey of frequent library visitors may not answer why non-visitors stay away. A test of a programme under intensive support may not estimate its performance under routine staffing. A demand estimate from last year may be poorly matched to a changed service environment.
These are not small caveats to add after valuing the research. They change what information the study provides about the uncertainty in the decision model. A study that answers a related question may still be useful, but the inferential bridge needs to be represented.
The general principles are developed in Comparative Systems Research and Causal Inference. For information valuation, the practical test is whether the evidence updates a quantity that can legitimately guide the target action.
A result does not acquire decision value simply because it is the best available source. Sometimes the best available source leaves the decisive uncertainty largely unresolved.
Time changes both the research cost and the benefit
Information arriving after an irreversible choice cannot improve that past choice. It may still improve future decisions, but those are a different benefit stream. An investigation’s completion date belongs alongside its expected precision.
In a constructed service decision, waiting for a long survey could postpone a beneficial change. A shorter investigation might resolve less uncertainty but arrive while both options remain available. The correct comparison includes what happens during the wait and how much the findings will still matter when they arrive.
A durable research question can support many later decisions. A transient question may lose value quickly. The Cambridge-hosted guide to information-value methods for research prioritisation illustrates the broader role of modelling in directing evidence collection. In the library setting proposed here, this motivates an explicit distinction between reusable knowledge and information tied to one expiring choice.
Do not multiply a per-decision gain by an unlimited future population. Account for the period during which the evidence remains relevant, the number of decisions that can actually use it and possible changes in the underlying system.
Implementation is part of the information pathway
The idealised EVSI calculation assumes that the decision-maker receives the result, understands it and follows the chosen policy. In a real institution, information can arrive in the wrong format, miss the decision meeting or reach someone without authority to act.
Before commissioning a study, specify who owns the decision, when it will be made and how different findings would change it. A report without a receiver may still contribute to public knowledge, but its claimed immediate operational benefit needs scrutiny.
There can also be a gap between choosing an option and delivering it. A recommendation to adopt a specialist service is not feasible if no qualified staff are available. The payoff table should represent feasible implementation, not an imagined institution with unlimited capacity.
This is why a useful research brief describes an action pathway. Evidence does not improve outcomes merely by existing. Its value for a decision depends on what the decision-maker can do differently because of it.
Ethics and public value cannot be hidden inside a convenient score
The fictional example begins after checking that both actions are admissible. In practice, research and decisions may involve privacy, consent, fairness, safety or obligations that should not be traded away simply because an aggregate score increases.
A high estimated information value is not permission to collect sensitive data without a legitimate basis. Nor does a low immediate operational value imply that preserving culturally important records is pointless. A library may have stewardship, access and educational purposes that extend beyond one decision model.
Represent the relevant objectives honestly. Where they can be placed on a defensible common utility scale, explain that scale. Where they cannot, retain a multi-criteria or constrained decision rather than inventing numerical comparability. Excluding an unacceptable option before optimisation is often clearer than assigning it a token penalty.
Value of information is an instrument for a defined decision. It is not a universal theory of everything worth knowing.
The model can be more uncertain than the calculation suggests
In our example, the payoffs, prior probabilities and survey performance were stipulated. In an actual analysis, those inputs would themselves need evidence or a transparent elicitation process. A result printed to two decimal places can conceal substantial uncertainty in the quantities being multiplied.
Test whether the research recommendation survives plausible changes. Does the study remain worthwhile if its performance is weaker, its cost higher or its delivery later? Does changing the current probability move the decision far from the threshold, as in the ninety-per-cent example?
Check numerical implementation separately from substantive assumptions. With a simple decision table, enumeration is preferable to an opaque simulation. More complex models may require simulation, but convergence does not demonstrate that the payoff model or prior beliefs are justified.
A useful analysis can end with a range or conditional recommendation. For the fifty-fifty example, the study has positive net value if its full cost stays below ten points, assuming its stipulated performance and the payoff model remain unchanged. At the assumed four-point cost, the net gain is six points. Keeping the gross benefit, cost threshold and net gain distinct prevents a correct calculation from becoming a misleading purchasing rule.
Existing information should be checked before buying new information
The baseline should include what the institution already knows or can reasonably retrieve. A study appears artificially valuable if the comparison assumes ignorance while relevant evidence sits in an accessible archive.
For a library, begin with a holdings check: existing reports, administrative records, prior studies, relevant external sources and documented limitations. Some apparent data gaps are discovery gaps. The necessary information exists, but people cannot find or interpret it.
Other gaps are genuine. A record may concern the wrong population, a past service version or an outcome no longer relevant. A retrieval process should therefore assess applicability, not merely return a document with matching words.
The lowest-cost useful action might be a better cross-reference, a targeted reanalysis or a small clarification from the original source. New collection should compete with these alternatives rather than receive automatic priority because producing something new is more visible.
A stopping rule for research that could otherwise continue forever
A practical stopping rule is to stop acquiring information for the immediate decision when the expected marginal improvement no longer justifies the marginal cost, delay and burden, subject to required duties and constraints. This does not mean stop learning forever. It means stop treating every unresolved curiosity as necessary before this particular action.
Ask what result would change the action, how likely the proposed study is to produce such a result, whether the resulting change would improve expected consequences and whether the information would arrive in time. Where these questions cannot be quantified credibly, they still structure an honest qualitative decision.
Do not continue solely because money has already been spent on earlier research. Prior expenditure may affect current resources, but it is not itself a benefit of collecting one more dataset. Compare the next feasible actions from the current state.
There may also be a useful intermediate action: proceed with a reversible design while collecting targeted information, or preserve an option until a decisive observation arrives. The choice is not always between complete certainty and reckless commitment.
A research commission that names its return
A strong commission can fit its essential logic into a short brief. State the decision and deadline; the feasible actions; the uncertainty that could change the choice; the current evidence; the study’s possible findings; the action planned after each important finding; and the full cost of obtaining and using the result.
For the fictional fifty-fifty case, the brief would say: choose between R and F; the current choice is R; a sufficiently positive signal would support F; the proposed survey has stipulated conditional accuracy of 0.8; its expected gross decision benefit is ten points; its full assumed cost is four; and the net expected gain is six. Every number has a defined role.
For the ninety-per-cent case, the brief should reject the same study for the immediate decision: both signals leave F preferred, so the gross EVSI is zero. A more discriminating study or a different research purpose would need a separate assessment.
That contrast is the central lesson. A good research question is not merely answerable. It is connected to a use, and the strength of that connection should be visible before resources are committed.
What a growing library should remember
Accumulation and learning are not identical. A library can add thousands of pages while leaving its readers’ decisive uncertainties untouched. It can also add one well-chosen explanation that prevents a recurring error across many fields.
Value-of-information thinking offers one way to prioritise that work. Ask what readers are trying to decide, which uncertainty blocks them, which existing owner already addresses it and what new evidence or explanation would materially improve their next step.
Keep other purposes visible too. Libraries preserve memory, enable discovery and support questions not yet imagined. Those values should not be erased simply because a single short-term decision table cannot capture them.
The right principle is not “collect only what pays immediately”. It is “know why this acquisition matters, what kind of return it can produce, and what assumptions support that expectation”. More information becomes better knowledge when its relation to evidence, judgement and action is understood.
Sources, scope and further reading
Source records and accessible methodological descriptions were checked on 5 September 2026. This article is an explanatory synthesis with original fictional calculations, not a systematic review, financial recommendation or evaluation of an actual service. The sources below primarily develop health-research applications; their inclusion supports the decision methods, not a claim that those applications can be copied unchanged into education.
- Fenwick and colleagues, Value of information analysis for research decisions—an introduction: Report 1, Value in Health, 2020;23(2):139–150.
- Rothery and colleagues, Value of Information Analytical Methods: Report 2, Value in Health, 2020;23(3):277–286.
- Heath and colleagues, Simulating Study Data to Support Expected Value of Sample Information Calculations: A Tutorial, Medical Decision Making, 2022.
- Jackson and colleagues, Value of Information Analysis in Models to Inform Health Policy, Annual Review of Statistics and Its Application, 2022;9:95–118.
- Jackson and colleagues, A guide to value of information methods for prioritising research in health impact modelling, published article held in the University of Cambridge repository.
Continue through the Library: Read Statistical Inference and Uncertainty for reasoning from evidence, Models and Simulations for model credibility, Data Economics and Valuation for the broader value of organisational data, and Strategic Decision-Making for choices under constraints. The Research Collections Directory provides the return route.