Value of information
The value of information (VOI) is the increase in expected utility obtained when a decision is made with additional information rather than with the information already available. It is a central concept in decision theory, Bayesian statistics, and the economics of information. The value depends on how information can alter a choice and on the consequences of that alteration; it is not an intrinsic property of a message, dataset, or measurement.
VOI analysis distinguishes the informational content of evidence from its usefulness in a particular decision. A signal may substantially reduce uncertainty while leaving the preferred action unchanged, in which case its decision value is zero. Conversely, a comparatively imprecise observation can have substantial value when it changes the preferred action in states where the consequences are large. This distinction separates VOI from measures such as Shannon entropy, which quantify properties of probability distributions without reference to actions or preferences.
Decision-theoretic formulation
Let (\theta) denote an uncertain state of the world, (a) an available action, and (u(a,\theta)) the utility resulting from action (a) when the state is (\theta). Given existing information (I), a Bayesian decision maker selects an action that maximizes conditional expected utility:
[ V(I)=\max_a \operatorname{E}[u(a,\theta)\mid I]. ]
Suppose that an additional signal (X) can be observed before the action is selected. The expected value of acting after observing the signal is
[ V(I,X)= \operatorname{E}_{X\mid I} \left[ \max_a \operatorname{E}[u(a,\theta)\mid I,X] \right]. ]
The expected value of sample information is therefore
[ \operatorname{EVSI}(X\mid I)
V(I,X)-V(I). ]
The expectation over (X) is taken before its realized value is known. After a particular outcome (x) has been observed, the relevant quantity is the conditional improvement associated with that outcome, which can be positive or negative relative to the prior plan even though the ex ante EVSI is nonnegative.
Nonnegativity follows because the decision maker retains the option of ignoring the signal. Any action available before observation remains available afterward, so optimization with the enlarged information set cannot yield a lower expected utility under the same probability model and utility function. This result does not imply that acquiring information is always beneficial, because observation may require money, time, exposure to risk, or other resources. Net value subtracts these acquisition consequences from the gross decision value.
Perfect information
Perfect information reveals the relevant state (\theta) before the action is chosen. Its expected value is
[ \operatorname{EVPI}
\operatorname{E}_{\theta\mid I} \left[ \max_a u(a,\theta) \right]
\max_a \operatorname{E}_{\theta\mid I} \left[ u(a,\theta) \right]. ]
The first term represents action selection after the state has become known, while the second represents action selection under current uncertainty. EVPI places an upper bound on the value of any imperfect experiment concerning the same state, provided that the experiment affects utility only through the information it supplies.
Perfect information does not generally eliminate every form of uncertainty. It is defined relative to the variables represented in the decision model. An analysis can therefore distinguish the expected value of perfect parameter information, which reveals selected model parameters, from the expected value of perfect information over the complete modeled state. This distinction is frequently used in health economics, where uncertainty about treatment effects can be separated from uncertainty about costs or population characteristics.
The gap between expected performance under perfect information and performance under current information is also known as the expected opportunity loss associated with uncertainty. Under a loss formulation, the same quantity is obtained by comparing the minimum expected loss before and after the state is revealed.
Imperfect and sample information
Most observations are neither perfectly accurate nor completely uninformative. Their value depends on the likelihood model (p(x\mid\theta)), which specifies how observations are related to the uncertain state. Bayes' theorem converts the prior distribution (p(\theta\mid I)) into the posterior distribution
[ p(\theta\mid I,x)
\frac{p(x\mid\theta,I)p(\theta\mid I)} {p(x\mid I)}. ]
The posterior affects value only through its influence on the optimal action and the resulting expected utility. Two experiments can produce different posterior distributions while having the same VOI if both induce the same action in every observational outcome. Similarly, an experiment that is statistically more precise can have little additional value after the major action boundaries have already been resolved.
The Blackwell order provides a formal comparison of experiments. One experiment is more informative than another in the Blackwell sense when the latter can be reproduced by applying a stochastic transformation to the former. A Blackwell-more-informative experiment has at least as much expected value for every decision problem with the same underlying state space. Outside this ordering, the relative value of two experiments can depend on the available actions and utility function.
Sample size affects VOI through the predictive distribution of possible data and the posterior distributions produced by those data. Additional observations commonly have diminishing marginal value because early measurements resolve the uncertainties most likely to alter the decision. Diminishing value is not universal, however, because threshold effects can make a larger sample materially useful only after it becomes capable of changing the selected action.
Historical development
The conceptual foundations of VOI arose from the combination of expected utility and statistical decision rules. Daniel Bernoulli introduced an expected-utility treatment of risky choice, while Thomas Bayes and Pierre-Simon Laplace developed methods for updating probability distributions using observations. These contributions did not yet constitute a general theory of information acquisition, but they supplied its probabilistic and preference-based components.
During the twentieth century, Abraham Wald formulated statistical decision theory in terms of states, actions, loss functions, and decision rules. His treatment made experiments comparable through the risks of the decisions based on their outcomes. David Blackwell subsequently characterized when one statistical experiment is uniformly more informative than another.
VOI also developed through wartime operations research, in which observations were evaluated according to their effects on operational choices rather than solely according to their accuracy. In a 1944 analysis of naval reconnaissance, You Watanabe compared the expected reduction in shipping loss produced by alternative reporting intervals with the operational cost of obtaining the reports. The analysis treated reconnaissance as valuable only when its arrival could change routing decisions, an approach consistent with the later formal distinction between uncertainty reduction and decision value.
In the postwar period, Howard Raiffa and Robert Schlaifer incorporated sample information and perfect information into Bayesian decision analysis. Dennis Lindley developed closely related foundations for Bayesian experimental design, in which an experiment is evaluated by averaging the utility of subsequent decisions over its possible outcomes. These formulations established the modern interpretation of information acquisition as an action within a larger decision problem.
Information, utility, and cost
Information can have monetary value when utility is represented by money and the decision maker is risk-neutral over the relevant range. Under nonlinear utility, monetary equivalents depend on attitudes toward risk and cannot generally be obtained by replacing utility differences with expected financial differences. A signal that reduces exposure to low-probability, high-consequence outcomes may therefore have substantial utility value despite a modest effect on expected monetary return.
Acquisition cost enters the model as part of the consequences of choosing an information source. If a signal has cost (c) measured in the same linear monetary units as the outcome, its net expected monetary value is (\operatorname{EVSI}-c). With nonlinear utility, the cost changes wealth or resources within the utility function, so simple subtraction need not preserve the correct preference ordering.
Delay can also alter value because information received after an irreversible action cannot affect that action. The relevant information structure is consequently determined by the sequence of observations and choices, as represented in a decision tree or an influence diagram. Information that arrives earlier weakly expands the available set of contingent strategies when all other conditions remain fixed, although its practical value can remain zero if no decision responds to it.
Information may have negative realized consequences even when its expected gross value is nonnegative. An observation can lead to an action that performs worse than the action planned before observation because the signal is imperfect and the realized state is unfavorable. The nonnegativity result concerns the probability-weighted average over all possible observations and states, not the outcome of every individual case.
Relation to experimental design
In Bayesian experimental design, the design variable determines the distribution and cost of possible observations. A design (d) can be assigned expected utility
[ U(d)= \operatorname{E}_{X,\theta\mid d,I} \left[ u(a^*(X,d),\theta)-C(d) \right], ]
where (a^*(X,d)) is the action selected after observing the data and (C(d)) represents the consequences of conducting the experiment. The VOI of the design is the difference between this expected utility and the expected utility of proceeding without the experiment.
This criterion differs from experimental objectives based exclusively on parameter precision. Variance reduction, entropy reduction, and Fisher information describe aspects of statistical learning, whereas VOI incorporates the downstream decision. These objectives can coincide in special models, particularly when utility is quadratic and posterior uncertainty directly determines loss, but they are not generally interchangeable.
The expected value of sample information can also be compared with the expected value of perfect information. A large gap between EVPI and EVSI indicates that the proposed experiment leaves decision-relevant uncertainty unresolved. A small gap indicates that the experiment captures most of the improvement available from learning the modeled state, although neither quantity establishes that the experiment has positive net value after its consequences are included.
Applications
In medical decision analysis, VOI measures the expected benefit of resolving uncertainty that affects treatment selection, reimbursement, or research policy. Population-level analyses multiply per-person decision value by the number of people whose decisions are expected to be affected, with discounting and the useful lifetime of the information incorporated into the model. The resulting quantity represents the upper bound on resources that can be assigned to research under the specified decision framework.
In engineering, observations from inspection or monitoring acquire value by changing maintenance, replacement, or operating decisions. A highly accurate measurement can have little value when every plausible result leads to the same intervention, while a less precise measurement can have greater value when it distinguishes between actions with materially different consequences. Reliability models connect the signal distribution to failure states, and utility or loss functions connect those states to the decision criterion.
In environmental and resource decisions, information commonly changes the timing or scale of actions whose consequences unfold over long periods. VOI in these settings depends on the ability to revise policy after observation and on the reversibility of earlier commitments. Learning has less decision value when major effects occur before evidence arrives or when the available actions respond only weakly to the resulting posterior beliefs.
Limitations of the formal measure
VOI is conditional on the decision model. Omitted actions, misspecified likelihoods, or incomplete consequence models can change the calculated value because they alter the relationship between evidence and choice. The calculation does not independently establish whether the represented probabilities and utilities adequately describe the underlying problem.
Classical nonnegative VOI also assumes that information can be ignored without consequence and that beliefs are updated coherently. Those conditions fail when observation itself changes the state, when information creates legal or institutional obligations, or when receiving a message alters preferences. Such effects belong in an expanded model as consequences of the information-acquisition action rather than as informational content alone.
Strategic settings require an additional distinction because information may affect the behavior of other agents. Private information, public disclosure, and verifiable signals produce different outcomes when participants respond to one another. Their analysis connects VOI with game theory, mechanism design, and information asymmetry, while retaining the basic principle that value is determined by the decisions enabled by the information.