| title | voiage: Value-of-information analysis for decisions about further research | |||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| tags |
|
|||||||||||||||||||||
| authors |
|
|||||||||||||||||||||
| affiliations |
|
|||||||||||||||||||||
| date | 26 July 2026 | |||||||||||||||||||||
| bibliography | paper.bib | |||||||||||||||||||||
| repository | https://github.com/edithatogo/voiage |
Decision-makers weigh acting on current evidence against commissioning a study. Value-of-information (VOI) analysis estimates the expected improvement in a decision from better information. Expected value of perfect information (EVPI) measures the value of resolving all modelled uncertainty; expected value of partial perfect information (EVPPI) measures the value of resolving selected inputs; and expected value of sample information (EVSI) measures the value of a proposed study. Expected net benefit of sampling (ENBS) compares population EVSI with research cost [@rothery2020voi].
voiage calculates EVPI and EVPPI from simulated net-benefit samples, estimates
EVSI from a declared study model, and combines study value and cost for ENBS. A
versioned record preserves option names, units, uncertain inputs, and data
sources with each result. Python, R, and Julia share the EVPI calculation. In
the fixed-seed health example, EVPI was estimated at 644 value units—a generic
scale—for each person affected by later decisions during the horizon. It supports
comparisons of
further-research choices.
A research team may build an uncertainty model in one programming language, calculate VOI in another, and prepare the report in a third. Moving an analysis requires more than copying a table of simulated results. Option names, units, uncertain quantities, population, time horizon, study design, and data sources affect what an estimate means. If these details are separated from the result, two identical-looking values may describe different choices, populations, or periods.
voiage validates a versioned description of competing options, uncertain
quantities, units, and input references, then links each calculation to that
description. It is designed for researchers and analysts using simulation
models to decide whether further evidence is worth collecting. The same
structure could describe a healthcare programme, policy intervention, business
investment, or marketing strategy when the analyst defines the alternatives,
uncertain outcomes, affected population, and value scale. The worked example
and current methodological validation are health-economic. voiage therefore
contributes a shared EVPI calculation and a portable record of decision
context, rather than the broader method and reporting coverage of established
specialist tools.
VOI analysis has a long methodological history in decision analysis and health
economics [@claxton1999irrelevance; @ades2004evsi].
The R package voi implements several approaches to EVPI, EVPPI, EVSI, and
ENBS [@voi_cran2024]. BCEA combines Bayesian cost-effectiveness analysis with
VOI and graphical reporting [@green2022bcea], while dampack supports analysis
and visualisation of decision-model outputs, sensitivity analysis, and VOI
[@dampack2024]. These packages provide methods and reporting features absent
from voiage's smaller R and Julia interfaces. The Sheffield Accelerated Value
of Information (SAVI) application provides web-based EVPI and
regression-based EVPPI [@savi2025; @strong2014evppi]. A review of web tools
describes differences in their methods and required inputs
[@tuffaha2021webtools]. Implementation-adjusted methods also estimate how much
value will be realised when new evidence does not fully change practice
[@andronis2016implementation].
Python alternatives address different parts of the problem.
value-of-information estimates the value of a noisy signal about one
uncertain option in a simplified binary decision
[@adamczewski2022valueofinformation]. The author's
trd-cea-toolkit places VOI within a disease-specific health-economic workflow
[@mordaunt2025trdcea]. voiage does not replace these tools or claim broader
method coverage. Its distinct contribution is one EVPI calculation and a
versioned decision record usable from Python, R, and Julia. Building that link
around one environment-specific package would either make that environment a
dependency or require another compatibility layer. The separate core lets
analysts keep their preferred modelling tools while sharing one calculation and
one description of the decision; specialist tools retain their broader methods
and reporting.
One Rust implementation calculates EVPI for Python, R, and Julia, reducing the risk that separately maintained versions produce different answers. Analysts can still prepare data, fit models, and plot results in their preferred language, and R and Julia users do not need Python. The trade-off is installation: each operating system needs a compatible compiled library. Python supports the broadest workflow; R also offers optional Python-backed EVPPI and EVSI, while Julia supports EVPI only.
Method tests compare specified inputs and outputs with separately implemented reference equations, and check repeatability and invalid inputs. A method is labelled stable only after its implementation, exposed language interfaces, documentation, release record, and supported platforms agree. The worked example uses a normal model in which an equally allocated two-group study updates uncertainty about average health gain. It assumes known outcome variability and a linear relationship between health gain and net benefit; the reproduction notes give the equations. Ades et al. and Rothery et al. describe the broader EVSI methodology [@ades2004evsi; @rothery2020voi].
Repository tests compare EVPI with hand-calculated cases in Python and Julia and with a native Rust-backed case in R. They compare analytical EVSI with a closed-form reference and reject invalid inputs. Continuous-integration workflows are configured to test Python wheels and the native R and Julia interfaces on Linux, macOS, and Windows.
The synthetic health example compares a programme with current practice using 10,000 simulated health-gain and programme-cost pairs. Health gain has mean 0.06 quality-adjusted life years (QALYs) and standard deviation 0.03 QALYs; cost is generated independently with mean 3,000 and standard deviation 650. Amounts are undated synthetic units with no jurisdictional or payer perspective. At 50,000 units per QALY, the generating distribution has an expected incremental net benefit of zero. The fixed sample mean is -15.9 units because a finite random sample does not reproduce that expectation exactly.
The algebraic illustration estimates the difference in mean QALYs between two equally allocated groups. Its normal model uses a known standard deviation of 1.0 QALY only to demonstrate the calculation; it is not a proposed trial. Total sample sizes range from 50 to 1,200. The study updates health gain, not programme-cost uncertainty.
The programme is preferred in 49.2% of simulations. Repeatedly resampling the 10,000 paired simulations gives 95% Monte Carlo intervals of 48.2% to 50.2% for programme preference, 624 to 658 for EVPI, 569 to 603 for health-gain EVPPI, and 229 to 265 for programme-cost EVPPI. These intervals measure numerical variation caused by a finite simulation sample; they do not represent uncertainty about whether the model is correct. Estimated EVPI is 644 units; EVPPI is 590 for health gain and 250 for programme cost. The regression is appropriate because net benefit is linear and the inputs were generated independently; it does not validate nonlinear or correlated applications. This comparison does not identify a preferred study, and EVPPI estimates should not be added together.
Under the stated assumptions, a 200-participant study has an EVSI of 124 units per eligible future person. ENBS multiplies that value across expected beneficiaries and subtracts study cost [@rothery2020voi]. The calculation assumes 1,300 eligible people annually for ten years, 3% annual discounting, a fixed study cost of 1.2 million, and 100 units per participant. Immediate, complete use changes ENBS from negative at 100 participants to positive at 200. With a two-year delay and 60% realisation, it changes from negative at 800 to positive at 1,200. These brackets do not identify the best sample size; delay and realisation are assumptions, not predictions of study conduct or practice change.
The sensitivity table varies assumed individual study-outcome standard deviation, population, study cost, delay, and value realisation. \autoref{fig:health-example} summarises the results. Panel C shows conditional scenarios, not uncertainty intervals.
{#fig:health-example}
A related health-economic project by the same author publishes versioned input
and expected-result formats designed to work with voiage
[@vop_poc_nz2026]. A released-package run calculated EVPI for its
HPV-vaccination model. This is completed
developer research use, not independent adoption or a policy estimate. The
fixed-seed example, equations, sensitivity data, and regeneration command give
another researcher material to check the calculations and assess the exchange
format for a real analysis. Development has been public since July 2025;
engagement by non-authors has not yet been documented.
Generative artificial intelligence (AI) tools assisted with this work. OpenAI Codex and Google Jules assisted with code and test generation, refactoring, documentation, workflows, and manuscript drafting and editing. The current preparation environment records Codex CLI 0.144.1 and Jules CLI 0.1.42. Codex used GPT-5-family models; Jules used Google-managed models whose identifiers were not exposed by the service. Exact model identifiers were not retained for every historical session. The human author remained the primary decision-maker, selected the research problem and architecture, reviewed, modified, and validated all retained AI-assisted outputs, and reran the reported tests and numerical checks. The author accepts responsibility for the software, manuscript, claims, citations, licensing, and submission. No AI system is an author.
This work received no external funding. The author declares no competing interests.
The source code, signed release 2.0.0, and synthetic worked-example materials are public [@voiage2026]. The Software Heritage snapshot includes the signed release [@voiage_software_heritage]. Repository manifests record release checksums and provenance, the software bill of materials, fixed random seeds, input and output hashes, and the command used to verify the worked example. The project provides the fixed-seed generation script, machine-readable inputs and outputs, and regeneration instructions; all worked-example data are synthetic. The manifests also identify execution environment details.