EVSI

Expected Value of Sample Information quantifies the expected benefit of a specified study before making the decision.

The public normal_normal_two_arm_evsi() function evaluates an equal-allocation two-arm study when the uncertain incremental effect has a normal prior, outcome variance is known, and incremental net benefit is linear in the effect. The calculation is Rust-owned and is tested against a prespecified analytical reference.

The built-in evsi(..., method="two_loop") contract uses one mean_<arm> parameter per trial arm and one finite, positive, fixed sd_outcome. It estimates a joint multivariate-normal prior from the PSA draws and applies the known-variance normal likelihood to the arm sample means. Current value, possible study results, and posterior value are all evaluated under that same fitted prior. Every correlated parameter is updated together rather than through independent marginal updates. The estimator uses genuine Gaussian prior and posterior draws. A diagnostic warning accompanies a negative estimate, and the raw value is returned so that convergence problems remain visible rather than being silently replaced by zero.

For another study model, supply two callbacks: one simulates trial data and one returns the requested number of joint posterior ParameterSet draws. A fixed seed gives the built-in path or both callbacks a reproducible local random stream.

The analytical function is the stable study-specific route. The generic two-loop interface remains developing because arbitrary economic models still require convergence assessment and method-specific validation. Increase both loop counts, compare repeated seeds, and verify that the estimate lies within simulation uncertainty of an independent reference.

The package also retains regression, efficient-linear, moment-based, seeded-bootstrap, adaptive, and network meta-analysis estimators. Their numerical implementations and software tests do not, by themselves, expose a complete validated study-model contract. The generic regression, efficient, and moment_based compatibility paths therefore emit a FutureWarning and are not represented as stable scientific estimators. Analysts need to validate any estimator against the sampling model for their decision problem.