Estimation diagnostics for a fit

Compute the estimation diagnostics available for a fit: Cook's distance, per-observation leverage, and observation influence (from the objective's Hessian at the fitted point) for fit_map() and fit_gmm() fits; leverage, PSIS-LOO Pareto-k, and prior influence for fit_bayesian() fits. Wraps the shared C++ Diagnostics layer (LeverageDiagnostics, InfluenceDiagnostics, PriorInfluenceDiagnostics).

Usage

fit_diagnostics(fit)

Arguments

fit

a corehydro_fit from fit_map(), fit_bayesian(), or fit_gmm(). fit_mle() fits are not supported: MaximumLikelihood has no posterior to diagnose.

Value

A named list. $cooks_distance and $leverage (one value per observation) and $observation_influence (an observations-by-parameters matrix) are populated for fit_map() and fit_gmm() fits. $pareto_k (one per observation), $max_pareto_k, and $prior_influence/$prior_influence_names (one per parameter) are populated for fit_bayesian() fits. Fields the fit's target does not support come back empty.

Details

Reruns the fit's own construct through the corresponding C++ diagnostics method rather than reusing anything cached on fit – for a fit_bayesian() result this reproduces the identical seeded chain (the construct carries the same explicit warmup/seed/... the original fit used), so the diagnostics agree with the fit they were computed from.

See also

fit_map(), fit_bayesian(), fit_gmm(), estimation_diagnostics().

Examples

peaks <- c(12500, 15300, 8900, 22100, 18700, 14200, 9800, 28500, 17400, 11600)
d <- fit_diagnostics(fit_map(model_univariate("Normal", peaks)))
d$cooks_distance
#>  [1] 0 0 0 0 0 0 0 0 0 0