Prior sensitivity analysis for an ML-UMR fit

Refit an mlumr() model across a grid of prior_beta scales (keeping the family, mean, and df fixed) and summarize how the posterior for the marginal treatment effects (delta_index, delta_comparator) moves. This is the workflow recommended by Vehtari et al.'s prior-choice wiki for judging how much of the posterior is driven by the data versus the prior.

Usage

prior_sensitivity(
  fit,
  prior_beta_scales = c(0.5, 1, 2.5, 5, 10),
  prior_beta_comparator_scales = NULL,
  probs = c(0.025, 0.5, 0.975),
  verbose = TRUE,
  ...
)

Arguments

fit

A fitted mlumr_fit object to re-fit under alternative priors.

prior_beta_scales

Numeric vector of scales for prior_beta. Default c(0.5, 1, 2.5, 5, 10). For a normal identity-link fit whose prior_beta is the package default or autoscaled, each scale is in units of the IPD outcome SD, as the fit's own prior is (see prior_normal()). The intercept and sigma priors are held at the fit's own.

prior_beta_comparator_scales

(Relaxed fits only.) Numeric vector of scales for prior_beta_comparator, paired elementwise with prior_beta_scales. NULL (default) sweeps the comparator prior in parallel with prior_beta_scales; the scale used is reported in the scale_comparator column. Ignored, with a warning, for SPFA fits.

probs

Quantiles for summarizing each posterior (default c(0.025, 0.5, 0.975)).

verbose

Logical; if FALSE, suppresses progress messages and final printed summary table.

...

Additional arguments forwarded to mlumr() on each refit (e.g. chains, iter, refresh). Sampling defaults otherwise inherit from the original fit.

Value

A data frame with one row per (prior scale, summarized parameter) pair, and columns scale, scale_comparator (dropped when the model has no comparator coefficient prior), parameter, effect, at_time (present only when the summarized effect has an evaluation time, so absent for every non-survival family and for survival scalars that carry none), mean, sd, and one column per requested quantile, named q followed by the percentage (the default probs give q2.5, q50, q97.5), matching marginal_effects(). Quantiles are columns, not a row dimension. Side effect: prints a summary table at the end when verbose = TRUE, which is the default; verbose = FALSE returns the same data frame and prints nothing.

Details

The design-matrix controls (center, qr) are taken from the original fit and replayed, so a refit reproduces the original parameterization instead of reverting to the defaults. A fit made with center = FALSE or qr = TRUE is a different parameterization, and replaying the defaults would vary the model as well as the prior.

Only the scale of prior_beta is varied; its family and mean, and every other prior and setting, come from the original fit. Each scale is applied to every coefficient, so the sweep reflects one level of prior informativeness per refit; an exponential prior_beta is swapped for prior_normal(0, scale). For a relaxed fit the comparator prior is swept alongside, because the index-population estimand is driven by the comparator coefficients; holding their prior fixed would report a flat curve for exactly the quantity most exposed to the prior.

See also

prior_summary() for a one-shot description of the priors on a fit; marginal_effects() for the posterior summary quantities this sweep tracks.

Examples

if (FALSE) { # \dontrun{
sens <- prior_sensitivity(fit_spfa, prior_beta_scales = c(1, 2.5, 5))
} # }