Run sensitivity analysis with different priors and outlier handling

Usage

run_sensitivity_analysis(
  data = NULL,
  base_model_formula = NULL,
  vary_priors = TRUE,
  vary_outliers = TRUE,
  family = "gaussian",
  priors = NULL,
  iter = 1000,
  chains = 2,
  model = NULL
)

Arguments

data: A data frame; with model, defaults to model$data.
base_model_formula: Formula string or formula object for the base model; ignored when model is supplied.
vary_priors: Logical; test different prior widths.
vary_outliers: Logical; add a Student-t refit.
family: brms family name of the base model (default "gaussian").
priors: Optional priors of the fitted base model, as a brmsprior or as the list returned by assemble_brms_priors(); prior arms replace only its treatment-coefficient prior (or its SD priors for prior_sd_small).
iter, chains: Sampler settings for each refit.
model: Optional fitted base brmsfit. Each arm then refits it with update(model, prior = ...) (and family = student() for the outlier arm), which keeps its full brms::bf() formula, including distributional submodels such as sigma ~ phase; its formula, family, and user-set priors replace base_model_formula, family, and priors. Without model, each arm is fitted from data and the mean formula only.

Value

A named list of brms::fixef() matrices, one per arm (prior_default, prior_narrow, prior_diffuse, prior_sd_small, student_t), with attributes treatment_sd (the default treatment-prior SD s of the arms) and treatment_coef (the targeted coefficient).

Details

The prior arms refit the model with the treatment-coefficient prior at the adaptive default SD (see generate_adaptive_priors()), at half of it (narrow), and at twice it (diffuse), each centered on 0 and targeted at the treatment coefficient. For a model with group-level terms the prior_sd_small arm keeps the base priors but sets every group-level SD prior to student_t(3, 0, s / sqrt(10)), where s is the default SD scale: a tenfold smaller variance scale, for the small individual-variation case where the variance priors drive the sensitivity. The outlier arm keeps every observation and refits with a Student-t likelihood, whose heavy tails downweight outlying points instead of deleting them; it runs only for Gaussian models, since a Student-t base fit already has that likelihood.

Examples

if (interactive()) {
dat <- process_scd_phases(simulate_scd_data(), list(), "AB")
run_sensitivity_analysis(dat, "outcome ~ time_A + level_AB + trend_AB + (1 | case)")
fit <- brms::brm(brms::bf(outcome ~ time_A + level_AB + trend_AB + (1 | case),
                          sigma ~ phase), data = dat, refresh = 0)
run_sensitivity_analysis(model = fit)
}