Bootstrap resampled parameter objects (without decomposition)

Stage 1 of a two-stage bootstrap. Runs B replicates of (resample -> fit GAMLSS -> ineqx_params) and returns the resampled ineqx_params objects together with the original (unresampled) fit and the resampling indices used. The standard errors come later, from decompose_boot_params.

Usage

bootstrap_params(
  data,
  formula_mu,
  formula_sigma,
  treat,
  group,
  time = NULL,
  post = NULL,
  ystat = "Var",
  estimand = "marginal",
  weights = NULL,
  B = 100L,
  parallel = FALSE,
  ncores = NULL,
  seed = NULL,
  verbose = TRUE,
  cl_type = NULL
)

Arguments

data: Data.frame, the original individual-level data
formula_mu: Two-sided formula for the mean (mu) equation
formula_sigma: One-sided formula for the log-SD (sigma) equation
treat: Character, treatment variable name
group: Character, grouping variable name
time: Character, time variable name. NULL for cross-sectional.
post: Character, pre/post indicator for DiD. NULL for simple diff.
ystat: Character, "Var" or "CV2"
B: Integer, number of bootstrap replicates. Default 100.
parallel: Logical, use parallel computation. Default FALSE.
ncores: Integer, number of cores. Default: all but one.
seed: Integer, random seed. Default NULL.
verbose: Logical, print progress. Default TRUE.
cl_type: Character, parallel cluster type, "fork" or "psock". Default NULL auto-selects (fork off Windows, psock on Windows). See boot_config for details.

Value

An object of class "ineqx_boot_params" with elements:

params_list: List of B (possibly < B if some replicates failed) ineqx_params objects from resampled GAMLSS fits.
boot_indices: List of B integer vectors of row indices used for resampling. Same RNG-derived indices as bootstrap_se() would have produced with the same seed.
B: Requested replicates.
B_successful: Replicates that produced a usable params object.
B_failed: Replicates that errored out.
type: "cross" or "longit".
seed: The random seed used.
ystat: The ystat the params were extracted at; informational only — decompose_boot_params() can override per call.

Details

Use this when you want to compute bootstrap SEs for several decomposition views of the same GAMLSS fit (e.g. different ystat, different ref) without paying the GAMLSS fit cost B times per view. The expensive work (GAMLSS fit + counterfactual prediction) runs once per replicate here; the cheap part (variance decomposition) runs as many times as you want via decompose_boot_params.

For the simple "one decomposition, one bootstrap" case, prefer the integrated bootstrap_se which wraps this two-stage flow.

Examples

if (FALSE) { # \dontrun{
# Two-stage workflow: one boot, many decompositions
bp <- bootstrap_params(
  data = mydata,
  formula_mu    = y ~ treat * group * time + age,
  formula_sigma =   ~ treat * group * time + age,
  treat = "treat", group = "group", time = "time",
  ystat = "Var", B = 500, seed = 42, parallel = TRUE
)

se_var_1980 <- decompose_boot_params(bp, ref = 1980, ystat = "Var")
se_cv2_1980 <- decompose_boot_params(bp, ref = 1980, ystat = "CV2")
se_var_2000 <- decompose_boot_params(bp, ref = 2000, ystat = "Var")
} # }