Skip to contents

Build the explicit one-arm NIMBLE bundle

build_nimble_bundle.Rd

build_nimble_bundle() is the detailed constructor behind bundle for one-arm models. It compiles the modeling plan into a self-contained object holding code-generation inputs, initialization rules, monitor policy, and stored MCMC defaults.

Usage

build_nimble_bundle(
  y,
  X = NULL,
  ps = NULL,
  backend = c("sb", "crp", "spliced"),
  kernel,
  GPD = FALSE,
  components = NULL,
  param_specs = NULL,
  mcmc = list(niter = 2000, nburnin = 500, thin = 1, nchains = 1, seed = 1),
  epsilon = 0.025,
  alpha_random = TRUE,
  monitor = c("core", "full"),
  monitor_latent = FALSE,
  monitor_v = FALSE
)

Arguments

y

Numeric outcome vector.

X

Optional design matrix/data.frame (N x p) for conditional variants.

ps

Optional numeric vector (length N) of propensity scores. When provided, augments the design matrix for PS-adjusted outcome modeling.

backend

Character; the Dirichlet process representation:

  • "sb": stick-breaking truncation

  • "crp": Chinese Restaurant Process

kernel

Character kernel name (must exist in get_kernel_registry()).

GPD

Logical; whether a GPD tail is requested.

components

Integer >= 2. Single user-facing truncation parameter:

  • SB: number of mixture components used in stick-breaking truncation

  • CRP: maximum number of clusters represented in the finite NIMBLE model

param_specs

Optional list with entries bulk and tail to override defaults.

mcmc

Named list of MCMC settings (niter, nburnin, thin, nchains, seed). Stored in bundle.

epsilon

Numeric in [0,1). For downstream summaries/plots/prediction we keep the smaller k defined by either (i) cumulative mass >= 1 - epsilon or (ii) per-component weights >= epsilon, then renormalize.

alpha_random

Logical; whether the DP concentration parameter \(\kappa\) is stochastic.

monitor

Character monitor profile:

  • "core" (default): monitors only the essential model parameters

  • "full": monitors all model nodes

monitor_latent

Logical; if TRUE, include latent cluster labels (z) in monitors.

monitor_v

Logical; if TRUE and backend is SB, include stick breaks (v) in monitors.

Value

A named list of class "causalmixgpd_bundle". Its primary components are spec, code, constants, dimensions, data, inits, monitors, and stored mcmc settings.

Details

The returned bundle encodes a finite approximation to a Dirichlet process mixture using either a stick-breaking ("sb") or Chinese restaurant process / spliced ("crp" / "spliced") representation.

For the bulk-only model, the target likelihood is the DPM predictive law $$f(y \mid x) = \sum_{k=1}^{K} w_k(x) f_k(y \mid x, \theta_k).$$ When GPD = TRUE, the bundle augments the bulk model with a threshold \(u(x)\) and generalized Pareto tail above that threshold, producing the spliced predictive distribution described in the manuscript vignette.

This function intentionally stops before model compilation and sampling. Use run_mcmc_bundle_manual or mcmc to execute the stored model definition.

The object contains:

  • compiled model spec

  • nimbleCode model code

  • constants, data, explicit dimensions

  • initialization function inits (stored as a function)

  • monitor specification

  • MCMC settings list (stored but not used for code generation)

See also

bundle, run_mcmc_bundle_manual, predict.mixgpd_fit, kernel_support_table, get_kernel_registry.

Examples

if (FALSE) { # \dontrun{
y <- abs(rnorm(60)) + 0.1
bundle <- build_nimble_bundle(
  y = y,
  backend = "sb",
  kernel = "normal",
  GPD = FALSE,
  components = 4,
  mcmc = list(niter = 200, nburnin = 50, thin = 1, nchains = 1, seed = 1)
)
bundle
} # }