Skip to contents

Apply the Bayesian subsampled bootstrap to a fitted tinyvamp model

Source: R/bootstrap_ci.R
bootstrap_ci.Rd

Compute bootstrap confidence intervals for the varying parameters of a fitted tinyvamp model using a Bayesian subsampled bootstrap procedure. Each bootstrap replicate refits the model using randomly generated observation weights, and confidence intervals are formed from the empirical quantiles of the bootstrap distribution.

Usage

bootstrap_ci(
  W,
  fitted_model,
  n_boot,
  m = NULL,
  alpha = 0.05,
  parallelize = FALSE,
  ncores = 5,
  seed = NULL,
  return_models = FALSE,
  verbose = FALSE,
  adjust = FALSE
)

Arguments

W

A numeric matrix of observed counts with \(n\) rows (samples) and \(J\) columns (taxa).

fitted_model

A fitted tinyvamp model object, typically returned by estimate_parameters(), containing the original fit and all inputs needed for refitting during bootstrap iterations.

n_boot

Integer; the number of bootstrap replicates.

m

Optional numeric subsample size parameter for the Bayesian subsampled bootstrap. If NULL, defaults to sqrt(n), where \(n\) is the number of rows of W.

alpha

Significance level used to construct two-sided confidence intervals. The default 0.05 yields 95% intervals.

parallelize

Logical; if TRUE, bootstrap replicates are fit in parallel using parallel::mclapply(). If FALSE, replicates are fit sequentially.

ncores

Integer; number of cores to use when parallelize = TRUE.

seed

Optional integer random seed. If NULL, a seed value of 0 is used.

return_models

Logical; if TRUE, return the fitted bootstrap model objects in addition to the confidence interval summary.

verbose

Logical; if TRUE, print progress messages during sequential bootstrap fitting and pass verbose output to the internal model fitting routine.

adjust

Logical; if TRUE, apply a finite-sample adjustment to the bootstrap deviations based on the number of non-"gamma" varying parameters.

Value

A list with components ci and bootstrapped_models. The former is a data frame with columns lower_ci and upper_ci. If return_models = TRUE, the latter is a list of fitted model objects from the bootstrap replicates; otherwise NULL.

Details

For each bootstrap replicate, the function generates a vector of Bayesian bootstrap weights from a gamma distribution, rescales them, and uses them as observation weights in a call to estimate_parameters(). The resulting bootstrap distribution is centered at the original fitted values and scaled by \(\sqrt{m}\). Confidence intervals are then obtained by inverting the bootstrap quantiles and scaling by \(1 / \sqrt{n}\).

When adjust = TRUE, the bootstrap deviations are multiplied by a correction factor $$\sqrt{\frac{nJ}{nJ - p}}$$ where \(p\) is the number of varying parameters whose param field is not equal to "gamma".