Parametric Monte Carlo draws from a TMB fit's asymptotic sampling approximation

Draws R samples of the internal-scale fixed-effect parameter vector from the asymptotic Gaussian sampling approximation $N(\hat\beta, \hat\Sigma)$ (a frequentist approximation to the MLE's sampling distribution; no prior is involved), where $\hat\beta$ is object$model$coefficients and $\hat\Sigma$ is vcov(object) (the TMB sdreport fixed-effect covariance, sdr$cov.fixed). The mean vector and covariance are positionally aligned because both derive from the optimizer's opt$par.

Usage

.tmb_parametric_draws(object, R = 1000L, seed = NULL)

Arguments

object: A beezdemand_tmb object.
R: Integer number of Monte Carlo draws.
seed: Optional integer seed. When supplied, the caller's RNG state is saved and restored so the global .Random.seed stream is left unperturbed.

Value

Numeric matrix of dimension R x p, with columns named after the internal coefficient vector (names(object$model$coefficients)).

Details

This is the shared primitive behind confint(object, method = "simulate") and is intended for reuse by derived-metric bootstrap helpers. Draws are fixed-effect-only on the internal scale; callers transform to the natural scale or evaluate derived metrics as needed.

The symmetric matrix square root is formed via an eigendecomposition. Eigenvalues that are negative only to within numerical tolerance are clamped to zero, matching MASS::mvrnorm's robustness to a near-semidefinite covariance without taking a dependency; a positive-semidefinite but singular covariance is therefore still usable (it defines a degenerate Gaussian). A materially indefinite covariance is refused instead of being silently repaired, because clamping a real negative eigenvalue would draw from a different distribution than the one requested (F-BD4-3). "Materially" is judged twice: against the largest absolute eigenvalue of the covariance, and against the largest absolute eigenvalue of its correlation-scaled form, so that a negative variance in a small-scale parameter is not masked by a large-scale one.