Expand a seed covariate frame to n_super rows via a Gaussian copula.

Usage

expand_covariates(
  seed_df,
  vars,
  n_super,
  seed = 20260903L,
  integer_vars = NULL,
  method = c("copula", "bootstrap"),
  latent_R = NULL
)

Arguments

seed_df: data.frame of the seed trial (only vars are used)
vars: covariates to expand (numeric or 0/1 binary)
n_super: rows to generate
seed: RNG seed
integer_vars: variables to round back to integers (auto-detected if NULL)
method: "copula" (default) or "bootstrap" (row resampling, as the package does by default – no novel covariate patterns)
latent_R: NULL (default), or a latent correlation matrix to use in place of a fitted one. It must be numeric, square, with row and column names equal (as sets) to vars; it is reordered to vars, and checked symmetric with unit diagonal and positive definite. A supplied matrix is used as given and no fit is attempted (latentcor is not called), so a DGM can be rebuilt on any platform from a stored dgm$mrct$latent_R. Ignored when method = "bootstrap".

Value

data.frame with n_super rows and columns vars

Details

Latent correlation (copula). With latent_R = NULL the latent correlation is fitted by latentcor::latentcor() (types "bin"/"con", method "original", nearPD) when latentcor is installed. When latentcor is not installed, or its call fails, the normal-score (van der Waerden) correlation of the seed rows, projected by Matrix::nearPD(), is used instead and a warning names the fallback and the reason. The two give different super-populations from the same seed, so a DGM built without latentcor is not the one built with it; pass latent_R to make the build platform-independent. The matrix used is returned as attr(, "latent_R").