Temporal-noise specification

Usage

noise_spec(
  struct = c("iid", "ar1", "ar2", "arp"),
  p = NULL,
  q = 0L,
  iter_gls = 1L,
  pooling = c("run", "global", "parcel"),
  shared_estimator = c("pooled_acvf", "mean_series"),
  parcels = NULL,
  voxelwise = FALSE,
  exact_first = FALSE,
  censor = NULL,
  shrink_c0 = 100L
)

Arguments

struct: One of iid, ar1, ar2, or arp.
p: AR order for struct = "arp".
q: Nonnegative moving-average order. Positive values request an ARMA(p, q) model, with the AR order determined by struct and p. The current MA-capable path supports runwise meta-estimation with run pooling, without censoring, parcel pooling, or voxelwise covariance estimation.
iter_gls: Maximum number of GLS refinement iterations. Standard design-corrected AR fitting uses the initial OLS residuals once and holds that estimate fixed for the GLS solve. Iterative refinement remains available to ARMA models (q > 0).
pooling: Temporal covariance pooling scope. With built-in shared AR estimation, "run" estimates one coefficient vector per run and "global" pools across runs.
shared_estimator: Spatial estimator for a shared temporal covariance. "pooled_acvf" (the default) pools residual autocovariances across voxels and therefore targets a typical voxel covariance. "mean_series" first averages residual values across voxels and targets the coherent spatial component; it is experimental, can be much more autocorrelated than an individual voxel. Because OLS projection is linear across response columns, the matching design correction remains valid after averaging.
parcels: Optional parcel labels for parcel pooling.
voxelwise: Estimate temporal covariance separately by voxel. In the built-in fitter this currently requires AR-only runwise meta-estimation, pooling = "run", iter_gls = 1, no censoring, and no volume weighting or soft-subspace projection. Robust fitting is supported, but robust AR re-estimation is not. Registered engines may define broader capabilities.
exact_first: Use exact first-observation AR scaling.
censor: Optional censor indices, logical mask, "auto", or "none". When left NULL a censor column carried by the dataset (as written by matrix_frame()) is used, which is what "auto" requests explicitly; "none" ignores such a column. Censoring feeds AR estimation and whitening only and does not drop volumes from the regression, so it has no effect when struct = "iid".
shrink_c0: Parcel shrinkage constant used by supporting engines.