Skip to contents

Post-smooth a fit's per-effect curves and add credible bands

Source: R/mfsusie_methods.R
mf_post_smooth.Rd

Four smoothing methods are dispatched by method:

Usage

mf_post_smooth(
  fit,
  method = c("TI", "scalewise", "HMM", "smash"),
  level = 0.95,
  threshold_factor = 1,
  wavelet_filter = 1L,
  wavelet_family = "DaubExPhase",
  halfK = 20L,
  overwrite_previous = FALSE,
  X = NULL,
  Y = NULL
)

Arguments

fit

a fit returned by mfsusie() or fsusie().

method

one of "scalewise", "TI", "HMM".

level

numeric in (0, 1), credible-band coverage. Default 0.95.

threshold_factor

numeric, multiplier on the universal sqrt(2 log T) threshold for method = "scalewise". Ignored by other methods.

wavelet_filter

integer, wavelet filter number for the TI stationary-wavelet transform. Default 1 (Haar).

wavelet_family

character, wavelet family for the TI stationary-wavelet transform. Default "DaubExPhase".

halfK

integer, half-grid size for the HMM fit_hmm helper. Default 20.

X

optional numeric matrix n x p, the original genotype matrix. Required when the fit was built with attach_smoothing_inputs = FALSE; ignored otherwise.

Y

optional list of length M of numeric matrices, the original per-outcome response matrices. Required when the fit was built with attach_smoothing_inputs = FALSE; ignored otherwise.

Value

the input fit with $effect_curves, $credible_bands, and $lfsr_curves populated. Scalar outcomes (T_basis[m] = 1) skip the wavelet step (smoothing is a no-op there).

Details

  • "scalewise" – per-scale soft-thresholding of the lead variable's wavelet posterior mean. Fast, no iterations, uses only the wavelet posterior moments. Suitable for quick visual cleanup.

  • "TI" (default) – cycle-spinning translation-invariant denoising (Coifman & Donoho 1995). For each effect, isolates the per-effect residual response (in position space), regresses onto the alpha-weighted X aggregate, applies the stationary wavelet transform row-by-row, scalewise ashr::ash shrinkage on wavelet coefficients, and inverts via cycle-spinning average. Produces tighter credible bands than scalewise.

  • "HMM" – hidden Markov denoising on per-position regression coefficients. Yields a posterior mean curve plus a mixture- posterior local false sign rate.

  • "smash" – smashr::smash.gaus empirical-Bayes wavelet shrinkage on the per-position regression estimate. Requires the smashr package (Suggests).

All four methods populate lfsr_curves. For TI, scalewise, and smash the lfsr is pnorm(-|mean| / sd) under the Gaussian-posterior approximation; for HMM it comes from the mixture posterior directly.

By default the smoother reads fit$Y_grid (post-remap Y) and fit$X_eff (per-effect alpha-weighted X aggregate), both attached by mfsusie() when attach_smoothing_inputs = TRUE. When the fit was built with attach_smoothing_inputs = FALSE you can pass the original X and Y here and the smoother will compute and attach them on the fly.