BSSplitLasso: Bootstrap-calibrated R-split method

In WaverlyWei/debiased.subgroup: Sharp Inference on Selected Subgroup in Observational Studies

Description

This method first obtains the estimate of β via repetitive splitting procedure (R-Split) through BB iterations. Then it calculates the calibration term \tilde{b}_{max} = (1-n^{r-0.5})(\tilde{β}_{max}-\tilde{β}_{j}). Through B iterations, it recalibrates the bootstrap statistic T_b. The bias-reduced estimate is computed as: \tilde{b}_{max}-\frac{1}{B}∑_{b=1}^B T_b.

Usage

BSSplitLasso(
  y,
  x,
  r = NULL,
  G = NULL,
  B = NULL,
  BB = NULL,
  alpha = 0.95,
  splitRatio = 0.6,
  fold = 2
)

Arguments

y	response
x	design matrix
r	tuning parameter
G	subgroup indicator
B	bootstrap number
BB	split number
alpha	level ## change other places
splitRatio	split ratio
fold	cross validation fold

Value

LowerBound	lower confidence bound
UpperBound	upper confidence bound
betaMax	bias-reduced maximum beta estimate
betaEst	debiased beta estimate for each subgroup
modelSize	selected model size for R-split
op	optimal tuning

WaverlyWei/debiased.subgroup documentation built on Jan. 27, 2021, 12:15 a.m.