NpregBand: Confidence band for nonparametric regression

In koohyun-kwon/HTEBand: Uniform confidence band for heterogenous treatment effects

Description

Constructs a minimax optimal honest confidence band for nonparametric regression function values under a specified function class.

Usage

NpregBand(
  y,
  x,
  C,
  level,
  fclass = c("L", "H"),
  n.eval = length(x)/5,
  eval = NULL,
  q.int = 0.025,
  n.sim = 10^3,
  kern = "tri",
  deg = 1,
  var.reg = "npr",
  seed = NULL,
  root.robust = FALSE,
  ng = 10,
  x.out = NULL,
  c.method = "supp2",
  print.t = TRUE
)

Arguments

y	vector of dependent variables.
x	vector of regressors.
C	bound on the first or the second derivative, depending on fclass.
level	coverage probability.
fclass	function class specification, "L" for the Lipschitz class, and "H" for the Hölder class.
n.eval	number of grid points to use when constructing confidence band; default is n.eval = length(x) / 5.
eval	grid points to use when constructing confidence band; if it is not specified, eval is automatically determined by n.eval and q.int.
q.int	parameter determining the distance from the boundary of the support of x. Confidence band is formed for points between the q.int quantile and the 1 - q.int quantile of x. q.int should be between 0 and 0.5. The default is q.int = 0.025. Setting q.int > 0 is recommended, and it is only relevant when eval is not specified.
n.sim	number of bootstrap samples to use when calculating quantiles of suprema of empirical processes; the default is n.sim = 10^3
kern	type of kernel used in estimation; currently, only the triangle kernel (kern = "tri") is supported.
deg	degree of local polynomial estimator used in the first-stage variance estimation; the default is deg = 1.
var.reg	nonparametric regression method used to calculate residuals; either "npr" (standing for nprobust package) or "locpol" (standing for locpol package). Default is var.reg = "npr".
seed	seed number for bootstrap random sample generation. The default is seed = NULL.
root.robust	if TRUE, the fuction conducts diagnostic test whether optimizaiton worked well; default is root.robust = FALSE. Currently, this feature is obsolete.
ng	the number of grids of quantile values over which the diagnostic test would be peformed if root.robust = TRUE; default is ng = 10. Currently, this feature is obsolete.
x.out	the grid of points the confidence band is evaluated; If x.out is not NULL, confidence band is calculated over x.out using linear interpolation. Default value is NULL.
c.method	method to calculate the optimal value of c_n. Currently, using the default value c.method = "supp2" is recommended and other options are obsolete.
print.t	if TRUE, print the progress of confidence band construction with the number of loop.

Value

data.frame object containing index set and corresponding confidence band values.

Examples

x <- seq(-1, 1, length.out = 500)
y <- x^2 + rnorm(500, 0, 1/4)
NpregBand(y, x, 2, 0.95, "L", n.eval = 25)
NpregBand(y, x, 2, 0.95, "H", n.eval = 25)

koohyun-kwon/HTEBand documentation built on Dec. 21, 2021, 7:42 a.m.