1\name{boot.rq} 2\alias{boot.rq} 3\alias{boot.rq.xy} 4\alias{boot.rq.wxy} 5\alias{boot.rq.pwy} 6\alias{boot.rq.spwy} 7\alias{boot.rq.mcmb} 8\title{ Bootstrapping Quantile Regression} 9\description{ 10These functions can be used to construct standard errors, confidence 11intervals and tests of hypotheses regarding quantile regression models. 12} 13\usage{ 14boot.rq(x, y, tau = 0.5, R = 200, bsmethod = "xy", mofn = length(y), 15 coef = NULL, blbn = NULL, cluster = NULL, U = NULL, ...) 16} 17\arguments{ 18 \item{x}{ The regression design matrix} 19 \item{y}{ The regression response vector} 20 \item{tau}{ The quantile of interest} 21 \item{R}{ The number of bootstrap replications} 22 \item{bsmethod}{ The method to be employed. There are (as yet) five 23 options: method = "xy" uses the xy-pair method, and 24 method = "pwy" uses the method of Parzen, Wei and Ying (1994) 25 method = "mcmb" uses the Markov chain marginal bootstrap 26 of He and Hu (2002) and Kocherginsky, He and Mu (2003). 27 The fourth method = "wxy" uses the generalized bootstrap 28 of Bose and Chatterjee (2003) with unit exponential weights, 29 see also Chamberlain and Imbens (2003). The fifth method 30 "wild" uses the wild bootstrap method proposed by Feng, He and Hu (2011). } 31 \item{mofn}{ optional argument for the bootstrap method "xy" that 32 permits subsampling (m out of n) bootstrap. Obviously mofn 33 should be substantially larger than the column dimension of x, 34 and should be less than the sample size.} 35 \item{coef}{coefficients from initial fitted object} 36 \item{blbn}{orginal sample size for the BLB model} 37 \item{cluster}{If non-NULL this argument should specify cluster id 38 numbers for each observation, in which case the clustered version of 39 the bootstrap based on the proposal of Hagemann (2017). If present 40 \code{bsmethod} is set to set to "cluster". If this option is used 41 and the fitting method for the original call was "sfn" then the 42 bootstrapping will be carried out with the "sfn" as well. This 43 is usually substantially quicker than the older version which 44 employed the "br" variant of the simplex method. Use of "sfn" 45 also applies to the "pwy" method when the original fitting 46 was done with "sfn". Finally, if \code{na.action = "omit"} and 47 \code{length(object$na.action) > 0} then these elements are also 48 removed from the \code{cluster} variable. Consequently, the 49 length of the \code{cluster} variable should always be the same 50 as the length of the original response variable before any 51 \code{na.action} takes place. } 52 \item{U}{If non-NULL this argument should specify an array of indices 53 or gradient evaluations to be used by the corresponding bootstrap 54 method as specified by \code{bsmethod}. This is NOT intended as 55 a user specified input, instead it is specified in \code{summary.rqs} 56 to ensure that bootstrap samples for multiple taus use the same 57 realizations of the random sampling.} 58 \item{...}{ Optional arguments to control bootstrapping} 59} 60\details{ 61Their are several refinements that are still unimplemented. Percentile 62methods should be incorporated, and extensions of the methods to be used 63in anova.rq should be made. And more flexibility about what algorithm is 64used would also be good. 65} 66\value{ 67A list consisting of two elements: 68 A matrix \code{B} of dimension R by p is returned with the R resampled 69 estimates of the vector of quantile regression parameters. When 70 mofn < n for the "xy" method this matrix has been deflated by 71 the factor sqrt(m/n). 72 A matrix \code{U} of sampled indices (for \code{bsmethod in c("xy", "wxy")}) 73 or gradient evaluations (for \code{bsmethod in c("pwy", "cluster")}) 74 used to generate the bootstrapped realization, and potentially reused 75 for other \code{taus} when invoked from \code{summary.rqs}. 76} 77\references{ 78[1] Koenker, R. W. (1994). Confidence Intervals for regression quantiles, in 79P. Mandl and M. Huskova (eds.), \emph{Asymptotic Statistics}, 349--359, 80Springer-Verlag, New York. 81 82[2] Kocherginsky, M., He, X. and Mu, Y. (2005). 83Practical Confidence Intervals for Regression Quantiles, 84Journal of Computational and Graphical Statistics, 14, 41-55. 85 86[3] Hagemann, A. (2017) Cluster Robust Bootstrap inference in 87quantile regression models, Journal of the American Statistical Association , 88112, 446--456. 89 90[4] He, X. and Hu, F. (2002). Markov Chain Marginal Bootstrap. 91Journal of the American Statistical Association , Vol. 97, no. 459, 92783-795. 93 94[5] Parzen, M. I., L. Wei, and Z. Ying (1994): A resampling 95method based on pivotal estimating functions,'' Biometrika, 81, 341--350. 96 97[6] Bose, A. and S. Chatterjee, (2003) Generalized bootstrap for estimators 98of minimizers of convex functions, \emph{J. Stat. Planning and Inf}, 117, 225-239. 99 100[7] Chamberlain G. and Imbens G.W. (2003) Nonparametric Applications of 101Bayesian Inference, Journal of Business & Economic Statistics, 21, pp. 12-18. 102 103[8] Feng, Xingdong, Xuming He, and Jianhua Hu (2011) Wild Bootstrap for 104Quantile Regression, Biometrika, 98, 995--999. 105} 106 107\author{ Roger Koenker (and Xuming He and M. Kocherginsky for the mcmb code)} 108\seealso{ \code{\link{summary.rq}}} 109\examples{ 110y <- rnorm(50) 111x <- matrix(rnorm(100),50) 112fit <- rq(y~x,tau = .4) 113summary(fit,se = "boot", bsmethod= "xy") 114summary(fit,se = "boot", bsmethod= "pwy") 115#summary(fit,se = "boot", bsmethod= "mcmb") 116} 117\keyword{ regression} 118