1\name{lmrob..M..fit} 2\alias{lmrob..M..fit} 3\title{Compute M-estimators of regression} 4\description{ 5 This function performs RWLS iterations to find an 6 M-estimator of regression. When started from an S-estimated 7 \code{beta.initial}, this results in an MM-estimator. 8} 9\usage{ 10lmrob..M..fit(x, y, beta.initial, scale, control, obj, 11 mf = obj$model, method = obj$control$method) 12} 13\arguments{ 14 \item{x}{design matrix (\eqn{n \times p}{n x p}) typically including a 15 column of \code{1}s for the intercept.} 16 \item{y}{numeric response vector (of length \eqn{n}).} 17 \item{beta.initial}{numeric vector (of length \eqn{p}) of initial 18 estimate. Usually the result of an S-regression estimator.} 19 \item{scale}{robust residual scale estimate. Usually 20 an S-scale estimator.} 21 \item{control}{list of control parameters, as returned 22 by \code{\link{lmrob.control}}. Currently, the components 23 \code{c("max.it", "rel.tol","trace.lev", "psi", "tuning.psi", "mts", "subsampling")} 24 are accessed.} 25 \item{obj}{an optional \code{lmrob}-object. If specified, this is 26 typically used to set values for the other arguments.} 27 \item{mf}{unused and deprecated.} 28 \item{method}{optional; the \code{method} used for \emph{obj} computation.} 29} 30\details{ 31 This function is used by \code{\link{lmrob.fit}} (and 32 \code{anova(<lmrob>, type = "Deviance")}) and typically not to be used 33 on its own. 34} 35\value{A list with the following elements: 36 \item{coef}{the M-estimator (or MM-estim.) of regression} 37 \item{control}{the \code{control} list input used} 38 \item{scale}{ The residual scale estimate} 39 \item{seed }{ The random number generator seed} 40 \item{converged}{ \code{TRUE} if the RWLS iterations converged, 41 \code{FALSE} otherwise} 42} 43\references{ 44 Yohai, 1987 45} 46\seealso{ 47 \code{\link{lmrob.fit}}, \code{\link{lmrob}}; 48 \code{\link[MASS]{rlm}} from package \CRANpkg{MASS}. 49} 50\author{Matias Salibian-Barrera and Martin Maechler} 51\examples{ 52data(stackloss) 53X <- model.matrix(stack.loss ~ . , data = stackloss) 54y <- stack.loss 55## Compute manual MM-estimate: 56## 1) initial LTS: 57m0 <- ltsReg(X[,-1], y) 58## 2) M-estimate started from LTS: 59m1 <- lmrob..M..fit(X, y, beta.initial = coef(m0), scale = m0$scale, method = "SM", 60 control = lmrob.control(tuning.psi = 1.6, psi = 'bisquare')) 61## no 'method' (nor 'obj'): 62m1. <- lmrob..M..fit(X, y, beta.initial = coef(m0), scale = m0$scale, 63 control = m1$control) 64stopifnot(all.equal(m1, m1., tol = 1e-15)) # identical {call *not* stored!} 65 66cbind(m0$coef, m1$coef) 67## the scale is kept fixed: 68stopifnot(identical(unname(m0$scale), m1$scale)) 69 70## robustness weights: are 71r.s <- with(m1, residuals/scale) # scaled residuals 72m1.wts <- Mpsi(r.s, cc = 1.6, psi="tukey") / r.s 73summarizeRobWeights(m1.wts) 74##--> outliers 1,3,4,13,21 75which(m0$lts.wt == 0) # 1,3,4,21 but not 13 76\dontshow{stopifnot(which(m0$lts.wt == 0) == c(1,3,4,21)) 77} 78## Manually add M-step to SMD-estimate (=> equivalent to "SMDM"): 79m2 <- lmrob(stack.loss ~ ., data = stackloss, method = 'SMD') 80m3 <- lmrob..M..fit(obj = m2) 81 82## Simple function that allows custom initial estimates 83## (Deprecated; use init argument to lmrob() instead.) %% MM: why deprecated? 84lmrob.custom <- function(x, y, beta.initial, scale, terms) { 85 ## initialize object 86 obj <- list(control = lmrob.control("KS2011"), 87 terms = terms) ## terms is needed for summary() 88 ## M-step 89 obj <- lmrob..M..fit(x, y, beta.initial, scale, obj = obj) 90 ## D-step 91 obj <- lmrob..D..fit(obj, x) 92 ## Add some missing elements 93 obj$cov <- TRUE ## enables calculation of cov matrix 94 obj$p <- obj$qr$rank 95 obj$degree.freedom <- length(y) - obj$p 96 ## M-step 97 obj <- lmrob..M..fit(x, y, obj=obj) 98 obj$control$method <- ".MDM" 99 obj 100} 101 102m4 <- lmrob.custom(X, y, m2$init$init.S$coef, 103 m2$init$scale, m2$terms) 104stopifnot(all.equal(m4$coef, m3$coef)) 105 106## Start from ltsReg: 107m5 <- ltsReg(stack.loss ~ ., data = stackloss) 108m6 <- lmrob.custom(m5$X, m5$Y, coef(m5), m5$scale, m5$terms) 109} 110\keyword{robust} 111\keyword{regression} 112