stepXR {XReg}R Documentation

Forward stepwise function for adding terms to a extreme regression model

Description

stepXR This functions calls builds extreme regression models in a step-wise fashion

Usage

stepXR(y, x, ypred = y, xpred = x, kstep = 5, varlist = NULL, outtemp = NULL, niter = 5, step.size = 0.2, 
	niterfull = 20, step.sizefull = 0.2, kbest = 1, limsize = 0.05, step.adapt = F, upper = max(ypred), lower
	 = min(ypred), predict = T, randinit=0,penalty = 2, update.all = T,survival=F,times=y,status=y)

Arguments

\ \

y response variable
x matrix of predictors
xpred xmatrix for predictions
ypred y values on test data set
varlist is a list specifying the form of the extreme regression model. EXAMPLE: varlist=list(1, c(2,3),c(3,4)) –> max (a1+b1x1,min(a2+b2x2,a3+b3x3),min(a4+b4x3,a5+b5x4))
lower minimum prediction
outtemp current output model
step.adapt If T the step size is adapted so that the sum of squares is reduced at each step. This slows down the algorithm.
predict make predictions on the xpred values
update.all update all coefficients as terms are added
upper maximum prediction EXAMPLE lower=min(y) and upper=max(y) predictions will always be in the range of y.
limsize the targeted smallest number of observations used to estimate a univariate function.
niter number of estimations steps
step.adapt If T the step size is adapted so that the sum of squares is reduced at each step. This slows down the algorithm.
survival If yes, the expoential survival model is used.
times If survival=T, time under observation
status If survival=T, survival status (1=dead, 0=alive)
randinit If randinit>0 some noise is added to the initial estimates. Randinit adds noise of standard deviation form randinit/sqrt(n). This avoids bad initial starts when the same variable is involved in multiple terms.
kstep the number of linear components to add
penalty penalty used in model selection
step.size step.size for each term considered for addition
niterfull number of steps after term has been selected
step.sizefull as above
kbest used in addeval function and is the number of variables considered for selection. It is the kbest variables with highest univariate correlation.

Value

a list containing:
stepcoef list of coefficients for the stepwise process
steplist list of variables for the stepwise process
stepfit fit for training data for each model
stepfitpred fit for test data (if given) for each model
bestgcv best model by gcv

Author(s)

Michael LeBlanc mleblanc@fhcrc.org

Examples

##
  # See XReg

[Package XReg version 1.0 Index]