Usage
cat_control(center = FALSE, standardize = FALSE, accuracy = 2, digits = 4,
g = 0.5, epsilon = 10^(-5), maxi = 250, c = 10^(-5), gama = 20, steps = 25,
nu = 1, tuning.criterion = "GCV", K = 5, cv.refit = FALSE,
lambda.upper=50, lambda.lower=0, lambda.accuracy=.01, scaled.lik=FALSE,
adapted.weights=FALSE, adapted.weights.adj = FALSE, adapted.weights.ridge =
FALSE, assured.intercept=TRUE,
level.control = FALSE, case.control = FALSE, pairwise = TRUE,
grouped.cat.diffs = FALSE, bootstrap = 0, start.ml = FALSE, L0.log = TRUE,
subjspec.gr = FALSE, high = NULL, ...)
Arguments
center
logical; if TRUE, all metric covariates are centered by their empirical mean
standardize
logical; if TRUE, the design matrix is standardized by its (weighted) empirical variances
accuracy
integer; number of digits being compared when setting coefficents equal/to zero
digits
integer; number of digits for estimates
g
step length parameter for the PIRLS-algorithm; out of )0,1( epsilon
small, positive, real constant; the PIRLS-algorithm is terminated when the (scaled, absolute) difference of the coefficients of the current iteration and the coefficients of the previous iteration is smaller than epsilon maxi
integer; maximal number of iterations in the fitting algorithm
c
small, positive, real constant; needed for the approximation of the absolute value function in the PIRLS-algorithm gama
positive number; tuning parameter for the approximation of the L0 norm
steps
integer; tuning parameter for path-plotting; minimal number of estimates employed for path-plotting
nu
optional weighting parameter
tuning.criterion
loss criterion for cross-validation; one out of "GCV" (generalized cross validation criterion), "deviance" (K-fold cross-validation with the predictive deviance as criterion)
K
integer; number of folds for cross-validation
cv.refit
logical; if TRUE, cross-validation is based on a refit of the selected coefficients
lambda.upper
integer; upper bound for cross-validation of lambda
lambda.lower
integer; lower bound for cross-validation of lambda
lambda.accuracy
numeric; how accurate shall lambda be cross-validated?; minimal absolute difference between two candidates for lambda
scaled.lik
if TRUE, the likelihood in the objective function is scaled by 1/n
adapted.weights
logical; if TRUE, penalty terms are weighted adaptively, that is by inverse ML-estimates; set to FALSE, if ML-estimates do not exist/are to close to zero; only for specials v, p, grouped, SCAD, elastic
adapted.weights.adj
logical; if TRUE, adapted weights of several categorical covariates are scaled such that they are comparable
adapted.weights.ridge
logical; if TRUE, adapted weights are based on aa estimate that is slightly penalized by a Ridge penalty
assured.intercept
logical; shall a constant intercept remain in the model in any case?
level.control
logical; if TRUE, the penalty terms are adjusted for different number of penalty terms per covariate
case.control
logical; if TRUE, the penalty terms are adjusted for the number of observations on each level of a categorical covariate
pairwise
experimental option; disabled if TRUE
grouped.cat.diffs
experimental option; disabled if FALSE
bootstrap
experimental option; disabled if 0
start.ml
logical; if TRUE, the initial value is the ML-estimate
L0.log
experimental option; disabled if TRUE
subjspec.gr
experimental option; disabled if FALSE
high
experimental option; disabled if NULL
...
further arguments passed to or from other methods