stepmpr {mpr}

Stepwise Selection Procedure for Multi-Parameter Regression (MPR) Models

Description

Applies a stepwise selection procedure to an object of class “mpr” to find the best model in the sense of AIC (or BIC).

Usage

stepmpr(object, scope = list(lower = ~ 1, upper = ~ .),
        comp = 1:(object$ncomp), direction = c("both", "backward", "forward"),
        joint = TRUE, jointonly = FALSE, aic = TRUE, trace = 3, ...)

Arguments

object

an object of class “mpr” which is the result of a call to mpr.

scope

either a single formula defining the upper (maximal) model or a list containing two formulae - the lower (minimal) and upper (maximal) models respectively. See “Details” for further information.

comp

a numeric vector indicating the regression component(s) to which the selection procedure should be applied. Note that “1” = \lambda, “2” = \gamma and “3” = \rho. For more information on the various components, see mpr and distributions.

direction

the mode of stepwise search, which can be one of "both", "backward", or "forward", with a default of "both".

joint

logical. If TRUE, the selection procedure carries out joint component (i.e., simultaneous) steps in addition to individual component steps. If FALSE, only individual component steps are carried out. See “Details” for more information.

jointonly

logical. If TRUE, the selection procedure only carries out joint component steps.

aic

logical. If TRUE, AIC is used as the basis for determining the best model among those considered. If FALSE, BIC is used.

trace

if positive, information is printed during the running of stepmpr. Larger values may give more detailed information.

...

additional arguments to be passed to internal methods.

Details

The function stepmpr uses repeated calls to addterm and dropterm and is based on the idea that variable selection should be applied to each component individually and to all components jointly (when joint = TRUE). As an example, consider the case where forward selection (direction = "forward") will be carried out in components 1 and 2 individually (comp = 1:2) and jointly (joint = TRUE). At a given iteration of the algorithm, the following single-term additions are then carried out:

individual step 1:

each term currently absent from component 1 will be considered.

individual step 2:

each term currently absent from component 2 will be considered.

joint step 1&2:

each term currently absent from both components 1 and 2 will be considered.

The reason for the joint step is to account for the possibility that a covariate may only appear significant when it is present simultaneously in both regression components. This situation can arise as the variance-covariance matrix for the estimated regression coefficients is typically not block diagonal with respect to the regression components and, in particular, coefficients for the same covariate in different components are typically highly correlated. Of course, the stepmpr function has the flexibility to carry individual steps only, joint steps only or individual steps in a particular component only as the end user prefers. See “Examples” below.

The set of models searched is determined by the scope argument which is either a single upper formula or a list whose elements are lower and upper formulae. The upper formula must contain each formula corresponding to the components under consideration (as indicated by comp) whereas the lower formula must be contained within each of these formulae. For more information on the use of lower and upper, see addterm.

If scope is missing, the lower model is simply the null model (i.e., a model with no covariates) and the upper model is formed using terms from the initial model; specifically, all terms from the regession components under consideration (as indicated by comp) are used in the upper formula.

Value

A final “best” mpr model as selected using the stepwise procedure.

Author(s)

Kevin Burke.

See Also

mpr, dropterm, addterm, update.mpr.

Examples

# Veterans' administration lung cancer data
veteran <- survival::veteran
head(veteran)

#######
mod0 <- mpr(Surv(time, status) ~ 1, data=veteran)
mod0 # family = "Weibull" by default

# the "upper" model formula (by default the lower will be ~ 1)
scope <- ~ trt + celltype

stepmpr(mod0, scope)
stepmpr(mod0, scope, direction="forward", aic=FALSE)

# individual steps only
stepmpr(mod0, scope, joint=FALSE)

# joint steps only
stepmpr(mod0, scope, jointonly=TRUE)

# component 1 only (and, hence, only individual steps)
stepmpr(mod0, scope, comp=1)

#######
mod1 <- mpr(Surv(time, status) ~ trt + celltype, data=veteran)
mod1

stepmpr(mod1)
stepmpr(mod1, scope = ~ .^2)

# "lower" model formula forces trt to stay in
stepmpr(mod1, scope = list(~trt, ~.))