pact (version 0.5.0)

pact.fit: Fits a predictive model to the full dataset

Description

pact.fit Fits a predictive model using data on all subjects. Currently supports Cox PH and logistic regression models for 'survival' and 'binary' response types respectively.

Usage

pact.fit(Y, Xf = NULL, Xv, Treatment, family = c("binomial", "cox"), varSelect = c("none", "univar", "lasso"), nsig = ifelse(varSelect == "univar", ifelse(nCovarv < 10, 3, 10), NA), cvfolds.varSelect = ifelse(varSelect == "lasso", 5, NA), which.lambda = ifelse(varSelect == "lasso", ifelse(nCovarv < 10, "min", "1se"), NA), penalty.scaling = ifelse(varSelect == "lasso", 0.5, NA))

Arguments

Y

Response variable. For family='binomial', Y should be a factor with two levels. For family='cox', Y should be a two-column matrix with columns named 'time' and 'status'. The latter is a binary variable, with '1' indicating death, and '0' indicating right censored.

Xf

An optional dataframe of the prognostic covariates that are not subject to variable selection and always fixed to remain in the model. Default is NULL (no variable).

Xv

The main dataframe of covariates that are to be used for predictive model development. Variable selection options affect only the variables in Xv. Xv cannot be NULL.

Treatment

The treatment assignment indicator. A factor with two levels. '0' indicating control (C) and '1' indicating experimental (E) treatment.

family

Type of the response variable. See above. Possible values are 'binomial' or 'cox'.

varSelect

The variable selection method. Possible values are "none","univar" or "lasso".

nsig

The number of covariates to use in the model for varSelect="univar". Defaults to 3 if the number of candidate covariates is less than 10, else defaults to 10.

cvfolds.varSelect

The number of folds in the internal cross-validation loop for variable selection with varSelect="lasso". Default is 5.

which.lambda

Used with variable selection with varSelect="lasso". Defaults to "min" if the number of candidate covariates is less than 10, else defaults to "1se". See Details.

penalty.scaling

Ratio of shrinkage applied for main coefficients to shrinkage applied for interaction coefficients. Used with varSelect="lasso". Default is 0.5. See Details.

Value

An object of class 'pact' which is a list with the following components:

reg

The fitted regression model

family

Type of the response variable

Y

The response variable used

Xf

The dataframe of prognostic covariates

Xv

The dataframe of candidate predictive variables

Treatment

The treatment assignment indicator used

nCovarf

The number of variables in Xf

nCovarv

The number of variables in Xv

varSelect

The variable selection method used

nsig, cvfolds.varSelect, which.lambda, penalty.scaling

The variable selection parameters used

call

The call that produced the return object

Details

A Cox proportional hazards (PH) or a logistic regression model is developed for data with survival and binary response respectively. Data from subjects in both 'experimental' (E) and 'control' (C) groups from a RCT is used for model development. Main effect of treatment, main effect of prognostic covariates, main effects and treatment by covariate interaction terms of candidate predictive covariates are considered in the model. Methods for variable selection can be optionally specified by the user for candidate predictive covariates (useful for high-dimensional covariates). Current options for variable selection include "univar" and "lasso". In the case of "univar", the number of predictive covariates (nsig) to be included in the model is specified by the user. A univariate selection procedure is applied to identify covariates that have the lowest treatment*covariate interaction p-values. The predictive model is then developed using the main effect of treatment, main effects of prognostic covariates, main effects of the nsig predictive covariates and treatment by covariate interaction terms for nsig predictive covariates.

In the case of "lasso", an internal cross-validation loop is used to find the penalty value that minimizes the cross-validated error. The user can choose either the value of the penalty 'lambda' as the penalty that minimizes the cross-validated error ("lambda.min") or the largest penalty for which the cross-validated error is within 1 standard error of the minimum ("lambda.1se"). Also, in the case of "lasso", differential shrinkage can be specified for main effect and interaction effect predictive coefficients by specifying a value for the ratio of shrinkage for main coefficients to shrinkage for interaction coefficients. Internally, 'lambda' is scaled using this ratio to allow for the differential shrinkage of main and interaction coefficients.The penalty factors affect only variables in Xv and not Xf.

Examples

data(prostateCancer)
Y <- prostateCancer[,3:4]
Xf <- prostateCancer[,7:8]
Xv <- prostateCancer[,c(5:6,9)]
Treatment <- as.factor(prostateCancer[,2])
pact.fit(Y=Y, Xf=Xf, Xv=Xv, Treatment=Treatment, family="cox",varSelect="univar")