PGEE: Function to fit penalized generalized estimating equations

Description

This function fits a penalized generalized estimating equation model to longitudinal data.

Usage

PGEE(formula, id, data, na.action = NULL, family = gaussian(link = "identity"), 
corstr = "independence", Mv = NULL, beta_int = NULL, R = NULL, scale.fix = TRUE, 
scale.value = 1, lambda, pindex = NULL, eps = 10^-6, maxiter = 30, tol = 10^-3, 
silent = TRUE)

Arguments

formula

A formula expression in the form of response ~ predictors.

A vector for identifying subjects/clusters.

data

A data frame which stores the variables in formula with id variable.

na.action

A function to remove missing values from the data. Only na.omit is allowed here.

family

A family object: a list of functions and expressions for defining link and variance functions. Families supported in PGEE are binomial, gaussian, gamma and poisson. The links, which are not available in gee, is not available here. The default family is gaussian.

corstr

A character string, which specifies the type of correlation structure. Structures supported in PGEE are "AR-1","exchangeable", "fixed", "independence", "stat_M_dep","non_stat_M_dep", and "unstructured". The default corstr type is "independence".

If either "stat_M_dep", or "non_stat_M_dep" is specified in corstr, then this assigns a numeric value for Mv. Otherwise, the default value is NULL.

beta_int

User specified initial values for regression parameters. The default value is NULL.

If corstr = "fixed" is specified, then R is a square matrix of dimension maximum cluster size containing the user specified correlation. Otherwise, the default value is NULL.

scale.fix

A logical variable; if true, the scale parameter is fixed at the value of scale.value. The default value is TRUE.

scale.value

If scale.fix = TRUE, this assignes a numeric value to which the scale parameter should be fixed. The default value is 1.

lambda

A numerical value for the penalization parameter of the scad function, which is estimated via cross-validation.

pindex

An index vector showing the parameters which are not subject to penalization. The default value is NULL. However, in case of a model with intercept, the intercept parameter should be never penalized.

eps

A numerical value for the epsilon used in minorization-maximization algorithm. The default value is 10^-6.

maxiter

The number of iterations that is used in the estimation algorithm. The default value is 25.

tol

The tolerance level that is used in the estimation algorithm. The default value is 10^-3.

silent

A logical variable; if false, the regression parameter estimates at each iteration are printed. The default value is TRUE.

Value

An object class of PGEE representing the fit.

References

Wang, L., Zhou, J., and Qu, A. (2012). Penalized generalized estimating equations for high-dimensional longitudinal data analysis. Biometrics, 68, 353--360.

Examples

Run this code

# Consider an example similar to example 1 
# in Wang et al. (2012).

# required R package 
library(mvtnorm)
# number of subjects
n <- 200
# number of covariates 
pn <- 10
# number of time points
m <- 4

# vector if subject ids
id.vect <- rep(1:n, each = m) 

# covariance matrix of (pn-1) number of continuous covariates 
X.sigma <- matrix(0,(pn-1),(pn-1))
{
for (i in 1:(pn-1))
X.sigma[i,] <- 0.5^(abs((1:(pn-1))-i))  
}

# generate matrix of covariates    
x.mat <- as.matrix(rmvnorm(n*m, mean = rep(0,(pn-1)), X.sigma))
x.mat <- cbind(rbinom(n*m,1, 0.5), x.mat)

# true values
beta.true <- c(2,3,1.5,2,rep(0,6))
sigma2 <- 1
rho <- 0.5
R <- matrix(rho,m,m)+diag(rep(1-rho,m))

# covariance matrix of error
SIGMA <- sigma2*R
error <- rmvnorm(n, mean = rep(0,m),SIGMA)

# generate longitudinal data with continuous outcomes
y.temp <- x.mat%*%beta.true
y.vect <- y.temp+as.vector(t(error))

mydata <- data.frame(id.vect,y.vect,x.mat) 
colnames(mydata) <- c("id","y",paste("x",1:length(beta.true),sep = ""))

###Input Arguments for CVfit fitting###
library(PGEE)
formula <- "y ~.-id-1"
data <- mydata
family <- gaussian(link = "identity")
lambda.vec <- seq(0.1,1,0.1)

## Not run: ------------------------------------
# cv <- CVfit(formula = formula, id = id, data = data, family = family,
# fold = 4, lambda.vec = lambda.vec, pindex = NULL, eps = 10^-6, maxiter = 30, 
# tol = 10^-3)
# 
# names(cv)
# cv$lam.opt
## ---------------------------------------------

lambda <- 0.1 #this value obtained through CVfit

# analyze the data through penalized generalized estimating equations

myfit1 <- PGEE(formula = formula, id = id, data = data, na.action = NULL, 
family = family, corstr = "exchangeable", Mv = NULL, 
beta_int = c(rep(0,length(beta.true))), R = NULL, scale.fix = TRUE, 
scale.value = 1, lambda = lambda, pindex = NULL, eps = 10^-6, maxiter = 30, 
tol = 10^-3, silent = TRUE)

summary(myfit1)

# analyze the data through unpenalized generalized estimating equations

myfit2 <- MGEE(formula = formula, id = id, data = data, na.action = NULL, 
family = family, corstr = "exchangeable", Mv = NULL, 
beta_int = c(rep(0,length(beta.true))), R = NULL, scale.fix = TRUE, 
scale.value = 1, maxiter = 30, tol = 10^-3, silent = TRUE)

summary(myfit2)