Fit a Mixed-Effects Cox Model

DESCRIPTION:

Returns an object of class "coxme" representing a fitted mixed-effects Cox model.

USAGE:

coxme(fixed, data, random, weights, subset, na.action,
      init, control, ties=c("efron", "breslow", "exact"),
      singular.ok=T, varlist, variance, vinit=.2,
      sparse=c(50, .02), rescale=T, refine.n=0, x=F, y=T, ...)

REQUIRED ARGUMENTS:

fixed
formula describing the fixed effects part of the model.
data
a data frame containing the variables.
random
a one-sided formula describing the random effects part of the model.

OPTIONAL ARGUMENTS:

weights
case weights for each observation
subset
an expression describing the subset of the data that should be used in the fit.
na.action
a function giving the default action on encountering missing values. It is more usual to use the global na.action system option to control this.
init
initial values for the coefficients for the fixed portion of the model, or the frailties followed by the fixed effect coefficients.
control
the result of a call to coxph.control
ties
the approximation to be used for tied death times: either "efron" or "breslow"
singular.ok
if TRUE, then redundant coefficients among the fixed effects are set to NA, if FALSE the program will fail with an error message if there are redundant variables.
varlist
variance specifications, often of class bdsmatrix, describing the variance/covariance structure of one or more of the random effects.
variance
fixed values for the variances of selected random effects. Values of 0 indicate that the final value should be solved for.
vinit
vector of initial values for variance terms. It is necessary that the initial variance matrix be symmetric positive definite. Normally, a simple sum of the varlist matrices will suffice, i.e., a vector of 1s; but not always.
sparse
determines which levels of random effects factor variables, if any, for which the program will use sparse matrix techniques. If a grouping variable has less than sparse[1] levels, then sparse methods are not used for that variable. If it has greater than or equal to sparse[1] unique levels, sparse methods will be used for those values which represent less than sparse[2] as a proportion of the data. For instance, if a grouping variable has 4000 levels, but 40% of the subjects are in group 1 then 3999 of the levels will be represented sparsely in the variance matrix. A single logical value of F is equivalent to setting sparse[1] to infinity.
rescale
scale any user supplied variance matrices so as to have a diagonal of 1.0.
pdcheck
verify that any user-supplied variance matrix is positive definite (SPD). It has been observed that IBD matrices produced by some software are not strictly SPD. Sometimes models with these matrices still work (throughout the iteration path, the weighted sum of variance matrices was always SPD) and sometimes they don't. In the latter case, messages about taking the log of negative numbers will occur, and the results of the fit are not necessarily trustworthy.
refine.n
the number of iterations to be used for iterative refinement of the Laplace approximation. This is a monte-carlo estimation of the error between the true integrated likelihood and the approximate one. Both the correction and the estimated variance of the correction are returned.
x
retain the X matrix in the output.
y
retain the dependent variable (a Surv object) in the output.

VALUE:

an object of class "coxme".

SEE ALSO:

.

EXAMPLES:

coxme(Surv(time, status) ~ rx, data=rats, random= ~1|litter)