coxinterval.control: Control model fit
In aboruvka/coxinterval: Cox-Type Models for Interval-Censored Data

Description Usage Arguments Details Value References See Also Examples

Set parameters controlling the model fit.

coxinterval.control(eps = 1e-7, iter.max = 5000, coef.typ = 1,
                    coef.max = 10, return.data = FALSE,
                    eps.norm = c("max", "grad"), armijo = 1/3,
                    var.coef = TRUE, trace = FALSE, thread.max = 1,
                    sieve = TRUE, sieve.const = 1, sieve.rate = 1/3,
                    risk.min = 1)

`eps`	threshold value for the norm used to measure convergence in the parameter estimates.
`iter.max`	maximum number of iterations to attempt. This ensures that the estimation routine will eventually exit, even when the convergence criteria are not met.
`coef.typ`	a scalar or vector of "typical" (absolute) values for the regression coefficient.
`coef.max`	a scalar or vector of probable upper bounds for the regression coefficient. This and the `coef.typ` arguments tune variance estimation via the curvature in the profile log-likelihood, following Boruvka and Cook (2015, Section 6).
`return.data`	a logical value indicating that the model object returned should contain an element `data` that generally gives the assigned support points and model matrix. Further details on the output is provided in the documentation for each model function.
`eps.norm`	a character string identifying the norm to use in the convergence criteria for `coxaalen`—either the maximum norm between the current and previous parameter values (`eps.norm = "max"`) or the absolute inner product between the current value and the score (`eps.norm = "grad"`).
`armijo`	a scale factor in (0, 1/2) for Armijo's (1966) rule—a line search used to ensure that each iteration in the estimation routine for `coxaalen` achieves an adequate increase in the log-likelihood. The model fit is typically not very sensitive to this value.
`var.coef`	a logical value indicating that `coxaalen` standard errors for the multiplicative regression coefficients should be estimated. This is done via profile likelihood—an approach that can require an inordinate amount of processing time under many regression coefficients and larger sample size.
`trace`	a logical value indicating that, on execution of `coxaalen`, CPLEX should print its results to the screen.
`thread.max`	maximum number of CPU threads for `coxaalen` to allocate to CPLEX. The default value disables multithreading. A value of zero allows CPLEX to set the number of threads automatically. The actual number of threads used is limited by the number of available processors and the CPLEX license.
`sieve`	a logical value indicating that the sieve rather than the semiparametric maximum likelihood estimator should be fit by `coxdual`. The default `TRUE` is recommended to avoid issues with support finding and convergence.
`sieve.const`	a constant factor that, in part, determines the sieve size. The factor can be made specific to the transition type with `sieve.const` a vector of length three. Indexing the states from zero, this vector's components correspond to the state 0 to state 1, 0 to 2, and 1 to 2 transition types, respectively.
`sieve.rate`	a scalar in (1/8, 1/2) determining the rate at which the sieve increases with the sample size.
`risk.min`	a positive integer giving the minimum size of risk set for support points defining the sieve.

For a given sample size n, the sieve for coxdual has size at most sieve.const*n^sieve.rate. Any reduction in size from this value is applied to ensure that each subinterval in the sieve's time partition captures at least one support point from the semiparametric maximum likelihood estimator based on the subsample with known progression status (Boruvka and Cook, 2016).

A list of the above arguments with their final values.

Boruvka, A. and Cook, R. J. (2015) A Cox-Aalen model for interval-censored data. Scandinavian Journal of Statistics 42, 414–426.

Boruvka, A. and Cook, R. J. (2016) Sieve estimation in a Markov illness-death process under dual censoring. Biostatistics 17, 350–363.

Armijo, L. (1966) Minimization of functions having Lipschitz continuous first partial derivatives. Pacific Journal of Mathematics 16, 1–3.

coxaalen, coxdual

if (is.loaded("coxaalen", "coxinterval")) {
  f <- Surv(left, right, type = "interval2") ~ prop(treat)
  coxaalen(f, data = cosmesis,
           control = coxinterval.control(iter.max = 2, trace = TRUE))
  coxaalen(f, data = cosmesis, iter.max = 2)
}

f <- Surv(start, stop, status) ~ cluster(id) + strata(from, to) +
     I(z * (to == 1)) + I(z * (from %in% 0 & to == 2)) +
     I(z * (from %in% c(NA, 1) & to == 2))
coxdual(f, data = dualrc,
        control = coxinterval.control(eps = 1e-5, sieve.rate = 2/5))
coxdual(f, data = dualrc, eps = 1e-5, sieve.rate = 2/5)