emfrail_control function

Control parameters for emfrail

emfrail_control(opt_fit = TRUE, se = TRUE, se_adj = TRUE,
  ca_test = TRUE, lik_ci = TRUE, lik_interval = exp(c(-3, 20)),
  lik_interval_stable = exp(c(0, 20)), nlm_control = list(stepmax = 1),
  zph = FALSE, zph_transform = "km", em_control = list(eps = 1e-04,
  maxit = Inf, fast_fit = TRUE, verbose = FALSE, upper_tol = exp(10),
  lik_tol = 1))

Arguments

• opt_fit: Logical. Whether the outer optimization should be carried out. If FALSE, then the frailty parameter is treated as fixed and the emfrail function returns only log-likelihood. See details.
• se: Logical. Whether to calculate the variance / covariance matrix.
• se_adj: Logical. Whether to calculate the adjusted variance / covariance matrix (needs se == TRUE)
• ca_test: Logical. Should the Commenges-Andersen test be calculated?
• lik_ci: Logical. Should likelihood-based confidence interval be calculated for the frailty parameter?
• lik_interval: The edges, on the scale of $\theta$, of the parameter space in which to search for likelihood-based confidence interval
• lik_interval_stable: (for dist = "stable") The edges, on the scale of $\theta$, of the parameter space in which to search for likelihood-based confidence interval
• nlm_control: A list of named arguments to be sent to nlm for the outer optimization.
• zph: Logical. Should the cox.zph test be performed at the maximum likelihood estimate?
• zph_transform: One of "km", "rank", "identity" or a function of one argument to be pased on to cox.zph.
• em_control: A list of parameters for the inner optimization. See details.

Returns

An object of the type emfrail_control.

Details

The nlm_control argument should not overalp with hessian, f or p.

The em_control argument should be a list with the following items:

• epsA criterion for convergence of the EM algorithm (difference between two consecutive values of the log-likelihood)
• maxitThe maximum number of iterations between the E step and the M step
• fast_fitLogical, whether the closed form formulas should be used for the E step when available
• verboseLogical, whether details of the optimization should be printed
• upper_tolAn upper bound for $\theta$; after this treshold, the algorithm returns the limiting log-likelihood of the no-frailty model. That is because the no-frailty scenario corresponds to a $\theta = \infty$, which could lead to some numerical issues
• lik_tolFor values higher than this, the algorithm returns a warning when the log-likelihood decreases between EM steps. Technically, this should not happen, but if the parameter $\theta$ is somewhere really far from the maximum, numerical problems might lead in very small likelihood decreases.

The fast_fit option make a difference when the distribution is gamma (with or without left truncation) or inverse Gaussian, i.e. pvf with m = -1/2 (without left truncation). For all the other scenarios, the fast_fit option will automatically be changed to FALSE. When the number of events in a cluster / individual is not very small, the cases for which fast fitting is available will show an improvement in performance.

The starting value of the outer optimization may be set in the distribution argument.

Examples

emfrail_control()
emfrail_control(em_control = list(eps = 1e-7))

See Also

emfrail, emfrail_dist, emfrail_pll