nlqmmControl: Control parameters for nlqmm estimation

In marco-geraci/nlqmm: Nonlinear Quantile Mixed Models

Description

A list of parameters for controlling the fitting process.

Usage

nlqmmControl(method = "Nelder-Mead", LL_tol = 1e-5, beta = 0.5,
	max_iter = 500,	analytic = FALSE, REoptimizer = "nlm",
	REcontrol = list(),	initialize = "nlme", verbose = FALSE)

Arguments

method	character vector that specifies the optimization algorithm in optim to maximize the log-likelihood. Only "Nelder-Mead" and "BFGS" allowed.
LL_tol	tolerance expressed as absolute change of the log-likelihood.
beta	decreasing step factor for smoothing parameter omega. See details.
max_iter	maximum number of iterations.
analytic	logical flag. If TRUE, the gradient and Hessian for computing the modal random effects with nlm are calculated analytically. Otherwise, they are computed with numerical differentiation (default).
REoptimizer	optimizer for the modal random effects. The options are "optim" for BFGS in optim or "nlm" (default) for Newton-type minimization in nlm.
REcontrol	a list of arguments to be passed to the optimizer for the modal random effects. See arguments in optim and nlm.
initialize	character specifying what algorithm should be used to initialize all nlqmm parameters. The options are "nls" for nls, "nlrq" for nlrq, and "nlme" (default) for nlme. If NULL, naive estimates are used as initial values.
verbose	logical flag.

Details

The parameters are initialized using one of "nls", "nlrq", or "nlme". These algorithms are started with the values given via the argument start in nlqmm. Only "nlme" provides estimates of variance-covariance parameters and random effects. Therefore, when using the other algorithms, these parameters are initialized with a naive estimate.

The parameter omega controls the quadratic approximation of the absolute deviation function at the kink 0 (Chen, 2007). In nlqmm, the starting value for omega is determined automatically and is not (currently) under the user's control. At each iteration, omega is decreased by a factor beta. The smaller omega at convergence, the smaller the approximation error. See details of the algorithm in Geraci (2017).

Value

a list of control parameters.

Author(s)

Marco Geraci

References

Chen C. (2007). A finite smoothing algorithm for quantile regression. Journal of Computational and Graphical Statistics, 16(1), 136-164.

Geraci M (2019). Modelling and estimation of nonlinear quantile regression with clustered data. Computational Statistics and Data Analysis, 136, 30-46.

See Also

nlqmm

marco-geraci/nlqmm documentation built on April 16, 2020, 12:37 a.m.