A list of parameters for controlling the fitting process.
lqmControl(method = "gs1", loop_tol_ll = 1e-5, loop_tol_theta = 1e-3,
check_theta = FALSE, loop_step = NULL, beta = 0.5, gamma = 1.25,
reset_step = FALSE, loop_max_iter = 1000, smooth = FALSE,
omicron = 0.001, verbose = FALSE)
character vector that specifies which code to use for carrying out the gradient search algorithm: "gs1" (default) based on C code and "gs2" based on R code. Method "gs3" uses a smoothed loss function. See details.
tolerance expressed as relative change of the log-likelihood.
tolerance expressed as relative change of the estimates.
logical flag. If TRUE
the algorithm performs a check on the change in the estimates in addition to the likelihood.
step size (default standard deviation of response).
decreasing step factor for line search (0,1).
nondecreasing step factor for line search (>= 1).
logical flag. If TRUE
the step size is re-setted to the initial value at each iteration.
maximum number of iterations.
logical flag. If TRUE
the standard loss function is replaced with a smooth approximation.
small constant for smoothing the loss function when using smooth = TRUE
. See details.
logical flag.
a list of control parameters.
The methods "gs1" and "gs2" implement the same algorithm (Bottai et al, 2015). The former is based on C code, the latter on R code. While the C code is faster, the R code seems to be more efficient in handling large datasets. For method "gs2", it is possible to replace the classical non-differentiable loss function with a smooth version (Chen, 2007).
Bottai M, Orsini N, Geraci M (2015). A Gradient Search Maximization Algorithm for the Asymmetric Laplace Likelihood, Journal of Statistical Computation and Simulation, 85(10), 1919-1925.
Chen C (2007). A finite smoothing algorithm for quantile regression. Journal of Computational and Graphical Statistics, 16(1), 136-164.