qrminfundtheorem: Minimizes the negative of log-likelihood in the OR1 model

In bqror: Bayesian Quantile Regression for Ordinal Models

Minimizes the negative of log-likelihood in the OR1 model

Description

This function minimizes the negative of log-likelihood in the OR1 model with respect to the cut-points \delta using the fundamental theorem of calculus.

Usage

qrminfundtheorem(deltaIn, y, x, beta, cri0, cri1, stepsize, maxiter, h, dh, sw, p)

Arguments

deltaIn	initialization of cut-points.
y	observed ordinal outcomes, column vector of size (n x 1).
x	covariate matrix of size (n x k) including a column of ones with or without column names.
beta	\beta, a column vector of size (k x 1).
cri0	initial criterion, cri0 = 1.
cri1	criterion lies between (0.001 to 0.0001).
stepsize	learning rate lies between (0.1, 1).
maxiter	maximum number of iteration.
h	change in each value of \delta, holding other \delta constant for first derivatives.
dh	change in each value of \delta, holding other \delta constant for second derivaties.
sw	iteration to switch from BHHH to inv(-H) algorithm.
p	quantile level or skewness parameter, p in (0,1).

Details

First derivative from first principle

dy/dx=[f(x+h)-f(x-h)]/2h

Second derivative from first principle

f'(x-h)=(f(x)-f(x-h))/h

f''(x)= [{(f(x+h)-f(x))/h} - (f(x)-f(x-h))/h]/h

= [(f(x+h)+f(x-h)-2 f(x))]/h^2

cross partial derivatives

f(x) = [f(x+dh,y)-f(x-dh,y)]/2dh

f(x,y)=[{(f(x+dh,y+dh) - f(x+dh,y-dh))/2dh} - {(f(x-dh,y+dh) - f(x-dh,y-dh))/2dh}]/2dh

= 0.25* [{(f(x+dh,y+dh)-f(x+dh,y-dh))} -{(f(x-dh,y+dh)-f(x-dh,y-dh))}]/dh2

Value

Returns a list with components

deltamin:	cutpoint vector that minimizes the log-likelihood function.
negsum:	negative sum of log-likelihood.
logl:	log-likelihood values.
G:	gradient vector, (n x k) matrix with i-th row as the score for the i-th unit.
H:	Hessian matrix.

See Also

differential calculus, functional maximization, mldivide

Examples

set.seed(101)
deltaIn <- c(-0.002570995,  1.044481071)
data("data25j4")
y <- data25j4$y
xMat <- data25j4$x
p <- 0.25
beta <- c(0.3990094, 0.8168991, 2.8034963)
cri0     <- 1
cri1     <- 0.001
stepsize <- 1
maxiter  <- 10
h        <- 0.002
dh       <- 0.0002
sw       <- 20
output <- qrminfundtheorem(deltaIn, y, xMat, beta, cri0, cri1, stepsize, maxiter, h, dh, sw, p)

# deltamin
#   0.8266967 0.3635708
# negsum
#   645.4911
# logl
#    -0.7136999
#    -1.5340787
#    -1.1072447
#    -1.4423124
#    -1.3944677
#    -0.7941271
#    -1.6544072
#    -0.3246632
#    -1.8582422
#    -0.9220822
#    -2.1117739 .. soon
# G
#    0.803892784  0.00000000
#   -0.420190546  0.72908381
#   -0.421776117  0.72908341
#   -0.421776117 -0.60184063
#   -0.421776117 -0.60184063
#    0.151489598  0.86175120
#    0.296995920  0.96329114
#   -0.421776117  0.72908341
#   -0.340103190 -0.48530164
#    0.000000000  0.00000000
#   -0.421776117 -0.60184063.. soon
# H
#   -338.21243  -41.10775
#   -41.10775 -106.32758

bqror documentation built on April 3, 2025, 9:36 p.m.