gbnlprobit1

Non linear probit regression

Usage

gbnlprobit [options] <function definition>

Overview

Non linear probit estimation. Minimize the negative log-likelihood

: sum_{i in N_0} log(1-F(g(X_i))) + sum_{i in N_1} log(F(g(X_i)))

where N_0 and N_1 are the sets of 0 and 1 observations, g is a generic function of the independent variables and F is the normal CDF. It is also possible to minimize the score function

w_0 sum_{i in N_0} theta(F(g(X_i))-t) +

: w_1 sum_{i in N_1} theta(t-F(g(X_i)))

where theta is the Heaviside function and t a threshold level. Weights w_0 and w_1 scale the contribution of the two subpopulations. The first column of data contains 0/1 entries. Successive columns are independent variables. The model is specified by a function g(x1,x2...) where x1,.. stands for the first,second .. N-th column independent variables.

options:

-O: type of output (default 0)

`0`	parameters
`1`	parameters and errors
`2`	<variables> and probabilities
`3`	parameters and variance matrix
`4`	marginal effects

-V: variance matrix estimation (default 0)

`0`	<gradF gradF^t>
`1`	< J^{-1} >
`2`	< H^{-1} >
`3`	< H^{-1} J H^{-1} >

-z: take zscore (not of 0/1 dummies)

-F: input fields separators (default " \t")

-v: verbosity level (default 0)

`0`	just results
`1`	comment headers
`2`	summary statistics
`3`	covariance matrix
`4`	minimization steps (default 10)
`5`	model definition

-g: set number of point for global optimal threshold identification

-h: this help

-t: set threshold value (default 0)

`0`	ignore threshold
`(0,1)`	user provided threshold
`1`	compute optimal only global
`2`	compute optimal

-M: estimation method

`0`	maximum likelihood
`1`	min. score (w0=w1=1)
`2`	min. score (w0=1/N0, w1=1/N1)

-A: MLL optimization options (default 0.01,0.1,100,1e-6,1e-6,5) fields are step,tol,iter,eps,msize,algo. Empty fields for default

step: initial step size of the searching algorithm

tol: line search tolerance iter: maximum number of iterations

eps: gradient tolerance : stopping criteria ||gradient||<eps

algo: optimization methods: 0 Fletcher-Reeves, 1 Polak-Ribiere, 2 Broyden-Fletcher-Goldfarb-Shanno, 3 Steepest descent, 4 simplex

-B: score optimization options (default 0.1,100,1e-6) fields are step,iter,msize. Empty fields for default

step: initial step size of the searching algorithm

iter: maximum number of iterations

msize: max size, stopping criteria simplex dim. <max size optimization method is simplex

Releases

5 releases since Debian 9 (Stretch); the text changed 3 times.