XploRe Help : glmest

Keywords - Function groups - @ A B C D E F G H I J K L M N O P Q R S T U V W X Y Z

Library:	glm
See also:	glmopt glminit glmcore glmstat glmest doglm

Quantlet:		glmest
Description:		glmest fits a generalized linear model E[y\|x] = G(x*b). This routine offers a convenient interface for GLM estimation. A check of the data is performed.

Reference(s):

McCullagh/Nelder, Generalized Linear Models, 1989.

Link:

Tutorial: GLM in XploRe

Usage:	`myfit = glmest(code,x,y{,opt})`
Input:
	`code`	text string, the short code for the model (e.g. "bilo" for logit or "noid" for ordinary PLM).
	`x`	n x p matrix, the predictor variables.
	`y`	n x 1 vector, the response variables.
	`opt`	optional, a list with optional input. "glmopt" can be used to set up this parameter. The order of the list elements is not important.
	`opt.weights`	string, type of weights. Can be "frequency" for replication counts, or "prior" (default) for prior weights in weighted regression.
	`opt.wx`	scalar or n x 1 vector, frequency or prior weights. If not given, set to 1.
	`opt.off`	scalar or n x 1 vector, offset. Can be used for constrained estimation. If not given, set to 0.
	`opt.shf`	integer, if exists and =1, some output is produced which indicates how the iteration is going on.
	`opt.norepl`	integer, if exists and =1, the data are assumed to have no replications in x. Otherwise, the data are searched for replications to fasten the algorithm.
	`opt.miter`	integer, maximal number of iterations. The default is 10.
	`opt.cnv`	scalar, convergence criterion. The default is 0.0001.
	`opt.fscor`	integer, if exists and =1, a Fisher scoring is performed (instead of the default Newton-Raphson procedure). This parameter is ignored for canonical links.
	`opt.pow`	scalar, power for power link. If not given, set to 0.
	`opt.nbk`	scalar, extra parameter k for negative binomial distribution. If not given, set to 1 (geometric distribution).
Output:
	`myfit.b`	p x 1 vector, estimated coefficients.
	`myfit.bv`	p x p matrix, estimated covariance matrix for coefficients.
	`myfit.stat`	list with components as computed by glmstat: serror (standard errors of coefficients), tvalue (t-values for coefficients), pvalue (p-values for coefficients), df (degrees of freedom), deviance (deviance), pearson (generalized pearson's chi^2), loglik (log-likelihood), dispersion (estimated dispersion =pearson/df), r2 ((pseudo) coefficient of determination), adr2 (adjusted (pseudo) coefficient of determination), aic (Akaike's AIC criterion), bic (Schwarz' BIC criterion), and it (number of iterations needed), ret (return code, 0 if everything went o.k., 1 if maximal number of iterations reached, -1 if missing values have been encountered), nr (number of replications found in x).

Example:

library("glm") ;========================== ; simulate data ;========================== n=100 b=1|2 p=rows(b) x=2.*uniform(n,p)-1 y=( 1./(1+exp(-x*b)).>uniform(n) ) ;========================== ; GLM fit ;========================== opt=glmopt("shf",1) lf=glmest("bilo",x,y,opt) b~lf.b lf.stat.deviance

Result:

A logit fit for E[y|x] is computed. The option "shf" for
information during the iteration is set. lf.b contains the
coefficients for the linear part. The example gives the
true b together with the GLM estimate lf.b. The deviance
of the fitted model is shown as well.

Author: M. Mueller, 20010228