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: gplm
See also: gplmopt gplminit gplmcore gplmstat glmest

Macro: gplmest
Description: gplmest fits a generalized partially linear model E[y|x,t] = G(x*b + m(t)). This macro offers a convenient interface for GPLM estimation. A preparation of data is performed (inclusive sorting).

Reference(s):

Link:
Usage: myfit = gplmest(code,x,t,y,h{,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 discrete predictor variables.
t n x q matrix, the continuous predictor variables.
y n x 1 vector, the response variables.
h q x 1 vector, the bandwidth.
opt optional, a list with optional input. The macro "gplmopt" can be used to set up this parameter. The order of the list elements is not important. Parameters which are not given are replaced by defaults (see below).
opt.wx scalar or n x 1 vector, prior weights. For binomial models usually the binomial index vector. If not given, set to 1.
opt.b0 p x 1 vector, the initial coefficients. If not given, all coefficients are put =0 initially.
opt.m0 n x 1 vector, the initial values for the nonparametric part. If not given, a default is used.
opt.wt n x 1 vector, weights for t (trimming factors). If not given, all set to 1.
opt.tg ng x 1 vector, a grid for continuous part. If tg is given, the nonparametric function will also be computed on this grid.
opt.m0g ng x 1 vector, the initial values for the nonparametric part on the grid. These values are ignored if direct update for nonparametric function is possible. Otherwise, if not given, it is approximated from m0.
opt.shf integer, if exists and =1, some output is produced which indicates how the iteration is going on.
opt.nosort integer, if exists and =1, the continuous variables t and the grid tg are assumed to be sorted by the 1st column. Sorting is required by the algorithm, hence you should switch if off only when the data are already sorted.
opt.miter integer, maximal number of iterations. The default is 10.
opt.cnv integer, 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.wtc n x 1 vector, weights for convergence criterion, w.r.t. m(t) only. If not given, opt.wt is used.
opt.off scalar or n x 1 vector, offset. Can be used for constrained estimation. If not given, set to 0.
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 coeff.
myfit.m n x 1 vector, estimated nonparametric part
myfit.mg ng x 1 vector, estimated nonparametric part on grid
myfit.stat list with the following statistics:
myfit.stat.deviance deviance,
myfit.stat.pearson generalized pearson's chi^2,
myfit.stat.loglik log-likelihood,
myfit.stat.r2 pseudo R^2,
myfit.stat.it scalar, number of iterations needed.

Example:
library("gplm")
;==========================
;  simulate data 
;==========================
n=100
b=1|2
p=rows(b)
x=2.*uniform(n,p)-1
t=sort(2.*uniform(n)-1,1)
m=cos(pi.*t)
y=( 1./(1+exp(-x*b-m)).>uniform(n) )
;==========================
;  semiparametric fit 
;==========================
h=0.6
sf=gplmest("bilo",x,t,y,h)
b~sf.b
pic=createdisplay(1,1)
show(pic,1,1,t~m,t~sf.m)
Result:
A generalized partially linear logit fit for E[y|x,t] is 
computed. sf.b contains the coefficients for the linear  
part. sf.m contains the estimated nonparametric part 
evaluated at observations t. The example gives the 
true b together with the GPLM estimate sf.b. Also, the  
estimated function sf.m is displayed together with the 
true fit m. 

Library: gplm
See also: gplmopt gplminit gplmcore gplmstat 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

Author: Marlene Mueller, 970523
(C) MD*TECH Method and Data Technologies, 28.6.1999