st: RE: AW: AW: GLM family and link (default)

["Nick Cox" <n.j.cox@durham.ac.uk>]

Very loosely, the kinship between the Gaussian and the binomial and the
possible merits of linear approximations to logit and probit lead one to
expect broadly similar performance in many cases. 

It's also true that a wrong model (meaning, arguably inappropriate) can
seem to fit better than a better model, at least in some respects: 

sysuse auto 
glm foreign weight
glm foreign weight, link(logit) f(binomial) 

Nick 
n.j.cox@durham.ac.uk 

Martin Weiss

I should also mention -cloglog- in this context:

*************
sysuse auto, clear
cloglog foreign length weight, nolog
glm foreign length weight, family(binomial 1) link(cloglog) nolog
*************

Martin Weiss

" Actually
this seems to work better than the probit command."

"Work better" is not an expression that conveys much to me. In which
respect
did it work better?

Note you can replicate the linear probability model, -probit- and
-logit-
via -glm-:

*************
sysuse auto, clear
reg foreign length weight
glm foreign length weight, family(gaussian) link(identity)  nolog

prob foreign length weight, nolog
glm foreign length weight, family(binomial 1) link(probit) nolog

logit foreign length weight, nolog
glm foreign length weight, family(binomial 1) link(logit) nolog
*************

mmolina@uniroma3.it

Looking at the glm help I found that the distribution of the dependent
variable -by default- is family(gaussian).

I am working with glm command, I did not specify any specific type of
family or link function, and I have a binary dependent variable.
Actually
this seems to work better than the probit command.

As I don't have continuous Gaussian responses but binary ones, which
should be the distribution family and link function underlying this
command?

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/