Stata The Stata listserver
[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

Re: st: [Stata8] Model building and stepwise

Subject   Re: st: [Stata8] Model building and stepwise
Date   Thu, 28 Apr 2005 23:27:59 EDT

It seems to me much more sensical to use Stata's -glm- command when  modeling 
grouped binomial data; e.g. grouped logistic regression.  You  can model:
 glm y x1 x2, fam(bin denom) eform  
to immediately obtain grouped logistic results parameterized as odds  ratios. 
-demon- is the binomial denominator. You also get AIC, BIC, and other  fit 
statsitics on the screen, plus you can request a number of different types  of 
robust variance estimators. You can also bootstrap or jacknife standard  errors 
-- and use -sw- for stepwise methods. Moreover, you can then reuqest  a 
number of residuals including the standards ones and Anscombe, likelihood  
residuals, etc. The Chi2 and deviance residuals come in raw, standardized  or 
studentized forms. Few of these statistics are available with glogit. 
Joe Hilbe
Ethan Corona wrote:

As part of an epidemiology class project, I have  to employ stepwise for
model selection. I can do that with no problems. My  problem lies with the
lack of output that -stepwise- generates compared to  the output that SAS
provides. Short of performing stepwise by hand, are there  any model building
ADOs that give more detailed output of each intermediate  model?
Specifically, I'm looking for type I and type III sum of squares as  well as
regression output for each intermediate model.

I looked at  -allpossible-, and while it provides all of the intermediate
models using the  -details- information, it doesn't provide the sum of
squares that I'm looking  for.

. . . 

Hello again,

I realized that I failed to  mention that I am doing logistic regression, so
-reg_ss- won't help me,  either.


What  do Type I and Type III sum of squares look like for logistic regression 

Is SAS using weighted least squares estimation on grouped  data, as in 
Nick Cox mentioned -allpossible-; you should be able  to do something with 
-allpossible glogit- after construction of the  indicator variables to get 
contrasts that you're looking for.  The  technique for constructing the 
difference contrast indicator variable to get  -regress- to match -anova , 
partial- (Type III sums of squares) is shown in  Sophia Rabe-Hesketh and 
Everitt, _A Handbook of Statistical Analyses  using Stata_ Second Edition. 
Raton: Chapman & Hall/CRC, 2000), pp.  73-75. The use of the technique with 
-glogit- will be directly analogous to  that with -regress-.  You can also 
-logit- with the same indicator  variables to get an analogous contrast to 
in -anova, partial-.  I  think that the SAS manuals variously call this a 
"Type 3 test" or "Type III  test," or something, outside the contexts of PROC 

Joseph Coveney

*   For searches and help  try:
*   For searches and help try:

© Copyright 1996–2017 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index