White standard errors

[kt_uwa1990]

Hi

I am trying to get white standard errors test in my logistic regression. I read a few articles on the internet and came up with the following code:

proc logistic data = LOG;

model log = lsize lprice lsales roa ato chscaled re_ta re_te bmratio indrev TAX86 TAX03/ acov;

run;

However this "/ acov" does not work. Any idea what I'm doing wrong?

[art297]

I think that would be correct if you were using proc reg.  Take a look at: http://pages.stern.nyu.edu/~adesouza/comp/sas.html

[kt_uwa1990]

These are all for PROC reg how do I what to use for the logistic code?

[kt_uwa1990]

This seems to work:

data log;

   set log;

   counter=_n_;

run;

proc surveyreg data=log;

   cluster counter;

   model log = lsize lprice lsales roa ato chscaled re_ta re_te bmratio indrev;

run;

Just checking that is this appropriate for the logistic model.?

[SteveDenham]

If the variable "log" is a binary variable, then you may want to look at PROC SURVEYLOGISTIC.  If it is a continuous variable, then the SURVEYREG code will be a good start.  Be aware, though, that it is NOT a logistic model.

Steve Denham

[kt_uwa1990]

This seems to work, just to confirm  my sample size is large, so do I put counter as my cluster variable?.

data log;

   set log;

   counter= year;

run;

PROC SURVEYLOGISTIC data=log;

   cluster counter;

   model log = lsize lprice lsales roa ato escaled re_ta re_te bmratio indrev;

run;

[SteveDenham]

Without knowing the data structure and sampling frame, I can't give a full answer.  It does seem like a logical choice for the cluster variable, however.

Steve Denham

[jpeipert]

Hi,

Just to confirm, is it correct then that proc surveyreg with a cluster variable gives White SE's? I can't find anywhere in the documentation for this procedure which actual statistic is used.

Thanks!

Devin Peipert

[SteveDenham]

The default variance estimation is a Taylor series expansion.  I don't think this meets the definition of White SE's.  The other methods are resampling (jackknife delete one) and Fay's method for balanced replication.  So far as I can tell, none of these are robust estimators of the type described by White.  However, I would trust the jackknife estimator more than I would a White estimator, as there are weaker assumptions, especially regarding independence.

Steve Denham

[jpeipert]

Thanks for your quick response, Steve! That is very helpful.

Regards,

Devin Peipert