Joana Quina <[email protected]> notes that the snippet of code that was
included in my earlier post about the Hausman test will not work if any of the
covariates are constant within panels,
> I used your code for the augmented Hausman test, but whenever I
> include time-invariant variables (or time dummies interacted with
> time-invariant variables), it does note work - Stata drops the mean
> for the time-invariant variable (or the diff for the interacted
> terms). [...]
Such variables are dropped from the fixed effects (FE) regression because they
are collinear with the effects and the code was not up to that. Eric's models
from some 3 years ago did not have this problem, but Joana's models do, and
such variables are quite common.
To make the code resistant to such variables, we just have to make sure that
they are not tested. We could do that by wrapping the tests in an additional
-if- that checks for the existence of differenced and mean variables in the
estimation results,
if colnumb(`b', "mean`i'") < . & colnumb(`b', "diff`i'") < . {
...
}
The final section of code then becomes,
------------------------------------ BEGIN code snippet -------------------
quietly test mean1 = diff1 , notest /* clear test */
local i 2
while "``i''" != "" {
if colnumb(`b', "mean`i'") < . & colnumb(`b', "diff`i'") < . {
if `b'[1,colnumb(`b', "mean`i'")] != 0 & /*
*/ `b'[1,colnumb(`b', "diff`i'")] != 0 {
quietly test mean`i' = diff`i' , accum notest
}
}
local i = `i' + 1
}
test
------------------------------------ END code snippet -------------------
Note that because we initialize the testing loop with
-quietly test mean1 = diff1 , notest-
the first variable in our <varlist> cannot be constant within panel.
Having said that, I like the version of the artificial (augmented) regression
that Mark Schaffer <[email protected]> and Carl Nelson <[email protected]>
use better that my own. Adapting theirs to loop over a <varlist> we get,
---------------------------------- BEGIN --- artreg0.do --- CUT HERE -------
local id idcode
local depvar ln_wage
local varlist ttl_exp union white
iis `id'
local i 0
foreach var of varlist `varlist' {
egen mymean`++i' = mean(`var'), by(`id')
}
xtreg `depvar' `varlist' mymean1-mymean`i' , re
testparm mymean1-mymean`i'
---------------------------------- END --- artreg0.do --- CUT HERE -------
It is much shorter and will work with variables that are constant within
panels. Again, replace -idcode-, -ln_wage-, and -ttl_exp union white- with
your own id variable, dependent variable, and independent variables.
Also note, that both of these snippets assume that you do not have any missing
values in you covariates, dependent variable, or id variable. If Joana has
missing data, she should take care to compute the means only over the
estimation sample. A better version, and one that does this is,
---------------------------------- BEGIN --- artreg.do --- CUT HERE -------
local id idcode
local depvar ln_wage
local varlist ttl_exp union white
iis `id'
xtreg `depvar' `varlist' , re
local i 0
foreach var of varlist `varlist' {
egen mymean`++i' = mean(`var'), by(`id')
}
xtreg `depvar' `varlist' mymean1-mymean`i' , re
testparm mymean1-mymean`i'
---------------------------------- END --- artreg.do --- CUT HERE -------
This is a different form of artificial regression than the one from my first
post and will differ slightly with small samples, but they are asymptotically
equivalent.
-- Vince
[email protected]
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
