Dear Mark,
thank you for your advise and the references. Wooldridge also cites the Verbeek and
Nijman (1992) paper, and actually, when I was writting my question I
had already tested whether the attriers characteristics are different
from that ones who stay in the sample. From this data analysis I
concluded that attrition is non ignorable... moreover, variable coding
whether person stays or drops out has highly significant coefficient
(which is, as I understand in the right way, more ore less what Verbeek
and Nijman had proposed as a test of "attrition ignorability")
so, better, my questions should be reformulated in another way...
if I know (checked by tests) that I have "non ignorable attrition"
and after some test (for example Wu-Hausman) I concluded that one of
the variables in my dataset is not exogenous...
WHAT IS THE FIRST STEP TO DO?
1) to do xtivreg2, trying to correct for endogeneity bias...and than
to try to test for the attrition bias (if so, should I put the
"stay-next-period"variable into the list of exogenous regressors???
(if I assume it could be related to the dependent variable of
interest, for example, through some mediating variable??? it is not
so exogenous..!!)
2) or I correct for the attrition bias firstly and than start the
instruments searching exercises to correct for the endogeneity of
that "bad" variable
THREE MORE QUESTIONS related to IV estimation (and particularly, by
xtivreg2):
I want to estimate -xtivreg2....,fe-
1) would it look reasonable to use as instruments lags of some of the
exogenous variables included into the model?
2) what if I guess that the dependent variable lag should be included
into the list of rhs variables.. (as I understand, -xtivreg2- allows
for including lags of the dependent variable into the model, yes??)
Ok,.. let say, I managed to estimates two models, one with the lag
(which is highly significant) and other without it... they both look
ok (and what whould, actually, tell me that they are "ok"???)
,.. is there some formal test to choose between "with" and "without
log" specifications? or I'm complitely wrong
with my startegy?
3) I have found your correspondence of some year ago with Susie
http://www.stata.com/statalist/archive/2006-05/msg01071.html
where more or less the order of actions when testing instruments is
discussed... Could you, please, correct me, whether I hve understood
everything in the right way...
Let's say, I've chosen a set of instruments for my ONLY endogenous
variable x... I checked whether these instruments are significant in my
"original" regression and they are not ..
than I do -xtivreg2..., fe gmm robust cluster(id)ffirst endog(x)-
than I see
- each instrument is individually significant in the first stage
regression (, and F-test of excluded instruments is significant,
moreover, F-statistics is much much greater than 14 for my case of
three excluded instruments .. I even conclude that my instruments
are not weak
- than all the three underidentificantion tests rejects the null
- as well as weak identificantio tests
- Weak instrument robust inference tests also reject the null
- Hansen J statistics doesn't reject the null
and now THE PROBLEM... I'm complitely lost whether it is Ok or not if
the -endog- option provided Endogeneity test should reject or accept
the null??
my observations were that when I was using "bad" instruments, which
were not significant in first stage, or were rejecting the null in
Hansen J test... this endogeneity test was strongly rejecting the H0
of exogeneity...
I've concluded that this test shows whether my endogenous x variable
is instrumentilized so well to pass for one exogenous one or not...
Am I right?
I would be very very grateful for any clarification,
thank you for you time again,
Ekaterina
> Ekaterina,
> Baltagi's book on panel data has a short but useful discussion of
> selection bias in panel data models. He cites Veerbeck and Nijman
> (1992, International Economic Review) on "ignorable" and "nonignorable"
> selection rules. In a nutshell, if you're lucky, you might have the
> "ignorable" case, in which case standard panel estimation methods are
> fine.
>> -----Original Message-----
>> From: [email protected]
>> [mailto:[email protected]] On Behalf Of
>> Ekaterina Selezneva
>> Sent: 24 May 2007 19:57
>> To: [email protected]
>> Subject: st: -xtivreg-, endogeneity and attrition
>>
>> > Just a couple of things to add to what Uli said above.
>>
>> > - If I'm not mistaken, dropping the singletons (one-observation
>> > groups) will affect the estimate of the constant. This is because
>> > Stata's xtreg,fe uses between-group variation to get the
>> constant (I
>> > find this approach a little odd, but maybe it's just me).
>>
>> > - If you want fixed effects estimates that drop singletons, and you
>> > don't need an estimate of the constant, you can use -xtivreg2-. As
>> > well as panel IV, -xtivreg2- will do fixed effects
>> estimates for the
>> > case where all regressors are exogenous.
>>
>> > Cheers,
>> > Mark
>>
>> Dear Mark,
>> thanks a lot for Your answer.
>>
>> Could I ask, please, for some other piece of advise?
>>
>> If I understand it in the right way, xtivreg2 is an
>> extension of xtivreg... the former giving access to wider
>> diagnostic techniques..
>>
>> Than, I've read that xtivreg "handles exogenously unbalanced
>> panel data"...
>>
>> In my model I have some variables that are suspected to be
>> endogenous.. moreover, analyzing attiers I've found that they
>> have characteristics different from those left in my panel..
>> so, attrition is also an issue to correct for..
> Hope this helps.
> Cheers,
> Mark
> Prof. Mark Schaffer
> Director, CERT
> Department of Economics
> School of Management & Languages
> Heriot-Watt University, Edinburgh EH14 4AS
> tel +44-131-451-3494 / fax +44-131-451-3296
> email: [email protected]
> web: http://www.sml.hw.ac.uk/ecomes
--
Best regards,
Ekaterina mailto:[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/
