Statalist The Stata Listserver


[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

Re: st: Fixed effects vs. Random effects


From   "Rodrigo A. Alfaro" <ralfaro76@hotmail.com>
To   <statalist@hsphsun2.harvard.edu>
Subject   Re: st: Fixed effects vs. Random effects
Date   Sat, 22 Jul 2006 13:40:38 -0400

Sorry to enter very late in the discussion, but let's make a difference
between Conditional Logit and FE-logit. For the first you need
changes in your dependent variable to identify the effect for a particular
cross-sectional unit. Otherwise the unit will be remove it from the 
analysis.
You say that yout T=3, it is very short but I suggest you to chech the
changes over time -table id year, c(N depvar). Also running -clogit- or
-xtlogit, fe- you will see this at the begin of the output.

FE-logit is directly logit+dummies. Incidental parameters generate a bias
in the estimation that can be adjusted by different methods (see Arellano
and Hahn 2005 "Understanding..." in Manuel Arellano's webpage).
Without the adjustment the bias can be huge (see for example in
William Greene's webpage "The bias of fixed-effects estimator in
nonlinear models" Table 2).

I prefer FE instead of RE. RE has the assumption of no-correlation with
other regressor which is very hard to believe. In addition, if your still 
have
some theoretical reason to do RE don't forget to check the quadrature.
If the model changes with the quadrature the approximation of very poor
and you can use the RE estimation.

Rodrigo.



----- Original Message ----- 
From: "Justin Smith" <smithjd5@univmail.cis.mcmaster.ca>
To: <statalist@hsphsun2.harvard.edu>
Sent: Friday, July 21, 2006 12:43 PM
Subject: Re: st: Fixed effects vs. Random effects


What i meant to say there is "there is NO magic number on the smallest
number of panels you would need for that"

Justin


On Fri, 21 Jul 2006 12:42:06 -0400
 Justin Smith <smithjd5@univmail.cis.mcmaster.ca> wrote:
> I don't think there is a minimum.  Since it is a within firm (in your
> case) regression, you could technically estimate a fixed effects
> model
> with one panel, as long as there is 2 or more time observations
> within
> that panel.
>
> Of course, you want a large number of panels for precise estimation
> of
> the parameters, but there is any magic number on the smallest number
> of
> panels you would need for that.  The more the better.  The size of
> your
> dataset seems plenty large enough to run the model you propose.
>
> Justin Smith
> smithjd5@mcmaster.ca
>
>
> On Fri, 21 Jul 2006 07:56:27 -0700 (PDT)
>  ILR School <ilrschool@yahoo.com> wrote:
> > Does anyone know where I can find out the minimum
> > number of panels one needs to run a fixed-effects
> > model?  I have a dataset of 800 firms with only 3 time
> > periods using a xtlogit regression model. I am
> > currently using a random effects model.
> >
> > Thank you,
> > Shon R.
> >
> >
> > __________________________________________________
> > Do You Yahoo!?
> > Tired of spam?  Yahoo! Mail has the best spam protection around
> > http://mail.yahoo.com
> > *
> > *   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/
>
> Justin Smith
> PhD Candidate
> Department of Economics
> McMaster University
> Phone: (905) 962-0353
> E-mail: smithjd5@mcmaster.ca
> *
> *   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/

Justin Smith
PhD Candidate
Department of Economics
McMaster University
Phone: (905) 962-0353
E-mail: smithjd5@mcmaster.ca
*
*   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/
*
*   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/



© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index