Re: st: Re: Error in stata manual on xtdpd?

["Brian P. Poi" <bpoi@stata.com>]

On Wed, 7 Oct 2009, Hewan Belay wrote:

> Dear Brian,
>
> Thanks a lot for your response. I also have another question about the 
> manual--not so much an error but something I think would be very 
> impoortant to include. I understand that one of the critical advantages 
> of the system GMM dynamic panel data estimator 
> (Arellano-Bover/Blundell-Bond) over the difference GMM estimator 
> (Arellano-Bond) is that the former is able to identify the effects of 
> time-invariant explanatory variables, which the latter can not. It is 
> not mentioned anywhere in the description of xtdpd or xtdpdsys how to do 
> this. After asking around, I found out that the additional command 
> -hascons- in the xtdpd line is the way to be able to include 
> time-invariant regressors. However, the description for -hascons- 
> doesn't say so. (However, please do let me know if the above is not a 
> correct representation of -hascons-)
>
> Many thanks!
> Hewan

-xtabond- does not allow for time-invariant regressors, because the 
Arellano-Bond estimator only uses the difference equation, and 
first-differencing will wipe out those time-invariant regressors.

-xtdpdsys- does allow for time-invariant regressors, because the 
Arellano-Bover/Blundell-Bond estimator uses both the difference equation 
and the level equation.  The coefficients for time-invariant regressors 
are identified by virtue of the fact that those variables still appear in 
the level equation.

> -hascons- in the xtdpd line is the way to be able to include 
> time-invariant regressors. However, the description for -hascons-

Yes, 'hascons' is needed with -xtdpd- to include time-invariant 
regressors.  To wit, you can see that

   . webuse abdata
   // make kbar time-invariant
   . bysort id: gen kbar = sum(k)/_N
   . by id: replace kbar = kbar[_N]
   . xtdpdsys n w kbar, lags(1)

and

   . xtdpd n L.n w kbar, div(w kbar) dgmmiv(n, lag(2 .))     ///
                lgmmiv(n, lag(1)) hascons

produce the same results.  Without the 'hascons' option, -xtdpd-, like 
-xtabond-, would eliminate kbar from the model because D.kbar would be 
identically zero in the difference equation and hence collinear with the 
constant term.

Generally, the best way to learn -xtdpd- is to first fit a simpler model 
using -xtabond- or -xtdpdsys-, then replicate that model using -xtdpd-. 
The instrument summary at the bottom of the output of the first two 
commands greatly helps in replicating results using -xtdpd-.  Then once 
you're sure a simpler form of your model is correctly specified with 
-xtdpd-, you can go ahead and modify the model with -xtdpd-.  -xtabond- 
and -xtdpdsys- are 'convenience' commands implemented on top of -xtdpd- 
that make fitting the most common DPD models easier.

  -- Brian Poi
  -- bpoi@stata.com

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/