Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down on April 23, and its replacement, statalist.org is already up and running.


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: RE: Error? xtdpdsys assigns explanatory power to fixed effects


From   Lara K <larasusannakrugman@gmail.com>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: RE: Error? xtdpdsys assigns explanatory power to fixed effects
Date   Thu, 10 Nov 2011 00:39:55 +0100

Gustavo your reply is really helpful! Thank you so much!!
I have another question nevertheless. In my application - in a year-country panel data set - I estimate the effects of a certain variable W on the dependent variable Y. I know that the effect is different depending on the type of country. Therefore I created dummies for mutually exclusive groups of countries (a & b). Now I interact the variable W with these dummies (W*a, W*b) to estimate the effect of W with regard to the type of country.
By estimating subsamples (a==1 or b==1) I know that W has a positive effect for a-countries but a negative effect for b-countries. If I estimate the whole sample using XTDPDSYS and including the interaction terms W*a and W*b I falsely obtain positive coefficients for both. (If I use XTABOND I find a positive and a negative coefficient as expected.)

Is this a regression that cannot be done using XTDPDSYS as it estimates explanatory power for the 'zeros' in (W*a) for the b-countries??
Thank you very much for your helpful explanation!

I hope I formulated the question clear enough for you to understand.
A variation of my first small example using stata data would be:

    clear
    use http://www.stata-press.com/data/r11/abdata
    gen a=0
    replace a=1 if id <= 7
    gen aW= a * w
    
    gen b=0
    replace b =1 if id > 7
    gen bW=b * w
    xtset id year

    xtabond n L(0/2).(aW bW k) yr1980-yr1984 year, vce(robust) 
    xtdpdsys n L(0/2).(aW bW k) yr1980-yr1984 year, vce(robust)

(This is not working as a very good example because the subsamples a and b are not very different from each other.)


On 09.11.2011, at 18:26, gsanchez@stata.com wrote:

> Lara <larasusannakrugman@gmail.com> included a dummy for a panel variable
> (time invariant variable) in her dynamic panel estimation using -xtabond-
> and -xtdpdsys-. The first command omits the dummy (because of collinearity)
> but the second produces a coefficient estimate for that variable. Lara
> states that:
> 
>> According to the idea of the estimator it should obviously not do this as 
>> observation entity specific fixed effects are an integral part of the
> estimator.
>> To make this point clear I have this example using Stata data:
>> 
>>     use http://www.stata-press.com/data/r11/abdata
>>     gen seven=0
>>     replace seven=1 if id==7
>>     xtset id year
>> 
>>     xtabond n L(0/2).(w k) yr1980-yr1984 year seven, vce(robust) 
>>     xtdpdsys n L(0/2).(w k) yr1980-yr1984 year seven, vce(robust)
> 
>> xtabond is doing what it should do and omits the entity specific 'seven'
>> xtdpdsys estimates some coefficient it shouldn't.
> 
> Then, Lara asks:
> 
>> Is this already a known problem? Or isn't it a problem at all?
> 
> 
> Both commands are handling the dummy variable properly. 
> 
> - Time invariant variables must be omitted from the Arellano/Bond estimation
> (with -xtabond-) because the model is fitted in first differences and,
> therefore, the fixed effects are removed from the estimation. In fact, the
> output for -xtabond- includes a note stating that 'seven' (the
> time-invariant variable in Lara's code) is omitted because of collinearity. 
> 
> - On the other hand, the Blundell-Bond/Arellano-Bover is a system estimator
> (implemented by -xtdpdsys-) with one equation in levels and one equation in
> first differences. Time invariant regressors are omitted for the equation in
> first differences (as expected) but they are still present in the equation
> in levels. Thus, getting coefficient estimates for time invariant regressors
> is correct in this second case. Notice that the output for the dynamic
> regression with -xtdpdsys- also indicates that 'seven' is (only) omitted for
> the difference equation. 
> 
> 
> --Gustavo
> gsanchez@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/


*
*   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/


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