Marco Sanfilippo <[email protected]>:
I think these models are guaranteed to have: various fixed/random
effects, spatially correlated errors, serially correlated errors,
lagged dependent variables, and endogenous regressors, plus an
inherently nonlinear form, i.e. imports are a nonlinear function of X
(think log-log but you need to include zeros, so you want -glm- with a
log link or equivalent). So I think you have to try a variety of
different models to address different concerns, since you cannot
address all at once. You might want to look up
http://www.stata.com/meeting/snasug08/drukker_spatial.pdf on spatially
correlated errors and
http://www.stata.com/meeting/snasug08/drukker_xtdpd.pdf on DPD.
-nlsur- can account for cross-equation correlation of errors, and the
-glm- style model. The -cluster- adjustment can address serial
correlation of errors. Instrumental variables would be useful for
addressing endogeneity (does GDP growth cause import growth or does
import growth cause GDP growth?). But you may want to stick with
whatever wrong way these models are estimated in the existing
literature, so you can produce a dissertation in finite time. But try
to convince people to stop calling them "gravity models" which implies
some kind of universal constant related to Euclidean distance
determining international trade which is just silly...
On Thu, Sep 25, 2008 at 11:20 AM, Marco Sanfilippo
<[email protected]> wrote:
> Dear all,
>
> I am trying to estimate a gravity model for my phd thesis where my dependent
> variable is the import (M) of country j from country i of product hs6
> (6-digit harmonized system) at time t:
>
> lMijths6=
> lGDPit+lGDPpcit+lGDPjt+lGDPpcjt+ldistij+landlockedi+comlangij+eijhs6t
>
> I have around 70 importers and 48 exporters, with exports ranging up to
> 5,000 products. Time period is 1995-2005.
> Beyond the difficulties of handling with such big dataset, the first issue
> I'm dealing with regards the identifier for my panel. I believe there is
> wide eterogeneity both at the product level and at the country level but, as
> far as I know, I have to choose only one id. In case I identify my panel
> with either the country-couple (i j) or the products (hs6) I incur in
> "repeated time values within panel" error. I wonder whether it is
> appropriate to combine country and product effects (egen panel=group(i j
> hs6)).
> The second question regards the most appropriate method to be adopted. I
> think that fixed effects would be proper given the presence of so many
> sources of unobservables. However, among my covariates, I have some time
> invarying standard-gravity-variables (distance-common language, land
> lockeness). Should I employ a pooled estimation or is it better a two-stages
> least squares (e.g. Hausman-Taylor) or, still, dynamic panel estimators?
>
> Any suggestion is more than welcome, pls let me know if you need more
> details.
*
* 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/
