Re: st: xtfrontier vs xtreg, re mle

[Scott Merryman <scott.merryman@gmail.com>]

Both are random effects models - one estimated using gls the other by
maximum likelihood.  The maximum likelihood approach involves making
stronger distributional assumptions for the u_{i} : half normal or
truncated normal.

Kumbhakar and Lovell (2000) "Stochastic Frontier Analysis" briefly
discuss whether or not the results are sensitive to the method of
estimation.  They concluded (for time invariant models) the approaches
(fixed effects, random effects, or maximum likelihood) generate
similar efficiency rankings.

Scott


On Fri, Apr 23, 2010 at 10:48 AM, Dean DeRosa <dderosa@adr-i.com> wrote:
> I am new to stochastic frontier (SF) analysis, and I have been experimenting
> with the xtfrontier command to estimate the parameters of a trade gravity
> model. I have been struck by the similarity of SF theory and its setup in
> Stata to the more common practice of estimating trade gravity models using
> the random effects (RE) method of estimation for panel data sets.
>
> On conducting side-by-side experiments using the two approaches in Stata
> (xtfrontier versus xtreg, re mle), I find that the gravity model estimates
> are virtually identical, except with regard to the intercept term --
> sensibly, the SF intercept estimate is greater in value than the RE
> estimate.
>
> Is this a common if not expected finding? Or, is this finding related to
> some problem with Stata? Any citations to similar findings would be welcome,
> as would a straightforward explanation for the finding.
>
> Finally, I would appreciate any citations to basic or introductory
> explanations of the diagnostic statistics, such as mu, gamma, etc.,
> generally reported by Stata and published studies featuring SF estimation
> results.
*
*   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/