Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.

# Re: st: Different result of Technical Efficiency with Frontier 4.1 and Stata

 From Federico Belotti To statalist@hsphsun2.harvard.edu Subject Re: st: Different result of Technical Efficiency with Frontier 4.1 and Stata Date Sat, 12 Oct 2013 01:19:37 +0200

Dear Vikash Vaibhav,

On Oct 4, 2013, at 9:19 AM, Vikash Vaibhav wrote:

> Dear Federico,
> Thanks for the reply and writing the sfpanel command.
> Earlier, i had used xtfrontier command (sfpanel is not supported by stata 11.1)
> I would like to make the following observations with the comments
> given by you and also about my experience with sfpanel in last two
> days:
>
> 1. When i used the same data (as attached by you) for sfpanel and
> xtfrontier, i got somewhat different result. my commands were:
> sfpanel y x1 x2, m(bc92) tech(dfp)
> xtfrontier y x1 x2, tvd tech(dfp)
> I dont know which one to use. If my understanding is correct both
> these models would be the same.
> sfpanel is preferable since its more general. but how to explain this?
>

The difference you observed is just a matter of differences in default optimization algorithm and convergence criteria between the two commands. Try using the following

sfpanel y x1 x2, m(bc92) iter(200) tech(dfp) difficult nrtol(1e-07)

xtfrontier y x1 x2, tvd tech(dfp)

> 2. When i 'made' the data supplied by you to be unbalanced (by
> deleting some of the observations) there were missing values in
> predicted efficiency. (i do not know whether i am justified in
> deleting the simulated data.)

I do not understand your point here. Please send the code you have used.

>
> 3. I am linking few of my comparison from my empirical data for one
> industry (56 firms, 18 years, total 704 obs).
> https://www.dropbox.com/sh/zi131lz3iz7r5l6/EFbH1IohY3
>
> I did not observe
> similarity in the efficiency scores between the one obtained by
> frontier and other by sfpanel command. I also noticed that there were
> many missing values in the predicted scores which is not the case with
> frontier program.
> I suspect it because of use of unbalanced panel data. there are two
> outputs attached one with mu (outmu) and other without mu (nomu). i
> had executed the instructions on the terminal so there might be some
> errors on instruction file (ins.txt)
>

The differences you are observing are due to differences in the optimization engine of the two softwares.
You can get very similar results in the BC92 normal-truncated normal model using

mat b0 = (0.68402974,0.42995639,-1.2135513)
mat sigma = .81661042
mat gamma = 0.72
mat mu = 0
mat eta = 0
sfpanel y k l, model(bc92) dist(tnormal) tech(dfp) iter(10) ///
svfront(b0) sve(mu) svsigma(sigma) svgamma(gamma) sveta(eta)

The same is true in the case of the BC92 normal-half normal model (i.e. constraining mu=0)

cons def 1 [Mu]_cons=0
sfpanel y k l, model(bc92) dist(tnormal) constr(1) tech(dfp) iter(100)

> One more observation is regarding the increasing efficiency over the
> years. Efficiency always increased in my dataset. i wonder is it
> because they really increase their efficiency or is it because of some
> assumptions in calculating them using statistical software.
>

The BC92 model assumes that inefficiency evolves over time following

u_it = u_i*exp(-eta*(t-T_i))

That is, the (in)efficiency will increase or decrease according to the sign of the \eta parameter.
Since the \eta parameter is positive and statistically significant (even if very small) in both models, there is evidence of increasing efficiency over time in your data.

Hope this helps.
Federico

> Looking forward to hearing a reply from you.
>
> ---
> Thanks & Regards,
> vikash vaibhav
>
>
> On Tue, Oct 1, 2013 at 4:35 PM, Federico Belotti <f.belotti@gmail.com> wrote:
>> Vikash Vaibhav <vikashv@igidr.ac.in> found dramatically different results comparing Stata with Tim Coelli's Frontier 4.1.
>>
>> First, I would like to ask to Vikash which Stata command has been used for the estimation.
>>
>> Second, in my experience the two softwares give exactly the same results. I compared the user-written command -sfpanel- (from SSC) and FRONTIER 4.1 for the Battese and Coelli (1995) model using a simulated dataset that can be downloaded at
>>
>> http://www.econometrics.it/data/FRONTIER41vsSFPANEL.zip
>>
>> Data has been simulated from the following normal - truncated normal production model
>>
>> y_it = 1 + 0.3 x1_it + 0.7 x2_it - u_it + v_it
>>
>> where u_it is truncated normal with mean
>>
>> mu_it = 1 + 0.5 zmu_it
>>
>> and unit variance and v_it is normal with mean 0 and variance = 0.25.
>>
>> To run the estimation in Stata you can use the data file "simd-dta.dta" and the do file "check.do".
>> To run FRONTIER 4.1 you can use the "dta.txt" and "ins.txt" files.
>>
>> I didn't find appreciable differences, considering also that the two log-likelihoods (the -sfpanel- and FRONTIER 4.1 ones) are differently parametrized. So, I'm wondering if you have some problems in variables and/or instruction file definition.
>>
>> On the other hand, I found appreciable differences using real data but only when the data has been transformed using, for instance, the log_var = log(var+0.1) trick to get rid of the log-zero problem. Usually this kind of tricks creates huge disproportions in the data which can be handled differently by the optimization engine of different software. A useful discussion on this issue can be found at
>>
>>
>> Hope this helps,
>> Federico
>>
>>
>> References
>> Battese, G., and T. Coelli. 1995. A model for technical ine ciency eff ects in a stochastic frontier production function for panel data. Empirical Economics 20: 325-332.
>>
>>
>>
>> On Oct 1, 2013, at 7:22 AM, Vikash Vaibhav wrote:
>>
>>> This is with respect to the dramatically different result that i got when i
>>> compared the results of technical efficiency estimated from Frontier 4.1 and
>>> Stata. (two more student here has the same problem and with a different
>>> dataset). I do not think that it could have been due to differing
>>> algorithms or some other things that vary across statistical packages.
>>> I would like to point out the following observations:
>>>
>>> 1. The results were different in case of cross section data. But they did
>>> not differ a lot.
>>> 2. The results for the panel data were hugely different with the two
>>> software.
>>>
>>> I have used Stata IC version11.1 (for windows).
>>> I have used Frontier 4.1 package freely available on CEPA website.
>>> Any help ??
>>>
>>> --
>>> Thanks & Regards,
>>> Vikash vaibhav
>>> *
>>> *   For searches and help try:
>>> *   http://www.stata.com/help.cgi?search
>>> *   http://www.stata.com/support/faqs/resources/statalist-faq/
>>> *   http://www.ats.ucla.edu/stat/stata/
>>
>> --
>> Federico Belotti, PhD
>> Research Fellow
>> Centre for Economics and International Studies
>> University of Rome Tor Vergata
>> tel/fax: +39 06 7259 5627
>> e-mail: federico.belotti@uniroma2.it
>> web: http://www.econometrics.it
>>
>>
>> *
>> *   For searches and help try:
>> *   http://www.stata.com/help.cgi?search
>> *   http://www.stata.com/support/faqs/resources/statalist-faq/
>> *   http://www.ats.ucla.edu/stat/stata/
>
>
>
> --
> Thanks & Regards,
> Vikash vaibhav
>
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/faqs/resources/statalist-faq/
> *   http://www.ats.ucla.edu/stat/stata/

--
Federico Belotti, PhD
Research Fellow
Centre for Economics and International Studies
University of Rome Tor Vergata
tel/fax: +39 06 7259 5627
e-mail: federico.belotti@uniroma2.it
web: http://www.econometrics.it

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