Re: st: Problem with sfcross - r(3200) error

[Federico Belotti <f.belotti@gmail.com>]

Dear Reut,

As expected, your problem was due to a not feasible starting values issue. The fact that now the optimization does not reach convergence might depend on a lot of things. See below for three of them.

On Sep 30, 2013, at 9:36 PM, Reut Levi wrote:

> Dear Federico,
> 
> First of all thank you for your comments. I was not aware that I need to explain where user written programs come from, but will definitely explain it in the future.  So, for any future commenters, as was already specified, the command, sfcorss, comes from the SSC.
> 
> Second, I’ve read about the rescale and restart options for the starting values, but I’m not sure I understand the reasoning behind it to determine whether I need it or not.  Nonetheless, I tried to add it to my code and I am not getting that error anymore (r 3200 – initial values not feasible).  However, it seems that convergence is not achieved because the outputs for Usigma, sigma_u and lambda are missing.  So when I try to predict my efficiency scores, using the syntax:
> predict double ine, u
> gen eff=exp(-ine)
> I get a single efficiency score for my entire data.


1) the inefficiency component might not be identified because of normality. So, try to test for residual normality after a simple constrained regression using the following syntax

cnsreg lntc lnp1-lnloanslndd,  constraints(1 2 3)
predict res, res
swilk res

2) Your translog cost frontier is not correctly specified. You can find an example of multiple output translog cost frontier in Kumbhakar and Lovell (2000) pag. 143.

3) Poor starting values may lead to convergence problems (here my second suggestion might help). 

> 
> As to the other suggestion you made, I don’t have previous constrained regression so I don’t think I can do that.  I’m not an expert in Stata, so if you don’t think I understood you suggestion, please clarity.

My suggestion was about the specification of better initial values for the coefficients using the option -svfrontier()-. Since you problem is due to poor starting values, you may try using the following syntax

cnsreg lntc lnp1-lnloanslndd, constraints(1 2 3)
mat b0 = e(b)
sfcross lntc lnp1-lnloanslndd, distribution(hnormal) cost constraints(1 2 3) svfrontier(b0) svu(-1) svv(-1)

As you can see, the options -svfrontier(b0) svu(-1) svv(-1)- have been added to the syntax. -svfrontier(b0)- specifies the values of the vector "b0" as starting values for the frontier coefficients, while -svu(-1) svv(-1)- specifies the value "-1" for the variances' parameters. Indeed, the evaluators of -sfcross- parametrize (regardless from the inefficiency distribution) the variances of inefficiency and idiosyncratic terms as

sigmau = exp(0.5 * lnsu)
sigmav = exp(0.5 * lnsv)

This means that a value of "-1" for "lnsu" and "lnsv" corresponds to a starting value for these two variances of about 0.6.

> 
> Third, as to you comment about not being able to understand if my translog cost frontier is correctly specified, can you please tell me what information should I give you so that you will able to determine that? Not every day I get to read suggestions from the actual person who wrote the command so I don’t want to pass on this opportunity!

Looking at your code, it seems to me that the squared outputs are not part of your model. As mentioned before, you can find an example of multiple output translog cost frontier in Kumbhakar and Lovell (2000) pag. 143.

References
Kumbhakar, S., and C. Lovell. 2000. Stochastic frontier analysis. Cambridge University Press.

HTH,
Federico

> I read the research paper series “Stochastic frontier analysis using Stata” that was written by you, but unfortunately, I was not able to follow some of the math. Also, this is the first time I’m using Stata for a research paper so, if you can, please try to simplify your response.
> 
> Thank you very much!
> I appreciate your time and input,
> Reut
> ____________________________________
> From: owner-statalist@hsphsun2.harvard.edu <owner-statalist@hsphsun2.harvard.edu> on behalf of Federico Belotti <f.belotti@gmail.com>
> Sent: Monday, September 30, 2013 5:05 AM
> To: statalist@hsphsun2.harvard.edu
> Subject: Re: st: Problem with sfcross - r(3200) error
> 
> Reut Levi <rlevi2@student.gsu.edu> have a starting values issue with the user-written command -sfcross-:
> 
> -sfcross- is from SSC and it is not an official Stata command. As Statalist FAQ clearly report,
> please always explain _where_ user-written programs you refer to come from.
> "This helps (often crucially) in explaining your precise problem, and it alerts readers to
> commands that may be interesting or useful to them."
> 
>> I have 10 groups of data, according to percentiles of asset size.
>> The command works for 3 out of my 10 groups.  However, for 7 of the groups, I’m getting the following error: r(3200) - “initial values not feasible”.  If there is a conformability error, why does the command work for 3 groups?
> 
> 
> I cannot understand from what Levi wrote if his translog cost frontier is correctly specified. Nevertheless, it seems to me that the -moptimize- conformability error is a consequence of the "initial values not feasible" issue.
> My suggestions to Levi are:
> 
> 1) Read carefully the  "starting values" section in the -sfcross- help file. A solution to his problem might be the use of options -restart- and -rescale-, i.e.
> 
> sfcross lntc lnp1-lnloanslndd, distribution(hnormal) cost constraints (1 2 3) restart rescale
> 
> 2) Try using starting values from a previous constrained regression, something like
> 
> cnsreg lntc lnp1-lnloanslndd, constraints(1 2 3)
> mat b0 = e(b)
> sfcross lntc lnp1-lnloanslndd, distribution(hnormal) cost constraints(1 2 3) svfrontier(b0)
> 
> Hope this helps,
> Federico
> 
> 
> 
> On Sep 30, 2013, at 2:44 AM, Reut Levi wrote:
> 
>> Dear Statalist members,
>> 
>> I am using the sfcross command to estimate inefficiency levels for the U.S. banking industry.
>> 
>> My syntax is: sfcross lntc lnp1-lnloanslndd, distribution(hnormal) cost constraints (1 2 3)
>> 
>> Where lntc is the natural log of total cost and lnp1-lnloanslndd are the natural logs for all of my input prices (3 types of input prices), outputs (2), and interaction terms.
>> Also, my constraints syntax is as follows:
>> 
>> constraint 1 [lnp1+lnp2+lnp3]=1
>> constraint 2 [lnp1lnp1+lnp1lnp2+lnp1lnp3+lnp2lnp2+lnp2lnp3+lnp3lnp3]=0
>> constraint 3 [lnp1lnloans+lnp1lndd+lnp2lnloans+lnp2lndd+lnp3lnloans+lnp3lndd]=0
>> 
>> Where, constraint 1 is for restricting the coefficients of input prices to 1.  Constraint 2 is for restricting the coefficients of price-price interaction terms to 0. And constraint 3 is for restricting the coefficients of price-output interaction terms to 0.
>> 
>> I have 10 groups of data, according to percentiles of asset size.
>> The command works for 3 out of my 10 groups.  However, for 7 of the groups, I’m getting the following error: r(3200) - “initial values not feasible”.  If there is a conformability error, why does the command work for 3 groups?
>> 
>> I tried a variety of things, including smaller groups, or smaller ranges for the variables; I also tried dropping all the data points for which input price or total outputs equal zero, but nothing seems to work.
>> 
>> Please let me know if you have any other suggestions or if you can explain the problem behind this error.
>> I would be happy to provide additional information about my data or my code, if needed.
>> Thank you very much for your help. Have a great week,
>> Reut
>> 
>> *
>> *   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/
> 
> 
> *
> *   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/