If you know R programming, there are very good libraries for SFA and DEA. R libraries designed for parametric and non-parametric efficiency analysis are:

Frontier

FEAR

Benchmarking

Documentation of these libraries come with full help to apply the methods....

Many thanks for the information provided. I am aware about the differences etc. and commands, but there is something missing when applying the same inputs and outputs and the results change depending on the order of variables for the same number of iterations. Best regards, Ane

I think there is something not correct. Theory of SFA and my experience in conducting SFA using STATA suggests, order of variables does not matter at all.  

Thanks Rupayan, but this is what science is about. I have been using the same variables in different order and the outputs from SFA is totally different in significant level and coefficients.  Hence why have I posed this question.

 I hope, you are not changing the order in a way that treats outputs as inputs, and vice-versa.

Thanks for the suggestion. No, the logs and iterations are not changed, neither the independent variable, etc. but I do not way for example when introducing environmental variables (zts) and if the order is a, b, c, and then I change it to b, c, a etc. the results are relatively different. The same when considering the order within the outputs or within the inputs. Additionally, it seems that there are not (or few) real empirical examples. I am applying SFA to a real industry. Not sure if you have any real case that may help. Ane

Do you have multiple outputs in the model? If yes, I think that there is some problem in estimating. 

Yes I do, this is a multi-output input model with all the iterations. It is based on a translog distance function. Many thanks

In order to estimate the multi-input multi-output distance function you have to "normalize" it dividing all inputs and outputs by an input. If you do not specify explicitly which the normalizing input is, STATA will use the first input you have included in your data file. I think that somewhere there is the solution of this strange situation

Thanks for your contribution Kostas, all the previous steps done!! not only normalising by one input, but previously all the variables by their arithmetic mean, and also I treied the geomean and the median (since it seems to push outliers)

Hi Ane, I suggest you read the method as specified in Kumbhakar et al(2015) A practitioner's Guide to stochastic frontier analysis using STATA, page 300 - 308

Hi Mrs.Ane, I suggest you use the Frontier 4.1 software (http://www.uq.edu.au/economics/cepa/frontier.php)

Thanks Franklin, I have also exchanged a discussion with Kumbhakar, but it seems that there is a lack of empirical research using SFA.

HI, Mrs. Ane, i actually have a question on stochastic cost frontier analysis. I'm checking in my panel data set for political determinants of public service distribution inefficiency. I have a total cost, many outputs, and a labor-cost input price. How do i normalize my data to be sure that the cost function behaves good(monotonic, homogenous, etc.)

Thank you in advance

Dear David, my apologies for my delay. I am really busy at this moment with several projects. Let's see. Within the SFA I have tried several normalisations to see if the way of normalising could influence the results of SFA (for example significance of parameters or expected signs, essentially of the basic variables). I have used arithmetic mean; median mean (since it seems to detect better the presence of outliers) and geometric mean. Based on literature review it seems that the best approach is using geometric mean. This is because it allows to estimate elasticities at the sample mean (Cuesta & Orea 2002; I think there is also another Cuesta 2009). The arithmetic mean has been demonstrated not to be a good approach when performing benchmarking (please see Fleming & Wallace, 1986). My results continue being not satisfactory. Therefore it may either the data I have or the model. Working on both issues now. I wish this helps. Ane

Dear Gürkan, apologies for my delay. I did not realise about your answer. Many thanks for the suggestion. I will have a look the link provided. Best regards, Ane

Hi hoaw to get a shadow price using the static software, from estimating the production function or cost?

Does anyone know

Please help. Thank you

HI ANE can I have distance function command in stata,how can I have environmental variables? what about metafrontier? any one can help me plz?my data are cross-section

When using the directional distance function in quadratic form, should I convert the variable values into a log like in translog form? Can anyone help?

The idea of using logs is for the interpretation, e.g. % It helps to interpret the coefficients. How many variables do you have?

three inputs, three outputs, and one undesirable output

Then, in a multi-input output context, I would recommend using a translog distance function. Therefore, log your distance function: to account for the iteration effects between inputs, outputs, and inputs and outputs. The log will help you interpret the effects. Otherwise, how would you interpret the variables squared (i.e., including the dependent)?

Hi

Can i get STATA code for distance function in Stochastic Frontier Model ?

You will have to prepare your distance function (excel or similar) before estimating SFA, in STATA. You may start for a translog and see if it works (i.e., with a dependent variable either overall costs or overall production). Otherwise, you may use a translog distance function, which implies dividing your inputs for an arbitrary input (if input distance function) or your outputs for an arbitrary output (if output distance function), and your dependent variable will be ln(1/arbitrary variable). The commands are the same no matter the functional form: i.e., Cobb-Douglas, CES or translog, or distance function. The commands depend on your assumptions. You will have to make assumptions regarding the distribution of the inefficiency term (u), hence choosing between half-normal, truncated, exponential or gamma. Each one has a different command, usually with distribution(). In the end, you need to choose the model that better presents the reality and is restricted to your data. Then there are different commands for time-varying or invariant models.

This paper may help for STATA commands for SFA https://journals.sagepub.com/doi/pdf/10.1177/1536867X1301300404