Hi  Yousuf, 

The easiest way is:

1) Open STATA

2) Choose "Statistics" from the menu

3) Choose "Linear models and related"

4) Choose "Frontier models"

5) Once you have opened it (Frontier models), you can see "Model 1" where you specify the production function variables, and "Model 2" you specify variables explaining the inefficiency.

Once you run the model, Technical Efficiency coeff are obtained with the command , predict te, te

Good luck!

Gordana

Thanks Gordana. It is very helpful

You can also use sfpanel (please see Federico Belotti's paper). This command refers to making different assumptions regarding the error term. It is assumed that a part of the error is explained by technical inefficiency and other part cannot be explained (noise, etc.)

You will generate u first (techical inefficiency) and then you can generate the TE based on the assumptions you choose for the inefficiency term.

For example, with this command you could predict TE following Battese & Coelli (1988; 1992; 1995) or Kumbhakar (1990)

It is really important to test the distribution of the inefficiency term (u) and not to assume that all the error is caused by inefficiency. In other words, managers may not always be able to reduce this.

I wish it helps

Ane

Thanks, Ane.

By testing of distribution of u, which mean I should consider the sigma-squared, gamma, and likelihood ratio. Isn't it?

The distribution of the namely error term u, is refer to different assumptions regarding Normal, exponential or gamma distributions. Where the expected value of u (called mu) could be zero or non-zero value, depending on which distribution you are assuming. Half-normal (kumbhakar, 1990)  or truncated normal (Battese & Coelli 1992, 1995). Again each distribution will have their respective  deviation (sigma-squared)

The idea is to consider as starting point, that not all the error is caused by the inefficiency of the firms, households etc. and therefore potentially to be reduced, control by the managers, but there is part of the inefficiency that cannot be reduced since it is not controlled, it is caused by external factors, e.g. in your case households could be macro politic applied by the Government or regional councils that may affect the regions differently. These factors are normally called ' environmental variables (Zis, where i is a specific firm, unit, household, etc.). For example in the household case they could be population, gdp per capita, employment rate, etc. of the area where a specific household is located that obviously are not controlled by the household and yet may affect its efficiency.

In my case I have use these models to test how much of the inefficiency of the airports may be reduced by the managers or not (noise). Firstly I have tested which model explains better the reality of the Spanish airports (based on the limited data I got). So I have used sfpanel and I have obtained all these models. Mainly by doing the contrast of mu (mu=0) you can decide which model explains better the inefficiency in your data. If the data is not good enough, you will see that the coefficients do not make sense although being significant, or other parameters such as gamma or lambda (overall these are referred to how much of the variance is explained by the inefficiency compared to the noise, kind of R2)

These models also consider time as key factor (panel data) to test if changes in efficiency are relevant or not. You can an invariant version which is Battese & Coelli (1988). Again, you could use the Greene models, etc.

This paper will provide you an overview of all the models that you could test in order to choose the one to estimate the technical efficiency (inefficiency) in your case,

http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2145803

So basically, the idea is not just to jump into calculating the efficiencies (assuming that all is controllable or potentially to be reduced or improved), but, to test firstly if your data is relevant enough to explain the efficiencies scores and secondly to choose the model that better explains the reality of your database.

I wish this is more clear now.

Best wishes,

Ane

Thank you, Ane. I use cross section data. But your explanation and suggested paper are helpful

I see, in this case you have the option for cross sectional data also. 

Good luck

You can buy the book "A Practitioner's guide to SFA using Stata" by Kumbharkar et al. (2015). It presents a powerful treatment about SFA from single equation to multiple equation modelling plus cross sectional and panel data settings. Best.

Thanks, John. I'll look for it

Dear Yousuf, you may have already the answer to your question, but in case it helps. The u is the technical inefficiency. In fact, the command states predict e (which is the error, u (which this is only the name you are giving to the error, you could call it ABC and then the command would be "predict e, ABC". In order to estimate the TE you must specify the command: generate TE (or other name) = exp(-u) as specified in SFA

I see. Thanks again, Ane. I'll try it

How to interpret the zero gama value in stochastic frontier analysis.

I agree with the comment Gordana .

How is a censored stochastic frontier model implemented in STATA

good question ,interstings

thanks

how can i run input or output distance function with cross section data in stata? plz help

Fati Ansari

you can simply run distance function using efficiency approaches.

how to estimate gamma value in stochastic frontier analysis on stata?

How to estimate technical progress and total factor productivity in SFA

What is the commond predict uhat_h normal predict in frontier model in stata?

thank you deepa.

I have a similar problem like Yousef except that i want to decompose the error term into four components namely: time-invariant inefficiency, time-varying inefficiency, the heterogeneity term, and the usual white noise. I cant find the syntax in Stata 15 which i am using on household panel data to estimate maximum likelihood procedure in a single stage. I need help. Thank you

what is the stata code for estimating efficiency of time series data by using SFA?

why there are more than 2000 iteration and it does not stops but continues to more iteration?

Hello, may I ask the TE predicted is decreasing year by year? Does it make sense?