sfpanel

[Anubha Dhasmana]

I have a quick query regarding the stochastic frontier models. I am running a stochastic frontier model with exponential distribution in STATA (sfpanel command). Now; the readings say that you cannot use the “emean” option with exponential distribution. But in my case, not only does STATA estimate sfpanel command with “emean” option but all the results are also reasonable (have expected signs and significance). Am I missing something?

Thanks for your help in advance.

[Jared Greathouse]

Let me say straight away I know nothing about this modeling approach.... but you don't post code or a sub-sample of your results. How can we find out what the problem is without being able to see what went wrong?

Anyways, what readings are you referring to? Where do they bar you from using emean, exactly? The help file, something else? I am sure they mention it in the citations or in the help file, even if it's hidden deep in the text.

[Anubha Dhasmana]

Sorry. The code is simple - sfpanel lnIK l.lnIK l.lnSalesK l.firm_vol, mode(tfe) distribution(exponential) emean(lnCF dumsize l.sa_debt_equity dumold) ort(o);


It gives the right results (no error):

Warning: only units with more than 1 time occasion will be considered

initial: Log likelihood = -36999.189
Iteration 0: Log likelihood = -36999.189
Iteration 1: Log likelihood = -31739.289
Iteration 2: Log likelihood = -30977.253
Iteration 3: Log likelihood = -30902.741
Iteration 4: Log likelihood = -30893.138
Iteration 5: Log likelihood = -30893.119
Iteration 6: Log likelihood = -30893.119

True fixed-effects model (exponential) Number of obs = 19488
Group variable: co_code Number of groups = 2290
Time variable: year Obs per group: min = 2
avg = 8.5
max = 29

Prob > chi2 = 0.0000
Log likelihood = -3.089e+04 Wald chi2(3) = 2055.61


lnIK Coef. Std. Err. z P>z [95% Conf. Interval]

Frontier
lnIK
L1. .0846493 .0071393 11.86 0.000 .0706565 .0986422

lnSalesK
L1. .656199 .018777 34.95 0.000 .6193968 .6930012

firm_vol
L1. -.0798561 .0297739 -2.68 0.007 -.1382119 -.0215003

Usigma
_cons -.7276494 .047535 -15.31 0.000 -.8208163 -.6344825

Vsigma
_cons -.0657597 .0205166 -3.21 0.001 -.1059716 -.0255478

sigma_u .695013 .0165187 42.07 0.000 .6633794 .7281551
sigma_v .9676548 .0099265 97.48 0.000 .9483935 .9873073
lambda .7182448 .0245424 29.27 0.000 .6701426 .7663469




The reading I am referring to is Belotti, Diadone and Ilardi (2013), page 732 -

4.1 Main options for sfpanel

True fixed- and random-effects models (Greene 2005a,b)

distribution(distname) specifies the distribution for the inefficiency term as halfnormal
(hnormal), truncated normal (tnormal), or exponential (exponential). The
default is distribution(exponential).

emean(varlist m , noconstant) may be used only with distribution(tnormal).
With this option, sfpanel specifies the mean of the truncated normal distribution
in terms of a linear function of the covariates defined in varlist m. Specifying
noconstant suppresses the constant in this function

[Andrew Musau]

sfpanel is from SSC (FAQ Advice #12). You are probably mistaken in believing that the option

emean(lnCF dumsize l.sa_debt_equity dumold)

is doing anything with an exponential distribution specified for the inefficiency. Compare your results with

sfpanel lnIK l.lnIK l.lnSalesK l.firm_vol, mode(tfe)

[Anubha Dhasmana]

Here are the results without emean -

sfpanel lnIK l.lnIK l.lnSalesK l.firm_vol, mode(tfe)

Warning: only units with more than 1 time occasion will be considered

initial: Log likelihood = -42360.739
Iteration 0: Log likelihood = -42360.739
Iteration 1: Log likelihood = -39242.497
Iteration 2: Log likelihood = -38436.744 (not concave)
Iteration 3: Log likelihood = -38426.879 (not concave)
Iteration 4: Log likelihood = -38418.989 (not concave)
Iteration 5: Log likelihood = -38412.678 (not concave)
Iteration 6: Log likelihood = -38404.653 (not concave)
Iteration 7: Log likelihood = -38398.766 (not concave)
Iteration 8: Log likelihood = -38389.386 (not concave)
Iteration 9: Log likelihood = -38374.535 (not concave)
Iteration 10: Log likelihood = -38363.246 (not concave)
Iteration 11: Log likelihood = -38296.649 (not concave)
Iteration 12: Log likelihood = -38205.076
Iteration 13: Log likelihood = -37449.124 (backed up)
Iteration 14: Log likelihood = -36593.065
Iteration 15: Log likelihood = -36515.505
Iteration 16: Log likelihood = -36512.64
Iteration 17: Log likelihood = -36512.616
Iteration 18: Log likelihood = -36512.616

True fixed-effects model (exponential) Number of obs = 22039
Group variable: co_code Number of groups = 2479
Time variable: year Obs per group: min = 2
avg = 8.9
max = 29

Prob > chi2 = 0.0000
Log likelihood = -3.651e+04 Wald chi2(3) = 3392.70


lnIK Coef. Std. Err. z P>z [95% Conf. Interval]

Frontier
lnIK
L1. .132896 .0067323 19.74 0.000 .119701 .146091

lnSalesK
L1. .6888306 .0170161 40.48 0.000 .6554797 .7221815

firm_vol
L1. -.1116538 .028586 -3.91 0.000 -.1676813 -.0556262

Usigma
_cons -.4725276 .0415192 -11.38 0.000 -.5539038 -.3911515

Vsigma
_cons .0258429 .0199516 1.30 0.195 -.0132616 .0649474

sigma_u .7895723 .0163912 48.17 0.000 .758091 .8223611
sigma_v 1.013005 .0101056 100.24 0.000 .9933912 1.033007
lambda .7794355 .0245498 31.75 0.000 .7313188 .8275523


.

Not only do I not get the coefficients of the efficiency equation, coefficients of the frontier equation are also different and log-likelihood is different as well.

Any help would be much appreciated.

[Andrew Musau]

I believe that all that is happening is that missing values in the variables specified within -emean()- are curtailing your sample. To get the same results, you should run

sfpanel lnIK l.lnIK l.lnSalesK l.firm_vol if !missing(lnCF) &!missing(dumsize) &!missing(l.sa_debt_equity) &!missing(dumold), mode(tfe)

Again, the description is not wrong here. The option -emean()- is not meaningful under the assumption of an exponential distribution for the inefficiency.

[Anubha Dhasmana]

Thanks. You are right and that clarifies my doubt. But how do I estimate inefficiency equation with exponential distribution then? Is there a command?

[Andrew Musau]

Get a hold of Kumbhakar's book - A Practitioner's Guide to Stochastic Frontier Analysis Using Stata-. It will give you a step-by-step guide to estimation of these models in Stata and assumptions that underlie a particular model.

[Anubha Dhasmana]

Thanks.

[Anubha Dhasmana]

Model selection for stochastic frontier.

Hello all.

Does anyone know of a good reference explaining model selection for stochastic frontier models using STATA?

Thanks.

[Akbar Muhammad]

Hi Anubha, i want to asking same question like you do. Have you solve it?

Im using TRE with command like this

sfpanel lny lnxi lnx2 lnx3 lnx4 lnx5 lndumx6, model(tre) distribution(hnormal) usigma(lnz1 lnz2 lnz3 lnz4 lndumz5) difficult rescale nsim(100) simtype(genhalton)

Is this true?

Thanks.