[Limdep Nlogit List] Different results in Frontier Analysis in Limdep and Stata

William Greene wgreene at stern.nyu.edu
Thu Aug 20 21:24:30 EST 2009

M. Karakaplan.
The difference in the number of iterations results because the
programs are using different algorithms - Stata uses Newton's
method and Limdep uses BFGS. You did not report the command you
used, so it is not clear that you actually were fitting the 
same model with the two programs. But, assuming you were, I suspect
the following. The estimator should converge in about 10 iterations
or so in Limdep, 5 or so with Stata. Since you went to 27 with 
Stata and never did actually converge with Limdep, I would guess
that neither program actually estimated a model. With this many
iterations, with this model, it is very likely that you have a
serious data problem. The answers produced by both programs are
probably nonsense.  In order to judge more completely, I would 
have to see your output.
One final point, Stata uses a different parameterization of the
model than Limdep. The asymmetry parameter that characterizes the
inefficiency is set up in Limdep according to the original Aigner,
Lovell and Schmidt specification. The same model is parameterized 
in Stata according to the model reformulated by Tim Coelli in his 
Frontier 4.1 program. It is a 1 to 1 reparameterization, so it 
should not affect the results, but optimization in an ill behaved 
data set can be affected by the normalization in the model.
/Bill Greene

----- Original Message -----
From: "M.U. Karakaplan" <karakaplan at tamu.edu>
To: limdep at limdep.itls.usyd.edu.au
Sent: Thursday, August 20, 2009 4:58:04 AM GMT -05:00 Columbia
Subject: [Limdep Nlogit List] Different results in Frontier Analysis in Limdep and Stata

I run exactly the same frontier model in Limdep and Stata, but I get 
different results:

The descriptive statistics are the same
The number of observations and parameters are the same.
The coefficients are very similar, but not the same.
The numbers of iterations are different: Limdep: 101 iterations; Stata: 
27 iterations.
When I set the maximum iterations in Limdep to be the same of Stata, the 
difference gets bigger.
When I set the maximum iterations to 200 in Limdep, it runs only 125 
iterations and the difference between the results gets very very small, 
but they are still not the same.

How can I get the same results? What is this difference between Limdep 
and Stata results due to? Is it related to confidence level? If yes, is 
there a way to change it in Limdep?

Limdep site list
Limdep at limdep.itls.usyd.edu.au

More information about the Limdep mailing list