[Limdep Nlogit List] Stochastic Frontier Analysis

[William Greene]

Dear Kathleen.  The JLMS formula for the estimates of E[u|v+u] can only be
computed to within the tolerance of the computation of the normal pdf. This
runs out at about +/-7  standard deviations, after which the accuracy of the
computation is less than the rounding error.  It is truncated at this value.
What you are seeing is a few extreme observations that have very large values
of (v+u)/sigma that fit this problem.  Unfortunately, it can't be fixed; it's a 
characteristic of your data.
/B. Greene

************************************************
Professor William Greene
Department of Economics
Stern School of Business
New York University
44 West 4th St., Rm. 7-78
New York, NY   10012
Ph. 212.998.0876
Fax. 212.995.4218
URL. http://www.stern.nyu.edu/~wgreene
Email. wgreene at stern.nyu.edu
************************************************

----- Original Message -----
From: Kathleen Carey <kcarey at bu.edu>
Date: Monday, August 13, 2007 6:15 am
Subject: [Limdep Nlogit List] Stochastic Frontier Analysis

> SFA users:
> 
> I am estimating a hospital cost function in LIMDEP 8.0 using 
> stochasticfrontier analysis (half-normal distribution assumption 
> for inefficiency
> component of the error).  I expect that the inefficiency scores 
> will be
> unique for each observation, but I am generating some peculiar 
> results.  Of
> 1032 observations, there are 29 observations that have the same 
> efficiencyscore, 1.9943.  These are the highest values in the 
> distribution. 
> My code is as follows:
> 
> frontier; lhs=lncost;
> rhs=one,lndis,lnopv,lnlos,scope1,aprdrg,beds,opvcmi,resrat,hhi,sysid,k1,k2,w
> age,psi07,psi15
> 
> ; Cost
> 
> ; eff=u $
> 
> create ; expu=exp(u)$
> 
> sort ; lhs=expu$
> 
> list ; expu$
> 
> 
> The dependent variable is the log of total hospital costs, and the
> independent variables include multiple outputs (also expressed in 
> logs) and
> several covariates.  Variations on the model and small adjustments 
> in the
> included observations produce a similar result, with a large 
> number of the
> exact same efficiency score, expu, (to 8 places after the 
> decimal).  These
> values are always at the top end of the tail, which otherwise 
> appears to be
> properly distributed.
> 
> Has anyone had this experience before, or have a suggestion for 
> what might
> be going on here?
> 
> Thanks in advance,
> 
> Kathleen Carey
> Boston University School of Public Health
> _______________________________________________
> Limdep site list
> Limdep at limdep.itls.usyd.edu.au
> http://limdep.itls.usyd.edu.au
>