[Limdep Nlogit List] Stochastic frontier with sample selection

[Risto.Herrala at bof.fi]

Hi all!

I am using the new 'stochastic frontier with selection' feature, and am constantly running into 'Error   806: (The log likelihood is flat at the current estimates.)'.

For example, a simple normal/half normal model without selection runs smoothly with the result:
--> FRONTIER;   If[T=1988 & DLI=1 & W#-999];Lhs=L;Rhs=one,W; $  /*M1 in converge...
Normal exit from iterations. Exit status=0.....
+--------+--------------+----------------+--------+--------+----------+
|Variable| Coefficient  | Standard Error |b/St.Er.|P[|Z|>z]| Mean of X|
+--------+--------------+----------------+--------+--------+----------+
---------+Primary Index Equation for Model
 Constant|    3.25073849       .07065629    46.008   .0000
 W       |     .27606220       .01554707    17.757   .0000   3.95278986
---------+Variance parameters for compound error
 Lambda  |    4.09604670       .25779771    15.889   .0000
 Sigma   |    2.07506813       .00098495  2106.774   .0000

The sample selection algorithm
--> Sample;     All $
--> PROBIT;     If[T=1988 & W#-999];Lhs=DLI;Rhs=one,W;Hold $
--> FRONTIER;   If[T=1988 & W#-999];Lhs=L;Rhs=one,W; Output=3;Selection $

runs into trouble after four iterations. I have tried different algorithms, start values, samples (from alternative years, or subsamples of specific years), alternative specifications for the selection probit and the frontier. The log (from Output=3) suggests that 1) likelihood is not improving after four iterations 2) the gradient even increases slightly at 4th iteration. I am not certaint that I understand correctly what is in the parameter vector but it looks to me like some of the parameters in the covariance matrix are pretty erratic. I cannot interpret the end result as a satisfactory outcome since different start values for the main parameters lead to very different outcomes at abort.


Any suggestions about how to proceed? Is this a sign that the sampling model may be incorrect, and in fact there is no sampling issue in the data? Can I test the non-sample specification against a sampling correction without actually estimating the sampling corrected frontier model?

Risto