[Limdep Nlogit List] Panel sample selection questions

[M.Dorotic]

Dear all,
I have panel data with 800 cross-sections and 183 time periods which I 
want to estimate using a panel sample selection model with random 
effects (app. 30% of observations are zeros). I have following two 
questions/issues to which I couldn't find straightforward answers in 
the book (EA) and the manuals (version 8).

1) (Non)convergence issue: I used the recommended three equations 
(probit, selection MLE and selection random effects). Selection MLE 
command gives the error 806 "Line search does not improve fn." Indeed 
the function is flat, the gradient value before the "crash" is .4d-4, 
after 50something iterations. LIMDEP however provides ML estimates 
with high face validity. Now, explanations of the error 806 on the 
back of the reference manual indicate this may not be an error, while 
explanations in the numerical optimization section (section 10) call 
for caution. My question is: can the obtained ML estimates be 
"trusted"  as the final results?
N.B. I have tried the recommended remedies (scaling (not a problem) 
and running a simpler model (the same occurs even with the simplest 
model), and selection models should be the right specification to use 
for this data.


2) Autocorrelation:  In Econometric Analysis prof. Green explains that 
panel data models often do not account for autocorrelation because 
time series are typically short (i.e. wide but short panels) (aside 
linear dynamic panels). After readings, it seems like I should not use 
Yt-1 as an Rhs in this nonlinear selection type of estimation? How 
else to account for serial correlation in the wide and long panels?

Thank you in advance.

Sincerely,
Matilda