[Limdep Nlogit List] Problem fitting Binomial Logit Model (with selection)

james benson jbenson at ssc.wisc.edu
Wed Feb 11 06:49:04 EST 2009

I am trying to fit the Binomial Logit Model (with a Binomial Probit 
Selection model) as described in Section E31.2.4 of the LIMDEP modeling 
guide (version 9.0, volume 2). I am concerned about the low level of 
prediction success for my Selectivity Corrected Logit Model. This 
(selectivity corrected model) predicts all cases to be zeros, and thus 
has a sensitivity of 0 percent (it does not successfully predict any 
ones). This exact same model, when run as a standalone Binomial Logit 
for the selected sample correctly predicts 56 percent of the ones. The 
companion Binomial Probit selection model that I am using is fairly 
effective: it correctly predicts 60 percent of the ones, and 72 percent 
of the ones and zeros in the full sample. 

I receive an error message (see below) indicating that the log 
likelihood is flat, and instructing me to check the derivatives. Before 
delving into the statistical issues associated with my modeling 
strategy, I want to make sure that I have correctly instructed LIMDEP to 
compute the model. In particular, I am concerned that N for the Binomial 
Logit (outcome model) is equal to the full sample size, not the selected 
sample size. Is this as it should be? Also, when I compare my output for 
the sample-selected Binomial Logit to output for a Binomial Logit run 
only for the selected sample, I notice that the results are similar in 
the sense that neither model correctly predicts any ones. That being 
said, the coefficients in these two models are different, with the 
coefficients in the Selectivity Corrected Binomial Logit being closer to 
what I would expect.

Any advice regarding correct setup of selection models will be 
appreciated. The commands that I am using to set up the model just 
described are included below.


* The first step instructs LIMDEP to use the full sample of students 
eligible to enter a community college or 4-year institution

* The PROBIT model predicts which students will enter a community 
college as compared to a 4-year

* The LOGIT model predicts which community college students will succeed 
(get a degree or transfer)

Error Message: Line search does not improve fn. Exit iterations. Status=3
Check derivatives (with ;OUTPUT=3). This may be a solution
if several iterations have been computed, not if only one.
  Error   806: (The log likelihood is flat at the current estimates.)


More information about the Limdep mailing list