Hi all,

I am trying to run a cmp model between three binary variables: new_bf23, new_bf3 and z, where new_bf3 is observed only for observations with new_bf23==1, i.e., there is selectivity between new_bf23 and new_bf3. z is not conditional on the other two variables and can be observed for all data points.

The pair-wise cross-tabulation between these variables are as follows:
new_bf3
new_bf23 0 1
0 129
1 131 163
z
new_bf23 0 1
0 123 6
1 264 30
z
new_bf3 0 1
0 250 10
1 137 26


I want to allow for the following correlations in the model:
corr( new_bf23, new_bf3)
corr( new_bf3 | new_bf23 , z)
corr( new_bf23 , z )


I get the error message when I specify the selectivity in the indicators as new_bf23*$cmp_probit.
i.e., for the specification
cmp (new_bf23= x12_1 x12_2 x12_3) (new_bf3= x2_1 x2_2 x2_3) (z = x3_1 x3_2 x3_3), ind ($cmp_probit new_bf23*$cmp_probit $cmp_probit)

error message "discontinuous region encountered cannot compute an improvement" is observed

However, the model converges when:
1. any two of the above variables are modelled jointly and
2. when I do not specify the selectivity condition. I.e., ind ($cmp_probit $cmp_probit $cmp_probit)


Could someone help me out with why this may be happening? And what would be a good work-around?


Thanks in advance,

Aravinda