Hello. I have been out of town.
For any of the optimization solvers, you should expect a certain amount of round-off error. This occurs because the underlying calculations use finite-precision arithmetic. There are several tolerance switches that can be changed (see options).
In general, 0.49999999999999999999999 and 0.5 are considered the same. Typically, you will see precision of 1.0e-6, no more.
As for (2), I am not sure what you are referring to with 1989 vs 1994 unique observations. Do you mean you are seeing 1994 observations in the solution set? And you expect 1989?
If you want me to look at this more carefully, please send me the code and data offline. You can contact me at matthew.galati@sas.com.
Thanks,
Matt