Thanks. Here is an example script: if I run it twice successively, the
estimates change as well as the size of the sample.
Your suggestion to introduce missing values seems to have stabilized things
(the problem seems to have occured though, but only once in several tests.)