Well, the big difference is in the distributions, so when you say "the only difference between my model and the example model is that the response variable in the example is a continues variable and has a normal distribution while in my model the response is a dichotomous variable and thus has a binary distribution", you are making a ginormous change. I'll through two things out, neither of which is going to get this to start converging ( I don't think). First, since the Monte Carlo part of the procedure depends on the parameter starting values, make sure that the variance and covariance estimates refer to the logits of the data. Using the values from the normal distribution won't suffice. Second, why? Why dichoomize the data when you have a very good analysis treating the response variable as continuous? Dichotomization throws away a LOT of information, and immediately reduces the power of the analysis. For example, see the paper presented by Dr. Stephen Senn at the 2011 FDA/industry Statistics Workshop, in the short course, Statistical Issues in Drug Development. From Section 3, slides 29 and 30: Slide 29: Dichotomisation Prospects for a Cure I am pessimistic Most physicians seem happy with dichotomisation Most statisticians seem happy to indulge them "That's not my dpartment' syndrome We have to bring home the following message: Slide 30 DICHOTOMISATION IS VERY SILLY! Your problem points out one of the drawbacks Dr. Senn was trying to make. Steve Denham
... View more
