Sooo TBH, I tried re-running the constrained optimization presented here and got awful parameter values.

This is because the initial values pointed the algorithm in the direction of pi=0.4 rather than 0.6, so pi ended up estimated on the boundary at pi=0.5, and the rest of the estimates were goofy as a result. I thought I was sooo clever too, haha!

The EM algorithm is so sensitive to initial values, it’s driving me mad! A common suggestion appears to be multi-start, i.e. using multiple start positions and then choosing the best one based on the highest likelihood within a few iterations.

Multistart just doesn’t seem satisfying to me. I wish there were a more robust, not terribly complicated way to set up the EM algorithm on the right foot.

If anyone has any thoughts, please share!

]]>