fatvegan

I've been using Stata's ml command for a while now. It's all quite nifty, but I couldn't figure out how to use the ml init part and I had a feeling that it would be a useful sort of command. When your likelihood function isn't globally concave there are no guarantees you're going to find the maximum you're looking for. The manual was pretty terse, and the book that explains it is $180 so I had just ignored it. But after running out of other options I started scouring the web for some example of it. And it's actually dead simple.

ml init beta:index=1 beta:_cons=0 /a=-4.566654 /b=-1.332323 /c=-0.54343

There is one equation and three parameters. All this command is saying that assume the coefficient on only regressor (mrw_index) is going to be in vicinity of 1 and the intercept in the vicinity of 0.

In this example the model failed to converge after 3 iterations with the hints and failed after 13 iterations without the hints. And that seems like a great improvement.