Hi, I am running some rather complex mixtures of beta regressions using the
betamix() command from the betareg package (V. 3.0-4). If I am doing
exploratory regressions with only one random starting value (nstart=1) I
obtain results which converge after about 100 iterations. However, if I run
regressions with nstart=100 random starting values I obtain solutions
amended with a warning that no convergence resulted after 200 iterations. I
conclude that 1) the initial regression converges to a local optimum, 2) the
likelihood function is flat. However, at least in theory the EM algorithm
should converge to an optimum. Thus, I am wondering about the meaning of the
warning: does it mean "no convergence at all, i.e. after 5000 iterations of
the BFGS algorithm" or convergence but after more than 200 iterations?
Moreover, is there a way out the flat trap? I have started to experiment
with the SANN algorithm within the betamix() function but it is slow (as
expected)...
Thanks for any hints
Chris
--
View this message in context:
http://r.789695.n4.nabble.com/convergence-warning-in-betamix-tp4693249.html
Sent from the R help mailing list archive at Nabble.com.
