Re: Question about optimization (maybe of interest to Marc Vontres...)

On Oct 30, 2:13 pm, "sci.stat.math" <fra...@xxxxxxxx> wrote:
On Oct 30, 2:21 pm, vontres...@xxxxxx wrote:

Thanks for the paper. I have a copy of it. I used it for the
confidence intervals. I suspect that the problem is that the
information matrix for the five parameter likelihood is nearly
singular, so different inversion methods embedded in the optimization
leads the routine in different directions toward an optimum. If you
use different starting values and arrive at the same value of the
objective function, even with different parameter estimates, then
those parameter estimates are okay. My experience so far has been that
p is usually near zero or one, so I've really just have one prior
gamma.

But let's just see what you get with this simulated dataset of 20
observed values (zero observed values excluded):
N=c(1, 1, 5, 2, 2, 2, 1, 1, 1, 1, 4, 1, 1, 1, 1, 1, 1, 6, 6)
and the corresponding expected values:
E=c(0.025, 0.15, 0.75, 0.1, 0.1, 0.1, 0.025, 0.05, 0.05, 0.025, 0.4,
0.025, 0.025, 0.025, 0.025, 0.025, 0.325, 1.95, 1.95)
with starting values: alpha1=02, beta1=0.1, alpha2=2, beta2=4, p=0.33
The parameters I'm obtaining are:

alpha1=1.3366504, beta1=0.0100000, alpha2=2.0816098 beta2=3.9549975
and p=0.7765547

Now the original authors obtain totally different values especially
for p. I am wondering what you will get? Thanks


Let's take this offline. What is the answer that the authors get?

Mark

.