Re: fminserch does not converge for every variable

yaoli <yaoli54@xxxxxxxxx> wrote in
news:260070.1214846233318.JavaMail.jakarta@xxxxxxxxxxxxxxxxxxxxxx:

Hi,

I'm doing a longitudinal simulation study to check if my modeling
approach could give the correct parameter estimate. To do this, I use
fminsearch function to find the MLE.

The code is:
options =
optimset('MaxFunEvals',100000,'MaxIter',100000,'TolFun',1.e-6,'TolX',
1.e-6); [x0new,fval]=fminsearch(@loglike,x0,options,Y,t);

where X0 is a vector of 17 parameters' true values and t is the time
points.

The problem is, for different generated data, either 14 parameter
estimates could be correctly found whereas the other 3 (always the
same 3) could not be, or the algorithm can not converge.

So could it because that my likelihood function is intrinsically too
hard to be handled by simplex?

Thanks in advance!

yaoli


17 variables is a huge amount to be trying to optimize, and even if you
successfully optimize your problem, you probably won't gain much insight
about your model. Look carefully at things like the condition number of
your problem, and other such signs of multicollinearity, do the
sensitivity analysis of your output with respect to every parameter, and
used the info to do all you can do to cut the numbers of parameters to
something reasonable.


--
Scott
Reverse name to reply
.