Maximum Likelihood

[bk]

How do I estimate using ML instead of the Bayesian approach?

Thank you for your help!

[MichelJuillard]

For each estimated parameters, you enter only the name of the parameter and the initial value for the optimization. The exact syntax is in the manual

Best

Michel

[bk]

Sorry, but I don't get it to work properly.

Do you know of any .mod files on the web that I could use as a point of departure?

[MichelJuillard]

Not at the tip of my finders. You rather send the *.mod file that isn't working

Best

Michel

[bk]

Here is the .mod file and the data. This is just a test model, so the parameter values, the parameter distributions and the data are more or less arbitrarily chosen.

[bk]

Here's an edited version of liten_estML.mod. I think I got it anyway?

[MichelJuillard]

Yes, it is correct now

Best

Michel

[glucke]

May I initiate one question? I'll use this .mod file and this data for illustrating.

It concerns that results of MLE depends on optimization algorithm. When we use “mode_compute=4” results are looked OK. But if another optimization algorithm is used (for example fminsearch) results became strange (likelihood functions is better, but few t-statistics are extremely large. It should mean that bound is near). It should be noted that problem of bound happens very often when I finished optimization.
Let’s describe an example. When we use “mode_compute=4” with all restrictions as at file, target (minus log-likelihood) functions value is larger then 79. All t-stats are reasonable. If we use “alternative search” with all restrictions as at file, target functions value is about 75.3. But parameters are equal to bound. So, t-stats are extremely high.
When we use “mode_compute=4” without parameters restrictions, target functions value is about “41,59”. “54,46” is highest value of t-stat. When we use “alternative search” without parameters restrictions, target functions value is about “41,56”. But “10097” is highest value of t-stat. So, we have extremely high difference at t-stat even with situation when likelihood function values are close.

It’s really interesting to hear any comments or ideas about this situation.

PS File “res.txt” contains results of optimization of described cases.