Hell all.
I¡¯m trying to estimate mixed logit model using MSLE.
In order to see that mixed logit model works better than simple logit model (
the logit model with fixed coefficient)
I simulated a dataset with random coefficients and tried to fit the data with
both mixed logit and simple logit model.
Because my mixed logit model contains analytically intractable integrations, I
applied simulated method (maximum simulated
Likelihood estimation) to estimate parameters.
I tested with both nlminb() and optim() function with ¡®L-BFGS-B¡¯ method. I
had to use them because my problem is constrained optimization
(I¡¯m trying to estimate underlying variance of random coefficients).
The thing is that when I used nlminb(), it ended up with ¡®false convergence¡¯.
When I used optim() with ¡®L-BFGS-B¡¯ method, I could get a set of parameter
estimates and likelihood estimates.
But the log likelihood value was worse than that from simple logit model (fixed
coefficient logit model).
(My conjecture was that mix logit model should give better result because the
data was generated from random coefficient model).
Here my question are that
1) Is it possible that mixed logit model gives poorer result than simple
logit model even the dataset is generated from random coefficient model and why?
2) If mixed logit model should give better result (in terms of likelihood
and other fitting criterion), I think there are some problems with programs. Is
there any function doing mixed logit analysis in R environment?
Thank you!
[[alternative HTML version deleted]]
______________________________________________
R-help@stat.math.ethz.ch mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
