ApolloChoiceModelling forum

Hi

I have always thought that a higher number of draws will improve the model fit and the log-likelihood (LL) value. However, running a mixed MNL model With 22 parameters, 18 normally Distributed and 4 lognormally Distributed, the LL value decreased from -1591 to -1596 when I increased the number of (Halton) draws from 500 to 2000.

Could this be because Halton is not a good Choice of draws With such a high number of parameters to be estimated? In case yes, what would be the best Choice of draws?

Hi

please have a look at the discussion at http://apollochoicemodelling.com/faq.html

More draws is always better in terms of approximating the real LL function with the simulated LL (SLL) function. That doesn't mean the SLL will be better, but just that the bias in the SLL will be less

Stephane

OK.
I have read the FAQ, which says:

FAQ: Can I at least use fewer draws if I use quasi-Monte Carlo draws?
In theory, yes, but again, more is better. Care is also required in deciding which type of draws to use. Halton draws are an excellent option for models with a low number of random parameters, but the colinearity issues with Halton draws means they should not be used with more than say 5 random components.

Having 22 random parameters, this probably means I should NOT use Halton draws? So what would be the better option?

Hi

there are many alternatives. Have a look in the manual, which lists some that you can use in Apollo

Best wishes

Stephane