ApolloChoiceModelling forum

Hi

I am running a MMNL model with 5 random parameters (all negative lognormal) on a panel dataset with 500,000 observations and 3200 individuals. I am running models on a 40 core machine, but the estimation (BHHH) crashes when I specify more than 150 draws (MLHS draws). I have set gradient comparison, Hessian calculation, and the scaling parameters to FALSE.

My question is whether 150 draws are enough for 5 random parameters and whether I should be using different draw types and/or estimation routines. Also, would bootstrapping the estimation on fewer draws be advisable or bad practice?

Thanks for your help!

Apologies for the slow reply.

This is an extremely large dataset and you are likely to run out of memory. 150 draws is a very small number of draws, and I would consider using sampling of decision-makers with a larger number of draws