ApolloChoiceModelling forum

Dear Sir,

I attended your incredible Apollo course last November in London. Now, I am using Apollo to analysis my data. However, I am confused about which parameters should be chosen as random parameters in MMNL model. Should I only choose those parameters which were statistically significant in MNL model as random ones while keep others fixed?

Thank you for your time.
Shan

Dear Shan

it is always impossible to know a priori which parameters are the right ones in terms of allowing for random heterogeneity. It is quite possible that random heterogeneity exists for most or all your coefficients. If you only allow for random heterogeneity in some of them, you will of course have an inferior model, but you may also get biased findings in terms of the heterogeneity for others, or indeed model parameters in general. But of course, it is often computationally difficult to use random heterogeneity for all parameters.

In general, a good approach is to first incorporate some deterministic heterogeneity, i.e. including socio-demographic effects, for example. This will reduce the scope for remaining random heterogeneity. But then in the next stage, you should really test for random heterogeneity in all the key parameters.

In terms of your comment about MNL and insignificant parameters in that model, it is not unheard of to have a case where a parameter is insignificant in MNL but becomes significant in a Mixed Logit model

I hope this helps, although I'm aware it raises more questions than answers!

Best wishes

Stephane

Hi Stephane,

Thank you for your detailed reply. As you suggest, I incorporated socio-demographic factors into the Mixed MNL model, while assuming all parameters were random. From the analysis results, I found that some of the parameters were statistically significant based on their estimated standard deviations. Also, I examined some combinations of random and fixed parameters. My question now is that could I only keep parameters that consistently showed significant standard deviation as random, while others fixed?

Best,
Shan

Hi Shan

yes, you should be able to make those parameters non-random where the standard deviations are not significant. Keep an eye on your model fit too to see if you're not losing fit by doing so

Best wishes

Stephane