Page 1 of 1

Defining model and likelihood function when using a (negative) lognormal distribution

Posted: 21 May 2025, 14:52
by Pea
Dear Hess and Palma

I have looked through the example "Mixed logit model on Swiss route choice data, uncorrelated Lognormals in preference space" in https://apollochoicemodelling.com/examples.html, and I am wondering if I understand the following correctly:

When using a (negative) lognormal distribution, the list of utilities should be just the same as when using e.g. a normal distribution - that is a transformation should not be made, so that a new variable logtt1 (database$logtt1<-log(database$tt1)) replaces tt1 in the following (correspondingly for tt2)

List of utilities: these must use the same names as in mnl_settings, order is irrelevant
V = list()
V[["alt1"]] = b_tt * tt1 + b_tc * tc1 + b_hw * hw1 + b_ch * ch1
V[["alt2"]] = b_tt * tt2 + b_tc * tc2 + b_hw * hw2 + b_ch * ch2


The only thing that should be changed is in apollo_randCoeff - and then the starting value - when using a (negative) lognormal distribution, is that correct?

But interpretation is different, right, the resulting number -1.9962 is the log coefficient and it has to be transformed like this: exp(-1.9962)*-1 to -0,1358505 which is the acutal beta mean (although it should probably be done using the delta method), is that also correct?

Best wishes
Pea

Re: Defining model and likelihood function when using a (negative) lognormal distribution

Posted: 27 May 2025, 11:37
by stephanehess
Yes, but the transformation afterwards is more complex than what you have put there. The mean of a lognormal is a function of both the mean and sd of the underlying Normal. You can find the formulae online quite easily

Re: Defining model and likelihood function when using a (negative) lognormal distribution

Posted: 07 Jun 2025, 15:55
by Pea
I see. Thank you for your help!