Re: Incorporate lognormal distribution into panel model

[Bierlaire Michel]

The lognormal distribution is not always ideal for numerical reasons. 

In Monte-Carlo, draws from the normal are generated, and used as an argument for the exponential. If the draws are large (for instance, when the mean and/or the standard error are large), the exponential may generate an overflow. 

If the problem occurs at the first iteration, start the estimation with small values for the mean (typically 0) and the standard error (e.g. 0.01, do not use zero). 

If the problem occurs in the course of the algorithm, consider another distribution. 

On 29 Apr 2022, at 20:57, Ning Siman <ning...@hotmail.com> wrote:

Hi Prof Bierlaire,

 

Hope you are doing well!

I have a question about the lognormal distribution model.

I use my own data, assuming the coefficient of a price variable being lognormal distribution. When plug in the model, I kept getting “Log likelihood (N = 1372): -1.797693e+308 Gradient norm:        inf Hessian norm:         inf “.  

However, when assumes normal distribution of the price variable, the model works fine.

Could you give me some suggestions on how to solve the inf Hessian norm problem? 

 

Thank you,

Siman