[Limdep Nlogit List] Request to help with estimating marginal effects

[Ashu Kedia]

Hi Professor William Greene,
I hope you are doing well.

I am Ashu Kedia, a PhD student in transportation engineering at the University of Canterbury, New Zealand.

I am having a few doubts regarding marginal effects calculation. It will be a great help if you could clarify my doubts.


1.       I am using Nlogit 5 version to carry out estimations, but have been following the Hensher et al. (2005) book to understand the process, which might have used Nlogit 3 version.

The book says (on pg. 398 and 399) that the Marginal effects as reported are ‘true marginal effects multiplied by 100’. So, we need to divide the reported values by 100 to obtain the true marginal effect value. Now, I presume that this is applicable to that particular version of the software (i.e. Nlogit 3 or 4).



However, I am using Nlogit 5 version, and it is not clear to me whether the above convention is applicable to the marginal effects results given by the Nlogit 5 version as well?

I am asking this because the marginal effects obtained in my study are already very small in magnitude. For example: they vary between 0.08 and 1.06. If I divide such small values by 100, it will be exceptionally small.

Is this not uncommon?

Could you please advise on this?



2.       The examples of marginal effects calculations given in the book, Hensher et al. (2005), are for the MNL model specification. Do the syntax remains the same if marginal effects are calculated for a mixed logit model?



3.       Hensher et al. (2005) book has shown two different methods to calculate marginal effects for continuous-level data and categorically coded data.

Details of attributes and levels adopted in my study are as follows:

Alternatives: 2

Attributes: continuous (e.g. cost, time, etc.) – 4 attributes with 3 levels in each; categorical (i.e. qualitative attributes; e.g. type of a facility, such as residential or commercial) – 3 attributes with 2 levels in each.

Model type: Mixed logit model with heterogeneity in mean

Marginal effect estimation method: ;Pwt (probability weighted average)



Question : I have estimated the marginal effects of the continuous level attributes, using the method given on page no. 397 (but applying changes to the syntax that are required to estimate a mixed logit model).

                     To estimate the marginal effects for the categorically coded variables, the book shows the ‘Simulation’ method to be used.



However, when we use these two different methods separately to estimate marginal effects (i.e. once for continuous variables and then for categorical variables), two sets of coefficient values (i.e. parameter estimates for the model) are obtained. But we need only one set of coefficients for the estimated model. Which one of those two sets of estimates are to be adopted for the final reporting?

It will be a great help if you clarify these questions. I am happy to provide more details, if needed.

Kind regards
Ashu

Ashu Kedia
PhD Student,
Department of Civil and Natural Resources Engineering,
University of Canterbury, Christchurch 8140
New Zealand
________________________________
From: Ashu Kedia
Sent: Tuesday, December 10, 2019 12:57 PM
To: limdep at limdep.itls.usyd.edu.au
Subject: Request to help with estimating marginal effects

Hi Professor William Greene,
I hope you are doing well.

I am Ashu Kedia, a PhD student in transportation engineering at the University of Canterbury, New Zealand.

I am having a few doubts regarding marginal effects calculation. It will be a great help if you could clarify my doubts.


1.       I am using Nlogit 5 version to carry out estimations, but have been following the Hensher et al. (2005) book to understand the process, which might have used Nlogit 3 version.

The book says (on pg. 398 and 399) that the Marginal effects as reported are ‘true marginal effects multiplied by 100’. So, we need to divide the reported values by 100 to obtain the true marginal effect value. Now, I presume that this is applicable to that particular version of the software (i.e. Nlogit 3 or 4).



However, I am using Nlogit 5 version, and it is not clear to me whether the above convention is applicable to the marginal effects results given by the Nlogit 5 version as well?

I am asking this because the marginal effects obtained in my study are already very small in magnitude. For example: they vary between 0.08 and 1.06. If I divide such small values by 100, it will be exceptionally small.

Is this not uncommon?

Could you please advise on this?



2.       The examples of marginal effects calculations given in the book, Hensher et al. (2005), are for the MNL model specification. Do the syntax remains the same if marginal effects are calculated for a mixed logit model?



3.       Hensher et al. (2005) book has shown two different methods to calculate marginal effects for continuous-level data and categorically coded data.

Details of attributes and levels adopted in my study are as follows:

Alternatives: 2

Attributes: continuous (e.g. cost, time, etc.) – 4 attributes with 3 levels in each; categorical (i.e. qualitative attributes; e.g. type of a facility, such as residential or commercial) – 3 attributes with 2 levels in each.

Model type: Mixed logit model with heterogeneity in mean

Marginal effect estimation method: ;Pwt (probability weighted average)



Question : I have estimated the marginal effects of the continuous level attributes, using the method given on page no. 397 (but applying changes to the syntax that are required to estimate a mixed logit model).

                     To estimate the marginal effects for the categorically coded variables, the book shows the ‘Simulation’ method to be used.



However, when we use these two different methods separately to estimate marginal effects (i.e. once for continuous variables and then for categorical variables), two sets of coefficient values (i.e. parameter estimates for the model) are obtained. But we need only one set of coefficients for the estimated model. Which one of those two sets of estimates are to be adopted for the final reporting?

It will be a great help if you clarify these questions. I am happy to provide more details, if needed.

Kind regards
Ashu