Partial effects for Log-Linear model

[azira muminovic]

Dear members,

I am currently working on my master thesis in empirics. My main goal is to investigate the newly proposed estimator for gravity equation (Flex estimator) and compare it with some other estimators. I collected data and did regression for 7 different models.

Currently I am having a problem trying to calculate the marginal effects for Log-Linear model (for other models it was quiet simple with incorporated Stata commands).

My dependent variable for Log-Linear model is ln(T_ij) where T_ij is number of sectors exporting from country _i to country _j (min 0; max 97), my independent variables are ln(distance) and 9 different dummies. I control using multilateral resistence terms (i.e. exporter and importer dummies).

I should calculate the average across the entire sample of the partial effects of each of the regressors on E(Tij∣xij) so that I need the partial effects on the exponential of the fitted values of lnTij, averaged over all observations.

I am not sure how to do this and if this particular specification of the partial effects for Log Linear model is to complicated for beginner like myself.

Thank you in advance for any comments, insights and suggestions.

Kind regards,

Azira

[Clyde Schechter]

If you show the actual code and output for this model, it is possible that someone can show you the way from there. While your description is fairly clear, it leaves relevant details to the imagination. To be sure you report exactly the code and output, please copy/paste directly from Stata's Results window or your log file into a code block on this Forum. (See FAQ #12 for information about code blocks.)

[azira muminovic]

This is code which I run to get regression results

global x "border island landlocked colony comcurr rta comlang bothin religion"
quietly reg lTij ldist $x exp_* imp_*, robust

These are the results

The question is basically how to get partial effects averaged over all observations.

[Joao Santos Silva]

Dear Azira,

I am delighted to see that you are interested in "Flex"

As for your question, to be able to compute the partial effects using a log-linear model you would need to make some (strong) assumptions about the distribution of the error term. That is known as the retransformation problem, and it is one of the big disadvantages of using models where the dependent variable is logged.

In the paper where we introduced the "Flex" we computed partial effects using the log-linear model making some standard (but unrealistic) assumptions about the error term (this is described in the paper), but I am not sure it is worth doing that because the results will be essentially meaningless.

Best wishes,

Joao

[azira muminovic]

Dear Mr. Santos Silva,

thank you for your enlightening informations. I am actually very excited to hear from you since I appreciate your work on trade-empirics a lot, and I have been studying it for a while.

Your answer has really clarified a confusion which I had with LogLin so that I will rather focus on other marginal effects for comparative purposes.

As for empirical application of Flex, I get approximately the same results as in your paper (I work with Dataset on 2-digit level of aggregation), so for an average graduate student it is simple enough to implement.

Thanks again and good luck with your future endeavors.

Azira

[Joao Santos Silva]

Thank you, Azira. Best of luck.

Joao