Panel Data Ordered Logit not able to calculate.

[Andreas Baltin]

Dear Statlisters,

Quick question regarding ordered logit. I have a panel dataset from China with individuals for 2 timesframes. My regressand is personal spending (ordinal - don't ask why :D) and amongst my regressors are for example homicide rate per city or economic characteristics of individuals.

I do

xtset ID year
xtologit Spending Homicide EconomicCharacteristics i.city, vce(cluster city)

--> I include dummies for cities in the prediction and cluster by cities.
However, there are 349 separate cities in the dataset, and I am unable to compute this. Is my computer just too slow? Any toughts on this?

Many thanks in advance!

[Carlo Lazzaro]

Andreas:
what if you cluster standard errors on -panelid- instead of -i.city-?

[Andreas Baltin]

Andreas:
what if you cluster standard errors on -panelid- instead of -i.city-?

Hi Carlo,

I have tried clustering on -panelid-. It is still running and has been running for over 30 minutes now, so that does not seem to work either.

Best
Andreas

[Carlo Lazzaro]

Andreas:
-xtologit- models are computationally demanding.
The usual recipe is to start with a more parsimonius model (ie, less predictors) and see if any convergence/computation issue comes alive from the first step already.
In your case, another option would be to run the model with default standard errors and see if the model still takes so long to throw an outcome table.

[Andreas Baltin]

Andreas:
-xtologit- models are computationally demanding.
The usual recipe is to start with a more parsimonius model (ie, less predictors) and see if any convergence/computation issue comes alive from the first step already.
In your case, another option would be to run the model with default standard errors and see if the model still takes so long to throw an outcome table.

Thanks for your reply.

The issue is the i.city regressor. If I run it without this regressor it computes in seconds. Furthermore, I encounter the same problem if I run the model with default standard errors.

[Carlo Lazzaro]

Andreas:
let's get rid of -i.city- and go smoother with the remaining predictors.
In your research report you can justify your choice of omitting -i.city- in the light of the computational problems (that in all likelihood mean convergence problems) encountered.

[Andreas Baltin]

Andreas:
let's get rid of -i.city- and go smoother with the remaining predictors.
In your research report you can justify your choice of omitting -i.city- in the light of the computational problems (that in all likelihood mean convergence problems) encountered.

Thank you. This is exactly what I read somewhere else, and I will proceed this way.

As always very appreciative of your help Carlo!

Best
Andreas