IV probit and fixed effects with large panel

[oliver wei]

Hi

I am currently running ivprobit with an unbalanced dataset. Assume that the panel has more than half a million observations, and I want to take user-level fixed effects into command; how can I achieve with speed?

ivprobit y (x = z) control1 control2 control3 i.timeperiod i.userid2, vce(cluster userid)

I have been trying to solve this problem for weeks. Simply put in this command with less than 10k observations is ok, but if there are too many observations, the computer takes forever to run.

Thanks in advance,

Oli

[oliver wei]

Any kind suggestions would be much appreciated!

[FernandoRios]

have you considered Correlated random effect model?
So instead of using individual fixed effects (which is not right due to incident parameter problem), you include persons average characteristics.
It works quite well with Linear models, and reasonably well with nonlinear ones.
F

[oliver wei]

have you considered Correlated random effect model?
So instead of using individual fixed effects (which is not right due to incident parameter problem), you include persons average characteristics.
It works quite well with Linear models, and reasonably well with nonlinear ones.
F

Thanks a lot. I will try that.

[oliver wei]

have you considered Correlated random effect model?
So instead of using individual fixed effects (which is not right due to incident parameter problem), you include persons average characteristics.
It works quite well with Linear models, and reasonably well with nonlinear ones.
F

what command would you use in this case? for iv and probit with CRE? xteprobit? For datasets with huge observations, can this work?

Thanks again

[FernandoRios]

I have a command I written for something like this.
First follow https://friosavila.github.io/chatgpt/fra_03_30_2023/ and install the command "fra"
then install -cre-

fra install cre

Then you can do this:

webuse laborsup
** with Fixed
ivprobit fem_work fem_educ i.kids (other_inc = male_educ)
** using CRE
cre , abs(kids):ivprobit fem_work fem_educ (other_inc = male_educ)

They will be slightly different, but lose enough for analysis
F