Probit/Logit regression with panel data.

[Jim Campasano]

I would like to run a panel regression. My y variable is binary and will be 0 most (>90%) of the time. To start, I have one X variable.
My dataset has weekly observations and the sample spans 20 years. I have approximately 700 to 2000 observations each week.

When I run this regression -- either probit Y X or xtprobit Y X -- does not converge. What can I do to rectify this?


Also, once I get this to work, I have other X variables I want to add. Which command should I use, given I have panel data.

Thanks so much!!

[Maxence Morlet]

Sounds like a large T small N panel.

One way could be xtlogit, but that would not address the enormous amount of time periods.

I would go LPM, with the community contributed command reghdfe.

Make your model as flexible as possible.

Also, if you have panel data, you will want to include unit and time dummies to address unobserved heterogeneity.

[Jim Campasano]

Correction. My original y variable was NOT binary. I was running Y X using reghdfe and clustering standard errors. I believe (based on your input) I can convert my Y to a binary variable and then rerun the regression. That should take care of it and I can interpret my coefficients.
Thanks!

[Carlo Lazzaro]

Jim:
As an aside to Maxence's insightful reply, please note that -xtlogit- was not developed for T>N panel dataset (rather the other way round).
In addition, due to parameter incidence bias, -xtlogit,fe- is a conditional -fe- estimator.
As far as convergence issue are concerned, with a single predictor only, it may be that the MLE came across a flate region in its way to the top and gasped.
Due to what above, I would stick with Maxence's suggestion about LPM.
As a sidelight, some more details on the origin of the massive amount of zeros that you mention may help with providing different hints, if feasible.