Panel data estimation with fractional dependent variable and two-way fixed effects

[Kamola Abdurasulova]

Hi fellow researchers,



I have a panel dataset with 67 counties and T=9. My dependent variable is a farm plan completion rate between 0 and 1 and I have zero completions too. If my goal is only to see the marginal effects, can I use a linear model with regdhfe two-way fixed effects (to control for time specific and county specific unobserved heterogeneity) instead of fractional logit or other fractional regressions?


Thank you beforehand.

#fractionallogit #panel #reghdfe #fractionaldependentvariable

[George Ford]

My vote is yes.

Your dataset is not so burdensome that you couldn't check it using i.county i.year in a fracreg or glm variant.

[Jeff Wooldridge]

You can always use a linear model for any outcome (provided there's no issue with actual data censoring). I recommend that. Putting in a full set of country dummies in fracreg will cause an incidental parameters problem. But you can include the time-averages of the time-varying controls to allow correlation between the heterogeneity and the x(i,t). This is the correlated random effects approach in Papke and Wooldridge (2008, Journal of Econometrics). The average partial effects are often similar to coefficients in the linear FE estimation.

With N relatively small (but bigger than T), I'd probably rely on the linear model mostly.

[Kamola Abdurasulova]

George Ford Thank you very much for your reply!

[Kamola Abdurasulova]

Jeff Wooldridge Thank you very much, professor!

[Kamola Abdurasulova]

Jeff Wooldridge Professor, which papers/books would you recommend to read in defense of using linear regression in this specific case?

[Marcelo Gonzalez]

Hello to all! I am brand new; I'll try to do not to ask the same. Although I have a similar question. I have a DV that is a proportion of SMEs that are offering green products (above the total of SMEs in a country).

I want to obtain the regression for 26 countries for the years 2012, 2013, 2015, 2017 and 2021. My goal is to obtain the results for fixed effects and random effects. Is it correct to use xtreg with the fe and re options? And then use the Hausman test to compare fixed and random effect models.


Many thanks in advance! Jeff Wooldridge Kamola Abdurasulova George Ford

#panel #xtreg #fractionaldependentvariable

[shem shen]

You can always use a linear model for any outcome (provided there's no issue with actual data censoring). I recommend that. Putting in a full set of country dummies in fracreg will cause an incidental parameters problem. But you can include the time-averages of the time-varying controls to allow correlation between the heterogeneity and the x(i,t). This is the correlated random effects approach in Papke and Wooldridge (2008, Journal of Econometrics). The average partial effects are often similar to coefficients in the linear FE estimation.

With N relatively small (but bigger than T), I'd probably rely on the linear model mostly.

If the N size is around 700 and T is around 10 (and the outcome var is bounded btw -1 and 1 but can be transformed to [0,1]), would it be strongly recommended to use the approach of Papke and Wooldridge (2008) instead of linear FE?

[Jeff Wooldridge]

Yes, I would use the correlated random effects fractional probit or logit in PW (2008). But you can compare the marginal effects with the linear model estimates.