xtivreg, the R-squared and F-stat issue

[Arjen Steiner]

Hello everyone,

I want to report the R-squared and the overall F-test in my regression table.

I use the command xtivreg, fe, as a panel data and some endogenous variables to instrument. Also, I account for heteroskedasticity, so i use vce (r) option.

• Problem 1 - the R-squared: Which one should I report? the Within, the Between and the Overall R-squared?
  
  Or should I use regress 2sls and report its R-squared?

• Problem 2 - the overall F-test: The overall F-statistic is missing. I need it at least for the 1st stage.

Is it fair to report the F-stat of the 1st stage without the robust option? (because i have it without vce (r))

Or both stages should account for heteroskedasticity?


Note: I use panel data of multiple firms i over time t.

My regression is the following: Y_i= β₀+β_{1 *} P_it+β_{1 *} S_it+β_{2 *} D_it+ FE +e_it (with S_it and D_it endogenous var.)

Thank you very much in advance for your help !!

[Andrew Musau]

• Problem 1 - the R-squared: Which one should I report? the Within, the Between and the Overall R-squared?
  
  Or should I use regress 2sls and report its R-squared?

You should probably avoid reporting the OLS R-squared as it has no clear interpretation (see https://www.stata.com/support/faqs/s...least-squares/ ). You may report the within-R-squared for the fixed effects model and define what it is.

• Problem 2 - the overall F-test: The overall F-statistic is missing. I need it at least for the 1st stage.

One reason that it is missing may be that you have fewer clusters than variables, in which case you should not be clustering at all as the bias from a small number of clusters outweighs any benefits you get from clustering. However, with no actual Stata output, it is difficult to advise you.

[Arjen Steiner]

Thanks for you answers ! They are very helpful!

I would like to know if it is also good to report the overall R-squared, as it takes into account both the wihthin and the between variation in the dependent variable?

I unsderstood that it assess how well the model fits the data. Usually, the papers using the fixed effect specification which report there R-squared do not precise which one the use unfortunately..

For the second problem, I detected heteroskedasticity so I have to use the command xtivreg, fe with the vce (robust) option for my estimates to be correct. Should I run this regression and after perform the testparm to ge the overall F-stat ?

Thanks

[Andrew Musau]

Thanks for you answers ! They are very helpful!

I would like to know if it is also good to report the overall R-squared, as it takes into account both the within and the between variation in the dependent variable?

which one the use unfortunately..
I unsderstood that it assess how well the model fits the data. Usually, the papers using the fixed effect specification which report there R-squared do not precise

Between variation does not matter in fixed effects models, so one will usually be interested in within-variation in these models. But at the end of the day, it's your paper and your call. I would assume that if the papers do not report what R-squared statistic that they are reporting, then they are reporting the conventional OLS R-squared. If you report the within R-squared, then you should make this explicit. More openness is good scientific practice. For your second question, I have already addressed it in #2.