Should I use a Fractional outcome regression?

[Dan Eran]

I have panel data with an outcome variable that ranges from 0-0.6 and theoretically can get up to 1 (but never does). t=20 and n=130.
The outcome variable is the number of positive sentences a person says divided by the number of all sentences a person says. There are, of course, countless combinations. e.g., 0/100, 6/13, 1/128, 240/4578, and so forth.
Can I use the "classical"

xtreg Y X  i.year, fe vce(cluster unit)

or should I use a fractional outcome regression?

Thanks,
Dan

[Fei Wang]

Yes, -xtreg- is a good start for your case. You may further use -fracreg logit- or -fracreg probit- with Mundlak device to incorporate fixed effects. Jeff gave an example code in #4 of https://www.statalist.org/forums/for...for-panel-data. Replacing "oprobit" there with "fracreg logit" or "fracreg probit" will do.

[Emmanuel De-Graft Quarshie]

if my regression has low r squared and the variables insignificant, can i rely on the module given the hypothesis? im confused and need assistance.

[Weiwen Ng]

if my regression has low r squared and the variables insignificant, can i rely on the module given the hypothesis? im confused and need assistance.

Is this related to the original question? If not, you are better off starting a new question - and it would help to provide some output and a quick description of your study and data. As a very general answer, low R^2 means that your independent variables aren't strongly related to the outcome. This is a separate problem from the p-values being over 0.05. It's quite possible to have a low R^2 and statistically significant and substantively meaningful associations (I work with some data like that quite regularly).

If your p-values are all > 0.05 and the magnitude of the associations is small, then assuming the data aren't erroneous and you're using a type of regression that's at least approximately correct for the outcome, then there's nothing to do about it per se. The question of can I rely on the result is the wrong one to ask. Your explanatory variables just aren't working. You can ask why they aren't, but there may not be anything to do about it except collect new data and measure things better (and put more thought into what is the correct thing to measure).

If you have a fractional outcome, then a fractional logit model is probably the best model to use. However, I think many researchers would accept the use of linear regression and its derivatives. It's unlikely that you would get non-significant effects with linear regression and significant ones with fractional regression (or vice versa), although you should feel free to try.