Dear listers,
I have estimated a set of regressions in which I relate the price each firm F charges in town T and year Y to a treatment that varies by firm and town and year, while controlling for Firm*Year FE:
Price(F,T,Y) = A + B*Treatment(F,T,Y) + FE(FY) + EPS(F,T,Y,)
On my understanding, which I'm happy to have corrected if I'm wrong, it is then appropriate to cluster standard errors by the panel dimension, which with Firm*Year FE should also be Firm-Years:
Luckily clustering by firm*year (or alternatively just by firm) does not seem to much affect statistical significance of my coefficients of interest here.
As a variation, I next wanted to try also regressions in which I weight each FTY observation with the fraction of the firm's sales in that year going to that town:
.
However, when I do so clustering costs a lot of significance.
That made me wonder whether I should be suspicious of my weighted regression results (or of all results), or whether there would be any reason to cluster differently when using weighted regressions, for example by the weighting variable or so?
Thank you so much!
PM
I have estimated a set of regressions in which I relate the price each firm F charges in town T and year Y to a treatment that varies by firm and town and year, while controlling for Firm*Year FE:
Price(F,T,Y) = A + B*Treatment(F,T,Y) + FE(FY) + EPS(F,T,Y,)
On my understanding, which I'm happy to have corrected if I'm wrong, it is then appropriate to cluster standard errors by the panel dimension, which with Firm*Year FE should also be Firm-Years:
Code:
reghdfe price treat, absorb(i.firm##i.year) cluster(firm##year)
As a variation, I next wanted to try also regressions in which I weight each FTY observation with the fraction of the firm's sales in that year going to that town:
Code:
reghdfe price treat [aweight=townsshare], absorb(i.firm##i.year) cluster(firm##year)
However, when I do so clustering costs a lot of significance.
That made me wonder whether I should be suspicious of my weighted regression results (or of all results), or whether there would be any reason to cluster differently when using weighted regressions, for example by the weighting variable or so?
Thank you so much!
PM
