Hello,
this is technically not a Stata-specific questions, but since there's a lot of expertise in this forum, I dare ask it. This is about a Diff-in-Diff regression in a high-dimensional panel dataset. Specifically, I have bank data which I match with firm data via a credit register to observe the loan volume from banks to individual firms. I'm using the reghdfe command to run the regression. Since my dataset is confidential, I cannot share any specific data (or results). It's a very general question, though.
I have annual data from 2011 to 2018 which I collapsed into two periods: control (2011-2014) and treatment (2015-2018). TREAT is a time dummy that is one in the treatment period. My explanatory variable (the 2011-2014 mean of share of bonds in total bank assets) is time-invariant at bank level. My dependent variable is time-variant at pair level: It's the loan volume from bank b to firm f, detrended by its own 2011-2014 trend, then divided by its own 2011-2018 mean. (hence, when I collapse the dataset, the control period value of this variable is zero or very close to zero). FIRM_ID and BANK_ID identify the firms respectively banks.
My issue is that the coefficient of the interaction term is exactly the same with and without the absorb(FIRM) option, down to the seventh digit! t-stat and p-value change slightly and R2 drastically increases when absorbing the firm FE. This startles my because even if the firm fixed effect and the X_b are highly uncorrelated, shouldn't there be at least some tiny effect, one way or another? When I change X_{b} from a time-invariant to a time-variant variable, absorb(FIRM_ID) has an impact on the interaction term.
Also, when I run a purely cross-sectional regression where the dependent variable is the delta between the control and the treatment period, absorbing the FIRM_ID does indeed affect the coefficient of X_{b}:
What am I missing here? Is the time dummy also absorbing the firm FE? I can't wrap my mind around it.
this is technically not a Stata-specific questions, but since there's a lot of expertise in this forum, I dare ask it. This is about a Diff-in-Diff regression in a high-dimensional panel dataset. Specifically, I have bank data which I match with firm data via a credit register to observe the loan volume from banks to individual firms. I'm using the reghdfe command to run the regression. Since my dataset is confidential, I cannot share any specific data (or results). It's a very general question, though.
Code:
reghdfe Y_{bft} = i.TREAT##c.X_{b}, absorb(FIRM_ID) vce(cluster BANK_ID)
I have annual data from 2011 to 2018 which I collapsed into two periods: control (2011-2014) and treatment (2015-2018). TREAT is a time dummy that is one in the treatment period. My explanatory variable (the 2011-2014 mean of share of bonds in total bank assets) is time-invariant at bank level. My dependent variable is time-variant at pair level: It's the loan volume from bank b to firm f, detrended by its own 2011-2014 trend, then divided by its own 2011-2018 mean. (hence, when I collapse the dataset, the control period value of this variable is zero or very close to zero). FIRM_ID and BANK_ID identify the firms respectively banks.
My issue is that the coefficient of the interaction term is exactly the same with and without the absorb(FIRM) option, down to the seventh digit! t-stat and p-value change slightly and R2 drastically increases when absorbing the firm FE. This startles my because even if the firm fixed effect and the X_b are highly uncorrelated, shouldn't there be at least some tiny effect, one way or another? When I change X_{b} from a time-invariant to a time-variant variable, absorb(FIRM_ID) has an impact on the interaction term.
Also, when I run a purely cross-sectional regression where the dependent variable is the delta between the control and the treatment period, absorbing the FIRM_ID does indeed affect the coefficient of X_{b}:
Code:
reghdfe s1.Y_{bft} = X_{b} if TREAT==1, absorb(FIRM_ID) vce(cluster BANK_ID)
What am I missing here? Is the time dummy also absorbing the firm FE? I can't wrap my mind around it.
Comment