DiD panel regression with one treatment group and one control group

[Guest]

Hello,

I am analyzing the impact of the Corporate Sector Purchase Program (CSPP) by the ECB on the firms debt capital structure. I have got firm characteristics such as total assets, current ratio etc. and also fixed effects as control covariates. The dependent variable is leverage (I use various proxies for the dependent variable leverage). My regression model is as follows:

reghdfe leverage CSPP##post firm chateristics,absorb(fixed effects)

CSPP is equals one if the firm received the treatment and is therefore is classified as an eligible firm (i.e., if the firm meet die ECBS eligibility criteria under the CSPP); and zero otherwise (noneligible firm). Post is equal to one for the period after the CSPP announcement. With the dummy variable "CSPP", I can see now how eligible firms differ in the change of their leverage compared to noneligible firms.

My question now is, how can I measure the effect of the CSPP for noneligible firms only? I adjusted the previous regression model with a modification of the dummy variable "CSPP". CSPP now is equal to one if the firm is classified as a noneligible firm, and zero otherwise. I ran the regression and got exactly the same results as with the first CSPP dummy variable but with a different sign. So the effects of the firms are always exact the same, but in a different direction (i.e., eligible firms increase their leverage by 2.34 pp and noneligible firms decrease their leverage by 2.34 pp).

How is this possible? Are these results to be expected because Stata measures always an absolute change and adjust the results by the sign of the respective group or could there be an error in my dataset? I also tried this regression:

reghdfe leverage Eligible##post Noneligible##post firm charateristics,absorb(fixed effects)

and both dummy variables were omitted due to collinearity.

I really do not know what to do about this.

Kind regards!

[Carlo Lazzaro]

Guest:
why not using -didregress- or -xtdidregress-?

[Guest]

Hi Carlo,

because I run different regressions, some of them should only include the dependent variable and the interaction term CSPP##post, without controlling for heterogeneity. So far as I understand these commands, group() would kind of violate this, right? And so far I do not know how to include my firm charateristics using didregress or xtdidregress.

However, the issue that I described above stays, I again receive the same results but with a different sign.

Kind regards,
Guest

[Jeff Wooldridge]

Guest: What do you mean by the effect on the "non-eligible" firms only? Those firms have to be used as the control group to estimate the effect of the eligible firms. You are, in effect, using the controls to create a counterfactual as to what would've happened to the eligibles if the intervention hadn't occurred. Are you talking about trying to see the differences over time on the control group? Because that's all you can do, and that would be the coefficients on the time dummies.

There is a parameter, the average treatment effect on the untreated, that can be estimated. But that is an estimate of the treatment effect in the control state. If you don't interact your controls with the treatment dummies then that estimate is the same (not the opposite sign) as the estimated average treatment effect on the treated.