Large standardised coefficients in DiD

[Henrik Dalriksson]

Hi,

I am examining whether a regulatory change affects profit margins (PM) and return on assets (ROA) using a difference-in-difference test.
Firms are divided into treatment (T=1) and non-treatment (T=0). The variable POST=0 before the adoption of the policy and POST=1 after the adoption.

Fixed effects used: Country-year, industry-year, firm
Standard errors are clustered on a firm-level

The problem is that my coefficients are very large in relation to the standard deviation. This is especially true for the non-DID estimators (T and POST).
Is this something that I need to worry about? Is it problematic that they are also statistically significant.

Please regression:

VARIABLES	PM
T	0.180***
POST	-0.726***
T*POST	-0.0199

Note that the standard deviation of PM is only 0.15!!!

VARIABLES	ROA
T	0.133***
POST	-0.345***
T*POST	-0.0146

Note that the standard deviation of ROA is 0.22.

Many thanks!

[Clyde Schechter]

Well, remember that the T coefficient represents the difference in expected outcome (PM or ROA, as the case may be) between the treatment and control groups before the regulation went into effect. Similarly, the POST coefficient represents the change in expected outcome following implementation of the regulation in the control group. So you have a pretty large difference between the treatment and control groups even in the absence of the regulation, and you have a strong secular trend apart from the regulation. It raises questions about whether you have selected an appropriate control group, and whether the regulatory effect may be too small to detect relative to other sources of ongoing change.

One thing I would do is graphically explore the outcome trends over time in both groups, especially during the pre-era. If the trends are more or less parallel, even though widely separated, then I wouldn't be terribly concerned.

[Henrik Dalriksson]

Many thanks @Clyde I will do that!

In this case there should not be a material difference between treatment and control in the baseline as they are matched pairs. The expected outcome following the implementation of the regulaton for the control should as well be small.