0
$\begingroup$

I read a paper using genenalized Difference-in-Difference

$Y_{it}$ = $\alpha$ + $\beta$ $(Leniency Law)_{kt}$ + $\delta$$X_{ikt}$ + $\theta$$_t$ + $\gamma$$_i$ +$\epsilon$$_{it}$

where i,k, and t index firms, countries, and years respectively. $X_{ikt}$ is a vector of the different firm, country, and industry control, while $\gamma$ and $\theta$ are firm and year fixed effects.

The variable of interest here is $(Leniency Law)_{kt}$. Dasgupta,2019,p.2597 documented that this variable equals 0 before the passage of the leniency law in country k, and 1 afterward.

From the Table IA2 in their Appendix, I am wondering how we can interpret the coefficient of some independent variables which are yellowed as below enter image description here

From my understanding, 0.0073 is explained as: asset growth in treatment countries increase 0.073% compared to the change in asset growth in the control country after anticollusion. I am wondering it is correct and how to explain the rest.

$\endgroup$

1 Answer 1

2
+50
$\begingroup$

So if I understand this question correctly you are asking about adding controls in the DD. First, let us consider why one would do this. There are essentially two reasons,

  1. Reduce Error Variance
  2. Aid in Identification

The first reason is simple enough. Adding in covariates helps to explain some of the variation in $Y$ and so reduces the variance in our residuals. This is desirable if we are after significance levels. The second reason is a little more tricky but essentially comes back to the same idea of explaining some of the variation in $Y$. Recall, that we need that the common trends assumptions holds between treatment and control. This is the key identifying assumption. It is possible that when we look at the pre-trends prior to controlling for a covariate we see a failure in common trends. It is possible by adding in the missing covariate we now control for the difference in slope. Notice, covariates only affect identification if they are at the group/time level. There are other more complicated cases to add covariates but I do not think this is relevant here (here is a paper if you are interested).

So now that we know why we do this, let us think about how. Well, we can simply add them to the regression. The main condition is that we really need these to be exogenous (unless of course, we are doing something more complicated like in that linked paper). Given that, if we are after (2) then we want something like,

$$Y_{igt}=\alpha + \beta D_{gt} + \gamma_g + \theta_t + X_{gt}\delta + e_{igt}$$

If we just want to reduce standard errors we can do,

$$Y_{igt}=\alpha + \beta D_{gt} + \gamma_g + \theta_t + X_{igt}\delta + e_{igt}$$

The within-group variation will have no effect on identification but may help lower the standard errors.

In terms of interpretation, we may interpret the $\beta$ as we usually do. The independent covariates do not really matter though unless we are interested in them. They do not affect the interpretation of the treatment effect, though they may allow us to claim better identification. If they are truly exogenous we may interpret them casually, otherwise we interpret them as we normally interpret multiple regression coefficients. So for example, we can interpret the coefficient on ROA as saying that if we held all other variables constant but increased ROA by 1 unit we would expect a -1.111 decline in the outcome, on average. However, in practice, I would caution against assigning too much weight to these coefficients because often the authors include them to reduce standard errors and are not the variables of interest. Finally, here are a set of notes by Pischke on DID that include adding covariates. And, here is another set by Wooldridge. Perhaps, these will be helpful for you.

$\endgroup$
7
  • $\begingroup$ Hi @Ariel, thank you for your help, my main focus is explaining the coefficients, (1)can you please help me, for example, explain the meaning of 0.073 in this picture then ? $\endgroup$ Jun 10, 2021 at 2:29
  • $\begingroup$ (2) So, the main difference between your two equations are the $X_{igt}$. meaning that adding the firm-level independent variables apart from group-level (country-level) independent variable to reduce std.err? And it seems that , adding firm-level independent variables mean reducing within-group variation? $\endgroup$ Jun 10, 2021 at 2:33
  • 1
    $\begingroup$ The coefficient on Leniency Law may be treated as you would normally treat your indicator of treatment. This is the ATT. Covariates refer to the variables denoted by $X$. The difference in the two equations is the level at which covariates are added. Adding covariates at the individual level has no effect on the identification of the treatment effect but may reduce standard errors which is desirable in most cases. The notes I linked by Pischke might be useful in understanding this. $\endgroup$
    – Ariel
    Jun 10, 2021 at 2:39
  • 1
    $\begingroup$ Not quite, it is the average difference between pre- and post for the treated population. Because the ATT is the effect on the treated population of the intervention. Remember, DID is using the common trends assumption to try to tell us what would have happened to the treated population had they not been treated. It then compares the realized treated outcomes to the predicted untreated counterfactual. $\endgroup$
    – Ariel
    Jun 10, 2021 at 2:50
  • 1
    $\begingroup$ Thank you so much, Ariel, it is clear to me now $\endgroup$ Jun 10, 2021 at 2:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.