[Omar Shaher]

Dear prof. Clyde,

I mean I can separate the observations for firms before the event and applying the following code:

 xtreg performance age size growth i. year, fe cluster (COMPANY)

And applying the same code for the firm's observations after the event.

In that way, I can compare the coefficients of the variables before and after the event.

so, my point is if the DID just tells me the impact on the relationship and does not tell me which variable increase or decrease because of event, then I can do the above-mentioned method!!

What do you think?

[Clyde Schechter]

There is a good chance that the coefficients you derive from running the code in #16 separately in the before-event observations and in the after-event observations will give you results that resemble the before and after event marginal effects you got from the -margins- command after the interaction regression. They will probably not be the same: if nothing else the i.year variables will be different, and that could throw things off. There is some chance, particularly if there are strong time trends in the data, that the differing values of i.year in these three regressions (the interaction one from before and the two proposed here) will lead to really large differences between these coefficients and the earlier marginal effects.

There is also the question of in what sense you plan to "compare" these results to the earlier ones. You can do that impressionistically, but there is no good way to statistically compare them. The -suest- command, which would ordinarily be used for this purpose, does not support -xtreg-.

Finally, I'm still left a bit confused by "

so, my point is if the DID just tells me the impact on the relationship and does not tell me which variable increase or decrease because of event, then I can do the above-mentioned method!!" I don't know what you're trying to accomplish. These two regressions also do not tell you which variables increase or decrease because of the event. If you want to contrast the before- and after- event levels of the variables in your model, do things like -by:Event, sort: summ age size growth-, or to focus on differences between before and after event variables regressions like -xtreg age i.Event, fe-. But it's still not clear to me why you want to do that. This would be interesting backkground information to report, but it is not part of your stated research goals, at least so far as I understand them.

[Omar Shaher]

Dear Prof. Clyde,

There is also the question of in what sense you plan to "compare" these results to the earlier ones.

I really don't want to compare the results to each other as much as I want to show that the two methods can achieve the same purpose.

Regarding the following code:
by:Event, sort: summ age size growth

The stata showed me the following message:

varlist required
r(100);

As for the following code:

xtreg Age i.Event, fe cluster (COMPANY)

It seems very interesting, what is the name of this method or this regression as I want to report it in my research?

In the following results, I have got a positive coefficient for the event, does it mean the event increases the size? If yes, also the same thing for negative value, I mean if I got a negative coefficient means that the decreases in the event is attributed to the event?

------------------------------------------------------------------------------
             |               Robust
centered_S~E |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
     1.Event |   .1912655   .2157138     0.89   0.405    -.3188167    .7013476
       _cons |  -.0717246   .0808927    -0.89   0.405    -.2630054    .1195563
-------------+----------------------------------------------------------------
     sigma_u |  .17970973
     sigma_e |  .50495073
         rho |  .11242209   (fraction of variance due to u_i)
------------------------------------------------------------------------------

But it's still not clear to me why you want to do that.

yes, you have understood my research very well. But my view is that it is really interesting to know if the event affects the relationship between the dependent variable and the independent by using the DID model, but after I found that before the event, the sign of the coefficients are positive, while after the event there are a lot of negative coefficients, so, I was curious to know the interpretation for these negative signals. Thus, I have resorted to the previous literature (they have not examined the moderating effect and this is one of my contributions), and I have found that their interpretations of the negative sign were somewhat mixed, some of them indicates that the negative signal means that there is a decrease in the independent variable that led to an increase in the dependent variable, and there are those who said the opposite, so I wanted to resolve the matter by testing whether the event is responsible for increasing or reducing the independent variables, that is, to do so only to complete the image.

Note such an example of regression that has been mentioned in the previous literature.

Xtreg size Age growth i. year, fe

In other words, if the performance is the dependent variable and SIZE is the independent variable and I have got positive coefficient for SIZE before the event and negative coefficient for SIZE after the event, that means the event decreases the slope of the relationship between SIZE and performance, but It would be interesting to know if the slope of the relationship decreases because of the effect of event on SIZE or on performance, so I think I will perform the above method that you have mentioned once on the dependent variable and once on the independent variable, to complete the image. And, the codes will be as below:

  xtreg SIZE i.Event, fe cluster (COMPANY)
xtreg performance, i.Event, fe cluster (COMPANY)

I don’t know if that makes sense or not!!

[Clyde Schechter]

Regarding the following code:

by:Event, sort: summ age size growth

The stata showed me the following message:

Code:
varlist required r(100);

Sorry, my typo. There shouldn't be a : between by and Event.

Code:
xtreg Age i.Event, fe cluster (COMPANY)
It seems very interesting, what is the name of this method or this regression as I want to report it in my research?

It doesn't have a special name. You're just regressing Age on an indicator for event with a fixed-effects regression. It's a generalization of a paired t-test to the situation where you have more than 2 observations per COMPANY.

Thus, I have resorted to the previous literature (they have not examined the moderating effect and this is one of my contributions), and I have found that their interpretations of the negative sign were somewhat mixed, some of them indicates that the negative signal means that there is a decrease in the independent variable that led to an increase in the dependent variable, and there are those who said the opposite, so I wanted to resolve the matter by testing whether the event is responsible for increasing or reducing the independent variables, that is, to do so only to complete the image.

It is certainly reasonable to observe that performance changed in some direction after the event, and to wonder what the changes in the predictor variables were that accompanied it. And it's reasonable to pursue that with the kind of analyses we have been talking about here. What isn't reasonable is to speak of whatever associations you find as if they are causal. They may or may not be; with observational data you can't really know.

[Omar Shaher]

Dear Prof. Clyde,
Thank you very much for your reply. Greatly appreciated.
So, if I am regressing SIZE on an indicator for the event, and I have found a positive coefficient for the event in the generalized paired t-test, my interpretation will be like as below:

I have observed an increase in the size values after the event.

While, if I am regressing the same model, and I have found a negative coefficient for the event, then my interpretation will be:

I have observed a decrease in the size values after the event.

please correct me if I am wrong.

What isn't reasonable is to speak of whatever associations you find as if they are causal. They may or may not be; with observational data you can't really know.

I am a bit curious to know then what kind of data that can be used for causality or is there any reference you recommend for me to read.

Thank you very much in advance.

[Clyde Schechter]

The interpretations you propose are correct.

To have a solid case for causality of the Event on a pre- and post-event difference the Event would need to have been assigned by a randomization mechanisms. That is, it would have to be a randomized experiment rather than observational data. I don't know what kind of event we're talking about here, but since the overall context seems to be finance, I'm guessing that randomization is either impossible or impractical here. In that case, you can make a weaker case for causality by using one of the common strategies used in finance and economics to try to estimate causal effects from observational data. The two that are, as far as I can see, most common, are difference-in-differences estimation, or the use of instrumental variables. Both of these, of course, rely on assumptions that are, in principle, not verifiable directly from the data being used (though they occasionally are verifiable from other kinds of data in a few cases), which is why they are only weak evidence of causality.

The literature on causal inference is very large, and there are many schools of thought and approaches in this area. I can't think of a compact, easily-readable overview to recommend.

[Omar Shaher]

okay, now I see your points.

Prof. Clyde, I can't thank you enough for your answers, for your prompt response, for your great efforts, and for valuable comments.
You have changed my understanding of statistics. I am so grateful to you.

It is really an honor for me to get in touch with a great person like you.

Thank you.

[Omar Shaher]

Dear Prof. Clyde,

I know it is been a while, but it is a quick question regarding #17, please.
As you have mentioned:

If you want to contrast the before- and after- event levels of the variables in your model, do things like -by:Event, sort: summ age size growth-, or to focus on differences between before and after event variables regressions like -xtreg age i.Event, fe-.

To tell you in brief my case:
I have unbalanced panel data for a group of firms since 2009-2015. There is an accounting standard (Event) issued and firms have started to adopt these standards non-simultaneously. For instance, firm A adopted in 2010. Firm B adopted in 2012, and so on. Thus, I wanted to observe if there is an increase in a specific variable after the event or not. Let's assume that I want to observe the leverage, then the code will be:

xtreg leverage i.Event, fe

And as you have mentioned that it's a generalization of a paired t-test to the situation where I have more than 2 observations per COMPANY.

My question is: can I add i.Year to the code for the same purpose:

 xtreg leverage i.Event i.Year, fe

I mean, if yes, what is the difference?

Many thanks in advance.