endogeneity

[Omar Shaher]

Dear researchers,
I am using the generalized DID model “Two way-fixed effect”. I have more than two groups and more than two periods. The period of my study is from 2000 to 2019. There is an event has been happened in 2005. And, I am applying the unbalance data for a set of firms across the years.

The main is to examine the effect of the event on the dependent variable.

The code is

Xtreg Depv Firm_Age firm_size Event Fim_age*Event Firm_size*Event i.year, fe cluster (COMPANY)

Where:
Event: dummy variable coded one in the year of adoption and for all subsequent years, and zero otherwise.

All the independent variables are continuous variables and time-varying variables.

The thing is I need to assess the endogeneity issue in the above-mentioned regression. I have resorted to the literature, and I have found that the lagged dependent variable technique has been used to assess the endogeneity. However, I have found that Paul Allison has mentioned that using the lagged dependent variable in the mixed model usually leads to severe bias. Here is the link:
https://statisticalhorizons.com/lagg...dent-variables

However, I have used it and I have got different results from those that I have got in the above-regression.

Thus, is there any method that you can recommend assessing the endogeneity for the DID?

Many thanks in advance.

[Omar Shaher]

Ok, I have found a paper has mentioned several techniques that have been used by previous studies to assess the endogeneity which are:

1. Graph dynamics of effect
2. To see if the effect is persistent
3. DDD
4. Include time trend specific to treated states
5. Look for effect prior to intervention
6. Include lagged dependent variable

And, here is the link
https://economics.mit.edu/files/750

I have tried hard to find a paper that used the second option “to see if the effect is persistent” but I couldn't find it. Thus, I am guessing that I should do the following points:

1. Identifying the year of the event.
2. Excluding observations in the year of the event to more cleanly compare the coefficient of the covariates before and after.
3. Excluding all firms that have not exposed to the event.
4. Firms should have more than one year of data before and after.
5. Then running the regression in the and comparing the results

I don’t know if this is correct or not?
Any suggestions???

Thank you in advance.