Difference in Differences model

[Yassir Alouhmy]

Hi Everyone,

I am trying to evaluate the 2001 Portuguese drug law with regards to crime, using a difference in differences approach. I have collected data for 13 european countries for the period 1993-2007.
For one of my equations, the dependent variable is harmful_acts_rate which is a rate of violent crime per 100.000 inhabitants. One of the issues arising is the related to the fact that most of my independent variables are in per-capita terms. Meaning that the coefficients are spuriously significant, driven by the common population component in the dependent and independent variable. How can i solve this issue ?

Your help would be much appreciated.

[Clyde Schechter]

Well, if you can get separate numerator and denominator data, you can avoid this problem. So, if you can get harmful_acts_count, and similar counts of whatever your predictor variables represent you can use a count-variable regression model such as -poisson- and include the population number as the -exposure()- variable in that model.

As an aside, the title of your post is poorly chosen. Your question has nothing to do with difference-in-differences modeling: the same question would arise with these variables regardless of the study design. By using a tangential title you may have caused some people to skip over your post--there are relatively few of us here who respond to questions about DID. Moreover, in the future, people looking for advice on the common denominator spurious correlation problem will not find your post in their searches and will be unable to learn from this thread. In the future, please use titles that accurately describe the question being asked.

[Yassir Alouhmy]

Thank you for your advice.
I do have a question on difference-in-differences model, if you dont mind.

My actual model includes 19 countries for the period 1993-2007. Considering i'm working with crime as my dependent variable, i'm concerned about serial correlation, especially that i'm using many time periods.

Does the robust option yield more consistent estimates in the presence of serial correlation in a DiD model ? Also, Does the inclusion of 1 lag of my dependent variable seem as a relevant procedure to implement ?

Thank you for your time.

Kind regards.

[Clyde Schechter]

I'm afraid I can't answer these questions. Serial correlation is an issue that almost never arises in my work. There are a lot of approaches to dealing with it, but I have only the most superficial familiarity with them. There are others on the forum who would be better able to assist you with this, and I hope one or more of them will chime in.

As for whether serial correlation is an issue with crime data, that is a substantive question that would best be answered by a criminologist.

[Yassir Alouhmy]

Thank you for the help. I Appreciate it.