Panel data: help on cohorts

[David Berselius]

Hi,

I am currently doing my bachelor thesis on grade inflation in upper secondary school with regard to a reform in Sweden. Just familiarising myself with STATA, I could need some guidance in what to do.

We have panel data on every grade of each school, with our "inflation variable", a "public/private dummy", a "post or prior to reform dummy (LroplanDUMMY)", year of entry, year, and some control variables.
The hypotheses we intend to test are:

1) Has the reform decreased grade inflation?
2) Has the reform affected private school to a greater extent than public schools?

We want to divide the data into cohorts based on school and year of entry in order to test our hypotheses.
As a first step, how do we make STATA identify our cohorts?

Secondly, what tests should we use for our hypotheses? We are right now considering a differences-in-differences method for the second hypothesis – do you agree?

Some background info: the reform has affected all students entering upper secondary school in 2011 and after. Prior research has found that private schools generally inflate grades more, and the reform is partly intended to deal with these issues.

Sorry for very basic questions and badly formatted attachments. Any help is very appreciated!


Best regards,
David

[Clyde Schechter]

First question:

egen cohort = group(school_id year_of_entry)

As for how to test your hypotheses, I would need to know more about the study design to answer that. Is this a purely pre-post design, or do you have some cohorts that were never exposed to the reform, and others that (at the same time) were? It makes a big difference. Also important to know: was the reform applied at the school level, the cohort level, or in some other way. And did the implementation of the reform begin at the same time in all schools, or did it start in different schools at different times?

[David Berselius]

Thanks for the quick reply!

All students entering before 2011 were not exposed to the reform, but everyone entering that year were for the rest of their studies. The reform was applied nationwide and for all schools. Upper secondary school in Sweden lasts for 3 years, so for 2011 and 2012 there was a mix between pre-reform and post-reform students. Before and after those two years only one "system" was in effect.

Hope that answers your questions. Again: big thanks!

[Clyde Schechter]

OK. So this is a difficult design to work with because you do not really have directly comparable treatment and control observations. Before 2011 everybody is in control condition. From 2013 on, everybody is in the treatment condition. In 2011 and 2012 you have a mix of treatment and control students. These are really the only data that provide direct information about the intervention effect. BUT, in those eras treatment/control status is confounded by grade level (1 to 3).

Also, it may be reasonable to suppose that the reason a new policy was even put in place is because there was concern over a rising trend in grade inflation over time. So it would be necessary to also adjust for calendar year.

And you have numerous schools observed over these time periods as well. So it looks like your code (in very broad outline) will look something like this:

xtset school_id
xtreg outcome i.treatment i.grade i.calendar_year, // re or fe might be suitable

There may be other covariates that are important to include here. I don't know what those might be.

To test whether the intervention effect is different in private schools, you would replace i.treatment by i.treatment##i.private in the command.

I'm sorry for not using your variable names, but it was easier for me to keep track of what is going on using names that are more suggestive (to me, at least) of what effects they represent.

As you say you are new to Stata, make sure you familiarize yourself with the [GS] manual. And you will also want to invest time learning about panel-data models (especially the -xtreg- command) in the [XT] manual. It will also be worth your while to learn about factor variable notation and the -margins- command. For a really clear introduction to those, I recommend Richard Williams' article in Stata Journal,which you can get at http://www.stata-journal.com/article...article=st0260.