Control variables for a difference-in-difference estimation

[Ashvinder Singh]

Dear Statalist community,

I have a general question today. I am doing a Difference-in-Difference (DiD) estimation on a panel dataset by country-year where the below is my equation:

 
 areg tariff i.treatment##i.pre_post, a(country)

However, I am wondering if any control variables should be included in the equation above? I have read some stuff online, but it seems inconclusive. Under what criteria should I include control variables in my equation, or in fact any DiD equation, generally speaking?

Thanks.

Ashvinder

[Clyde Schechter]

There are two different reasons that covariates are sometimes added to a model (whether it's a DID model or any other kind of regression model.)

One is to reduce confounding (omitted variable) bias. Confounding bias can only happen when the covariate is associated both with the outcome variable and with the main variable(s) in the study (in your case the treatment variable). Depending on whether we are dealing with continuous or discrete variables, cross tabulations or correlation coefficients are usually good ways to see that.

The other reason covariates are sometimes added is to reduce residual variance. The circumstance here is one where the covariate is an important determinant of the outcome variable. (It may or may not be related to the main variables.) In that circumstance, including it can allow for a more precise prediction of the outcome, which in turn increases the precision of your model parameter estimates.

[Ashvinder Singh]

Great - that clears it up. Thank you, Clyde.