Omitted variables and multicollinearity

[Nikos Korompos]

Hi! I am relatively new on Stata, so I have a question.

I am trying to estimate a model based on repeated cross-sections, in which I use a dummy which distinguishes two different time periods (0 if 2004-2005, 1 if 2007-2012). I was thinking to control for some year dummies as well. When I control for the both of them, the stata output omits not only 2004 (the base category) but also the last year.

Could I have your opinion on this? Do you think I should avoid controlling for both of them, due to potential multicollinearity? I thought that the omission of variables (done by Stata) solves the problem..

Thank you in advance.

Best regards,

Nikos

[Jesse Wursten]

If you include year dummies, then the dummies distinguishing time periods are redundant. I.e. if you already control for it being the year 2004 and it being the year 2005, then there's no added value in trying to control for it being either 2004 or 2005. The results should be the same regardless of whether you include these time period dummies, it is only the interpretation of the year dummies (and potentially the constant) that gets trickier.

[Nikos Korompos]

Thank you Jesse! I want to check some pre- and post-period effects, but I was thinking that it might be a good idea to control for year dummies as well. In general, do you think that controlling for both us them creates problems to my model due to multicollinearity?

(totally agree that the interpretation changes)

[Jesse Wursten]

Well yes, they are perfectly collinear, so you cannot estimate both at the same time. I think you're better off restricting yourself to the year dummies and then using appropriate tests to see whether they are significantly different from each other in both periods. I'm sure there's techniques for this.