Suppressor Variables Are Sweet.

[anan_ab]

This week has been unbelievably busy. Between the three research projects running, another research proposal, and a massive stats project, I haven't had much time to breathe. The bulk of my time has been dedicated to figuring out this stats project and I have become quite enamored by suppressor variables and suppression within multiple regression

Comments

[hostirad]

Supressor variables are awesome. And rare. In real life they just don't turn up that often.

[anan_ab]

Their occurrence is funny. My professor says she didn't see one for years and then oddly (and unintentionally) they arose in each of the 7 problems from our last homework assignment. I am unsure about their actual presence in reality because of the relevant articles I have read, none have agreed on a common definition of a suppressor variable and some have been so bold to say that researchers often assume something is wrong with their research project if suppression emerges in their output

[hostirad]

Your definition of a suppressor variable is dead-on. What the suppressor does is to literally subtract out the variance in another predictor that is not relevant to the dependent variable. Suppressors will always have a negative weight in the regression equation.

[anan_ab]

Now, I'm still a little hazy on some of the more rare cases. What if a variable is not significantly correlated with the DV and then becomes positively correlated after partialling out another variable(/set of variables)? Would that most likely be some kind of a spurious relationship with some other unincluded variable driving the effect?

I also had one interesting case where I was calculating a regression equation to predict graduation rates where I had two variables that could act in concert to make a third variable change signs (and hence fit most definitions of suppressor variables). The tricky part there was that neither variable could actually suppress that variable to the point of changing signs alone. Both variables had to be accounted for first to elicit that effect. How would you classify an instance like that?