I wanted to ask whether it can make sense to include a variable in a regression model that is coded in the following way?

  • -1: something bad

  • 0: something neutral

  • 1: something good

The alternative of using a "good" and a "bad" dummy predictably runs into multicollinearity problems. Also, I am wondering what model would be appropiate in case I use such a variable as dependent variable? Thank you very much!

Best wishes, Felix


You are talking about defining your latent variable in terms of sign function. If it "makes sense" it depends on what exactly do you want to do and with what data, but there is no reason why you shouldn't use it in your model. You can start with the Why is gender typically coded 0/1 rather than 1/2, for example? thread that discusses different kinds of coding of variables.

People commonly use two kinds of coding of variables:

  • 0 vs 1; that leads to $\beta \times 0 = 0$ and $\beta \times 1 = \beta$ estimates of effects for both levels,
  • -1 vs 1; that leads to $\beta \times -1 = -\beta$ and $\beta \times 1 = \beta$ estimates of effects for both levels.

In your case your variable would lead to decrease by constant $-\beta$ for the "bad" level, would have no effect with the "neutral" level and would increase the predictions by constant $\beta$ for the "good" level.

  • $\begingroup$ Dear Tim, a belated thank you for your thoughts. I concluded that it made sense in my research setting. $\endgroup$ – Felix Jun 11 '18 at 15:46

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.