0

In a regression, when you get negative coefficient which you know should be positive, why it is necessary to include possible omitted variable that is likely to have positive coefficient and negatively correlated with the variable already in the model? How will this change the sign of the variable?

For example, if the dependent variable is car price, and the independent variable is fuel economy (with the wrong negative sign). Why it is advised to include perhaps curb weight which is negatively correlated with fuel economy and positively with car price?

Quirik
  • 673
  • 2
  • 6
  • 19
  • I think this needs expansion in one or more of the following ways: A detailed example. Reference to the advice you (think you) are following. Otherwise partial answers could be of the form: Perhaps your idea (or the prevailing theory) is wrong. The "wrong" sign could be a side-effect of the way variables are interacting. For example, opposite signs on linear and squared terms in some $x$ could be consistent with a monotonic dependence over the range of the data on the corresponding variable $x$. – Nick Cox Jun 25 '15 at 09:57
  • @Nick Cox I have edited my question. – Quirik Jun 25 '15 at 10:03
  • Sorry, but I don't think the edits help much. I can't follow what advice you are following or its rationale. Watching out for the effects of the predictors on each other and worrying about how far they are standing proxy for other predictors not in the model is a large part of the art of regression. If your question arises from a specific dataset can you post that? There isn't much in your post beyond "I am puzzled by my results". – Nick Cox Jun 25 '15 at 10:20
  • What's "curb weight"? car weight? kerb weight? – Nick Cox Jun 25 '15 at 10:21
  • @Nick Cox I read this in "Oh no! I got the wrong sign! What should I do?" by P. Kennedy. I am not performing regression at the moment on my datase, only trying to understand rationale behind this. – Quirik Jun 25 '15 at 10:25
  • References should be complete enough that the reader can find them directly, i.e. in the form that you would expect in an academic text or journal. I understand your concern, but can't see here a question that is both new and easy to answer. Look around for other threads e.g. http://stats.stackexchange.com/questions/34151/positive-correlation-and-negative-regressor-coefficient-sign – Nick Cox Jun 25 '15 at 10:29
  • Read all about this phenomenon in various contexts by searching our site for [regression change sign](http://stats.stackexchange.com/search?q=regression+change+sign). – whuber Jun 25 '15 at 13:33

0 Answers0