0
$\begingroup$

I'm working with some real world data and the regression models are yielding some counterintuitive results. Normally I trust the statistics but in reality some of these things can not be true. The main problem that I am seeing is that an increase in one variable is causing an increase in the response when, in fact in reality, they must be negatively correlated.

Is there a way to force a specific sign for each of the regression coefficients in SPSS?

$\endgroup$
  • 5
    $\begingroup$ Despite the focus on SPSS, I see a statistical question here. So, you don't believe some of your regression results. Some broad answers are: your expectations might be wrong; you may be fitting an inappropriate model and need to look more closely at the data (e.g. you may need transformations or interaction terms); implausible signs for regression coefficients may be side-effects of fitting an over-complicated model. There isn't an ethical or even a technical solution that does what you ask, and better answers lie elsewhere. More information on your data and results would help. $\endgroup$ – Nick Cox Feb 24 '16 at 10:26
  • 2
    $\begingroup$ I agree fully with @NickCox - the first two things that sprung to mind were to check your standard errors/confidence intervals (it doesn't mean a lot if an insignificant coefficient has a counterintuitive sign - if you expect it to be significant one way or the other, you're most likely using too small a sample to detect the effect, or you may have a multicollinearity issue inflating your standard errors) or your coefficient, even if significant, may be affected by Omitted Variable Bias (do you lack data on possible confounders? Is your functional form correct?). $\endgroup$ – Silverfish Feb 24 '16 at 11:28
  • $\begingroup$ Without some idea of your data, what your results look like, or what you expected, it is hard to be more specific. $\endgroup$ – Silverfish Feb 24 '16 at 11:29
  • $\begingroup$ @NickCox In the absence of further information from the OP, I wonder if you might be persuaded to post that comment, perhaps a little fleshed out, as an answer. $\endgroup$ – Silverfish Feb 25 '16 at 12:21
  • $\begingroup$ @Silverfish I'd say that you are at least as well qualified to do that. Feel free to take any or all of my comment and no need to acknowledge. I think we are both making standard points. But I'd search briefly for a duplicate before investing time on this. $\endgroup$ – Nick Cox Feb 25 '16 at 12:26
1
$\begingroup$

The most common reason for "wrong signs" is multicollinearity, where you have two variables competing for the same effect. Looking at the VIF available as part of the regression output can help you detect this.

$\endgroup$
0
$\begingroup$

Another term you might see related to this is a "suppressor variable." This question can provide you with some background information on the topic.

$\endgroup$
0
$\begingroup$

I've had this same problem with smallish datsets - to my knowledge, there's no way to force a coefficient sign, but I agree that there are legit (i.e., causal) reasons for wanting to do so with some equations. All you can do is make sure the data are appropriate. Check for outlier influence, or try removing single variables one at a time from the equation to see if one variable is screwing the other ones up. Good luck!

$\endgroup$

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.