I'm calculating multiple regression with R and trying to decide which predictors to keep and which to drop. I realized that when I use the lm.beta function I'm not sure whether the significance levels (presented with the typical stars) relate to the estimates/coefficients or the standardized coefficients.

This made me wonder whether I should report significance for the "normal" coefficients or for the standardized ones. Is the significance different between them?

  • 1
    $\begingroup$ Why do you want to drop variables? $\endgroup$ – Michael M Jul 26 '19 at 19:49

First, you should not use this as a method of model building and variable selection. You should use substantive knowledge. If you must use an automated method (i.e. you are saying "I don't really know much about this") then LASSO is good.

Second, the significance level is the same for standardized or unstandardized variables.

  • $\begingroup$ Thanks for your helpfulll answer. $\endgroup$ – Mobody Jul 29 '19 at 8:16
  • $\begingroup$ Of course I selected a set of potential predictors in relation to theoretical considerations and research already done in that field (if that's what you mean with 'substantive knowledge'). But to make the model more consistent for my particular sample (better r² and only variables with an relevant effect) I want do drop predictors with low coefficients and/or which are insignificant. $\endgroup$ – Mobody Jul 29 '19 at 8:22
  • $\begingroup$ That is not a good strategy. First, it will lower $r^2$ not raise it (although maximizing $r^2$ isn't a good goal. Second "relevant" is not the same as "statistically significant". Third, variables may be important to include even if their effects are small. $\endgroup$ – Peter Flom - Reinstate Monica Jul 29 '19 at 10:46

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.