0

Just have a quick question.

You have to know by looking at the residual plot and the normal qq plot that the residual should be distributed as normal, average of residuals should be 0 and residuals should not have a discernible trend and residuals should have constant variance.Then, from these points, you know that the OLS is a good or bad choice.

But, let's say from the residual plot, you see that the residuals have inconsistent variance meaning that as the fitted value increases, you can see higher and higher residuals. Then, one option is by applying log transformation by reducing the phenomenon where the residuals grow in variability.

My question is this: if you apply log transformation, you hope to see that the new residuals should be now following the normal distribution as well? thanks

user398060
  • 53
  • If you apply a log (or any other) transformation before running your linear regression, then you are changing your model. You need to be satisfied that any model you choose run makes theoretical sense in terms of whatever you are modelling. – Henry Oct 12 '23 at 16:06
  • One can hope for anything. For discussions of the usual objectives of a transformation, please see our threads at https://stats.stackexchange.com/questions/298, https://stats.stackexchange.com/questions/24227, https://stats.stackexchange.com/questions/582106, https://stats.stackexchange.com/questions/35711, etc. – whuber Oct 12 '23 at 16:30
  • @whuber Thank you for the follow-up. Just wondering if looking at qq plot, you don't see that many points do not follow the normality. In the first link, when the relationship is exponential, you should apply logarithmic function. Does that mean you should look at qq plot to judge if the graph looks exponential? thx – user398060 Oct 12 '23 at 17:35

0 Answers0