I have trying to improve a multiple regression model (reducing RMSE further), and have found heteroscedasticity evidence for 2 variables. I have found 2 options for reducing heteroscedasticity in linear regression: first is to get square root Y (response variable), and second is to use box-cox transformation (as per https://www.r-bloggers.com/how-to-detect-heteroscedasticity-and-rectify-it/). All examples I have found show single input variable solutions. I'm trying to figure out how this will play out in multiple linear regression analysis. In analysis of each of the 3 relationships (predictors and responses) individually, only 2 show evidence of heteroscedasticity. How do I apply to only 2 out of 3? I am playing around eg square root of y to entire multiple regression.

In R, below, I've created a new column which is squareroot of output variable y (using caret library):

  model2 <- train (ysqrt ~ x1 + x2 + x3, trainX, method = "lm", trControl = trainControl (method="cv", number=10))

This just doesn't seem right though.

Thanks in advance.


What about using White robust standard errors?


Take a look at this to apply this method with R:

Replicating Stata's "robust" option in R

  • $\begingroup$ Thanks, I gave this a go. It's the best info I've seen so far. $\endgroup$ – user637251 Nov 20 '17 at 7:01
  • $\begingroup$ I'm glad info was useful for you. Sorry if I don't give in depth explanations. Sometimes it's hard for me to write in english. $\endgroup$ – Alejandro Carrera Nov 22 '17 at 5:57

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.