I have a model based on a dataset that respects all linear model assumptions except for homoscedasticity. When I just ignore the problem of heteroscedasticity, the p-value, for the interaction with group, in my model is <.00001. I definitely know that there is something as per my previous studies and the literature in this field. However, I would like to be honest regarding my analyses and assumptions. Is this assumption really needed if the other 3 main ones are respected (independence, linearity, absence of collinearity) for the interpretation of the p-value in the mixed effects models?

When I run the following on my lmer model called mod:


enter image description here

I get a cone shape distribution. I then try to log transform it, and recheck the model, for the interaction with group in my model the p value goes to .40. Quite a jump! My data comes brain activity from patients and healthy individual, just to clarify.

This is my model:

lmer(value ~ dist*group + (1|patientnumber), dat1)

This is how I obtained the p-value:


Kindly advise.

  • $\begingroup$ Not enough information here to advice regarding transformations or use of a GLMM. However, there is also the option of modeling variance in dependence of a variance covariate with package nlme. $\endgroup$
    – Roland
    Feb 17 '15 at 18:44

(Posting as an answer so it gets carried over during migration rather than disappearing with comments.)

Inferences about interactions are extremely sensitive to log-transformation. It completely changes the meaning of the interaction, from additive to multiplicative. You will indeed get better advice regarding this on CrossValidated, although I would strongly recommend that you try to post your data, or at least an informative picture of your data, when asking a question there (or editing the version of this that gets migrated there).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy