How to address heteroskedasticity in a regression model?