The use of (generalized) linear regression is ubiquitous in genetic association studies of complex phenotypes. In this area, homoscedasticity is often assumed, namely, the variances of modelling errors are assumed to be invariant with respect to the effects being modelled.