It is the very popularity and ready accessibility of
Osborne and Waters’ article that prompts us to pen
this response more than a decade after its original
publication. Our concern is that Osborne and Waters’
article contained two fairly substantial misconceptions
about the assumptions of multiple regression. These
misconceptions are that multiple regression requires the
assumption of normally distributed variables; and that
measurement error can only lead to under-estimation
of bivariate relationships. Misconceptions about
distributional assumptions can have serious
consequences, including the expending of effort on
checking unnecessary assumptions, the performing of
problematic transformations and “corrections”, and the
neglect of the actual assumptions of the analysis being
used. In this paper we correct the misconceptions
contained in Osborne and Waters’ article, making use
of simple computer simulations to illustrate our points.
We also provide a brief corrected summary of the
assumptions of multiple regression. For simplicity, our
examples are restricted to the bivariate or “simple”
regression case—i.e., just one predictor and one
response variable. Our statements nevertheless apply to
both multiple and simple linear regression, and indeed
can be generalized to other instances of general linear
models with a single dependent variable such as
between-subjects ANOVA and ANCOVA, and
independent samples t-tests. Comments are restricted,
however, to models in which the estimation method is
ordinary least squares (OLS)—as is usually the case.