Osborne and Waters (2002) do mention briefly the
assumption of normality of errors, but say that
regression is robust to this assumption and do not give
it any further discussion. The assumption of normally
distributed errors is useful because when it holds true,
we can make inferences about the regression
parameters in the population that a sample was drawn
from, even when the sample size is relatively small.
Such inferences are usually made using significance
tests and/or confidence intervals. However, when the
sample is small, and errors are not normally distributed,
these inferences will not be trustworthy. Normality
violations can degrade estimator efficiency in at least a
technical sense: When errors are normally distributed,
OLS is the most efficient of all unbiased estimators
(White & MacDonald, 1980), whereas in the presence
of non-normal errors it is only the most efficient in the