In the first step of the simulation, we simulate a
single sample of 30 participants, with 15 participants in
each of the two subsamples formed by the X variable.
This allows us to offer a visual depiction of the
distribution of the response variable Y. In Figure 1 a
histogram shows that due to the strong influence of the
dichotomous predictor variable X, the response
variable Y is bimodal. Its non-normality is also clear in
a normal q-q plot, where the quantiles of the
distribution do not match those of a normal
distribution, as indicated by the straight line. A Shapiro-
Wilk normality test also provides evidence to reject a
null hypothesis of a normal distribution for this
variable, W = 0.910, p = .015. On the other hand, if we
regress Y on X and then calculate the residuals, there is
no evidence to reject a null hypothesis of normality for
the marginal distribution of the errors, W = 0.971, p =
.562. In sum, despite the presence of normally
distributed errors, the response variable in this
simulated example is clearly not normally distributed