Notice that, for g positive, the degree of heteroskedasticity
is the same for h = g and h = −g/(g + 1).
13 However, the two situations are quite
different in that h = g implies that the observations have the larger variance with
probability 1−p, whereas when h = −g/(g + 1) the probability of the larger variance
is p. Therefore, the designs with h = g and h = −g/(g + 1) will have very different
implications for the performance of the estimators.