The random error is generally distributed normally with a zero mean and constant variance 2, and the parameters are estimated by the ordinary least squares method. Nevertheless, to take the heteroscedasticity into account, a variance function that involves the predicted values or other explanatory variables must be considered
(Fortin et al., 2007). The estimation of the parameters of this variance function can be observed as a weighted regression; the weight of which is parameterised instead of being fixed arbitrarily. An example of this function is based on the power of the predicted value, as indicated in Eq. (2).
Var(εi)
where 2 is the residual variance, yi is the predicted value of fruits for plant i and is a parameter to be estimated using the generalised least squares method (Pinheiro et al., 2011). This means that the random error continues to be distributed normally with a zero mean but with nonconstant variance 2y2
Another way to control the variance is by means of variable transformation. The transformed variables model is defined according to Eq. (3).