Estimation of the harvest index and the relative water content – Two examples of composite variables in agronomy

Abstract Composite variables are variables derived from measurable traits. They are commonly used in agronomy: two well-known examples being the harvest index and the relative water content. There are two approaches for finding estimated averages of such variables that are derived from direct measurements: They can be found either based on a calculation using individual measurements (“pre-processing”) or from a calculation using averages or estimates (“after-fitting”). The former needs to be done prior to fitting a statistical model whereas the latter is carried out after a statistical model has been fitted to the original measurements. We show that the commonly used pre-processing approach results in biased estimates. Moreover, the bias depends on both the correlation between and the uncertainty associated with the variables used for the composite variable. This finding is shown in two examples and a simulation study.