Development and evaluation of general or transfer models of relationships between wheat yields and fertilizer rates in southern Australia

Abstract

General or transfer models have been developed relating soil, weather and management variables to wheat yields and responses to nitrogen and phosphorus fertilizers, for a wide area of southern Australia, covering a diversity of soils and growing conditions. The models are based on sets of regressions relating polynomial fertilizer-yield functions to site variables, including dummy variables for effects of location in geographic region and kind of soil. The contributions of the dummy variables to the regressions provide measures of lack of transferability. The empirical procedures employed to evolve regression equations for the general models may produce serious statistical biases leading to exaggerated impressions of statistical significance and the reliability of predicted yields. Evidence of omission bias is provided by the need for dummy variables, the substantial residues in analyses of variance, and doubts concerning causal effects by some of the predictor variables. Selection bias is expected from the use of the same set of data both to select variables and to estimate parameters, and by the lack of a theoretical basis to define the form of the regression models. Until improvements in these respects have been provided by basic research, some protection from the effects of these biases may be sought in large samples. General models offer advantages in this respect over soil or locality specific models.