Variability in regional wheat yields as a function of climate, soil and economic variables: Assessing the risk of confounding

Mechanisms that explain spatial variability and trends in agricultural productivity at the regional scale are not well understood. Statistical approaches may be used to relate crop yields and trends in crop yields to changes in the economic and bio-physical environment. However, potential yield-explaining variables tend to confound at the regional scale due to strong correlations between these variables, which complicates the interpretation of such empirically derived relationships. In this paper, we assess relationships between different physical and economic variables and yields and trends in yields at the regional scale along a climatic gradient in Europe. We assess the extent of confounding (i.e. confusing the roles of different variables due to strong correlations) among these variables and the associated risk for explaining yield variability and trends in yields. We analyze regional wheat yield data at NUTS3 and NUTS2 level for most of the EU-countries. Soft wheat (Triticum aestivum L.) is chosen as an indicator crop as it is grown over a wide climatic gradient, and its physiology has been subject to many agronomical studies. Time series were used to derive trends in yields. Data of important climatic, soil and economic variables were also derived at NUTS3 and NUTS2 level. Correlation coefficients were calculated between these variables and yields and yield trends. Confounding was assessed by comparing the R2 values of the regression models with and without (groups of) variables. Soft wheat productivity could be described to a very satisfying level. High R 2 values were obtained, partly due to aggregation of spatially autocorrelated in- and output data. High correlations were found between all variables, which indicates a risk of misinterpretation of results from statistical models when only few (groups of) variables are considered in the analysis. At a higher aggregation level (NUTS2) both the model fit and the risk of confounding increase. Validation by an independent dataset does not lead to exclusion of confounding. This paper serves as a basis for further research on spatial variability and trends in agricultural productivity at the regional scale, indicating both the possibilities and the risks of such research