Sorting through Multiple Predictors in Analyses of Field Experiments.

Metadata are critical for gaining more knowledge from observations collected in agronomic field experiments. Multiple predictors or covariates collected as part of the metadata are often observed at at two or more different scales: within-field and field level. For example, soil test measurements, soil type attributes, disease rating or plan density counts collected in grid patterns or targeted locations could be examples of within-field level covariates. Monthly rainfall and field-level management such as tillage, planting timing or crop genetics (where it is uniform) could be examples of field-level covariates. A unique challenge for researchers is to quantify how multiple predictors of interest observed at different scales or levels affect crop yield or crop yield response to different inputs, treatments or new technologies. This presentation will show examples of how multiple predictor hierarchical (mixed effects) linear models, ridge and lasso regressions with both frequentist and Bayesian interpretations of their parameters can be used to quantify effects of multiple predictors. Methods for dealing with missing data, multicollinearity problems, and advantages for using standardized vs non-standardized predictors will be also demonstrated. The presented examples are from analyses of on-farm replicated strip trials conducted in Iowa by farmers to study different seed treatments, plant population densities or row spacing on soybean (Glycine max) and to quantify the impact of reducing or increasing normal nitrogen fertilizer rates in corn (Zea Mays L.). The speaker will also provide several practical hints for using different packages and dealing with program code issues.