Sifting and winnowing: Analysis of farmer field data for soybean in the US North-Central region

Field trials are commonly used to estimate the effects of different factors on crop yields. In the present study, we followed an alternative approach to identify factors that explain field-to-field yield variation, which consisted of farmer survey data, a spatial framework, and multiple statistical procedures. This approach was used to identify management factors with strongest association with on-farm soybean yield variation in the US North Central (NC) region. Field survey data, including yield and management information, were collected over two crop growing seasons (2014 and 2015) from rainfed and irrigated soybean fields (total of 3568 field-year observations). Fields were grouped into technology extrapolation domains (TEDs) that accounted for soil and climate variation and 9 TEDs were selected based on the number of fields needed to detect yield differences due to management as determined using power analysis. Average yield ranged from 2.5 to 5 Mg ha(-1) across TEDs, with field yield distributions in half of the domains having a distributional peak that was close to maximum yields. Conditional inference trees analysis was chosen among 26 statistical procedures as the approach that best combines ability to detect and rank factors (and their interactions) with greatest influence on on-farm yield and relatively easy interpretation of results. Survey data from ca. 150 fields in each of the nine TEDs allowed us to identify key management factors influencing yields for an agricultural area that includes ca. 7 million ha sown with soybean. In five of the nine TEDs, highest yields were observed in early-sown fields. Other factors explaining on-farm yield variation were maturity group, and in-season foliar fungicide and/or insecticide application, but, in some cases, their influence on yield depended upon sowing date and water regime. While the approach proposed here cannot establish cause-effect relationships conclusively, it can certainly provide a focus to replicated field experiments in relation to which management factors to investigate. We believe that future agronomic studies based on farmer survey data can greatly benefit from ex-ante identification of most important TEDs (relative to crop area and production) as well as determination of minimum number of farmer survey data that needs to be collected from each of them based on expected yield differences and variability. The approach is generic enough to be applied in other crop producing regions as long as farmer data and associated climate and soil databases are available.