Treatment effects are traditionally quantified in controlled experiments. However, experimental control is often achieved at the expense of representativeness. Here, we present a data-driven reciprocal modelling framework to quantify the individual effects of environmental treatments under field conditions. The framework requires a representative survey data set describing the treatment (A or B), its responding target variable and other environmental properties that cause variability of the target within the region or population studied. A machine learning model is trained to predict the target only based on observations in group A. This model is then applied to group B, with predictions restricted to the model's space of applicability. The resulting residuals represent case-specific effect size estimates and thus provide a quantification of treatment effects. This paper illustrates the new concept of such data-driven reciprocal modelling to estimate spatially explicit effects of land-use change on organic carbon stocks in European agricultural soils. For many environmental treatments, the proposed concept can provide accurate effect size estimates that are more representative than could feasibly ever be achieved with controlled experiments.
Keywords: association; causal inference; causation; correlation; land-use change; machine learning; soil organic carbon; statistical modelling.
© 2021 The Authors. Global Change Biology published by John Wiley & Sons Ltd.