Translating causal models into environmental practice

Many decisions in agriculture and environmental management now rely on digital information including satellite indicators, reanalysis climate datasets, in-situ sensors and analytics from digital farm platforms. These data are used in predictive models to forecast yields, detect crop stress or classify land use. Prediction is useful, but it does not answer a central question in many decision-making contexts: what would have happened if we acted differently?

Causal machine learning has been proposed as a way to address this gap (Sitokonstantinou et al., 2025). Instead of predicting outcomes, causal ML aims to estimate the effects of policies, management practices or climate shocks and to support decisions about interventions. In my own work, ranging from estimating the impact of humanitarian aid on food security to evaluating the heterogeneous effect of crop practices and digital agricultural advisory services on ecosystem services, causal ML offers a structured way to work with these questions.

At the same time, causal ML raises ethical and epistemic issues that are common across environmental data science. The causal questions that can be asked and the actions that appear reasonable, depend strongly on how socio-ecological processes are translated into variables, interventions and mechanisms. This contribution examines this process of translation in causal ML for environmental and agricultural applications and shows how it is shaped by ontological choices, data availability and institutional priorities.

Ontological choices affect how causal entities are defined. For example, in evaluations of digital agricultural advisory services, “adoption of advice” is often treated as a binary variable. This framing reduces complex farmer decision making, interpretation, partial use, experimentation and risk management, into a single model variable. As a result, the causal effect being estimated reflects the model’s definition of adoption rather than farmers’ actual behaviour.

Data availability further limits what can be studied causally. In analyses of crop diversification or rotation effects, Earth observation metrics such as vegetation indices are often used as proxies for management practices because detailed field level data are unavailable. Consequently, estimated treatment effects capture only the practices that leave a detectable signal in the data, while excluding important management choices that cannot be observed from space.

Institutional priorities also shape causal models. Agricultural research programs and policy initiatives often focus on certain crops or regions that are politically or economically prioritized, leaving smallholder farms or minor crops underrepresented. This means that the causal interventions included in the model reflect institutional focus rather than the full range of agronomic or environmental processes that may be important.

These modelling choices are not mistakes; they reflect real constraints in data and governance. However, they influence how causality, responsibility and intervention are understood. I argue for causal modelling practices that make these translation choices explicit and that pay closer attention to context, plurality and responsibility so causal ML can better support environmental decision-making.

 

Reference

Sitokonstantinou, V. et al. (2025). Causal machine learning for sustainable agriculture. NeurIPS 2025 Workshop: Tackling Climate Change with Machine Learning. https://openreview.net/forum?id=CE5T6BPFBk