Why Causality Is the Holy Grail of Agricultural Economics
In the tangled world of agricultural economics, understanding what truly causes what is like trying to find a needle in a haystack — except the haystack is made of data, and the needle is hidden behind layers of confounding factors, measurement errors, and reverse causality. Researchers want to know: does a policy increase crop yields? Does a price change reduce pollution? But unlike a lab experiment where you can control every variable, most agricultural data comes from the messy real world, where farmers, weather, markets, and policies all interact in complex ways.
A team of economists from the University of Copenhagen, Wageningen University, University of Bonn, and University of Hohenheim — led by Arne Henningsen and colleagues — have taken a deep dive into this challenge. Their 2025 paper lays out a thoughtful roadmap for how agricultural and applied economists can better estimate causal effects using observational data, which is data collected without controlled experiments. Their work is a guidebook for navigating the minefield of assumptions and pitfalls that come with trying to tease out cause and effect from the wild data of the real world.
The Mirage of Causality in Observational Data
Imagine you notice that farms using a new irrigation system tend to have higher yields. Is the irrigation system the cause? Or are wealthier farmers more likely to adopt it and also have better yields for other reasons? This is the classic problem of confounding variables — hidden factors that influence both the treatment (irrigation) and the outcome (yield).
Randomized controlled trials (RCTs) are the gold standard for causal inference because they randomly assign treatments, breaking these confounding links. But in agriculture, RCTs are often impossible or unethical — you can’t randomly assign tariffs or withhold food aid from needy regions just for the sake of an experiment.
So researchers turn to observational data, but this requires careful strategies to mimic the conditions of an experiment. The paper emphasizes that without transparent discussion and testing of the assumptions behind these strategies, causal claims can be misleading. In fact, previous studies have shown that common methods like ordinary least squares (OLS) regression or matching can overstate effects by up to 80% compared to experimental benchmarks.
Tools for Untangling Cause and Effect
The paper walks through several powerful econometric tools that help researchers get closer to causality:
Selection on Observables: This approach assumes that all confounding factors are observed and controlled for. Methods like OLS regression and propensity score matching fall here. But the assumption is strong and often unrealistic because some confounders remain hidden. The authors recommend using Directed Acyclic Graphs (DAGs) — visual maps of assumed causal relationships — to clarify which variables to control for and which to avoid, preventing the common mistake of controlling for variables that lie on the causal path itself.
Instrumental Variables (IV): When unobserved confounders lurk, IV methods can help. An instrument is a variable that influences the treatment but affects the outcome only through that treatment. For example, rainfall variation might serve as an instrument for irrigation use if it affects irrigation but not yields directly. However, finding valid instruments is notoriously difficult, and weak or invalid instruments can make estimates worse than simpler methods. The paper stresses rigorous testing of instrument strength and validity, including theoretical justification and statistical tests.
Fixed Effects and Difference-in-Differences (DID): These methods exploit panel data — repeated observations over time — to control for unobserved factors that don’t change over time. DID compares changes over time between treated and untreated groups, assuming their trends would have been parallel without treatment. The paper highlights recent advances that address complications like staggered treatment timing and heterogeneous effects, which can otherwise bias results.
Synthetic Control Method: This innovative technique constructs a weighted combination of untreated units to create a “synthetic” control group that closely matches the treated unit before treatment. It’s especially useful for evaluating policies affecting a single region or country. The authors note its underuse in agricultural economics despite its potential.
Regression Discontinuity Designs (RDD): When treatment assignment hinges on a cutoff — say, farms below a certain size get a subsidy — RDD compares observations just above and below the threshold. This local comparison can reveal causal effects under certain assumptions. The paper also discusses Difference-in-Discontinuity designs that combine RDD with DID for even stronger inference.
Why This Matters Beyond Academia
At first glance, these might seem like dry statistical tools, but their implications ripple far beyond academic journals. Agricultural policies influence food security, environmental sustainability, and rural livelihoods worldwide. Decisions based on shaky causal claims can lead to wasted resources, ineffective programs, or unintended harm.
The authors remind us that policymakers, NGOs, and agribusinesses rely on credible evidence to make choices that affect millions. For example, if a subsidy program is claimed to boost yields but the evidence is biased, funds might be misallocated, leaving farmers worse off. Conversely, robust causal evidence can guide smarter interventions that truly improve lives and ecosystems.
The Art of Causal Humility
One of the most refreshing aspects of the paper is its call for humility and transparency. The authors caution against casually using causal language without carefully justifying assumptions. Instead of saying “X causes Y,” it might be more honest to say “X is associated with Y, conditional on these assumptions.” This subtle shift acknowledges uncertainty and invites critical scrutiny.
They also advocate for using multiple methods and sensitivity analyses to triangulate findings. If different approaches with different assumptions converge on similar results, confidence grows. If not, it signals caution.
Machine Learning: A Double-Edged Sword
The paper touches on the growing role of machine learning in agricultural economics. While machine learning excels at prediction, it doesn’t automatically solve causal puzzles. Naively applying machine learning can even worsen bias by dropping important confounders. However, when integrated thoughtfully into causal frameworks — so-called “causal machine learning” — these tools can flexibly model complex relationships and uncover heterogeneous treatment effects.
Final Thoughts
Henningsen and colleagues have crafted a vital compass for navigating the tricky terrain of causal inference in agricultural economics. Their guidelines don’t promise easy answers — because there are none — but they do offer a path toward more credible, transparent, and useful research.
In a world where data is abundant but truth is elusive, their work reminds us that the journey from correlation to causation demands rigor, creativity, and above all, honesty about what we can and cannot claim. For anyone interested in how we understand and improve the systems that feed us, this paper is a must-read.
