This is our final blog post for our Job Market Paper Series blog for 2025-2026.
Anirudh Sankar is a PhD student at Stanford University, specializing in development and behavioral economics. He is interested in how learning and knowledge-building shape development.
People in developing countries make decisions across complex domains—for example, agriculture, health, and finance—where the right choice is often hard to identify. A smallholder farmer, for instance, must decide how to purchase and deploy specialized inputs on his farm, how to prevent and treat serious illness, and how to navigate intricate insurance and credit schemes. These decisions are so difficult that there are entire fields of specialized knowledge—agronomy, medicine, and economics—devoted in part to identifying (and often developing) good choices.
To share the fruits of these fields with lay decision-makers, governments and development programs invest heavily in information interventions to promote good choices. Most of these efforts, however, are prescriptive and “black box”: they present results (“as you can see, this choice worked well for others before you”) without explaining why they work—the agronomy behind an agricultural input, the immunology behind a vaccine, or the economics behind an insurance product. The implicit assumption is that understanding the underlying mechanism is unnecessary or burdensome; people only need the result, not the reasoning.
When opacity blocks adoption
Yet the absence of a mechanistic understanding may itself be a key barrier to adopting good choices. Take agriculture, where smallholder farmers are often said to be under-adopting high-return technologies—such as improved seed varieties or synthetic fertilizers. Too few farmers adopt these technologies, and many of those who do use them suboptimally.
One reason may be that such technologies are encountered as opaque products developed in distant laboratories. Fertilizers, for example, appear as colored pellets that somehow make plants grow faster or yield more when placed near the soil. Technologies like these are often demonstrated through their results, yet farmers attending such demonstrations often discount what they see, remaining skeptical (Alidaee 2023).
Their caution is reasonable: returns to agricultural technologies often vary sharply across soil types, weather patterns, and microclimates, and many can recall neighbors whose inputs failed or, conversely, whose crops were destroyed through improper use. Explaining why a technology works—in the case of fertilizers, the underlying soil–nutrient–plant mechanisms—can build trust and, crucially, equip farmers to adapt recommendations to their own conditions rather than copy them wholesale or dismiss them entirely. In this way, mechanistic explanations can be essential for enabling confident, context-specific decision-making.
This approach is especially motivated by insights from psychology and cognitive science. An influential body of work in these literatures suggests that mechanisms are central to human cognition, arguing that from early childhood, humans seek to learn the mechanisms underlying the phenomena they observe and feel deprived when they do not learn them (Ahn et al. 1995; Keil 2022).
Testing the idea in agricultural demonstrations
To test this idea, my job market paper embeds a randomized controlled trial (RCT) within government agricultural demonstrations for tomato farmers in Eastern Uganda. Extension officers showed nearly 800 farmers the impact of a recommended fertilizer recipe by displaying side-by-side farming plots with and without the fertilizer recipe. We randomized the type of explanation that accompanied the demonstrations.
In control demonstrations, the officer simply described the recipe and its results. In treatment demonstrations, the officer went further—explaining why the recipe worked in terms of nutrient–plant–soil interactions.
We conducted a lab-in-the-field study with farmers at the demonstration site to see how they used this information in applied decision problems, such as substituting between fertilizers, adapting recipes across soil types, or diagnosing production problems. Crucially, these tests did not assess recall or rote learning. Farmers were free to apply, ignore, or reinterpret what they learned as they solved realistic, often incentivized scenarios.
We then followed up with the same farmers in two subsequent growing seasons to study how these updated beliefs translated into real-world fertilizer use, yields, and profits on farmers’ own fields.
Unpacking mechanistic explanations
To understand what farmers were actually learning from these explanations, it helps to unpack what a mechanistic explanation consists of. A mechanistic explanation of fertilizers gives farmers both a language and a structure for reasoning about their choices.
The language names the key conceptual parts of the system—in this case, macronutrients such as nitrogen (N), phosphorus (P), and potassium (K)—and encourages farmers to think in terms of these invisible (since nutrients cannot be “seen”) but causally meaningful building blocks. For example, in the fertilizer recipe shown on the left of Figure 1, what really matters is the pattern of macronutrient additions shown on the right: the “active ingredients” behind plant growth.
The structure, meanwhile, describes how these elements interact in the soil and plant, providing a model of nutrient response. Once understood, the shape of the nutrient distribution (Figure 1, right) begins to make intuitive sense. Phosphorus, for instance, is recycled within the plant, so it is most effective when applied early.
Together, this language of macronutrients and structure of nutrient response can enable two forms of generalization. The language makes any data or correlations more informative, and the structure helps optimization within and adaptation across contexts. In soils with higher clay content, for example —something farmers can readily observe— more nutrients can be applied before losses occur, because clays hold on to them and reduce leaching.
What we found
In our lab-in-the-field experiment, we find evidence for both language- and structure-based generalization. Farmers who received mechanistic explanations were better able to substitute across fertilizers based on their nutrient content and perform what we call “nutrient arbitrage”: choosing fertilizer bundles that delivered the same nutrients at lower cost.
These are crucial capabilities, since farmers frequently face stockouts or price hikes of their preferred fertilizers, and they provide direct evidence of reasoning in the language of macronutrients.
To test for structure-based reasoning, we designed an intuitive, literacy-friendly digital “fertilizer application game,” where farmers could configure any fertilizer recipe and instantly receive feedback on its simulated profits. They played multiple rounds of this game, mimicking how farmers choose fertilizers season after season, and were incentivized based on performance—capturing the real-world tradeoff between exploring new strategies and exploiting known ones.
Farmers who received mechanistic explanations achieved higher simulated profits across two soil types: the demonstration soil—one which they directly observed the impact of the fertilizer recipe– and a counterfactual soil with different observable characteristics (for example higher clay content) and a different nutrient response. This pattern indicates that mechanistic explanations improved farmers’ understanding of nutrient response structure and its applicability across heterogeneous conditions. The treatment especially helped them minimize losses and increase the frequency of moderate profits, while also modestly raising the share of very high-profit outcomes (Figure 2).
Follow-on impacts
We then followed up with many of the same farmers at their own plots over subsequent growing seasons to track real changes in input use, yields, and profits based on their self-reports. Treated farmers’ use of nitrogen and potassium was more agronomically sound, and their improved timing of potassium—an increased emphasis at the fruiting stage—reflected the same decision logic that generated higher simulated profits we observed in the lab-in-the-field game. Treated farmers also had 14% higher yields.
Conclusion
Given the light-touch nature of the intervention these results already suggest meaningful learning and behavioral change with positive downstream impact. Together, they show that simply adding mechanistic explanations—clarifying why an otherwise opaque technology works—can have distinctly instrumental benefits. This raises broader questions. What is the role of additional data—for example, collecting many soil tests across an entire village—relative to mechanistic explanations? To what extent do mechanistic explanations substitute for data, and to what extent do they complement its interpretation? How deeply should mechanistic explanations be communicated—for instance, in the case of fertilizers, should training stop at macronutrients, or extend to micronutrients, clay–nutrient electrostatics, or nutrient speciation? While our complementary theory work offers some higher-level predictions based on the statistical structure of the problem, there are also practical considerations, including the risk of cognitive overload and the activation of non-Bayesian learning strategies. Understanding these tradeoffs could inform the design of information interventions across many domains.
Revisiting paternalism in development
Besides our academic and policy contributions on mental models, human capital, and technology adoption, this work also raises a deeper question about paternalism in development. Nearly a century ago, the journalist Walter Lippmann and the philosopher John Dewey debated how citizens should navigate the growing complexity of modern life.
Lippmann believed that complexity should be managed by experts who distill it into simple prescriptions; Dewey countered that people must engage with complexity directly—drawing on expertise, but also peering into and participating in its formulation.
Our findings offer an instrumental justification for Dewey’s view. In complex domains like agriculture, showing people what lies behind the curtain of expertise—helping them see why a recommendation works rather than simply asking them to follow it—can directly improve learning and adaptation.
