Causal Misperception and Persistent Fertilizer Overuse in China

We study interventions to improve fertilizer management among rice farmers in China by providing time-specific and plot-specific information and decision support that helps farmers rebalance nitrogen, phosphorus, and potassium use without reducing yields.

In March 2020, we implemented three interventions: (i) plot-specific soil-testing information (T1); (ii) soil-testing information plus customized, plot- and time-specific fertilizer recommendations delivered through a mobile application (T2); and (iii) the same app-based recommendations plus in-person training by agricultural extension agents, with emphasis on the roles of phosphorus and potassium in improving yields and profits (T3).

In November 2020, we introduced a leaf color chart intervention. The chart is a low-cost visual decision aid that helps farmers compare rice leaf greenness with calibrated reference colors when deciding whether additional nitrogen fertilizer is needed.

In the late-August 2026 long-term follow-up, we implement a short randomized in-survey decision-aid experiment among relevant follow-up respondents. Some respondents receive a standardized reminder and demonstration on how to use the leaf color chart before completing incentivized decision tasks, while others complete the same tasks without this reminder.

2020-04-03

2026-11-01

Our primary outcomes are:
1. Fertilizer use: amounts and extensive-margin use of urea/nitrogen, compound fertilizer, phosphorus fertilizer, and potassium fertilizer.
2. Fertilizer timing and mix: whether farmers apply fertilizer at recommended stages and whether they rebalance away from excessive nitrogen toward phosphorus and potassium.
3. Gaps between actual fertilizer use and recommended fertilizer use.
4. Beliefs about production: especially beliefs about the relationship between rice greenness and yield, and beliefs about the yield effects of increasing or decreasing urea use.
5. Understanding of fertilizer nutrients: knowledge of the roles of nitrogen, phosphorus, and potassium.
6. Agricultural outcomes: rice yields, revenues, fertilizer/input costs, and profits.
7. Phase III task outcomes: accuracy in leaf-color interpretation and incentivized photo-based yield-ranking tasks.

Fertilizer quantities will be converted to kilograms and normalized by cultivated area where appropriate. We will separately measure compound fertilizer, urea/nitrogen fertilizer, phosphorus fertilizer, and potassium fertilizer. When the N-P-K formula of compound fertilizer is available, we will also construct nutrient-content measures.

Belief outcomes will be coded as indicators for correct responses. For example, the correct belief about greenness and yield is that yield first increases and then decreases with greenness, rather than “the greener, the better”. Nutrient knowledge outcomes include whether farmers correctly identify nitrogen as affecting greenness, phosphorus as affecting flowering and root development, and potassium as affecting grain filling or grain density.

Yield will be measured as harvested rice output divided by harvested area. Profits will be constructed as revenue minus reported production costs. For Phase III decision tasks, we will construct task-level and respondent-level accuracy.

Secondary outcomes include:

1. Long-run use and memory of the leaf color chart.
2. Whether the leaf color chart had a lasting influence on fertilizer decisions.
3. Willingness to receive and pay for personalized fertilizer advice.
4. Willingness to receive and pay for a leaf color chart.
5. Farmer experimentation with fertilizer use on their own plots.

1. Long-run use and memory of the leaf color chart.
2. Whether the leaf color chart had a lasting influence on fertilizer decisions.
3. Willingness to receive and pay for personalized fertilizer advice.
4. Willingness to receive and pay for a leaf color chart.
5. Sources of fertilizer information, including extension agents, fertilizer sellers, mobile tools, media, and peers.
6.Farmer experimentation with fertilizer use on their own plots.
Attention to crop characteristics such as greenness, roots, flowering time, lodging risk, and grain fullness.
Peer discussion and diffusion of information about formula fertilizer, leaf color charts, and optimal greenness.
Other agricultural practices, including pesticide use, labor, mechanization, subsidies, insurance, and adaptation to weather or seasonal timing.

We partner with local agricultural extension services to study fertilizer management among rice farmers in China. The original experiment surveyed 1,200 households in 200 villages. Villages were randomly assigned to four arms: control, soil-testing information, soil-testing information plus app-based fertilizer recommendations, and soil-testing information plus app-based recommendations plus extension-agent training.

After the first follow-up survey, we introduced a leaf color chart intervention to help farmer to for nitrogen top-dressing decisions in 2020 November. In the 2026 long-term follow-up, we revisit the original respondents and conduct a short randomized in-survey experiment. Some respondents receive a standardized leaf-color-chart reminder before completing incentivized decision tasks, while others complete the same tasks without the reminder. We measure fertilizer behavior, beliefs, nutrient knowledge, yields, profits, and decision-task performance.

Not available

Randomization is done by computer. Phase I was randomized at the village level. Phase III A/B questionnaire assignment is randomized by computer before fieldwork, using a fixed random seed and balancing within historical treatment groups and villages where feasible.

Phase I randomization is at the village level. Phase II leaf color chart exposure follows the original village-level treatment assignment. Phase III A/B questionnaire assignment is at the individual respondent level.

Yes

200 villages in the original experiment. The main Phase III historical comparison sample includes approximately 100 villages from the original soil-testing-only and control groups.

1,200 households in 200 villages in the original experiment. In Phase III, we plan to re-contact the original named respondents. The main A/B subexperiment includes up to approximately 600 households from the historical soil-testing-only and control groups, subject to attrition.

Sample Size by Treatment Arms

Phase I:
1. Control: 50 villages, approximately 300 households.
2. Soil-testing information only: 50 villages, approximately 300 households.
3. Soil-testing plus app-based recommendations: 50 villages, approximately 300 households.
4. Soil-testing plus app-based recommendations plus extension-agent training: 50 villages, approximately 300 households.

Phase II:
1. Historical treatment villages received the leaf color chart. Main comparison: original soil-testing-only group versus original control group, approximately 300 households each.

Phase III
Vension A: LCC reminder/demonstration: approximately half of the historical comparison sample.
Vension B: No LCC reminder: approximately half of the historical comparison sample.
Expected allocation: about 300 respondents per A/B version, subject to attrition.