Can AI technologies increase farmer’s resilience to climate change? Impact evaluation of Croppie

To evaluate the impact of Croppie on small-scale farmers, we use a randomized controlled trial (RCT) in Antioquia, Colombia.
In the first stage, we will conduct a one-to-one baseline survey, which will allow us to characterize farmers' digital skills, agricultural practices, and socioeconomic conditions. The farmers will be selected using the list of cooperative members of the association. Specifically, we will interview a sample of 1500 farmers from the association. The 150-minute survey will be conducted at farmers' houses.

The visits will be pre-announced, and convenient times will be agreed upon. During the initical visit, participants will be informed about the project's objectives, the benefits they can expect, and how their data will be managed. Participants can withdraw from the project at any time without incurring any further consequences.

In the second stage, 1500 farmers will be randomized into the following five groups:

T1: Training AI Platform Individual
T2: Training AI Platform Family
T3: Cropppie
T4: Platform no training
T5: Pure Control Group


Farmers in T1 and T2 will participate in three one-day training sessions. The sessions will be conducted at a farm nearby, and transportation will be provided. The training sessions aim to prepare farmers for the use of sustainable agricultural practices following European Union Regulation on Deforestation-free Products (EUDR).\footnote{The regulation aims to reduce carbon emissions caused by EU consumption by avoiding deforestation and forest degradation in the EU and globally.} The sessions use a participatory methodology where farmers learn practical skills in group activities. In the sessions, farmers discuss sustainable agricultural practices for the protection of soil, water, forest, and biodiversity. They learn how to document good practices using an App and practice commercialization and negotiation in different markets. In T1 the sessions are conducted with one household member, while in T2, they receive the training accompanied by another household member. This part of the project aims to test whether technology adoption is more efficient when more members in the household are trained.

Farmers in T3: Croppie, are the main treatment group in this project. Farmers in this group receive a three-hour practical training on the use of Croppie. The training is conducted at their farms by an agricultural extensionist. During the session, the agricultural extensionist will explain the protocols to capture and transfer images using Croppie App. They will also learn how to interpret the information provided by the App and receive training on how the App helps them to improve financial planning.

Farmers in T4: will have the option to install the App, but will not receive any type of agricultural training. They will also not receive an explanation on Croppie. This group is used as a reference to evaluate the impact of T1 to T3.

Farmers in T5 (Control group) will have access to the app one year later. The staggered implementation would allow us to validate the impact of the app without limiting the possibilities for the control group to benefit from it. This group works as a reference to evaluate the impacts of T3 and T4.


In the last stage, we will conduct a one-to-one end-line survey to evaluate the impact of the treatments and, in particular, Croppie. As before, the survey will be conducted at farmers' houses. The visits will be pre-announced and agreed upon. The impacts of the training program will be assessed at the household level by comparing the changes in welfare and resilience to climate change among households in the treated and control groups. The main outcomes of interest are yield estimates, use of precautionary measures, access to credit, demand for insurance, and use of complementary income sources (non-agricultural employment).

2025-09-26

2025-11-30

The project evaluates the impact of Croppie on resilence towards climate change and financial planning. We measure the impact of the training of the following outcomes:

- Coffee yields and coffee price estimates.
- Annual Income
- Relative wealth
- Diversification of income measured as proportion of non-farm labor income and as indicator of number of household members working outside the farm.
- Financial Security
- Social Protection and Risk Sharing
- Mental health
- Food security
- Impact of shocks
- Investments in coffee cultivation (labor and capital) and use of more sustainable práctices
- Impact of negative income shocks

Yield estimates will be measured building the expected distribution of productivity at the farm level. We first ask participants for the mean expected kg produced in the farm. When then elicit the minimum and maximum expected production. Based on the extreme values we ask for the likelihood in a 1 to 10 scale that the production is in each quantile of the distribution. This information is used to estimate the mean expected value and its standard deviation.

Price estimates will be measured constructing the expected distribution of prices per 125kg. When then elicit the minimum and maximum expected prices. Based on the extreme values we ask for the likelihood in a 1 to 10 scale that the price is in each quantile of the distribution. This information is used to estimate the mean expected price and its standard deviation.

Annual Income estimated as the sum of labor income by household members plus farm income (reported value of own production of coffee and other agricultural products.

Relative wealth is a subjective measure on the standing in the social ladder, where the poorest are in the first ladder and the richest in the 10th ladder.

Diversification of income is measured as an index that considers standarized measures on the proportion of non-farm labor income, share of household members working outside the farm.

Financial security considers access and use of financial services. This indix includes an indicator of use of precautionary savings, value of savings (formal and informal), access to credit (extensive and intensive margins), access to crop insurances and other social protection mechanisms (e.g. Sisben) and value of assets. This includes titles over properties and estimated value of assets (land, vehicles, machines and other assets)

Risk Sharing is measured as the extend of participation and support from social network. This measure accounts the value of transfers received from social networks to support the household.

Social Protection is measured as number of organization in which the farmer actively participantes and the frequency of participation in associations and extend to which households receive support from governamental and non-governamental institutions.

Mental health uses an index that considers farmers perceptions on their perceived capacity to organize agrobusiness (autonomy), perceived control over their lives (freedom of choice), perception of whether their fate is determined by their own actions (locus of control) and a
composite index that considers (anxiety, depression, confidence and enjoyment of life).

Food security is a index that considers the frequency in which households members where worried about not having enought to eat or experienced food shortages.

Impact of natural dissasters and economic shocks. Captures the extent to which there were economic or personal impacts due to shocks and values the magnitude of the economic effects. This measure considers the extent of use of ex-ante and ex-post risk copping mechanisms to deal with shocks.

Investments in coffee cultivation considers the extent and the value invested in coffee production. This include investment in labor and capital. We also construct a measure on the use of more sustainable práctices in the coffee cultivation (use of organic fertilizer, efficient water use, reciclying of material).

- Digital literacy
- Agricultural entrepreneurial image
- Resilience capacity
- Marketing strategies
- Risk and time preference
- Financial literacy

Digital literacy is measured using a composite index that considers: frequency of use of mobile phone (e.g. chat, listen music or watch tv, financial transactions, communication, etc.), perceived ability to handle different applications, perceived ability to adquire and select information, digital awarenes and knowledge of different Apps. The index is standarized based on meand and standard deviations for the control group.

Agricultural entreprenerial image is a perception measure on willigness to innovate, take risk and lead the community. Questions are normalized so a higher value indicates more entreprenerial skills.

Resilience capacity is a indicator of the perception on the ability to cope with difficulties and stress.

Marketing strategies consider the transformation to higher value added in coffee production chain by delivering a more elaborated product (e.g. class A dry coffee) or accessing new markets.

Trust in local organization. We measure trust in coffe cooperative, local administration, national goverment and police.

Risk aversion, patience, impulsiveness combine self-reported measures in a 1-10 scale with behaviroral question on hypothetical scenarios that imply intertemporal decisions or risky investments. The measures are not incentivized.

Financial literacy measures knowledge of financial system

The experimental design uses a between subject desing. Based on baseline characterictics of the population, we create 4 treatment groups and a control group. The groups vary the access to a mobile App to document agricultural practices (access vs. no access), the modules included in the app (Al Invest vs Croppie) and the type of training (individual or with family member).

T1: Access to the APP, individual trainning on basic modules AL invest
T2: Access to the APP, group trainning on basic modules AL invest
T3: Access to the APP, individual trainning on Croppie module
T4: Access to the APP, no training on Croppie module
Pure control: No Access to the APP, no training

Not available

The randomization is done applying the Mean Squared Error (MSE) approach developed by Schneider (2021). The method uses baseline sample characteristics to obtain balanced treatment groups by an interative process that forms groups such that it minimized the mean squared error of the baselien observable characteristics. Link: https://www.sebastianoschneider.com/publication/schneider-schlather-2021/

The unit of randomization is the individual.

No

1500 farmers

1500 farmers

450 farmers in the control, 300 farmers in croppie trainning, 450 farmers in app only, 150 farmers AL Invest individual and 150 farmers AL Invest Group.

The Minimum Detectable Effect size is 0.3 standard deviations, considering a take-up of 80 percent, an attrition level of 5 percent, and a power of 0.8 with a 5 percent significance level.