The financial well-being of fruit farmers in Chile and Tunisia depends more on social and geographical factors than on climate change

The study investigates whether and how climate change affects the financial well-being of fruit farms, focusing on cherry growers in Chile and peach growers in Tunisia. Prior work shows climate change, particularly rising temperatures, can reduce yields of major crops, yet evidence on fruit crops and on farm-level financial outcomes in North Africa and South America is scarce. The authors aim to link farmers’ experiences of climate change (temperature increase, reduced precipitation, drought frequency, extreme weather) with self-perceived financial well-being, and to compare the importance of climatic versus non-climatic (social, human, biophysical, economic, and regional) factors. To address complex, high-dimensional data and maintain interpretability, the study uses a hybrid approach combining generalized linear models with boosting-based variable selection. The research addresses three questions: (1) Do climate change experiences affect financial well-being? (2) If not, what other factors matter? (3) Do interactions among factors modulate these effects?

The literature has largely concentrated on climate impacts for staple crops (wheat, rice, maize, soybean), often using simulation and economic models such as the Ricardian approach, with limited attention to fruit crops and to farm-level financial well-being in regions like North Africa and South America. Prior studies document temperature-driven yield declines and impacts of extreme events, as well as crop- and region-specific vulnerabilities. Methodologically, classical econometric models can struggle with high-dimensional, interacting determinants of farm outcomes, motivating the integration of machine learning for variable selection and prediction. Machine learning has shown promise in agricultural big data analyses, including yield prediction and climate impact assessment, but interpretability can be limited. The authors position a hybrid, interpretable boosting framework to bridge the gap between predictive performance and explanatory insight for financial outcomes of fruit farms.

Study areas comprised four regions: two in Tunisia (Mornag—Northern Tunisia; Regueb—Central Tunisia) and two in Chile (Rengo—Central Chile; Chillán—Southern Chile), with contrasting Mediterranean to semi-arid climates. Data were collected via face-to-face surveys of 801 farmers (401 peach farms in Tunisia; 400 cherry farms in Chile) after harvest (Tunisia: fall 2018; Chile: spring 2019). Sampling frames came from national Ministries of Agriculture; randomly selected farmers had to own, manage, and work the farm and derive >70% of income from farming. The survey, prepared in English, translated and back-translated into Tunisian Arabic and Chilean Spanish, used multiple-choice, open-ended, Likert, and Yes/No items concerning experiences over 2009–2018, adaptive measures, and farm characteristics. Independent variables included both individual items and grouped assets: natural (region), human (education, age, gender, knowledge), social (information use; trust in information sources, community, science, religion), biophysical/manufactured (farm size, water systems, crop diversity, adaptive measures), economic (debt, performance, reliance on orchard income), climate experience (changes in temperature, precipitation, drought, extreme weather), and income damage (financial impacts attributed to specific climate factors). The dependent variable was self-assessed farm financial well-being: Doing well/very well; Neutral; Not well/not well at all. Two binary outcomes were constructed: High well-being (well/very well vs. neutral+not well) and Low well-being (not well vs. neutral+well/very well). Analysis strategy: (1) Logistic regression estimated odds ratios (ORs), p-values, and CIs for climate experiences and income damage effects on high and low well-being. (2) To identify non-climatic predictors amidst many variables and groups, model-based boosting with sparse group boosting (sgb) was used for variable selection and effect estimation, allowing the algorithm to select individual variables or grouped assets. Cross-validation (25-fold) controlled boosting iterations and regularization; ridge shrinkage moved small effects toward zero. (3) Pairwise interactions among variables were explored using component-wise boosting to assess relative importance, acknowledging the high-dimensional search space. (4) Benchmark machine learning models (random forest, gradient boosted trees, neural networks) and generalized linear models were compared on predictive performance. Data were split 70/30 into training/test sets; performance was evaluated via AUC and accuracy across thresholds. All analyses were performed in R; visualizations used ggplot2.

- Sample and context: 801 farmers (401 Tunisia; 400 Chile). Farmers’ 10-year perceptions of temperature and precipitation changes generally matched 30-year meteorological trends. - Climate experiences and financial well-being: - Decreasing rainfall was associated with reduced odds of high financial well-being when pooling Chile and Tunisia (OR≈0.635, p=0.020). Increased temperatures showed a weaker, marginal association with reduced odds of high well-being (OR≈0.751, p=0.11). - Increasing drought frequency and extreme weather experiences were not significantly associated with high or low well-being across regions. - Country differences: Negative effects of reduced rainfall and increased temperature were more discernible in Chile than Tunisia. - Income damage linked to specific climate factors: - For farms doing well financially, rainfall-reduction-related income damage was significantly associated with lower odds of high well-being (Chile+Tunisia OR≈0.568, p=0.002; Chile OR≈0.434, p<0.001). - For farms not doing well, temperature-related income damage was associated with higher odds of low well-being in Chile (OR≈2.119, p=0.021). Increased drought frequency was associated with higher odds of low well-being (Chile+Tunisia OR≈2.457, p<0.001; Chile OR≈2.623, p=0.003; Tunisia OR≈2.385, p=0.006). Decreasing rainfall contributed somewhat to explaining low well-being, particularly in Chile. - Variable importance (sparse group boosting): - Predictors unrelated to climate experiences were generally more important for predicting financial well-being. - High well-being: Social assets (use/trust of information sources, community/science/religion) and biophysical assets (farm size, water systems, crop diversity) were key in both countries. Years of farm ownership tended to have a negative effect. Regional differences (natural assets) strongly predicted outcomes in Chile (Central Chile > Southern Chile). In Tunisia, prior family ownership and human assets (education, age, gender, knowledge) were important. - Low well-being: Regional differences, income impact, and economic assets were most influential in both countries; higher debt and heavier reliance on orchard income increased low well-being odds. Tunisia-specific factors included length of ownership, drought, social and biophysical assets, and varieties grown (the latter three linked to reduced low well-being). Chile-specific individual factors included use of wells (reduced low well-being) and years of farm management (increased low well-being). - Some factors predicted both high and low well-being with opposing effects (e.g., well usage in Chile increased high and reduced low well-being; prior family ownership in Tunisia reduced high and increased low well-being). Biophysical assets reduced odds of both high and low well-being, indicating greater utility for farmers with low well-being. - Interactions: - Interactions among social and human assets—especially information use and trust—were salient. Interactions involving adaptive measures, current climate change acceptance, and education also mattered. - Use of newspapers buffered the negative impacts of increasing temperature and decreasing precipitation on high well-being; trusting industry showed similar buffering, particularly in Tunisia. In Chile, trust in industry sometimes intensified temperature-related negatives. - Trust in media combined with industry information use yielded the highest probability of high well-being; trust in industry synergized with trust in experts and government. - Education interacted positively with media use in Chile (not in Tunisia) to improve financial well-being probabilities. - Predictive performance: Random forest generally outperformed other models across settings, though the sparse group boosting and model-based boosting provided strong, interpretable performance. Generalized linear models using only climate experiences and their financial impacts had lower predictive power, underscoring the need to include broader asset and regional variables.

The analysis shows that while specific climate change experiences—particularly reduced precipitation and, to a lesser extent, increased temperatures—are associated with reduced farm financial well-being (especially in Chile), climate-related factors are not the dominant predictors of financial outcomes when considered alongside social, human, biophysical, economic, and regional determinants. For farms already performing well, social assets and biophysical capacities are decisive, suggesting that strengthening information access, trust networks, and on-farm systems can sustain financial resilience. For farms struggling financially, climate-related income damages (notably from drought and heat) and economic constraints (debt, reliance on orchard income) become more salient, indicating targeted support is needed. Interactions reveal that trust and use of information sources can buffer adverse climatic effects, highlighting the importance of communication strategies and credible intermediaries (industry, experts, government). Regional heterogeneity is crucial: location is the strongest predictor in Chile, while prior family ownership and human capital are more influential in Tunisia. These findings address the research questions by demonstrating (1) climate change has measurable but secondary effects on financial well-being compared to social/geographical factors; (2) non-climatic assets and regional context are key drivers; and (3) interactions—especially involving trust and information—modulate outcomes and can mitigate climatic risks. Policy implications include prioritizing social capital, tailored regional interventions, economic risk management, and effective information ecosystems alongside climate adaptation measures.

The study introduces a hybrid, interpretable analytical approach that integrates generalized linear models with boosting-based variable selection to predict farm financial well-being using high-dimensional survey data. Empirically, it identifies that social assets, biophysical capacities, and regional context outperform climate experiences in predicting fruit farm financial well-being in Chile and Tunisia, although reduced precipitation and higher temperatures still pose risks—especially for farms already under financial strain. Interactions underscore the buffering role of trust and information use. Policymakers should strengthen information access and trust networks, support economic and water management measures, and tailor interventions regionally. Future research should balance theory-driven parsimony with data-driven complexity, collect actual financial performance data to validate self-assessments, consider larger and more granular regional samples, and further examine interaction structures and non-linearities while maintaining interpretability.

Key limitations include reliance on self-reported financial well-being rather than audited financial data, which may introduce perception and recall biases; a moderate sample size (n=801) constrained subnational comparisons and increased the risk of false selections in a high-dimensional variable space; practical constraints (one-hour interviews) limited broader sampling and variable depth; and while the hybrid approach improves interpretability over black-box models, classical inference (e.g., t-tests) is not valid post data-driven selection. Generalizability may be limited to similar regions and crop systems.