Socio-Economic and Institutional Drivers of Willingness to Pay for Sustainable Arsenic Mitigation in Rural India

Technical Summary: Socio-Economic and Institutional Drivers of Willingness to Pay for Sustainable Arsenic Mitigation in Rural India

Problem Statement
Despite the widespread recognition of severe public health risks associated with arsenic-contaminated groundwater, the adoption of mitigation technologies in affected regions, particularly in rural India, remains persistently low. This study addresses the critical disconnect between awareness of arsenic risks and actual household-level adoption behavior. While traditional approaches often emphasize awareness and perceived health risks as primary drivers, this research posits that adoption is shaped by a complex interplay of socio-economic constraints, institutional trust, social hierarchy, and experiential burdens related to water access. The study focuses on rural Bihar, a major hotspot where groundwater arsenic levels frequently exceed WHO guidelines, to investigate the determinants of Willingness to Pay (WTP) for sustainable mitigation technologies.

Methodology
The study employs an integrated analytical framework combining traditional statistical analysis with interpretable machine learning (ML) techniques to model household WTP behavior.

• Data Collection: Primary data were gathered from 340 households across three arsenic-affected villages (Suarmarwa, Rampur Diara, and Bhawani Tola) in the Maner block of Patna district, Bihar. The survey captured demographic characteristics, livelihood conditions, water access burdens, institutional trust, social participation, and risk perceptions.
• Variable Definition: The dependent variable, WTP, was dichotomized into "Low/No WTP" (≤ ₹25) and "Higher WTP" (> ₹25) based on the observed distribution of responses. Independent variables included socio-demographics (caste, education, occupation), economic indicators (income, landholding), water access metrics (distance, time, labor), institutional trust (government, NGOs, PRIs), and risk perceptions (health, economic, social).
• Analytical Approach:
  1. Statistical Analysis: Descriptive statistics, contingency tables, and Chi-Squared tests were used to assess associations. Spearman's rank correlation was employed to evaluate monotonic relationships, accounting for the ordinal nature of the data.
  2. Machine Learning Modeling: A suite of classification algorithms was evaluated, including Logistic Regression (LR), Support Vector Machines (SVC, NuSVC, Linear SVC), Random Forest (RF), Gradient Boosting, Decision Trees, K-Nearest Neighbors, Naïve Bayes, and Multi-layer Perceptron (MLP). Models were trained on 75% of the data and validated on a 25% holdout set.
  3. Performance Metrics: Models were evaluated using accuracy, precision, recall, F1-score, Area Under the ROC Curve (AUC), and Cohen's Kappa. The F1-score served as the primary metric for model selection to balance false positives and false negatives.
  4. Explainability: SHapley Additive exPlanations (SHAP) were applied to the benchmark Logistic Regression model to quantify the magnitude and direction of feature contributions, ensuring the ML outputs aligned with statistical inference and offered policy-relevant interpretability.

Key Contributions

1. Integrated Framework: The study develops a novel framework that combines statistical inference with interpretable ML and SHAP-based explainability to dissect the multidimensional nature of WTP behavior.
2. Re-evaluation of Drivers: It systematically challenges the prevailing assumption that awareness and perceived health risks are the dominant drivers of adoption. Instead, it provides evidence that socio-economic feasibility, social structure, and institutional trust play more substantial roles.
3. Methodological Rigor: By comparing linear and non-linear models, the study demonstrates that interpretable linear models (specifically Logistic Regression) can perform as robustly as complex non-linear classifiers in this socio-economic context, offering a balance between predictive power and policy interpretability.

Results

• Statistical Findings: WTP exhibited significant heterogeneity across villages and was strongly associated with caste, occupation, income, and institutional trust. Notably, variables such as gender, age, marital status, and basic arsenic awareness showed no statistically significant association with WTP. Perceived economic risk was a significant predictor, whereas perceived health risk was not, reinforcing the finding that economic constraints often outweigh health concerns in decision-making.
• Model Performance: Logistic Regression emerged as the most robust and interpretable model, achieving a test accuracy of 0.753, an F1-score of 0.687, an AUC of 0.743, and a Kappa of 0.495. While non-linear models like NuSVC and Random Forest showed competitive performance, they exhibited signs of overfitting (e.g., Random Forest training accuracy of 1.000 dropping to 0.741 on test data). The moderate performance across all models (AUC ~0.70–0.74) suggests that WTP behavior is inherently complex and context-dependent, with significant overlap between adopters and non-adopters.
• SHAP Interpretation: The SHAP analysis identified caste, agricultural landholding class (AgLHC), perceived social risk, and trust in NGOs as the most influential predictors. Crucially, the analysis revealed mixed directional effects: while higher perceived economic risk and asset ownership generally increased WTP, higher levels of certain institutional trust variables and participation indices sometimes correlated with lower WTP, suggesting that structural constraints and institutional inefficiencies can act as barriers even when engagement is high.

Significance and Claims
The paper claims that its findings underscore the insufficiency of awareness-driven approaches alone to improve the adoption of arsenic mitigation technologies. The study concludes that effective interventions must address affordability, institutional credibility, and socio-economic vulnerabilities. By demonstrating that WTP is deeply embedded within social hierarchies and governance dynamics rather than just health risk perception, the research advocates for context-specific, trust-driven intervention strategies. The study positions itself as a step toward designing economically viable and socially inclusive mitigation strategies for resource-constrained settings, emphasizing that sustainable solutions require addressing the broader socio-environmental challenges beyond mere technological provision. The authors maintain a modest tone, acknowledging limitations such as the cross-sectional nature of the data and the use of stated-preference measures, and call for future validation using longitudinal data and revealed-preference measures.