Scholarlink

<p>In Pakistan’s evolving agricultural landscape, smallholder farmers are increasingly exiting farming, not by choice but under structural pressure. This study investigates the underlying associations using an integrated machine learning and econometric approach, based on novel data from 500 current and former farmers. We employ CatBoost, a high-performance gradient boosting algorithm designed to handle categorical data efficiently, in combination with recursive feature elimination and cross-validation to identify seven key features of exit. CatBoost demonstrates robust performance on both imbalanced and balanced datasets. SHapley Additive Explanations (SHAP) analysis reveals that reliance on full credit, high debt, distant markets, and natural hazards significantly increase exit risk. Conversely, larger landholdings, non-farming income, and livestock ownership serve as protective factors. Logistic regression confirms these associations, with marginal effects indicating that reliance on credit rather than cash is associated with a 9.5% higher probability of exit, and a 22.6% increase for high debt. Market distance is associated with an 18.3% increase, whereas each additional unit of land size reduces exit risk by 15.5%. Subgroup analysis reveals exit rates of 72.6% among credit-dependent farmers and 56.0% among those with less than five acres, compared to none among large landowners. By framing exit as structural exclusion, this study offers a scalable, explainable AI tool for inclusive, faith-sensitive rural reform aligned with food security, poverty reduction, and multiple Sustainable Development Goals.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Exit without choice: interpretable machine learning unlocks the structural drivers of smallholder dispossession in Pakistan

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月5日

查看全文
  • Ghazi Abbas,
  • Zhou Ying,
  • Chi Goutai,
  • Qamar Abbas,
  • Khalil El Hindi

摘要

In Pakistan’s evolving agricultural landscape, smallholder farmers are increasingly exiting farming, not by choice but under structural pressure. This study investigates the underlying associations using an integrated machine learning and econometric approach, based on novel data from 500 current and former farmers. We employ CatBoost, a high-performance gradient boosting algorithm designed to handle categorical data efficiently, in combination with recursive feature elimination and cross-validation to identify seven key features of exit. CatBoost demonstrates robust performance on both imbalanced and balanced datasets. SHapley Additive Explanations (SHAP) analysis reveals that reliance on full credit, high debt, distant markets, and natural hazards significantly increase exit risk. Conversely, larger landholdings, non-farming income, and livestock ownership serve as protective factors. Logistic regression confirms these associations, with marginal effects indicating that reliance on credit rather than cash is associated with a 9.5% higher probability of exit, and a 22.6% increase for high debt. Market distance is associated with an 18.3% increase, whereas each additional unit of land size reduces exit risk by 15.5%. Subgroup analysis reveals exit rates of 72.6% among credit-dependent farmers and 56.0% among those with less than five acres, compared to none among large landowners. By framing exit as structural exclusion, this study offers a scalable, explainable AI tool for inclusive, faith-sensitive rural reform aligned with food security, poverty reduction, and multiple Sustainable Development Goals.