Baul Tushi, Karlan Dean, Toyama Kentaro, Vasilaky Kathryn
Playgig, United States of America.
Northwestern University, Innovations for Poverty Action, and M.I.T. Jameel Poverty Action Lab, United States of America.
J Dev Econ. 2024 Jun;169:103267. doi: 10.1016/j.jdeveco.2024.103267.
Providing agricultural advice at scale poses operational challenges. Technology may help if repeating content reinforces learning for recipients and thus improves adoption, but risks reducing efficacy given limited customization and human interaction. We tested videos shared with female farmers in India as a supplement to standard human-provided extension services promoting a climate-smart practice, System Rice Intensification. The average treatment effects are large but imprecise because of non-normally distributed outcomes, specifically fat right tails. Weighted quantile regressions show that the imprecision in estimating an average treatment effect comes from farmers with output or yields in the upper quantiles. Both quantile regressions of the 25% and 50% quantiles and a Bayesian hierarchical model (robust to several priors) reveal positive treatment effects, and two subtreatments, one that reinforces information on labor costs from adoption and a second that presents role models to motivate adoption, lead to even higher estimated treatment effects on output.
大规模提供农业建议面临操作挑战。如果重复内容能增强接受者的学习效果从而提高采用率，技术或许会有所帮助，但鉴于定制化和人际互动有限，存在降低效果的风险。我们测试了与印度女性农民分享的视频，作为推广气候智能型做法——强化栽培稻作系统的标准人工提供的推广服务的补充。由于结果呈非正态分布，特别是右侧长尾，平均处理效应较大但不精确。加权分位数回归表明，估计平均处理效应的不精确性来自产量处于上分位数的农民。25%和50%分位数的分位数回归以及贝叶斯分层模型（对多种先验稳健）均显示出积极的处理效应，并且两个子处理，一个强化采用后的劳动力成本信息，另一个展示榜样以激励采用，导致对产量的估计处理效应更高。
