Estimating impact with surveys versus digital traces: Evidence from randomized cash transfers in Togo
研究能否用数字痕迹数据和机器学习替代传统调查来估计现金转移项目的影响，发现仅用预测结果估计不显著，但结合事后数据可得到类似结果。
We study whether program impacts can be estimated using a combination of digital trace data and machine learning. In a randomized controlled trial of cash transfers in Togo , endline survey data indicate positive treatment effects on food security , mental health , and perceived economic status. However, estimates of impact based solely on predicted endline outcomes (generated using trace data and machine learning, which do successfully predict baseline poverty) are generally not statistically significant. When post-treatment outcome data are used in conjunction with predictions to estimate treatment effects, predicted impacts are similar to those estimated using surveys.