由于充分随机试验在实际使用的局限性，非充分随机化的观测研究（Observational studies）是统计应用中常用的研究方法。它可以有效无偏地在备择假设下检测到对照样本和实验样本的区别。但观测研究需要对由于没有进行充分随机化的数据样本所带来的选择偏差进行去除。目前的观测研究还主要是针对独立数据并使用传统的回归分析方法来调整混杂干扰因素。本项目将以计量经济学、环境统计学、统计理论中的交叉问题为背景，提出针对时间空间数据的观测研究方法的几个重要问题展开研究，使用现代统计学习方法研究观测研究中基于时空数据的统计推断方法和算法。主要研究内容包括（1）构造针对时间空间数据的度量试验效应统计指标和统计估计框架；（2）通过修正随机森林方法以适应时空数据的特征，给出基于随机森林及其它统计学习方法的试验效应统计估计；（3）研究动态观测实验模型和动态实验效果的估计方法。

Existing observational studies are largely conducted for independent or time series data with conventional tools for regression estimation for the conditional average outcome given some covariates. The project will investigate several important issues of statistical inference for observational studies with spatial-temporal data, which are commonly collected these days in economic, public health and environmental related problems in the econometric, environmental and statistical type of interdiscipline studies. We will propose (i) valid statistical measures for treatment effects (TEs) and (ii) utilize the random forrest and other statistical learning methods to estimate the measures for the TEs. Renovations of the random forrest procedures to suit spatial-temporal data features will be conducted. We will also study inferences on the TEs for dynamic treatment actions as a generalization to the conventional time-invariant actions, as well as introducing variable selection mechanism in quasi-experiments to enhance the match rates of the observational studies.