处理效应模型在医学研究领域一直以来都是研究药物、手术等治疗效果的重要模型，后来该模型逐渐拓展到其他自然和社会科学领域。随着研究的深入和领域的拓展，利用观测数据进行处理效应的估计越来越受到人们的重视。但是，由于“选择偏差”的存在，利用观测数据进行处理效应估计是存在一定困难的。同时观测数据中总是存在着数据缺失的问题，因为实际问题中总是会由于各种人为或者未知的客观因素导致一定的缺失数据。所以本文主要讨论的问题就是利用存在缺失的观测数据进行平均处理效应的估计。 本文主要研究的是多维协变量存在“选择偏差”的情况下，若因变量存在随机缺失的平均处理效应估计。由于协变量是多维的，所以会导致非参数估计方法的失效。为了使用非参数方法就需要进行维数的降低，同时为了处理“选择偏差”需要利用倾向得分，而倾向得分是一个一维变量。所以，可以利用倾向得分将多维协变量的函数投影到倾向得分的空间中，从而实现协变量维数的降低。此时，可以将模型转化为部分线性模型，运用B-样条方法拟合关于倾向得分的一维函数。为了处理因变量的数据缺失，对估计的得分函数进行逆概率加权，进而得到平均处理效应的估计。随后，本文对平均处理效应的估计量的渐进正态性进行了证明。最后，利用数值模拟的方法说明了该方法的优势，并展现了估计结果的渐进正态性。

Treatment effect model has been used to evaluate the effect of medicine in the medical domain for a long time. Later the model was introduced to other domains. With the development of the model, estimation of treatment effect with observed data be-comes much more popular. However, it’s difficult to estimate treatment effect with observed data because of the “selection bias”. Meanwhile, some parts of the data are always missing in observed data for various reasons, such as human factor or unknown natural factor. Therefore, the question discussed in this paper is estimating average treatment effect with missing observed data. In this paper, we focus on estimating treatment effect using multivariate predictors with “selection bias” and missing outcomes. Because the predictors are multivariate, the nonparametric method will suffer the “curse of dimension”. Therefore, we must reduce the dimension to use the nonparametric method. Meanwhile, the propensity score is used to deal with the “selection bias”. Therefore, we can project the function of multivariate predictors to the propensity score to reduce the dimension. The model will be changed to a partially linear model which can be handled with B-spline. In order to deal with the missing outcomes, we using the IPW to weight the score function. Then we proved the asymptotic normality of the estimator. Lastly, we show the advantage and the asymptotic normality of the estimator with simulation study.