统计模型的“不确定性”问题与倾向值方法  

Uncertainty of Statistical Models and Propensity Score Methods

作　　者：胡安宁[1] HU Anning

机构地区：[1]复旦大学社会学系

出　　处：《社会》2017年第1期186-210,共25页Chinese Journal of Sociology

基　　金：国家社科基金青年项目(15CSH030);上海市教育委员会科研创新项目(15ZS001);复旦大学"卓学人才计划"项目的支持~~

摘　　要：量化社会学研究往往基于特定的统计模型展开。近十几年来日益流行的倾向值方法也不例外,其在实施过程中需要同时拟合估计倾向值得分的"倾向值模型"与估计因果关系的"结果模型"。然而,无论是其模型形式还是系数估计,统计模型本身都具有不可忽视的"不确定性"问题。本研究在倾向值分析方法的框架下,系统梳理和阐释了模型形式不确定性与模型系数不确定性的内涵及其处理方法。通过分析"蒙特卡洛模拟"数据与经验调查数据,本文展示了在使用倾向值方法进行因果估计的过程中,研究者如何通过"贝叶斯平均法"进行多个备选倾向值模型的选择,以及如何通过联合估计解决倾向值模型与估计模型中的系数不确定性问题。本文的研究也表明,在考虑倾向值估计过程的不确定性之后,结果模型中对于因果关系的估计呈现更小的置信区间和更高的统计效率。Quantitative sociological research has always employed certain specific statistical models. Over the past several decades, the focus on causal relationship in sociological studies has led to a wide spread application of propensity score methods. Using an explicit estimation of the probability of being subject to a specific treatment or intervention, sociologists are able to mimic random experiments to predict causal effects. In practice, propensity score methods require an estimation from two models： one predicts propensity scores and the other estimates causal effects. However, the model structure and coefficient of both contain considerable uncertairity. This study offers a systematic review of the model structure and coefficient uncertainty in propensity score methods as well ascertain strategies to tackle the issue. By analyzing Monte Carlo＇s simulated data along with empirical survey statistics, the paper demonstrates how researchers can use Bayesian Model Averaging to select multiple backup models and deal with possible model-coefficient uncertainty with the joint maximum likelihood estimation in propensity score methods. The paper also finds that after taking into account of various sources of uncertainty, the estimated causal effects display a narrower confidence interval but a higher level of statistical efficiency.

关 键 词：模型形式不确定性 模型系数不确定性 贝叶斯平均 倾向值方法 统计效率 

分 类 号：C8[社会学—统计学]