Page 203 - 《社会》2022年第3期

社会·2022·3

causal graphs. Then，it discusses the opening and blocking of causal pathways between
variables and the three sources of bias that can mislead the identification of true causal
relationships，namely confounding bias，over鄄control bias and endogenous selection
bias. The article further introduces the D鄄separation rule used to determine which
variables in a causal inference should be controlled. On this basis，various empirical
examples are brought in to interpret four endogenous problems of omitting variable bias，
sample selection bias，self鄄selection bias，and simultaneity bias through causal graphs.
Graphic expressions and implementing conditions of several causal inference methods
are also identified，including multiple regression and matching，proxy，experiments，
instrumental variable，and panel models. In addition，this article attempts to clarify two
common misconceptions in causal inference：conditioning on a post鄄treatment variable
does not necessarily lead to bias and conditioning on a pre鄄treatment variable may cause
deviation. Finally，it is suggested that the application of the causal graph method can
help standardise causality research and facilitate the teaching and dissemination of
causal inference knowledge.
Keywords：causal graph，non鄄parametric causal inference，confounding bias，over鄄
control bias，endogenous selection bias

一、导言

因果推断是社会科学实证研究中的核心问题。 自十八世纪的休谟
开始，哲学层面对因果性相关问题已进行了丰富的探讨，中文文献中王
天夫（ 2008）、彭玉生（2011）等对此进行过系统的总结。 现代社会科学对
因果关系的探索则构建于唐纳德·鲁宾（ Donald Rubin）提出的反事实框
架（counterfactual framework）之上（H觟fler，2005；Rubin，2011），并发展出包
括实验与准实验 （ experiment and quasi鄄experiment）、 匹配和倾向值匹配
（matching/propensity score matching）、工 具 变 量（instrumental variable）、 倍
差法（ difference in differences）、断点回归（regression discontinuity）等适 用
于不同情境的因果推断方法。
国际社会学界对因果推断在实证研究中的应用研究起步较早 （如
Sobel，1996；Winship and Morgan，1999；Winship and Sobel，2004；Morgan
and Winship，2007）。 2010 年以来，相关问题在中国社会学界逐渐得到关
注。 如陈云松与范晓光（2010，2011）、胡安宁（2020）系统介绍了影响因

· 196·