因果推论与效应评估：区段识别法及其於「选制效应」之应用

Causal Inference and Treatment Effect Evaluation: Partial Identification Approach and Its Application to Electoral System Effect

社会科学中涉及效应评估的问题，都无法回避因果推论。以观察研究进行因果推论之所以棘手，症结在於比较研究的组别，往往取决於因和果之间的内部因素，也就是所谓「内因性」(endogeneity)，造成平均因果效应的识别问题。一般分析因果效应的参数模型(parametric models)，虽有考虑内因问题，但多建立在很特定的函数形式及变数分布等假定之上。如果研究的主题及资料的确符合这些假定，自可充分运用；但社会科学研究也常常会碰到与假定不符的情况，此时Manski的无母数局部识别法(nonparametric partial identification)最爲适合，因爲这个方法从无假定出发，逐步带入不同强度的假定，检视其对於参数区段的影响，将假定与推论之间的关系完全透明化，避免爲了达到「定点识别」而强加或暗藏与实际不符的假定，导致过当的推论。本文从「反事实因果模型」(counterfactual model of causality)的角度，以最基础的逻辑与机率论，探讨Manski的区段识别法，及各种学理假定与「平均因果效应」之上下限的关系，并以2008年立委选举台联提名区域立委对其政党票得票率之影响爲例，将区段识别法应用於分析混合选制中所谓之「污染效应」(contamination effect)。

In social science we routinely ask questions of the form: What is the effect of X on Y? Attempts to answer these questions unavoidably involve causal inference. However, social scientists relying on observational studies are often plagued by the endogeneity problem. That is, the treatment and control groups are not randomly assigned by researchers but formed spontaneously by some factors related to the causal variable of interest. Some existing parametric models, such as the popular Heckman's treatment-effects model, do take account endogeneity problem but are built upon quite stringent functional and distributional assumptions such as linearity and bivariate Normal distribution. Powerful as they are in point identifying causal parameters, their assumptions are not always met in reality. When these assumptions are violated, a better alternative is to adopt Charles F. Manski's nonparametric partial identification approach. This uncommon approach promotes forthright acknowledge of ambiguity in social science research and discredits misplaced certainty of point identification at the cost of imposing strong and yet incredible assumptions. Relying on available data and weak but credible assumptions, partial identification theory reveals the causal effect parameter that lies in a set that is smaller than the logical range of the parameter but lager than a single point. Yet it makes transparent the relationship between maintained assumptions and causal inference.Starting from the counterfactual model of causality, this article introduces Manski's partial identification theory and examines its implications on the upper and lower bounds of the average treatment effect (ATE). We then illustrate the approach by applying it to the case of Taiwan's 2008 Legislative Yuan election and examining whether Taiwan Solidarity Union's nomination in 13 single-member districts had any ”contamination effect” on its party list vote shares.