Causal effect sizes may vary among individuals and they can even be of opposite directions. When there exists serious effect heterogeneity, the population average causal effect (ACE) is not very informative. It is well-known that individual causal effects (ICEs) cannot be determined in cross-sectional studies, but we will show that ICEs can be retrieved from longitudinal data under certain conditions. We will present a general framework for individual causality where we will view effect heterogeneity as an individual-specific effect modification that can be parameterized with a latent variable, the receptiveness factor. The distribution of the receptiveness factor can be retrieved, and it will enable us to study the contrast of the potential outcomes of an individual under stationarity assumptions. Within the framework, we will study the joint distribution of the individual's potential outcomes conditioned on all individual's factual data and subsequently the distribution of the cross-world causal effect (CWCE). We discuss conditions such that the latter converges to a degenerated distribution, in which case the ICE can be estimated consistently. To demonstrate the use of this general framework, we present examples in which the outcome process can be parameterized as a (generalized) linear mixed model.
翻译:当存在严重的影响异质性时,人口平均因果关系(ACE)的分布情况并不十分丰富。众所周知,单项因果关系(ICES)无法在跨部门研究中确定,但我们将表明,在某些条件下,可以从纵向数据中检索ICE。我们将提出个人因果关系的一般框架,在这个框架内,我们将研究个人潜在结果的联合分布情况,其条件取决于所有个人的事实数据,随后的跨世界因果关系的分布。我们将讨论以下条件,即后者与退化分布相融合,可以与潜在变量即接受性系数相参照。可检索接受性系数的分布,它将使我们能够研究个人在定置性假设下的潜在结果的对比。我们将在框架内研究个人潜在结果的联合分布情况,以所有个人的事实数据为条件,然后研究跨世界因果关系的分布。我们将讨论后者与退化分布相融合的条件,在这种条件下,ICE的分布可以与潜在变量(接受性系数)相匹配。为了显示这一总框架的使用情况,我们提出了可以将结果过程作为线性参数混合的参数。