Graded Matching for Large Observational Studies
针对大规模观察性研究中匹配网络过密导致计算困难的问题，提出一种分级匹配策略，优先使用高质量配对，逐步引入低质量配对，在保证匹配质量的同时大幅降低计算量，并通过医学和经济学实例验证。
Observational studies of causal effects often use multivariate matching to control imbalances in measured covariates. For instance, using network optimization, one may seek the closest possible pairing for key covariates among all matches that balance a propensity score and finely balance a nominal covariate, perhaps one with many categories. This is all straightforward when matching thousands of individuals, but requires some adjustments when matching tens or hundreds of thousands of individuals. In various senses, a sparser network—one with fewer edges—permits optimization in larger samples. The question is: What is the best way to make the network sparse for matching? A network that is too sparse will eliminate from consideration possible pairings that it should consider. A network that is not sparse enough will waste computation considering pairings that do not deserve serious consideration. We propose a new graded strategy in which potential pairings are graded, with a preference for higher grade pairings. We try to match with pairs of the best grade, incorporating progressively lower grade pairs only to the degree they are needed. In effect, only sparse networks are built, stored and optimized. Two examples are discussed, a small example with 1567 matched pairs from clinical medicine, and a slightly larger example with 22,111 matched pairs from economics. The method is implemented in an R package RBestMatch available at https://github.com/ruoqiyu/RBestMatch. Supplementary materials for this article are available online.