Sensitivity Analysis for Matched Observational Studies with Many Ordered Treatments
提出一种在匹配对研究中，当处理有多个水平或剂量时，展示置换推断对未观测协变量敏感性的方法，推广了此前针对两个处理组的程序。
Permutation inferences based on randomization distributions are not generally applicable in observational studies, that is, studies in which treatments are not randomly assigned to experimental subjects. Treated and control subjects who appear similar based on recorded measurements may none the less differ in ways that have not been observed and recorded. It is possible to display the sensitivity of permutation inferences to a range of assumptions about unobserved covariates. Studies vary considerably in their sensitivity: small unobserved covariate differences suffice to alter the qualitative conclusions in some studies, while extremely large differences would be required to do so in others. A method is proposed for displaying the sensitivity of permutation inferences in matched pairs when the treatment occurs at several levels or doses, perhaps infinitely many; the method generalizes a procedure previously proposed for two treatment groups. The method uses a semiparametric model for the distribu- tion of dose levels given observed and unobserved covariates. The model is a member of the exponential family, with an infinite dimensional nuisance parameter eliminated by conditioning, and a scalar sensitivity parameter that is varied systematically to display the senstivity of the inference to unobserved covariate differences. As special cases, the model includes: (i) logit models for two doses; (ii) linear models for normally distributed doses; (iii) certain log-linear models for doses taking several discrete values; and (iv) other dose distributions within the exponential family.