Presentation: Evaluating Estimators when Modelling Assumptions Fail
Speaker: David Whitney, Graduate Student, UW Biostatistics
Abstract: In public health and other research areas, the difference in means, relative risk, odds ratio, hazard ratio, and other summary measures are frequently reported to describe the association between an outcome and exposure. Estimation of these measures of association often follows by assuming the data is generated from a simple parametric or semiparametric regression model, appealing to the idea that “all models are wrong, but some are useful.” We propose the use of techniques from the study of regular asymptotically linear estimators to derive the large-sample interpretation and behavior of an estimator in a general nonparametric model. This framework is illustrated through case studies of commonly used regression models, including linear regression, partial linear regression, and proportional hazards regression.