Estimation for Regression Analysis with Validation Sampling

Abstract

Validation sampling, where a primary sample with crude data is obtained first and then a validation subsample with exact data is further selected, has been adopted as a cost-effective strategy for regression analysis in epidemiological studies. Owing to the nature of the study design, regression problems with data from validation sampling are essentially those for measurement error (with validation data) and missing data, including (a) how to extract information from data incompletely observed, and (b) how to take into account the random mechanism generating the validation data. In this talk, I will present two of my representative works from 1999 to 2002 in this area. The first (JRSSB 2002, pp. 51-62) focuses on Cox models in cohort studies, where the problem (a) is the major concern; and the second (JASA 2003, to appear) addresses both problems (a) and (b), which, relaxing an important assumption required by other methods, particularly works out a solution allowing the restricted design where the validation sample is restricted to subjects with certain outcome values.