| Takumi Saegusa, Department of Mathematics, University of Maryland, College Park Hosted by Yanqing Sun |
| Title: Statistical Methods for Two-phase Stratified Sampling |
| Abstract: Two-phase stratified sampling is a sampling technique for cost reduction and improving efficiency. Examples includes stratified case control study and exposure-stratified case cohort study. A main theoretical challenge is a dependent and biased sample due to sampling without replacement in each stratum. A biostatistical approach to this issue is to approximate design by stratified Bernoulli sampling and focus on few specific models including the Cox model, resulting in paucity of research in general methodology such as bootstrap. An approach from sampling theory is to impose general conditions regardless of designs, leading to implicit asymptotic distributions and inefficient statistical methods applicable for any design. Our approach, which extends empirical process theory to two-phase stratified sampling, explicitly obtains asymptotic distributions, and yields general methodology tailored to two-phase stratified sampling. In this talk, we consider three statistical problems, model selection, improving efficiency and variance estimation, arising from the RV144 case control study. Our approach illustrates inadequacy of existing methods in these problems, and naturally introduces our proposed methods. Finite sample properties are investigated in simulation studies using the logistic regression model and the Cox model. |