Practical Issues in Resampling

• Activity Number: 362
• Type: Invited
• Date/Time: Wednesday, August 11, 2004 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Tim C. Hesterberg*+
• Companies: Insightful Corporation
• Keywords: bootstrap ; standard errors ; central limit theorem ; finite population

Abstract

I'll focus on some practical issues in resampling. The first is that resampling is really only practical using software that makes it easy. I'll demonstrate S-PLUS software, suitable for use by introductory statistics students on up. The graphical diagnostics included give intro stat students a better concept of sampling variability and distributions, and demonstrate to more advanced statisticians that traditional inferences based on normal approximations are often woefully inaccurate. A second practical issue is the number of bootstrap samples used; in particular you need a lot for Bca confidence intervals, many fewer for bootstrap tilting intervals. A third practical issue is that resampling should mimic how the data were collected. This is not always straightforward. For example, the naïve way of doing finite-population bootstrapping can result in greater variance than sampling without replacement. The final issue is that bootstrapping underestimates standard errors, substantially in the case of stratified sampling with small strata. There are four simple remedies.