Bootstrap and Jackknife Techniques Applied to Complicated Data Structure Contaminated with Censored Outcomes

• Activity Number: 288
• Type: Contributed
• Date/Time: Wednesday, August 14, 2002 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section*
• Author(s): Philimon Gona*+ and Jose Becerra and Martin Larson
• Affiliation(s): PPD Development and Centers for Disease Control and Prevention and Boston University
• Keywords: Semi-Markov Models ; Censored Outcome ; Bootstrap and Jackknife ; EM Algorithm

Abstract

Semi-Markov models do not always yield stable covariance matrices when data are censored or in highly multivariable situations. Jackknife and bootstrap procedures avoid having to derive formulae via difficult analytical arguments for solving complex inferential problems in exchange for accurate estimates without the need to impose a convenient statistical model that may not have strong scientific basis. We applied jackknife and bootstrap methods as alternative approaches for assessing errors in a semi-Markov setting with some censored outcomes. The goal was to estimate transition probabilities for two possible outcome states and state-specific survival parameters using parametric models. We obtained jackknife and bootstrap estimates and compared them to EM estimates using USA completion of TB surveillance data set in which 15% of records had censored TB completion status (completed/died).

Bias and variance of jackknife and bootstrap estimates were assessed relative to EM estimates. EM estimates were not perfect but not very biased. Jackknife estimates were numerically closer to EM estimates than bootstrap estimates. EM did not did not produce serious underestimation of variance.