Abstract
Estimation bias arising from local model uncertainty and incomplete data has been studied by Copas & Eguchi (2005) under the assumption of a correctly specified marginal model. We extend the approach to allow additional local uncertainty in the assumed marginal model, arguing that this is almost unavoidable for nonlinear problems. We present a general bias analysis and sensitivity procedure for such doubly misspecified models and illustrate the breadth of application through three examples: logistic regression with a missing confounder, measurement error for binary responses and survival analysis with frailty. We show that a double-the-variance rule is not conservative under double misspecification. The ideas are brought together in a meta-analysis of studies of rehabilitation rates for juvenile offenders.
Original language | English |
---|---|
Pages (from-to) | 285-298 |
Number of pages | 14 |
Journal | Biometrika |
Volume | 99 |
Issue number | 2 |
Early online date | 26 Feb 2012 |
DOIs | |
Publication status | Published - 1 Jun 2012 |
Externally published | Yes |