Use this URL to cite or link to this record in EThOS:
Title: Correction for classical covariate measurement error and extensions for life-course studies
Author: Bartlett, Jonathan William
ISNI:       0000 0004 2702 0570
Awarding Body: London School of Hygiene and Tropical Medicine (University of London)
Current Institution: London School of Hygiene and Tropical Medicine (University of London)
Date of Award: 2010
Availability of Full Text:
Full text unavailable from EThOS.
Please contact the current institution’s library for further details.
Measurement error in the covariates of regression models is a common problem in epidemiology, generally causing bias in estimates of covariate effects. In the first part of the thesis we examine the effects of, and methods to allow for, classical covariate measurement error in regression models for continuous, binary, and censored time-to-event outcomes. We describe the most commonly used estimation methods, including regression calibration (RC), maximum likelihood (ML), the conditional score method, multiple imputation (MI), and moment reconstruction. We demonstrate how, for continuous and binary outcomes, MLEs for particular parametric specifications can be obtained by fitting a simple linear mixed model to the error-prone measurements. We also illustrate how MLEs for certain parametric specifications can be obtained by the Monte-Carlo Expectation Maximization (MCEM) algorithm. This includes a novel proposal for multiply imputing the covariate measured with error for Cox proportional hazards outcome models using rejection sampling. Simulations are used to compare the performance of the methods. In the second part of the thesis we consider the extension of these methods to lifecourse studies. We show that our proposal for ML estimation in the case of classical covariate measurement error and the MCEM algorithm can both be extended to this more general setting. In applications we typically do not know which aspects of a longitudinal trajectory influence the outcome of interest, leading us to fit a number of different models. We show that naive application of RC gives biased parameter estimates when important aspects of the longitudinal trajectory are omitted from the outcome model. We show how multiple imputations, which are a by-product of the MCEM algorithm, may be used to obtain consistent estimates in such situations. In the final part of the thesis we use data from the Framingham Heart Study to illustrate the application of RC and our proposed estimation approaches
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available