Assessment of record linkage and measurement error in cohort mortality studies.

摘要

Epidemiological studies are designed to investigate disease causation through observational study of human population groups. The overall statistical goal of an epidemiological study is accuracy and precision in estimation; that is, to estimate the value of the parameter of interest with as little error as possible. In addition to random errors, epidemiologic studies may involve systematic error which undermines their validity. Record linkage error and measurement error are two common errors that occur in epidemiological studies.; In this thesis, we study the impact of linkage errors on estimates of various epidemiological indicators of risk such as relative risk regression model parameters and standard mortality ratios. We describe a method to adjust the biases in the indicators of risk in the presence of linkage errors. To explore the impact of record linkage errors in cohort mortality studies in more detail, we conduct a simulation study based on actual data from the National Dose Registry of Canada being linked with the Canadian Mortality Data Base.; The proportional hazards regression model is commonly used to model survival data as a function of covariates. Typically, these covariates are measured with error and are not directly observable. The Cox proportional hazards model is applied in the Harvard Six Cities Study to estimate the mortality rate ratios for exposure to different air pollutants in individuals, with and without other risk factors as covariates. There are several methods in the literature for adjusting for measurement errors. In this thesis, we consider two recent methods, namely RCAL and SIMEX, and evaluate their effectiveness in correcting measurement errors in the Cox proportional hazard model, using real data from the Harvard Six Cities Study. We use computer simulation to evaluate the accuracy and precision of both RCAL and SIMEX, and to assess the robustness of RCAL to mis-specification of the measurement error distribution.