3^rd December, 2-5pm at RSS Errol Street

 

1. Covariate measurement error - concepts, issues and methods for correction

Dr. Jonathan Bartlett, LSHTM

Abstract: In this overview I will begin by outlining the problem of measurement error in continuous covariates of regression models and the impact of ignoring such errors. I will then describe some of the most popular approaches for allowing for covariate measurement error, including method of moments, regression calibration, and multiple imputation.

2. Categorizing error-prone exposures results in bias bue to misclassification: Methods for correction

Dr. Ruth Keogh, Department of Medical Statistics, LSHTM

Abstract: To investigate the association between a continuous exposure and an outcome it is common to categorize the exposure, in particular to investigate non-linearity in the association. If the continuous exposure is measured with error, this results in misclassification when the exposure is categorized, which in turn results in bias in the estimates of the exposure-outcome association. Although categorization of exposures is common, little work has been done on this problem. In this talk I will outline methods for measurement error correction in this situation, which include use of multiple imputation and moment reconstruction. Examples will be given from a study in nutritional epidemiology.

3. Cost-efficient designs to correct for measurement error.

Dr. Roseanne McNamee and Dr. Evridiki Batistatou, Centre of Biostatistics, Institute of Population Health, University of Manchester.

Abstract: Although many exposures are measured poorly and the effects of exposure measurement error on estimates of exposure –disease relationships are well documented, in practice, correction for measurement error remains uncommon - perhaps due to perceptions about the cost implications of internal reliability studies.  The cost –efficiency of epidemiological study designs which include reliability assessment is therefore an important concept: design A is said to be more cost- efficient than study design B if, for a fixed overall study cost,  V(β_B) /V(β_A) > 1 where β is the chosen measure of effect of exposure.  We will review the literature on study design, concentrating on 2-phase designs, including our work which emphasises the cost-efficiency aspect.