Repeat measurement of case-control data: correcting risk estimates for misclassification due to regression dilution...

In a case-control study to determine the role of lipids as risk factors for ischemic cerebrovascular disease, 105 cases of transient ischemic attacks and minor ischemic strokes were compared with 241 controls. Recruitment to the study took place in Oxford, United Kingdom, in 1986 and 1987. In a random subset of 100 of the original controls, repeat assays of total cholesterol, low density lipoprotein cholesterol, (LDL cholesterol), and high density lipoprotein cholesterol (HDL cholesterol) were performed. With maximum likelihood theory, the repeat data were used to produce odds ratio estimates of relative risk corrected for mismeasurement. Uncorrected odds ratios associated with total cholesterol of more than 6 mmol/liter, LDL cholesterol of more than 3.5 mmol/liter, and HDL cholesterol of more than 1.2 mmol/liter were, respectively, 2.06 (95% confidence interval (CI) 1.26-3.37), 2.02 (95% CI 1.24-3.30), and 0.46 (95% CI 0.28-0.74). Corrected for mismeasurement, the corresponding odds ratios were 2.90 (95% CI 1.42-5.93), 2.57 (95% CI 1.24-5.32), and 0.36 (95% CI 0.17-0.71), respectively. The attributable risks changed from 34 to 48% for total cholesterol, from 34 to 43% for LDL cholesterol, and from 35 to 45% for HDL cholesterol. The maximum likelihood procedure described here corrects for substantial underestimation of the strength of an association and indicates the need to obtain repeat measurements on variables which are subject to random fluctuation.