Reducing classification errors in cohort studies: The approach and a practical application

Publisher Website
Google Scholar
Abstract
Classification errors of dependent variables can distort the results of observational studies. To reduce misclassification from our multicentre observational study of abortion complications, we extended the methodology of Lawrence and Greenwald9 for use in situations of unequal sample sizes and implemented both an office review and a field review. We reabstracted 424 reported complications and a random sample of 364 reported non‐serious cases from 12 institutions participating in our study. In total, 30 per cent of the reported serious complications turned out to be misclassified: the office review detected 74 per cent of the total number of misclassifications with the remainder found in the field review. Because, with our particular data base, we estimated expending only 15 per cent of the total resources with our office effort, this represented the most cost‐efficient approach to reducing classification errors. By eliminating the false positives from our study, we forced the specificity to 1·00 which produced both an unbiased estimate of the relative risk and an increase of 4 per cent to 63 per cent in the power of our study.

This publication has 12 references indexed in Scilit: