A review of methods for misclassified categorical data in epidemiology

Abstract

Misclassification introduces errors in categorical variables. This paper presents a review of methods for misclassified categorical data in epidemiology. Different sampling schemes for a 2 x 2 x 2 table and methods of analyses will be discussed first. A misclassification matrix is defined, and the usual misclassification models will be shown to be a subclass of log-linear models. Well-known results on a 2 x 2 table with misclassification and recent results on a 2 x 2 x 2 table are then reviewed. Finally two methods of adjusting for misclassification will be given. The first method assumes a known misclassification matrix, and the second method uses subsampling to estimate the misclassification matrix. The analysis is based on a recursive system of log-linear models: first determine a misclassification model, then select a model for the correctly classified variables. The methods are illustrated by data from traffic safety research on the effectiveness of seatbelt use in reducing injuries.