The effect of misclassification on the estimation of association: a review

✍ Scribed by Michael Höfler

• Publisher: John Wiley and Sons
• Year: 2005
• Tongue: English
• Weight: 155 KB
• Volume: 14
• Category: Article
• ISSN: 1049-8931
• DOI: 10.1002/mpr.20

✦ Synopsis

Introduction

What one measures in empirical sciences is usually not the 'reality' or the phenomenon of interest itself but a more-or-less biased 'projection' of it. In medical research, behavioural sciences and other fields the measurement of any characteristic is subject to potential error. Especially in health sciences, measurement error is common not only for psychological variables (Cohen, 1988) but also for physiological phenomena (Brenner, 1996). One reason for this is that error-free measurements or measurements considered as 'goldstandard', respectively, are often much more expensive than error-prone measurements. In the case of categorical variables such as binary diagnoses, measurement error is referred to as 'misclassification'.

In psychiatric research, misclassification is an important issue, for instance, when deriving diagnoses of mental disorders if the best diagnostic standards are not available (Cohen, 1988; Ts uang and To hen, 2002). For example, diagnoses derived from the full Composite Diagnostic Interview (CIDI; WHO, 1997) are often regarded as gold-standard diagnoses, and symptoms assessed, for example, with the Composite International Diagnostic Screener (CID-S; Wittchen et al., 1999) can be regarded as crude measurements of diagnoses derived from the full CIDI. In general, misclassification can arise from choosing instruments, informants or their perspective or an inadequate combination or mixture of these aspects when designing a study (Kraemer, Measelle, Ablow, Essex, Boyce and Kupfer, 2003).

In many medical applications concerning diagnosis, however, several measurements exist, none of which is accepted as gold standard. In such cases, the terms and methods to address misclassification described in this paper do not apply. Instead, by using