Misclassification bias

misclassification bias occurs when a study participant is categorised into an incorrect category altering the observed association or research outcome of interest.

Background

Correct classification of individuals, and of exposures and participant characteristics, is an essential element of any study. Misclassification occurs when individuals are assigned to a different category than the one they should be in. This can lead to incorrect associations being observed between the assigned categories and the outcomes of interest.

Non-differential misclassification occurs when the probability of individuals being misclassified is equal across all groups in the study.  Differential misclassification occurs when the probability of being misclassified differs between groups in a study (Porta et al. 2014).

Included studies in a systematic review could use different classification systems, potentially causing misclassification bias when the studies are pooled in a meta-analysis.

Example

A meta-analysis of body size and development of prostate cancer found that the criteria used to define nonaggressive and aggressive prostate varied between cohorts which may have lead to misclassification bias. (Xie et al. 2017).

Impact

In measuring relationships between exposures and disease risk, misclassification bias can have unpredictable effects, i.e. it could increase or decrease an observed association.

Flegal and colleagues investigated misclassification bias in hazard ratios (estimates of risk) in studies looking at the relationship between body mass index (BMI) and mortality (Flegal et al 2017).

Misreporting at higher BMI categories tended to bias hazard ratios upwards for some categories, but that effect was counterbalanced or even reversed by misreporting in other BMI categories, in particular, those that affected the reference category.

For example, among healthy male never-smokers, misclassifications affecting the overweight category and the reference categories changed significantly the hazard ratio for overweight from 0.85 with measured data to 1.24 with self-reported data.

Both the magnitude and direction of bias varied according to the hazard ratios with the measured data. Because of misclassification effects, self-reported weight and height could not reliably indicate the lowest-risk BMI category. Where an association between a category of body size and a health outcome is found, misclassification bias may have influenced that observation, sometimes increasing a risk estimate, sometimes decreasing it. This is important because understanding the relationship between obesity and underweight and health is a key factor in public health.

The study also highlights that the underlying hazards influence the way that misclassification affects risk estimates in each study, and the necessity to understand misclassification bias within the specific group or population under study and its effect on outcomes.

Preventive steps

Prevention of bias from misclassification includes using the most accurate measurements available and thinking carefully about the categorisation of individuals or data points into groups.

Where misclassification bias is suspected, some statistical techniques exist to deal with the bias.

(Van Walraven 2017) investigated two methods to help account for misclassification bias. In a study of two conditions in separate cohorts: severe renal failure and Colles’ fracture, true disease prevalence and relationship of the disease with other factors were measured and compared with results when disease status was determined using diagnostic codes.

“Differences (‘misclassification bias’) were then adjusted for using two methods: quantitative bias analysis (QBA) with bias parameters (code sensitivity and specificity) of varying accuracy; and disease status imputation using bootstrap methods and disease probability models.”

The authors report that using quantitative bias analysis will not necessarily decrease bias. QBA is dependent upon the accuracy of the data when addressing bias. They recommend using values that are actually measured on the population used in the study (or ones that are similar to that in the study). The accuracy of recording each condition was less when using database diagnostic codes than using hospital biochemical data, and measures of disease association with covariables were also substantially affected. The use of one of two statistical approaches can, but does not always, reduce bias from misclassification.