Methods to account for outcome misclassification in epidemiology

Edwards, Jessie

Download PDF

Request Version for Screen Reader

Last Modified

March 22, 2019

Creator

Edwards, Jessie
- Affiliation: Gillings School of Global Public Health, Department of Epidemiology

Abstract

Outcome misclassification occurs when the endpoint of an epidemiologic study is measured with error. Outcome misclassification is common in epidemiology but is frequently ignored in the analysis of exposure-outcome relationships. We focus on two common types of outcomes in epidemiology that are subject to mismeasurement: participant-reported outcomes and cause-specific mortality. In this work, we leverage information on the misclassification probabilities obtained from internal validation studies, external validation studies, and expert opinion to account for outcome misclassification in various epidemiologic settings. This work describes the use of multiple imputation to reduce bias when validation data are available for a subgroup of study participants. This approach worked well to account for bias due to outcome misclassification in the odds ratio and risk ratio comparing herpes simplex virus recurrence between participants randomized to receive acyclovir or placebo in the Herpetic Eye Disease Study. In simulations, multiple imputation had greater statistical power than analysis restricted to the validation subgroup, yet both provided unbiased estimates of the odds ratio. Modified maximum likelihood and Bayesian methods are used to explore the effects of outcome misclassification in situations with no validation subgroup. In a cohort of textile workers exposed to asbestos in South Carolina, we perform sensitivity analysis using modified maximum likelihood to estimate the rate ratio of lung cancer death per 100 fiber-years/mL asbestos exposure under varying assumptions about sensitivity and specificity. When specificity of outcome classification was nearly perfect, the modified maximum likelihood approach produced estimates that were similar to analyses that ignore outcome misclassification. Uncertainty in the misclassification parameters is expressed by placing informative prior distributions on sensitivity and specificity in Bayesian analysis. Because, in our examples, lung cancer death is unlikely to be misclassified, posterior estimates are similar to standard estimates. However, modified maximum likelihood and Bayesian methods are needed to verify the robustness of standard estimates, and these approaches will provide unbiased estimates in settings with more misclassification. This work has highlighted the potential for bias due to outcome misclassification and described three flexible tools to account for misclassification. Use of such techniques will improve inference from epidemiologic studies.

Date of publication