Selection and Misclassification Biases in Longitudinal Studies

Area of Science:

• Epidemiology
• Biostatistics
• Diagnostic Test Evaluation

Background:

• Imperfect diagnostic tests can introduce bias in disease frequency and association measures.
• Previous research focused on cross-sectional or single-time-point misclassification in cohort studies.
• Longitudinal studies face combined selection and misclassification biases from errors at multiple time-points.

Purpose of the Study:

• To evaluate the combined impact of selection and misclassification biases on incidence and risk ratio estimates in longitudinal studies.
• To assess the relative influence of diagnostic test sensitivity and specificity on bias magnitude.
• To investigate how disease prevalence and incidence affect bias in epidemiological measures.

Main Methods:

• Simulated 1,000 hypothetical cohort studies (1,000 observations each) with varying test characteristics and disease contexts.
• Analyzed bias in disease incidence and risk ratio due to misclassification at baseline and follow-up.
• Assessed diagnostic test sensitivity (0.7-1.0) and specificity (0.8-1.0) against disease prevalences (5%, 20%) and incidences (0.01, 0.05, 0.1).

Main Results:

• Bias in disease incidence and risk ratio estimates primarily depends on test specificity, disease prevalence, and incidence.
• High test specificity is critical for minimizing bias, particularly for prevalent diseases.
• Divergence from perfect specificity rapidly leads to over-estimation of disease incidence when prevalence is high and incidence is low.

Conclusions:

• Test specificity is more important than sensitivity for minimizing bias in incidence and risk ratio estimates in longitudinal studies.
• Accurate diagnostic testing, especially high specificity, is essential for reliable epidemiological measures, particularly for prevalent diseases.
• Bias is minimized with high sensitivity and specificity in low-prevalence, high-incidence diseases, but prevalent diseases show substantial bias towards the null even with near-perfect tests.