Methods for calculating sensitivity and specificity of clustered data: a tutorial

Area of Science:

• Biostatistics
• Diagnostic Test Evaluation
• Medical Informatics

Background:

• Diagnostic test performance is typically measured by sensitivity and specificity.
• Clustered data, with multiple observations per patient, are common in diagnostic test evaluations.
• Ignoring within-patient correlation can lead to inaccurate performance metrics.

Purpose of the Study:

• To discuss and illustrate statistical methods for calculating sensitivity and specificity with clustered data.
• To address the correlation between multiple observations within a single patient.
• To guide the selection of appropriate statistical methods based on clinical context.

Main Methods:

• The study presents four common statistical methods for analyzing clustered diagnostic test data.
• Methods include: (a) ignoring correlation, (b) variance adjustment, (c) logistic random-effects models, and (d) generalized estimating equations.
• These methods adjust for the potential correlation between observations within each patient.

Main Results:

• The choice of statistical method depends on the clinical decision problem and the relevance of multiple observations per patient.
• Ignoring correlation can lead to biased estimates of sensitivity and specificity.
• Methods that account for correlation provide more accurate performance measures.

Conclusions:

• When multiple observations per patient are relevant, the correlation between them must be considered in the statistical analysis.
• Appropriate statistical methods ensure reliable diagnostic test performance evaluation.
• The level of reporting should align with the clinical decision-making process.