Logistic regression when the outcome is measured with uncertainty

Area of Science:

• Epidemiologic research
• Biostatistics
• Medical diagnostics

Background:

• Logistic regression is a common tool in epidemiology for estimating odds ratios.
• Imperfect sensitivity and specificity of diagnostic tests can introduce misclassification bias, leading to inaccurate results.
• Existing methods often ignore or inadequately address this diagnostic test error.

Purpose of the Study:

• To present a method for incorporating known diagnostic test sensitivity and specificity into logistic regression models.
• To develop an Expectation-Maximization (EM) algorithm for unbiased estimation of odds ratios and their variances.
• To explore the method's utility when sensitivity and specificity are unknown or vary among subjects.

Main Methods:

• An EM algorithm is described to fit logistic regression models with misclassified outcomes.
• The method accounts for known sensitivity and specificity of the diagnostic test.
• Extensions are discussed for nondifferential misclassification and scenarios with unknown accuracy.

Main Results:

• The proposed method yields unbiased estimates of odds ratios and their variances, unlike standard logistic regression.
• Estimates from this method are typically farther from the null but have increased variance compared to ignoring misclassification.
• The approach is robust and can be adapted for various assumptions about test accuracy.

Conclusions:

• Incorporating diagnostic test characteristics into logistic regression provides more accurate epidemiologic estimates.
• The EM algorithm offers a practical solution for handling outcome misclassification bias.
• This method enhances the reliability of findings from studies using imperfect diagnostic measures.