Flexible regression models for ROC and risk analysis, with or without a gold standard

Area of Science:

• Biostatistics
• Medical Diagnostics
• Epidemiology

Background:

• Estimating medical test accuracy often relies on a gold-standard procedure, which is not always feasible.
• Many real-world scenarios lack a gold-standard test due to cost or invasiveness, leading to missing disease status data.
• Covariate information can be leveraged to infer disease status and evaluate test performance in the absence of a gold standard.

Purpose of the Study:

• To develop a novel semiparametric regression model for evaluating covariate-specific accuracy of continuous medical tests or biomarkers.
• To address scenarios where gold-standard diagnostic procedures are unavailable or impractical.
• To enable quantification of covariate effects on both disease probability and test accuracy.

Main Methods:

• Developed a semiparametric regression model incorporating 'disease covariates' to predict disease probability.
• Modeled test/biomarker outcomes based on 'test covariates' to assess accuracy.
• Utilized flexible semiparametric distributions for test outcomes and proved model identifiability under mild conditions.

Main Results:

• The proposed model allows for inference on covariate-specific test accuracy and disease probability.
• Demonstrated the model's utility through simulation studies.
• Applied the model to evaluate the accuracy of soluble epidermal growth factor receptor as a lung cancer biomarker in men, adjusting for age.

Conclusions:

• The novel semiparametric model provides a robust framework for assessing medical test accuracy in the common scenario of unknown disease status.
• It effectively utilizes covariate information to improve diagnostic evaluations and understand factors influencing test performance.
• The methodology is applicable to various continuous medical tests and biomarkers, with provided SAS code for implementation.