How to deal with double partial verification when evaluating two index tests in relation to a reference test?

Area of Science:

• Medical Statistics
• Diagnostic Test Evaluation
• Reproductive Medicine

Background:

• Diagnostic accuracy research often faces challenges due to partial verification, where not all patients undergo reference testing or receive all index tests.
• Existing adjustment techniques for missing data assume verification is 'missing at random', which may not hold true in practice.

Purpose of the Study:

• To develop and demonstrate methods for reducing verification bias in diagnostic accuracy studies with partial and potentially 'missing not at random' (MNAR) verification.
• To analyze the diagnostic values of the chlamydia antibody test and hysterosalpingography using a reproductive medicine clinical example.

Main Methods:

• Plotting all possible sensitivity and specificity combinations for index tests within ignorance regions.
• Constructing models with varying assumptions for the verification process, including missing index tests and MNAR mechanisms.
• Analyzing joint data from multiple diagnostic tests used in the same population.

Main Results:

• Methods were developed to adjust for verification bias arising from omitted data in partially tested patients.
• The study explored the influence of patient characteristics and potential MNAR mechanisms on accuracy estimates.
• In a clinical study where over half the patients lacked reference tests, the impact of an MNAR verification process was found to be limited.

Conclusions:

• Joint analysis of data from diagnostic tests within the same population is preferable for accurate evaluation.
• The developed methods help mitigate bias in diagnostic accuracy research, even with complex missing data patterns.