Sensitivity to imputation models and assumptions in receiver operating characteristic analysis with incomplete data

Area of Science:

• Statistics
• Biostatistics
• Medical Informatics

Background:

• Missing data present challenges in statistical analyses, particularly in medical research.
• Receiver operating characteristic (ROC) analysis is a widely used method for evaluating diagnostic tests and biomarkers.
• The impact of missing data assumptions on ROC analysis requires further investigation.

Purpose of the Study:

• To evaluate the performance of multiple imputation (MI) inference within ROC analysis.
• To investigate parametric and non-parametric MI techniques under various missingness mechanisms.
• To assess the influence of model specification on the accuracy of ROC analysis with incomplete data.

Main Methods:

• Utilized multiple imputation (MI) techniques, including parametric and non-parametric approaches.
• Simulated data with different missingness mechanisms to test imputation strategies.
• Performed ROC analysis on datasets with imputed values to assess inference performance.

Main Results:

• MI generally yields well-calibrated inferences in ROC analysis when missingness mechanisms are ignorable.
• The coherency between the imputation model and the data generation mechanism is crucial for accurate results.
• Performance can vary depending on the specific imputation technique and missingness patterns.

Conclusions:

• Multiple imputation is a viable approach for handling missing data in ROC analysis.
• Careful consideration of imputation model assumptions is necessary for reliable diagnostic test evaluations.
• Further research can explore advanced imputation methods for complex missing data scenarios in biostatistics.