Cox regression is robust to inaccurate EHR-extracted event time: an application to EHR-based GWAS

Area of Science:

• Genomics
• Biostatistics
• Health Informatics

Background:

• Logistic regression models are common in genomic studies with electronic health records (EHRs).
• These models do not fully utilize time-to-event data present in EHRs.
• Cox regression can account for data complexities like left truncation and right censoring in EHRs, potentially increasing power for genotype-phenotype associations.

Purpose of the Study:

• To evaluate the relative performance of Cox regression and logistic regression models.
• To assess the impact of delayed event times (positive errors in event time) on statistical power.
• To compare model sensitivity in detecting genotype-phenotype associations under various event time accuracy scenarios.

Main Methods:

• Compared one Cox model against three logistic regression models.
• Utilized extensive simulations and a genomic study application.
• Investigated scenarios with varying degrees of delayed event time.

Main Results:

• Logistic regression models were more sensitive to delayed event times than Cox regression.
• Excluding patients diagnosed before their entry time was crucial.
• Cox regression demonstrated similar or modestly improved statistical power over logistic models.
• Cox models consistently showed higher sensitivity in detecting known genotype-phenotype associations across all tested scenarios.

Conclusions:

• Cox regression is a more robust approach for analyzing time-to-event data in EHRs, particularly when event time accuracy is a concern.
• The findings underscore the importance of accounting for event time accuracy in genomic association studies.
• Cox regression provides a reliable method for enhancing the power to detect genotype-phenotype associations in EHR data.