PIE: A prior knowledge guided integrated likelihood estimation method for bias reduction in association studies using...

Area of Science:

• Biostatistics
• Health Informatics
• Epidemiology

Background:

• Electronic Health Records (EHRs) are valuable data sources for research.
• Phenotype misclassification in EHR data can introduce significant bias in association estimations.
• Existing methods for correcting such bias are often limited in their effectiveness.

Purpose of the Study:

• To introduce a novel Prior knowledge guided Integrated likelihood Estimation (PIE) method.
• To correct estimation bias stemming from misclassified binary phenotypes in EHR data.
• To evaluate PIE's performance against established methods.

Main Methods:

• Simulation studies and real-world EHR data analysis (diabetes, Kaiser Permanente Washington) were conducted.
• PIE was compared against a method ignoring errors and a maximum likelihood method with misspecified sensitivity/specificity.
• PIE utilizes prior information on phenotyping accuracy to form a prior distribution for sensitivity and specificity, integrated for bias reduction.

Main Results:

• Simulation studies and real data confirmed PIE effectively reduces estimation bias compared to current methods.
• PIE performance approached the gold standard when prior information was accurate.
• EHR data analysis showed PIE estimates closely matched gold standard, reducing bias by 60%-100%.

Conclusions:

• The proposed PIE method effectively mitigates estimation bias in EHR-derived data.
• Incorporating prior information via integrated likelihood is key to PIE's success.
• PIE offers a robust solution for accurate association studies using imperfect EHR data.