Logistic regression with a continuous exposure measured in pools and subject to errors.

Area of Science:

• Biostatistics
• Epidemiology
• Reproductive Health

Background:

• Biomarker measurement in logistic regression can be costly.
• Sample pooling offers a cost-effective alternative for continuous exposures.
• Assay errors, including measurement error (ME) and processing error (PE), can affect the validity of pooled data analysis.

Purpose of the Study:

• To develop and evaluate a likelihood-based inference method for logistic regression with pooled biomarker data.
• To account for both measurement error (ME) and processing error (PE) in pooled samples.
• To assess the impact of replicates on study design and error correction.

Main Methods:

• Developed a logistic regression model for poolwise data incorporating ME and PE.
• Assumed independent, normally distributed ME and PE components.
• Employed likelihood-based inference and compared with a discriminant function approach.
• Utilized a reproductive health dataset with pooled samples (size 2), individual samples, and replicates.

Main Results:

• The model incorporating both ME and PE showed a lower AIC compared to ME-only models.
• The adjusted log-odds ratio for MCP-1 and spontaneous abortion risk was slightly higher when accounting for errors.
• Simulations confirmed method validity and demonstrated that ME and PE reduce the efficiency of pooling designs.
• Replicates were shown to improve model stability in the presence of both ME and PE.

Conclusions:

• Likelihood-based inference effectively handles ME and PE in pooled biomarker data for logistic regression.
• Accounting for both types of errors is crucial for accurate risk estimation and understanding pooling design efficiency.
• Incorporating replicates into study designs enhances the robustness and stability of analyses with pooled samples and assay errors.