A two-part model for reference curve estimation subject to a limit of detection

Area of Science:

• Biostatistics
• Pediatric Endocrinology
• Reproductive Health

Background:

• Reference curves are crucial for identifying developmental stages and extreme values in clinical variables.
• Estimating reference curves is challenging when measurements are censored by a limit of detection (LOD), particularly for hormones during puberty.
• Existing methods for handling LOD data often rely on untestable assumptions.

Purpose of the Study:

• To propose and validate a novel two-part statistical model for estimating reference curves in the presence of a limit of detection (LOD).
• To address the specific challenges of analyzing reproductive hormone levels in boys during puberty, where LOD is common.
• To provide a flexible and implementable method for calculating relative standing of measurements above the LOD.

Main Methods:

• Developed a two-part model: one component models the probability of exceeding the LOD as a function of age, and the other models the conditional distribution above the LOD.
• Combined information from both model parts to estimate the observable portion of the reference curve.
• Applied the method to reproductive hormone data from the Third National Health and Nutrition Examination Survey.

Main Results:

• The proposed two-part model effectively estimates reference curves for hormone levels above the LOD in pubertal boys.
• The model provides a reliable way to determine the relative standing of individual measurements.
• Demonstrated the model's flexibility and ease of implementation with real-world data.

Conclusions:

• The two-part model offers a robust solution for estimating reference curves with left-censored data due to LOD.
• This approach avoids untestable assumptions and is practical for clinical and research applications.
• Facilitates more accurate assessment of pubertal development and hormone status.