A stratification approach using logit-based models for confounder adjustment in the study of continuous outcomes

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Confounding control is crucial in observational studies for causal inference.
• Existing methods like matched designs (e.g., conditional logistic, stratified Cox) are established for binary and survival outcomes.
• Extending these methods to continuous outcomes without functional form assumptions remains a challenge.

Purpose of the Study:

• To adapt logit-based regression models for controlling confounding in continuous outcome observational studies.
• To evaluate the performance and diagnostic utility of the proposed method.
• To demonstrate the application of the unified regression approach across different outcome types (binary, survival, continuous).

Main Methods:

• Extension of regression models from matched designs (binary/survival) to continuous outcomes using logit-based models.
• Comparison of maximum likelihood estimators via simulation studies.
• Development of a heuristic for residual-based model diagnostics.
• Application to real-world datasets (mammographic density, inpatient blood glucose).

Main Results:

• The proposed stratification approach is robust to model misspecification.
• Estimated residuals offer useful diagnostics for moderate-sized strata.
• Identified associations between parity/menopausal status and mammographic density.
• Demonstrated variations in blood glucose levels between hospital wards.

Conclusions:

• A unified regression modeling framework can effectively adjust for confounding across binary, time-to-event, and continuous outcomes.
• The proposed method provides a flexible and robust tool for epidemiological research.
• This approach enhances causal inference capabilities in observational studies.