Training a Smoking Status Probabilistic Model Using Cotinine Levels in a Large Claims Database.

Area of Science:

• Epidemiology
• Health Informatics
• Biostatistics

Background:

• Smoking status is a critical confounder in epidemiological studies.
• Administrative claims data often lack comprehensive smoking status documentation.
• Existing probabilistic models for smoking status typically rely on self-reported data.

Purpose of the Study:

• To develop and validate a probabilistic model for predicting smoking status using cotinine values from a large claims database.
• To leverage objective cotinine measurements for a more reliable proxy of smoking status in real-world data.

Main Methods:

• Beneficiaries with cotinine measurements were included.
• Current smokers were defined by specific cotinine thresholds (serum/plasma ≥5ng/mL, urine ≥30ng/mL).
• A logistic regression model with stepwise forward selection was employed, using predictors from the year prior to cotinine assessment.

Main Results:

• The model achieved an area under the receiver operating characteristic curve of 0.77 (95% CI: 0.75-0.78) with good calibration.
• Key predictors included diagnosis codes for smoking and drug abuse, and medication counts.
• The model demonstrated high specificity but lower sensitivity at probability cutoffs ≥ 0.2.

Conclusions:

• A validated probabilistic model for smoking status in claims data was successfully developed using cotinine values.
• The model, utilizing 26 predictors, offers a simplified approach for application in other claims databases.
• External validation is recommended for future epidemiological research applications.