A predictive algorithm to identify ever smoking in medical claims-based epidemiologic studies.

Area of Science:

• Epidemiology
• Health Informatics
• Biostatistics

Background:

• Estimating smoking prevalence is crucial for public health research.
• Administrative claims data offer a vast resource but lack direct smoking status information.

Purpose of the Study:

• To develop and validate an algorithm for estimating the probability of ever smoking using administrative claims data.
• To assess the utility of this algorithm in large-scale epidemiologic analyses.

Main Methods:

• A logistic regression model was developed using demographic and claims data from Medicare beneficiaries.
• The model's performance was validated using a "gold standard" based on tobacco-specific diagnosis codes and cancer codes.
• Spearman's rho was calculated to correlate the algorithm's predicted probability with smoking status in prior Parkinson disease studies.

Main Results:

• The predictive model incorporated 23 variables, including demographics and health conditions.
• The Area Under the Curve (AUC) for predicting ever smoking was 67.6%.
• The full algorithm achieved a Spearman's rho of 0.82 when compared to prior smoking assessments.

Conclusions:

• Administrative claims data can be utilized to approximate "ever smoking" status.
• The developed algorithm provides a probabilistic variable suitable for epidemiologic analyses.
• This method enhances the ability to study smoking's impact using large datasets.