Validation of Veterans Affairs Electronic Medical Record Smoking Data Among Iraq- and Afghanistan-Era Veterans

Research using the Veterans Health Administration (VA) electronic medical records (EMR) has been limited by a lack of reliable smoking data. To evaluate the validity of using VA EMR Health Factors data to determine smoking status among veterans with recent military service. Sensitivity, specificity, area under the receiver-operating curve (AUC), and kappa statistics were used to evaluate concordance between VA EMR smoking status and criterion smoking status. Veterans (N = 2025) with service during the wars in Iraq/Afghanistan who participated in the VA Mid-Atlantic Post-Deployment Mental Health (PDMH) Study. Criterion smoking status was based on self-report during a confidential study visit. VA EMR smoking status was measured by coding health factors data entries (populated during automated clinical reminders) in three ways: based on the most common health factor, the most recent health factor, and the health factor within 12 months of the criterion smoking status data collection date. Concordance with PDMH smoking status (current, former, never) was highest when determined by the most commonly observed VA EMR health factor (kappa = 0.69) and was not significantly impacted by psychiatric status. Agreement was higher when smoking status was dichotomized: current vs. not current (kappa = 0.73; sensitivity = 0.84; specificity = 0.91; AUC = 0.87); ever vs. never (kappa = 0.75; sensitivity = 0.85; specificity = 0.90; AUC = 0.87). There were substantial missing Health Factors data when restricting analyses to a 12-month period from the criterion smoking status date. Current smokers had significantly more Health Factors entries compared to never or former smokers. The use of computerized tobacco screening data to determine smoking status is valid and feasible. Results indicating that smokers have significantly more health factors entries than non-smokers suggest that caution is warranted when using the EMR to select cases for cohort studies as the risk for selection bias appears high.