Approach to addressing missing data for electronic medical records and pharmacy claims data research

Objective: The complete capture of all values for each variable of interest in pharmacy research studies remains aspirational. The absence of these possibly influential values is a common problem for pharmacist investigators. Failure to account for missing data may translate to biased study findings and conclusions. Our goal in this analysis was to apply validated statistical methods for missing data to a previously analyzed data set and compare results when missing data methods were implemented versus standard analytics that ignore missing data effects.

Design: Using data from a retrospective cohort study, the statistical method of multiple imputation was used to provide regression-based estimates of the missing values to improve available data usable for study outcomes measurement. These findings were then contrasted with a complete-case analysis that restricted estimation to subjects in the cohort that had no missing values. Odds ratios were compared to assess differences in findings of the analyses. A nonadjusted regression analysis ("crude analysis") was also performed as a reference for potential bias.

Setting: Veterans Integrated Systems Network that includes VA facilities in the Southern California and Nevada regions.

Patients: New statin users between November 30, 2006, and December 2, 2007, with a diagnosis of dyslipidemia.

Main outcome measure: We compared the odds ratios (ORs) and 95% confidence intervals (CIs) for the crude, complete-case, and multiple imputation analyses for the end points of a 25% or greater reduction in atherogenic lipids.

Results: Data were missing for 21.5% of identified patients (1665 subjects of 7739). Regression model results were similar for the crude, complete-case, and multiple imputation analyses with overlap of 95% confidence limits at each end point. The crude, complete-case, and multiple imputation ORs (95% CIs) for a 25% or greater reduction in low-density lipoprotein cholesterol were 3.5 (95% CI 3.1-3.9), 4.3 (95% CI 3.8-4.9), and 4.1 (95% CI 3.7-4.6), respectively. The crude, complete-case, and multiple imputation ORs (95% CIs) for a 25% or greater reduction in non-high-density lipoprotein cholesterol were 3.5 (95% CI 3.1-3.9), 4.5 (95% CI 4.0-5.2), and 4.4 (95% CI 3.9-4.9), respectively. The crude, complete-case, and multiple imputation ORs (95% CIs) for 25% or greater reduction in TGs were 3.1 (95% CI 2.8-3.6), 4.0 (95% CI 3.5-4.6), and 4.1 (95% CI 3.6-4.6), respectively.

Conclusion: The use of the multiple imputation method to account for missing data did not alter conclusions based on a complete-case analysis. Given the frequency of missing data in research using electronic health records and pharmacy claims data, multiple imputation may play an important role in the validation of study findings.