Mind the Gaps: Literature Survey Reveals Shortcomings in Handling Missing Data in Clinical Practice Research Datalink (CPRD), a UK Primary Care Health Records Database.

Background: In the United Kingdom (UK) primary care electronic health records (EHR), key demographic, clinical, and lifestyle variables such as ethnicity, social deprivation, body mass index and smoking status are often incomplete. This incompleteness can compromise research validity by introducing bias and reducing statistical power. There are a number of frequently used approaches to handling missing data, including complete records analysis (CRA), missing indicator method and multiple imputation (MI), however it is not clear to what extent these are used in primary care EHR analyses or whether their use is appropriate. This study examines current practice for applying methodologies and reporting of missing data, in one of the largest UK primary care EHR databases, the Clinical Practice Research Datalink (CPRD). Methods: A random ~10% sample of observational studies from the CPRD bibliography, published between 01 January 2013 and 31 December 2023, was selected. Article screening and data extraction for each paper was completed by two reviewers, who used pre-prepared pro-forma to independently extract reporting and methods for handling missing data. Results: From 2,481 publications during the study period, a random 220 were selected for detailed review. Missing data were reported in 163 (74%) studies. CRA was applied in 50 studies (23%), missing indicator method was used in 44 studies (20%), MI in 18 studies (8%), and alternative methods such as reclassification and mean imputation, in 15 studies (6%). Conclusion: Many studies fail to follow published best practice, often relying on flawed methods like the missing indicator method. Greater transparency, rigorous missing data techniques, and clearer reporting are needed. Improved guidance with practical examples would enhance research quality. Without methodological consistency and scrutiny, the risk of bias and misinterpretation remains high, making it essential to integrate missing data considerations into study design and analysis.