Representativeness and optimal use of body mass index (BMI) in the UK Clinical Practice Research Datalink (CPRD)

• Published in: BMJ open 2013-09, Vol.3 (9), p.e003389-e003389
• Main Authors: Bhaskaran, Krishnan, Forbes, Harriet J, Douglas, Ian, Leon, David A, Smeeth, Liam
• Format: Article
• Language: English
• Subjects: Age groups; Body mass index; Calendars; Clinical medicine; Epidemiology; Lifestyles; Medical records; Mortality; Obesity; Population; Researchers; Studies
• ISSN: 2044-6055; 2044-6055
• DOI: 10.1136/bmjopen-2013-003389

Objectives To assess the completeness and representativeness of body mass index (BMI) data in the Clinical Practice Research Datalink (CPRD), and determine an optimal strategy for their use. Design Descriptive study. Setting Electronic healthcare records from primary care. Participants A million patient random sample from the UK CPRD primary care database, aged ≥16 years. Primary and secondary outcome measures BMI completeness in CPRD was evaluated by age, sex and calendar period. CPRD-based summary BMI statistics for each calendar year (2003–2010) were age-standardised and sex-standardised and compared with equivalent statistics from the Health Survey for England (HSE). Results BMI completeness increased over calendar time from 37% in 1990–1994 to 77% in 2005–2011, was higher among females and increased with age. When BMI at specific time points was assigned based on the most recent record, calendar–year-specific mean BMI statistics underestimated equivalent HSE statistics by 0.75–1.1 kg/m2. Restriction to those with a recent (≤3 years) BMI resulted in mean BMI estimates closer to HSE (≤0.28 kg/m2 underestimation), but excluded up to 47% of patients. An alternative strategy of imputing up-to-date BMI based on modelled changes in BMI over time since the last available record also led to mean BMI estimates that were close to HSE (≤0.37 kg/m2 underestimation). Conclusions Completeness of BMI in CPRD increased over time and varied by age and sex. At a given point in time, a large proportion of the most recent BMIs are unlikely to reflect current BMI; consequent BMI misclassification might be reduced by employing model-based imputation of current BMI.