Combining observational data from multiple databases: Comparison of individual patient data and aggregate data meta-analysis in the CARING study

Background: Combining multiple databases is valuable for analysing rare exposures and outcomes. In the CARING (CAncer Risk and INsulin analoGues) project, data from National Health Registers (NHR) in Denmark (DK), Finland, Norway (NO) and Sweden (SE) were combined with data from the United Kingdom Clinical Practice Research Datalink. Objectives: To evaluate the use of individual patient data (IPD) as compared with aggregate data (AD) meta-analysis combining three of these databases (DK, NO and SE). Methods: Population-based cohort studies of incident insulin users aged 18+ with no history of cancer were conducted using NHRs in DK, NO and SE. Based on the first dispensing, patients were classified as exposed to human insulin (N=98,154), glargine (N=12,529), detemir (N=5,252) or other insulin types (N=57,601). Poisson regression was used to estimate incidence rate ratios (IRR) of colorectal cancer, breast cancer and prostate cancer, comparing different insulins. Analyses were performed on a common dataset with IPD from all countries (adjusted for common covariates) and separate datasets for each country (adjusted for all available covariates, country-optimized), pooling estimates using fixed and random effects models. Results: In IPD analyses of colorectal, breast and prostate cancer, IRR (95% CI) for glargine vs. human insulin were 0.86 (0.66 to 1.14), 0.87 (0.59 to 1.30) and 1.07 (0.83 to 1.38), for detemir vs. human insulin 0.76 (0.47 to 1.22), 0.40 (0.16 to 0.98) and 0.91 (0.57 to 1.44). Fixed and random effects meta-analyses gave similar results. The AD meta-analysis did not include the NO cohort for most comparisons because of few outcome events whereas the IPD meta-analysis used all available data. Country-optimized adjustment was comparable to adjustment for common covariates. Conclusions: No consistent differences in risk of the cancers investigated between different insulins were found. Low power in individual cohorts and uniform distribution of available covariates between cohorts favored the use of IPD over AD meta-analysis in this specific case.