One-stage and two-stage meta-analysis of individual participant data led to consistent summarized evidence: lessons learned from combining multiple databases

OBJECTIVE: Combining multiple healthcare databases allows comparing the effects of a wide variety of healthcare services. There is a growing interest in methods for combining the results from multiple databases. We attempted to learn lessons about the performance of one-stage and two-stage approaches from the re-analysis of data drawn from two studies of pharmacoepidemiology based on multiple databases. STUDY DESIGN AND SETTING: Two nested case-control studies were carried out for estimating the tricyclic antidepressants-arrhythmia and etoricoxib-heart failure associations, respectively from the Italian I-GrADE and the European SOS programs. The associations of interest were modelled by conditional logistic regression for matched case-control sets, fitting fixed-effect and random-effect models with both one-stage and two-stage approaches. RESULTS: One-stage and two-stage approaches gave very similar results, showing uncertainty of tricyclic antidepressants-arrhythmia association (random-effect odds ratios, 95% confidence interval, 1.26, 0.71-2.24, and 1.30, 0.66-2.55, respectively) and statistical evidence for etoricoxib-heart failure association (fixed-effect odds ratios, 95% confidence interval, 1.53, 1.41-1.66, and 1.54, 1.42-1.66, respectively). CONCLUSION: Our study offers further evidence that two-stage approach generates estimates very similar as those from one-stage approach, even in the case of between databases exposure heterogeneity and when several covariates must to be concurrently considered. As current rules limit the free movement of electronic health data, our findings open the door of treating data within the country where they are generated, and then to apply conventional techniques for summarizing estimates, that is the two-stage approach for meta-analysis using individual participant data