Enabling the analysis of patient-level data across jurisdictions for research use: a real-world exploration of a federated likelihood approach.

Introduction: The disclosure of health data is governed by strict privacy regulations which significantly restrict the transfer of data across jurisdictions. These limitations restrict the scope of health research, particularly the ability to conduct studies that span multiple jurisdictions (e.g. provinces, states, or countries). One common practice is to use meta-analysis to pool jurisdiction-specific estimates. However, this approach relies on combining aggregate-level data, which may overlook important nuances. Therefore, alternative methods are needed. This study investigates the potential of a federated likelihood approach to analyse health data across jurisdictions while keeping the data inside the jurisdiction. Objectives: (1) Evaluate the effectiveness of the meta-analysis technique in pooling jurisdiction-specific results to produce overall estimates (2) assess the utility of a federated likelihood approach for interjurisdictional health data analysis; and (3) compare the performance of a federated likelihood method with the conventional meta-analysis approach. Methods: Log-likelihood functions from separate jurisdictions were directly combined to form a single overall log-likelihood function, and overall parameter estimates. We show that these parameter estimates are mathematically equivalent to those obtained from analysing the pooled individual-level data. Using data from the Canadian Primary Care Sentinel Surveillance Network, and using chronic kidney disease as a case study, estimates from five jurisdictions are combined using meta-analysis and the federated likelihood approach. These results are compared to the overall estimates from a global model. Result: The likelihood method demonstrated strong performance in analysing federated data, yielding the lowest relative percent absolute bias at 1.34%. This was notably lower than the bias observed with the common effect meta-analysis (4.09%). Conclusion: The federated likelihood method offers a promising approach to interjurisdictional health data analysis. However, further investigation is needed to assess its feasibility across different model types and model complexity to fully understand the scope of its utility.