Bayesian hierarchical approaches for multiple outcomes in routinely collected healthcare data

Background: Routinely collected healthcare data provides a rich environment for the investigation of drug performance in the general population, while also offering the possibility of assessing rare outcomes. The statistical analysis of this data poses a number of challenges. The data may be biased and lack the structure and balance provided by the drugs’ clinical trials. Outcomes are often modelled individually with an associated lack of control for multiple comparisons, as well as a difficulty in assessing multiple risks.

Methods: Bayesian models provide methods for analysing multiple clinical outcomes, using relationships between outcomes and handling the types of multiple comparison issues which may occur when using multiple single-variate approaches. Lack of balance within the data may be catered for by dividing the population into clusters with similar characteristics, allowing within cluster inferences to be made. A Bayesian hierarchical model for multiple outcomes is proposed and applied to data from a safety and effectiveness study of direct oral anticoagulants (DOACs) in Scotland 2009 – 2015.

Results: The Bayesian modelling results were comparable to the results from the original safety and effectiveness study, with the additional benefit of balancing patient clusters and controlling for relationships in the data.

Conclusion: Bayesian hierarchical models are a suitable approach for modelling routinely collected healthcare data. There is the possibility of moving to an integrated Bayesian approach, with the inclusion of treatment relationships; uncertainty regarding cluster membership; and treatment allocation in the model, eventually leading to more reliable treatment decisions.