Monitoring clinical outcomes across healthcare providers is increasingly important in the UK National Health Service. Neonatal care is no exception, but care is centralised, so is delivered via co-ordinated networks of neonatal units (NNUs), with sicker infants treated in larger centres. This results in frequent transfers, but it is unclear how to attribute outcomes of transferred infants. Hierarchical regression is recommended for assessing performance of healthcare providers, but existing studies have either excluded transferred patients or assigned them to a single provider. In this thesis, hierarchical Bayesian multiple membership (MM) models are used to evaluate NNU mortality of very preterm infants, attributing outcomes to all NNUs providing care.
Data for all singleton infants born 2011-2013 below 32 weeks gestation and admitted to neonatal care in England are obtained from the National Neonatal Research Database. Using established methods, a series of Bayesian hierarchical models with two (infants within NNUs) and three levels (infants within NNUs within networks) are developed for non-transferred infants. This approach is extended to include transferred infants using MM models. A variety of weightings, some specified using Beta distributions, are used to allocate outcomes of transferred infants. In contrast to other applications of MM models, results differ across weightings due to transfer patterns. The recommendation is that transferred patients are allocated equally to all providers, regardless of duration or intensity of care, accompanied by sensitivity analyses using alternative weights.
Methodologically, this thesis demonstrates a statistically principled way of accounting for transfers when evaluating provider-specific outcomes, and presents a new application of MM models with novel weightings. Clinically, the variation attributable to providers is low, and for each NNU an estimate of risk-adjusted mortality compared with similar NNUs is obtained. Practical implications extend beyond neonatal medicine as centralisation and electronic patient data become integral to improving healthcare.