posted on 2011-04-04, 09:53authored byBradley Neil Manktelow
Aims: Since 1990 data have been collected by the Trent Neonatal Survey (TNS) on neonatal intensive care activity within the area of the former Trent Regional Health Authority. While TNS is a unique data set, no systematic investigation had previously been undertaken to ensure that the most appropriate statistical methods were applied to its analysis. In this thesis, methods for the analysis of in unit mortality rates were reviewed, critically appraised and, where appropriate, developed in order to identify the most suitable methods. Methods: Statistical methods were illustrated using data from infants born in the years 2000 to 2002, at 32 completed weeks gestational age or less, admitted to one of the sixteen neonatal intensive care units (NICUs) within the area. The methods were discussed and risk adjustment methods were explored to allow for differences in disease severity between the units.
Results: Simple descriptive approaches and statistical models are presented. In particular, summary statistics derived from logistic regression models were explored, including odds ratios and statistics from both direct and indirect standardization. In the final approach, logistic regression models were applied to obtain estimated standardized mortality ratios (SMRs) for each NICU. Proposed methods to estimate confidence intervals for the SMR were investigated through a simulation study and by application to the TNS data, with the method proposed by Hosmer and Lemeshow (1995) applied in the final models. The use of Bayesian methods was proposed and a model developed allowing the appropriate estimation of all uncertainty. Conclusions:
The use of SMRs was proposed for the reporting of mortality in future TNS annual reports. The advantages of a Bayesian approach, with the ability to make probability statements about the SMR, were also emphasised. Further work is required into the effect of specification of prior distributions before this method can be recommended routinely.