1. ¹Northern and Yorkshire Cancer Registry and Information Service, St James's Institute of Oncology, St James's University Hospital, Level 6, Bexley Wing, Beckett Street, Leeds LS9 7TF, UK
2. ²Cancer Epidemiology Group, Division of Epidemiology and Biostatistics, University of Leeds, St James's Institute of Oncology, St James's University Hospital, Bexley Wing, Beckett Street, Leeds LS9 7TF, UK
3. ³Biostatistics Unit, Division of Epidemiology and Biostatistics, University of Leeds, Worsley Building, Leeds LS2 9JT, UK
4. ⁴Biostatistics Unit, South African Medical Research Council, Private Bag X385, Pretoria 0001, South Africa

Correspondence: L Fairley, E-mail: lesley.fairley@nycris.leedsth.nhs.uk

Received 9 June 2008; Revised 25 September 2008; Accepted 1 October 2008; Published online 4 November 2008.

Abstract

Primary Care Trust (PCT) estimates of survival lack robustness as there are small numbers of deaths per year in each area, even when incidence is high. We assess PCT-level spatial variation in prostate cancer survival using Bayesian spatial models of excess mortality. We extracted data on men diagnosed with prostate cancer between 1990 and 1999 from the Northern and Yorkshire Cancer Registry and Information Service database. Models were adjusted for age at diagnosis, period of diagnosis and deprivation. All covariates had a significant association with excess mortality; men from more deprived areas, older age at diagnosis and diagnosed in 1990–1994 had higher excess mortality. The unadjusted relative excess risks (RER) of death by PCT ranged from 0.75 to 1.66. After adjustment, areas of high and low excess mortality were smoothed towards the mean, and the RERs ranged from 0.74 to 1.49. Using Bayesian smoothing techniques to model cancer survival by geographic area offers many advantages over traditional methods; estimates in areas with small populations or low incidence rates are stabilised and shrunk towards local and global risk estimates improving reliability and precision, complex models are easily handled and adjustment for covariates can be made.