Spatio-temporal Modelling of Cancer Data in Queensland Using Bayesian Methods
Cancer is the leading contributor to the disease burden in Australia, accounting for almost one-fifth of the total burden. Broad geographical inequalities in cancer outcomes were known to exist within Australia, but few small-area cancer analyses had been conducted, and none within Queensland. Challenges include the small population dispersed over vast distances, yet Bayesian hierarchical models are able to accommodate sparse numbers while allowing for spatial correlation between small areas.
This research aims to develop and apply Bayesian hierarchical models to facilitate an investigation of the spatial and temporal associations for diagnostic and survival outcomes for Queenslanders diagnosed with cancer. The key objectives are to document and quantify the importance of spatial inequalities, explore factors influencing these inequalities, and investigate how spatial inequalities change over time.
Data on all primary invasive cancers diagnosed from 1996 onwards were obtained from the Queensland Cancer Registry. Patient residence at diagnosis was provided as one of 478 Statistical Local Areas (median population of 6,390 in 2011). All models allowed for local and global smoothing via spatially structured and uncorrelated heterogeneity components, respectively. Spatial smoothing in all analyses used an intrinsic conditional autoregressive prior based on first-order contiguity.
The first objective, and the foundation for further analyses, was to identify cancers with evidence of spatial inequalities. Cancers tending to have higher incidence rates in more urban areas included breast, prostate, non-Hodgkin lymphoma, male kidney and bladder. In contrast, cervical, male lung and male oesophageal cancers had higher incidence rates in more remote areas. For survival spatial inequalities, a consistent pattern of lower survival among remote areas and higher survival among urban areas was observed for non-Hodgkin lymphoma, lung, colorectal, female breast, male leukaemia, male stomach and prostate cancers.
Next, the influence on diagnostic spatial patterns by area-level factors such as remoteness, socioeconomic disadvantage and Indigenous proportion of the population was considered. Due to the complex interplay between these influences, a classification and regression tree analysis was applied to Bayesian modelled incidence estimates. The remoteness of an area was found to be a key influence on spatial incidence inequalities for several cancers, while Indigenous ethnicity was an important influence only for cervical cancer. Socioeconomic disadvantage interacted with remoteness for melanoma, breast (females), cervical, lung and prostate cancers.
Small-area changes over time were investigated for lung cancer incidence and a modelled estimate of its risk factors via a spatio-temporal shared component model. The modelled shared component appeared to reflect past trends in tobacco smoking, and found consistent changes across time over all small areas. This suggests that spatial inequalities have largely remained consistent, with the same areas remaining at higher risk. Limitations of survey-based data meant it had not been possible to look at small-area tobacco smoking prevalence changes over time previously.
Small-area survival inequalities were also further investigated. The influence of tumour stage at diagnosis is an important prognostic influence, so was included in the Bayesian additive risk model with piecewise constant hazards examining spatial relative survival inequalities for breast and colorectal cancers. Much of the lower survival observed for breast cancer patients residing in remote areas resulted from a greater proportion of advanced tumours diagnosed in these areas. An estimated 640 breast and colorectal cancer deaths resulted from spatial inequalities in cancer survival in Queensland during 1998-2007.
When survival was predicted by cancer stage, localised breast cancer had quite similar survival across all statistical local areas. However, 5-year relative survival varied between areas by up to 7% for advanced breast cancer, with more remote areas tending to have poorer survival. In contrast, even localised colorectal cancers showed maximum differences in predicted survival of almost 5% between areas, and up to 14% for advanced tumours, with survival generally decreasing as remoteness increased.
Spatio-temporal changes in breast and colorectal cancer survival by tumour stage were also examined. Larger survival improvements were observed between 2002-2006 and 2007-2011, than between 1997-2001 and 2002-2006. Nonetheless, during the entire time period of 1997-2011 all small areas showed improvements in survival for both localised and advanced cancers, with the median 5-year relative survival improvement ranging from 2% for localised breast cancer to 8% for advanced colorectal cancer.
Important methodological contributions resulted from this project. A fully Bayesian approach to quantify premature deaths from spatially structured variation in cancer survival inequalities was developed. The advantages of this include obtaining measures of uncertainty, the ability to adjust for prognostic influences, and excluding deaths considered to result from random variation.
A spatial flexible parametric relative survival model was also introduced, and further expanded to provide the first spatio-temporal flexible parametric relative survival model. Benefits over previous spatial relative survival models include the ability to predict smooth survival functions, the ease of including continuous variables, and the capacity to use individual-level input data.
Practical benefits for Queenslanders diagnosed with cancer also directly resulted from this project. The Patient Travel Subsidy Scheme, which offsets some of the costs associated with travelling for medical treatment, was increased after lobbying using our results. Additional Cancer Council Queensland regional support staff positions were created in response to the demonstrated survival inequalities. Results were used by Queensland Health to formulate cancer health service strategies for the next decade, with a focus on reducing variations in cancer outcomes throughout the state.
This detailed and comprehensive analysis of small-area inequalities in cancer outcomes clearly demonstrated the versatility of Bayesian hierarchical models in cancer control. Existing Bayesian hierarchical models were refined, new models and methods developed, and tangible benefits obtained for cancer patients in Queensland.