TL;DR
This paper tackles the challenge of understanding cancer risk factor prevalence at a highly localized level across Australia. By developing a novel Bayesian small area estimation method, the researchers provide unprecedented spatial resolution for these factors, enabling more targeted public health interventions and a clearer picture of health disparities.
What Problem Does This Paper Address?
Cancer risk is known to vary geographically, but detailed, small-area data on the prevalence of modifiable risk factors (like smoking, diet, and physical activity) has been lacking in many countries, including Australia. Existing national atlases provide valuable insights but often lack the fine-grained resolution needed for precise local interventions. This gap hinders the development of targeted prevention strategies, especially in remote areas where data is even sparser. The authors aim to bridge this gap by creating high-resolution, reliable estimates of eight key cancer risk factors across Australia's smallest statistical areas (SA2s).
Methods
The core of this research is a Bayesian two-stage small area estimation (SAE) methodology. This approach combines survey data with auxiliary information (like census data) to produce estimates for areas with limited or no direct survey data.
Here's a breakdown of the key methodological components:
- Two-Stage Modeling:
- Stage 1 (Individual Level): A Bayesian pseudo-likelihood logistic mixed model is used to model individual survey responses. This stage accounts for individual characteristics and incorporates random effects to capture unobserved heterogeneity.
- Stage 2 (Area Level): A Bayesian spatial Fay-Herriot model is employed. This stage aggregates the Stage 1 predictions to the area level (SA2) and incorporates spatial smoothing and hierarchical modeling to borrow strength across neighboring areas. This helps to stabilize estimates, especially in areas with small sample sizes.
- Incorporation of Auxiliary Data: A wide array of auxiliary data, including census information (socioeconomic status, education, occupation, etc.), remoteness classifications, and even previous estimates from the Social Health Atlases of Australia (SHAA), are integrated into the models. Principal Component Analysis (PCA) is used to reduce the dimensionality of the census data.
- Validation: The estimates are rigorously validated through both internal (Bayesian benchmarking) and external methods (comparison with SHAA and other Australian health surveys). This ensures the reliability and credibility of the generated estimates.
What's novel here is the specific application of a recently developed Bayesian two-stage SAE methodology, tailored to address the sparsity and spatial nature of the data, and its comprehensive application to a wide range of cancer risk factors at a fine spatial resolution.
Datasets
- 2017-2018 National Health Survey (NHS): This is the primary individual-level survey data used, collected by the Australian Bureau of Statistics (ABS). It provides information on various health behaviors and risk factors.
- 2016 Australian Census Data: Used to derive auxiliary variables related to socioeconomic status, demographics, and other characteristics at the SA2 level.
- Social Health Atlases of Australia (SHAA) Estimates: Used as an external latent field in the Stage 2 model to borrow strength from previously published estimates.
- Australian Statistical Geography Standard (ASGS) 2016: Used for defining the geographical areas (SA2, SA3, SA4).
- Accessibility and Remoteness Index of Australia (ARIA+): Used to classify areas by remoteness.
- ABS Socio-Economic Indexes for Areas (SEIFA): Specifically, the Index of Relative Socio-Economic Disadvantage (IRSD) was used.
Key Insights and Ideas
- Unprecedented Spatial Resolution: The study demonstrates that by applying advanced SAE techniques, it's possible to generate reliable cancer risk factor estimates at a much finer spatial scale (SA2) than previously available, revealing more granular patterns of health disparities.
- Remote Areas Highlighted: The method successfully provides estimates for very remote areas, which were previously excluded from national surveys. This is crucial because these areas often exhibit higher prevalence of unhealthy behaviors and greater health burdens.
- Socioeconomic and Remoteness Link: The findings reinforce the well-established link between lower socioeconomic status, increased remoteness, and higher prevalence of many unhealthy behaviors, a critical insight for targeted public health interventions.
- Risky Alcohol Consumption Anomaly: Interestingly, risky alcohol consumption showed a different pattern, with higher prevalence in less disadvantaged areas and major cities, contrasting with other risk factors. This highlights the nuanced nature of health behaviors and the importance of specific analysis for each factor.
- Validation is Key: The rigorous internal and external validation process underscores the reliability of the developed methodology and the generated estimates, building confidence in their use for policy and research.
Key Results
- The study generated prevalence estimates for 2221 small areas (SA2s) across Australia for eight cancer risk factor measures.
- The new estimates provided improved spatial resolution and reach compared to previous publications like the SHAA.
- Higher prevalence of unhealthy behaviors was observed in more remote areas and areas of lower socioeconomic status, a trend consistent with previous research.
- The uncertainty (CV) for estimates in very remote areas was significantly higher (over 3 times higher than in major cities), a crucial consideration for interpretation.
- The modelled estimates showed considerable stability improvements over direct estimates, with a reduction in variability by an average factor of 3.3.
- The reliability of the estimates was high, with approximately 97% possessing coefficients of variation (CV) below 25%.
Limitations and Future Work
- Self-Reported Data: A critical limitation is that most risk factor measures are based on self-reports, which are susceptible to biases. Additionally, the survey questions sometimes captured behavior from different timeframes (e.g., last week vs. usual week), which could affect comparability.
- Ecological Fallacy: The estimates are at the area level, meaning they cannot be used to establish individual-level associations between risk factors and cancer incidence.
- Modifiable Areal Unit Problem (MAUP): The choice of SA2 as the small area unit is one of many possible spatial partitions, and conclusions drawn are inherently tied to this choice.
- Exclusions from NHS: The study acknowledges that the exclusion of very remote areas, discrete Aboriginal and Torres Strait Islander communities, and non-private dwellings from the NHS means the impact of these exclusions on the modelled estimates cannot be fully assessed.
- Lack of Sex-Specific Estimates: Due to data sparsity at the SA2 level, sex-specific estimates could not be generated, which might lead to inadvertent smoothing toward the mean, as health behaviors can depend on sex.
- Future Work: The authors suggest developing distinct models for each risk factor, potentially allowing for different covariates and structures. They also propose exploring multivariate models that jointly model multiple risk factors. Further research could also involve linking multiple surveys to increase sample sizes and enable higher resolution estimates. The findings will be made available in the Australian Cancer Atlas 2.0.
Summary updated