Methods

LEARN MORE WITH THE

Australian Cancer Atlas 2.0 Technical report

The Australian Cancer Atlas 2.0 reveals how the impact of cancer varies across Australia.

Below, you will find an overview of the methods used to create the Atlas. If you would like more detailed information about the methods, technology, data sources or visualisations, please refer to our Technical Report (e-book).

View Technical ReportView Technical Report

Sex

Many of the measures included in the Australian Cancer Atlas 2.0 are reported separately for Males, Females and Persons. The information about a person’s sex is sourced from the original data custodians.

Geographical Areas

The Australian Cancer Atlas 2.0 divides the country into specific areas using something called Statistical Area 2 (SA2), a classification system by the Australian Bureau of Statistics (ABS). SA2s are the smallest areas with yearly population data. The ABS says they’re meant to show communities that interact socially and economically.

These small areas can also be grouped into broader regions defined by the Australian Bureau of Statistics. Within the Australian Cancer Atlas 2.0, summaries of the spatial patterns within four broader regions are provided:

Remoteness areas

Categorises small areas across Australia into five categories of remoteness based on their relative access to services.

Socio-economic index

A measure of area-level socioeconomic disadvantage, specifically the Index of Relative Socio-economic Advantage and Disadvantage (IRSAD).

States/Territories

New South Wales (NSW), Queensland (QLD), South Australia (SA), Tasmania (TAS), Victoria (VIC), and Western Australia (WA), the Australian Capital Territory (ACT) and the Northern Territory (NT).

Capital city areas

Geographical areas designed to represent the functional extent of each of the eight state and territory capital cities.

Methods of presenting results on maps

In the Australian Cancer Atlas 2.0, you’ll find three types of estimates displayed on maps, but the available types vary for each health indicator. For instance, for cancer diagnoses, you can see all three estimates, while for cancer risk factors, only Geographical Patterns (average) is accessible.

Geographical patterns (average)

Provides a snapshot of geographical patterns for the latest available time period (up to 10 years). Information is available for both relative rates (compared to the Australian average) or absolute modelled counts (such as the number of cancers diagnosed).

Geographical patterns for each time period

Shows how geographical patterns across Australia have changed over different time periods (typically, but not always, single years). This is available for relative estimates only, where area-specific rates for each time period are compared to the Australian average for that time period.

Changes over time for each geographical area

Shows how the rates for each geographical area have changed over a combined time period. This is also available for relative estimates only, where the area-specific rates for each time period are compared to the Australian average over the combined time period. These changes reflect the national trends over time in many cases.

Dataset sources

Cancer registry

Information about cancer diagnosis and survival across Australia comes from the Australian Cancer Database. This database is managed by the Australian Institute of Health and Welfare (AIHW). The Cancer Data and Monitoring Unit at AIHW brings together data from each of the eight state and territory cancer registries in Australia to create the Australian Cancer Database.

Hospital treatment data

Hospital admission data for any of the three treatments were extracted from the National Hospital Morbidity Database and the Australian Institute of Health and Welfare.

National Health Survey

Data from the 2017-18 National Health Survey conducted by the ABS were used to estimate the prevalence of cancer risk factors. The survey aimed to collect health information from one adult and, if possible, one child in selected households. The survey covered 76% of SA2s across Australia, with a median sample size of 8 and an interquartile range of 5 to 13 at the SA2 level.

Statistical Methods

Spatial models (all indicators except risk factors)

The Australian Cancer Atlas 2.0 used Bayesian spatial (small area) models to estimate the geographical patterns (relative to Australian average) over aggregated time periods for each included indicator. There are different types of such models. All models used belong to a category termed “Conditional Autoregressive (CAR) models (Cramb et al.,2016).

These models readily incorporate the spatial correlation between areas in a natural manner as part of the prior information, recognising that adjoining geographical areas are likely to have some similar characteristics. These models also provide a probabilistic description (posterior distribution) for the estimated parameters, which is helpful in quantifying and understanding the extent of uncertainty around the spatial estimates.

All models included spatial smoothing. Models for cancer diagnosis, screening or testing ,and treatment were based on aggregated count data. The observed counts for each indicator were modelled as a Poisson process (except for BreastScreen data where a binomial model was used), offset by the corresponding expected counts which were based on the national age-specific rates for the analysed time period. For cancer survival, a piecewise Poisson spatial survival model was used, that included various covariates (broad age group, sex and cancer type) and follow-up time.

Spatio-temporal models

Spatio-temporal data are stratified by both geographical area and time. Spatio-temporal models for cancer diagnosis in the Australian Cancer Atlas 2.0 were based on the model of Bernardinelli et al, 1995 (Bernardinelli et al 1995).

The observed counts of cancer diagnoses were modelled as a Poisson process offset by the expected counts, for each combination of area and time. Expected number of diagnoses were calculated using the age distribution of the area at each time and the age-specific rates of cancer diagnoses.

Models described the interaction between time and space by including a national temporal trend, area-specific effects, and area-specific deviations from the national trend.

Survey estimation (cancer risk factors)

Because the risk factor data came from survey data, a newly developed Bayesian two-stage small area modelling approach was used to obtain risk factor estimates by SA2.

In the first stage, we used a hierarchical Bayesian logistic model on individual-level National Health Survey data. This model included various health, and socio-demographic factors, in addition to SA2-level socioeconomic status. The probabilities from the first stage model were used to calculate weighted prevalence estimates for SA2 areas with survey data.

In the second stage, a Bayesian spatial model was used to predict prevalence estimates for all SA2 areas. This model included spatial smoothing, random effects at the Statistical Area Level 3 (SA3) level, SA2-level covariates (such as socioeconomic data), and previously published risk factor prevalence estimates from the Social Health Atlas of Australia (Social Health Atlas).