Premature deaths

Quantifying the magnitude, variation and the causes of premature death, deaths before the age of 75, have implications for public health policy in terms of priority setting, monitoring and policy development. The loss in the potential years of life for premature death in Australia has been estimated at 871,807 years or 37 years per 1,000 population in 2018 (AIHW, 2020). The impacts of premature death on society are therefore substantial with significant economic consequences. Carter et al., 2017 estimated that in Australia about 284,000 working years and $13.8 billion in lifetime income was lost due to premature deaths occurring in 2003. They report that deaths from cancer and cardiovascular disease accounted for more than half the lifetime income lost while injuries and mental disorders were associated with the highest average loss per death. These significant costs are also apparent internationally, compelling the United Nations to set a key international public health goal ( Sustainable Development Goal 3.4) to reduce premature mortality by a third for non-communicable diseases by 2030.

The premature deaths indicator

The premature death indicator refers to deaths among people younger than 75 years. This cut-off age group produces conservative estimates of premature mortality because it represents deaths that occur at an age lower than the median age of death (78.8 years for males and 84.8 years for females (ABS, 2020) and life expectancy at birth (80.9 years for males and 85.0 for females between 2017 and 2019) (ABS, 2020a). Premature deaths represented just over 54,000 deaths in 2018, which equated to 34% of all deaths in Australia (n=158,493). Cancer, diseases of the circulatory system and the combined external causes of accidents, poisonings and violence were the main causes of premature death in Australia in 2018 (AIHW, 2020a).

Premature deaths are reported in numerous ways, as a single category (by persons), by sex (male and female) and also by specific cause of death. Within the Social Health Atlas(PHIDU, 2021), PHIDU releases premature death data on cancer deaths as a total or by specific types of cancer; colorectal, lung, and female breast cancer. Other causes of death are reported in relation to circulatory system diseases, ischaemic heart disease, cerebrovascular disease, diabetes, respiratory system disease, chronic obstructive pulmonary disease, external causes, road traffic injuries, and suicide and self-inflicted injuries.

Potentially avoidable deaths

Potentially avoidable deaths, also referred to as avoidable mortality, are a subset of the premature death category (deaths under 75 years). They are used as a performance indicator for the effectiveness of the health system, including hospital, primary and community care (Falster and Jorm, 2017). While a primary measure for health system effectiveness, it is also used for measuring improvements to health standards and primary care which lead to fewer deaths.

Over 26,000 potentially avoidable deaths occurred in Australia in 2018 at an age-standardised rate of 100.3 deaths per 100,000 persons. Around 50% and 46.9% of all premature deaths for males and females were classified as potentially avoidable in 2018 (AIHW, 2020).

In their publication “A guide to the potentially avoidable deaths indicator in Australia” (Falster and Jorm, 2017), report that potentially avoidable deaths can be used for comparisons between geographic regions, be broken down by cause of death and by population subgroups, and used to show trends over time. They report that this information can be used to identify priorities for targeted policy interventions as well as to monitor for improvements or identify emerging problem areas in the health system.

The potentially avoidable death indicator

A potentially avoidable death is a death from conditions that are potentially preventable (avoidable) through individualised care and/or treatable through existing primary or hospital care i.e. avoidable in the context of the present health system (AIHW, 2020b). They reflect a formal indicator measure using nationally agreed definitions based on the cause of death within Australian National Healthcare Agreement PI 16 – Potentially avoidable deaths, 2020 (AIHW, 2020a). Cause of death are defined by ICD-10 codes/specifications laid out in the current 2020 definition.

These codes are revised regularly by a panel of clinicians, policymakers and data experts to ensure that the measure remains relevant to Australian policy priorities, are reliable in its measurement, and are comparable between regions and over time. Conditions are reviewed according to both the relative volume of deaths and the presence of contemporary interventions shown to be capable of reducing mortality within five years (Falster and Jorm, 2017).

Data

Death records

Death data was based on deaths occurring between 2009 to 2018 and was taken from the Causes of Death Unit Record Files supplied by the Australian Coordinating Registry and the Victorian Department of Justice, on behalf of the Registries of Births, Deaths and Marriages and the National Coronial Information System.

Population

Persons, male and female populations were taken from the Australian Bureau of Statistics (ABS) Estimated Resident Population (ERP) for June 2009 to June 2018 on the 2016 Australian Statistical Geography Standard geography at the Statistical Area Level 2 level (SA2).

Pre-processing of the data

Premature Deaths

Premature deaths, defined as deaths between zero and 74 years of age were extracted. Data were extracted for total premature deaths by males, females and persons as well as for specific causes of death. These causes were: all cancers, colorectal cancer, lung cancer, female breast cancer, diabetes, circulatory system diseases, ischaemic heart disease, cerebrovascular disease, respiratory system disease, chronic obstructive pulmonary disease, external causes, road traffic injuries, and suicide and self-inflicted injuries.

Potentially Avoidable Deaths

Potentially avoidable deaths, defined by ICD-10 cause of death codes within the National Healthcare Agreement: PI 16 – Potentially avoidable deaths 2020 (AIHW, 2020b) were extracted. These were: avoidable cancers, colorectal cancer, female breast cancer, diabetes, circulatory system diseases, ischaemic heart disease, cerebrovascular disease, respiratory system disease, chronic obstructive pulmonary disease, external causes, suicide and self-inflicted injuries, other external causes and transport accidents. Total avoidable death data was extracted for males, females and persons.

Setting the geography of the death data to the 2016 Australian Statistical Geography Standard

For the years 2009 to 2015, the residential location of the individual who died was coded to the Australian Bureau of Statistics (ABS) 2011 Australian Standard Geographical Classification geography at the Statistical Areas Level 2 (SA2) area unit. The number of deaths by SA2 for each death category was proportioned to the 2016 Australian Statistical Geography Standard using the ABS SA2 geographic correspondence file.

For the years 2016 to 2018, the published geography of the Causes of Death Unit Record Files was geocoded to the 2016 Australian Statistical Geography Standard. We calculated the total number of deaths for each SA2 for each premature and potentially avoidable death category.

The SA2s for all years and death categories were then corresponded to the PHIDU Population Health Area geography (PHA). PHAs are comprised of a combination of whole ABS defined SA2s and multiple (aggregates of) SA2s. Nearly 40% of PHAs are combined of one SA2, another 38% have two SA2s with the other 22% of PHAs comprising of multiple combinations of SA2s.

Extraction of death data by age-group

In order to directly age-standardise the rates of death, we broke the death datasets for each premature and potentially avoidable death category into seven age-groups categories (0-14 years old, 15-24 years old, 25-34 years old, 35-44 years old, 45-54 years old, 55-64 years old, 65-74 years old). The majority of these classification represent a ten-year time period. This ten-year time period has been deemed acceptable in the standardisation process with very little difference between age-standardised rates using either five-year or ten-year age groups (AIHW, 2011).

Merge with ABS Estimated Residential Population

Yearly death data by PHA geography and age-group for each premature and potentially avoidable death category was merged with the corresponding yearly ABS ERP by age group.

Using the Empirical Bayes process to calculate small area death rates

The calculation of crude or age-standardised mortality rates for small areas has the potential to produce large unreliable rates of mortality. This occurs when deaths are rare occurrences, such as for a specific cause of death or when they are collected over small time periods such as annually. The corresponding populations used in the process also pose an issue, with small age-specific populations within each PHA inflating the estimate of the age-standardised rate. Benach et al., 2003 suggests that the large heterogeneity in population size across areas leads to various levels of precision of area specific mortality risk estimates and that age-distribution differences across areas result in mortality risk estimates that are influenced by age effects. The outcome of these issues is that the wide variation of rates make comparison of rates across small areas difficult. For example, when locations of high population density are compared to ones with lower population densities, the calculated age-specific ratios do not reflect the relative sizes and relatively equal rates from these areas may in reality be quite different from each other (Türkan et al., 2020). Additionally, Yasaitis et al., 2015 suggests that there are calculation issues when an investigator wishes to explore local health phenomena where the populations are too small to create statistically stable estimates.

To minimize the potential fluctuations in mortality rates, the Empirical Bayes (EB) estimator has been proposed. Early studies (Tsutakawa et. al, 1985; Manton et al., 1989) used the EB methodology in small and moderate sized areas to obtain adjusted rates in cancer morality, reporting that these were more stable for comparison. Later studies have used extensions of this methodology to investigate geographic variation in premature mortality (Fukuda et al., 2004), area based standardised mortality ratios for rare occurrences of death such as suicide (Yamaoka et al., 2020), cardiovascular conditions (Rodrigues et al., 2015; Darikwa et al., 2019), Chagas’ disease (Martins-Melo et al., 2012), schistosomiasis-related mortality (Pinheiro et al., 2020), infant mortality (Türkan et al., 2020) and cancer (Agovino et al., 2018). The use of EB in the creation of atlases to map the geographic variation of mortality has also been significant. For example, Benach et al., 2003 used EB to estimate the small area age-adjusted relative risk for each cause/gender/age group combination of mortality to create the Spanish Atlas of Mortality. Pickle, 2009 describes its use for making small area maps of direct age-standardised death rates for multiple cancers across the United States, while Athens et al., 2013 used the methodology to rank the health status of counties by premature mortality and other population health measures in the United States. These atlases are an important tool for examining geographic patterns of mortality risk, identifying specific high-risk areas and providing an evidence base for health policy planning.

When we implement the Empirical Bayes methodology, the process weighs the information of the small area of interest together with the information from all areas under investigation (Benach et al., 2003). This involves the creation of a prior distribution of relative risks in the form of a probability density function where the prior distribution reflects prior beliefs about the distribution of relative risks and this is parameterized by hyperparameters (Leyland et al., 2005). The use of the hyperparameters smooths or shrinks the observed value, improving the estimate by borrowing strength from the values of all observations in order to produce more stable estimates. In practice, those areas with smaller populations will have their rates smoothed more or shrunk towards the mean more than areas with larger populations.

One criticism of the EB approach is that there is over shrinkage towards the mean with the result that the variance of a collection of EB estimates will underestimate the variance of the true rates (Leyland, 2005). The amount of shrinkage is determined by the extent of unreliability in the original data and the model assumed for the prior distribution of relative risk (Congdon 1994). The author suggests that suicide and less common causes, especially for areas with small populations are contexts where this type of analysis is appropriate. For larger populations and more common causes of death, shrinkage towards the mean diminishes and the need for adjustment by the EB methodology is reduced. Ugarte et al., 2009 qualifies this statement demonstrating with sensitivity analysis that while smoothing is the desirable property of the model, an excess of smoothing may hinder detection of high-risk areas. The reliability of the estimates is mostly dependent on the magnitude of the underlying risk and population. If the underlying risk is low and the population is sparse then smoothing may reduce the identification of true positives. Later analysis (Yasaitis et al., 2015) showed similar results where EB tended to adjust the values to the global mean with more sparsely populated areas adjusted more dramatically towards that mean. These studies suggest that the EB methodology is a very conservative approach to estimate mortality rates.

Worked examples of the EB calculation to estimate age-standardised death rates for the years 2009 to 2018

Step 1 Estimation of the prior probabilities

Step 1 Estimation of the prior probabilities

We ran the ebbr package on each of the seven age-group categories for the Premature – Persons death category. Figure 3 to Figure 5 show the Beta distributions and density plots for the 0-14, 15-24 and 25-34 years of age categories. In these younger age-categories where the rates of death are lower when compared to older ages, the dynamics of the Beta distributions can be characterized as tall and peaked with low EB fitted mean crude death rates (the dashed line). These dynamics highlight the existence of a small number of PHAs with either single or a small number of multiple deaths, large ten-year aggregated populations, and the existence of small variations in death rates between PHAs.

At older age-groups, higher numbers of deaths occur within sometimes smaller aggregated ten-year populations. This generates a greater magnitude and statistical range of death rates within and between PHAs generating flatter and wider Beta distributions (Figure 6 to Figure 9).

The prevalence of death rates differs by age-group. The differences between EB fitted crude death rates and the crude death rates are shown in Figure 3 to Figure 9 (right panels). The darker points represent PHAs with higher populations while the lighter coloured points representing PHAs with lower populations. The solid line is the 1:1 ratio between the EB fitted crude and crude death rate estimates while the dashed line represents mean age-group specific death rate. A similar pattern occurred across all figures, darker points are apparent near the 1:1 line demonstrating minimal shrinkage to the mean death rate. Lighter coloured points, representing PHAs with higher crude rates (to the right of solid line) show greater shrinkage to the mean death rate (the dashed line). PHAs with high crude rates and high populations are also shrunk to the mean but by less than those with lower populations. Comparison between figures shows that the shrinkage towards the mean was higher for younger age groups than for older age groups.

The estimated α and β parameters for the ten years of Premature – Persons death category for each age-group were extracted and then used in Step 2.

Step 2 Use of the ten-year average hyperparameters to create annual EB fitted estimates

Step 2 Use of the ten-year average hyperparameters to create annual EB fitted estimates

For each age-group, the calculated ten-year average hyperparameters are used to adjust the annual age-specific crude rates of death for each PHA using the equation:

EB fitted rate = (Deaths in PHA+ α)/(Population in PHA + α+ β)

Where α, β are the estimated hyperparameters for each age-specific for Premature – Persons death category over the ten years.

Using Premature -Persons deaths in the 35-44 age-group category as an example, the estimated hyperparameters, α and β were 5.9 and 5259. Substituting these values into the above equation, with zero death and population numbers gives a mean death rate of 0.00112 or 112 premature deaths per 100,000 persons for the 35-44 age group category. Table 2 shows the effect of the hyperparameters in shrinking the annual crude death rate towards the mean. For low populations, the calculated crude rates for one or multiple deaths are very high. Applying the hyperparameters, shrinks the two high crude rates towards the mean death rate while keeping the relative magnitude between them. As populations increase, rates are adjusted less towards the mean. However, there is some bias to the mean with larger populations and low numbers of deaths. Importantly though, the relative differences are still maintained. Table 2 also shows the mathematical limits imposed by the hyperparameters on the resultant age-specific EB fitted rates, limiting the variation of EB fitted death rates in the calculation. Here, the EB fitted death rates for one death occurrence in populations that increase from 200 to 1,000 to over 10,000 persons, sees a decrease in rates from 126 to 110 to 45 deaths per 100,000 persons. This change is different in the calculation of crude rates with rates declining from 500 to 100 to 9.8 deaths per 100,000 persons. The EB method also places additional constraints on the estimation of death rates with multiple occurrences in the same age-group category. The estimated rates of one occurrence in small populations (126 deaths per 100,000 persons) is similar to the rate to three occurrences in a larger population (123 deaths per 100,000 persons). This compares to the large relative difference when using the crude rates which were 500 to 150 deaths per 100,000 persons.

Table 2: An example of the effect of the hyperparameters (Alpha and Beta) on the calculation of EB fitted death rates and comparison with the Crude death rate for the 35-44 years of age category for the Premature – Persons death category.

Deaths	Population	Alpha	Beta	Crude EB fitted death rate (per 100,000 persons)	Crude death rate (per 100,000 persons)
1	200	5.9	5,259	126	500
6	330	5.9	5,259	213	1,818
1	1,000	5.9	5,259	110	100
15	1,200	5.9	5,259	323	1250
3	2,000	5.9	5,259	123	150
23	4,527	5.9	5,259	295	508
1	10,187	5.9	5,259	45	9.8

 

The effect of hyperparameters on the death rates for Premature – Persons deaths in the 35-44 years old age group is shown graphically in Figure 10 (left panel). The points within the figure represent each year of the ten years for each PHA. Similar to Table 1, those PHAs with low population (darker colour) are shrunk to the mean death rate (dashed line). Those PHAs with higher populations (lighter colour) are shrunk less to the mean. Those PHAs near the1:1 vertical line have similar values to the crude rates. EB fitted death rate values below the mean EB fitted death rate show the imposed bias placed on the estimation, with higher EB fitted death rates when compared to the crude death rates. The branching out of estimates graphically along a fairly linear pattern represent PHAs with the same number of death occurrences (the bottom line of values in Figure 10 being one death) with decreasing populations. Therefore, for one death occurrence and decreasing populations (darker coloured points), the crude death rates increase (along the x-axis), as populations get smaller. This issue is minimised with the EB fitted death rate, as rates are shrunk to the mean (dashed line).

This example shows the hyperparameters for the most prevalent cause of death category. For causes of death that are less prevalent, such as deaths from Premature Cancer in the same age category (35-44 years old), the effect of the hyperparameters on the estimation of the EB fitted death rates will be more dramatic (Table 3). All crude death rates are shrunk towards the mean which was 28.6 deaths per 100,000 persons. The lack of prevalence of Premature Cancer deaths in this age category means that the majority of PHAs will record only one or a small number of multiple death occurrences annually. This is shown graphically in Figure 10 (right panel) as a stepwise linear increase representing an increase in death occurrences from one to seven deaths per PHA annually. This linearity occurs as population decrease (darker coloured points) and the death rates are shrunk towards the mean. The resultant distribution of EB fitted death rates is small between 19 to 59 deaths per 100,000 persons because of the lack of prevalence of Premature Cancer deaths for this age group.

Table 3: An example of the effect of the hyperparameters (Alpha and Beta) on the calculation of death rates for the 35-44 years of age category for the Premature Cancer death category.

Deaths	Population	Alpha	Beta	Crude EB fitted death rate (per 100,000 persons)	Crude death rate (per 100,000 persons)
1	100	3.8	13,393	36	1,000
2	223	3.8	13,393	43	897
1	1,000	3.8	13,393	34	100
2	1,000	3.8	13,393	41	200
3	1,000	3.8	13,393	47	300
1	2,000	3.8	13,393	31	50
1	5,000	3.8	13,393	26	20
1	10,000	3.8	13,393	21	10

 

Table 3 and Figure 10 represents an example of the estimation of EB fitted death rates with a low prevalence of death in a younger age group category. Figure 11 shows a comparative example of the estimation of EB fitted death rates for the 65-74 years old age group which has a greater prevalence of death for the Premature – Persons (left panel) and Premature Cancer (right panel) death category. The figure shows similar trends to Figure 10 with death rates for lower populated PHAs being shrunk towards the mean (dotted line) but with less dramatic effects than in Figure 10. Once again, the shrinkage to the mean was more apparent for Premature Cancer (right) than for the Premature death category (left) because of the lower ten-year prevalence of this cause of death in this age-category.

Comparison of EB fitted age-standardised annual death rates for cause of death with a low and high number of deaths

Comparison of EB fitted age-standardised annual death rates for cause of death with a low and high number of deaths

When the number of deaths is small, for example, for deaths caused by COPD (n=19,286), the mean rates of each age-specific prevalence will be small. This means that annually, the majority of deaths per PHA will also be small with around 56% of all PHAs recording between one to three COPD deaths. This distribution coupled with the age-standardisation process will impose issues in the calculation of the age-standardised death rate. Figure 12 shows the comparison between the EB fitted age-standardised rate and the age-standardised rate for COPD deaths for the ten years of data. For PHAs recording only one death from COPD in a year (top-left), the shrinkage of rates to their age-specific ten-year average is apparent. The five-tier structure in the graph represent five different age groups, the lowest step representing COPD deaths for the 35-44 years old age group and the highest being deaths for the 65-74 years old age group. In comparison to the age-standardised death rates, the rates are much smaller and there now exists little variation in the EB fitted age-standardised death rates. The only significant variation is the differences in the mean EB fitted age-standardised death rates between age groups. It is important to note, that these rates are shrunk significantly to their ten-year age-group average reflecting the overall prevalence of the cause of death in the PHAs over the ten years. Also of note, are the points along the 1:1 line which represent the substitution of the age-standardised death rates for deaths in age-groups where the numbers of deaths were too small (age groups 0-14, 15-24 and 25-34 years old) to be implemented using the ebbr algorithms.

The next three graphs in Figure 12 show the additive nature of the construction of the EB fitted age-standardised death rates for each PHA. Rates for two or more deaths in a PHA annually are calculated in relation to deaths corresponding to either one specific age group or across multiple age groups. If there are multiple deaths in one age-specific category, as in Table 2 and Table 3, there is only a small rise in death rates between one and multiple occurrences. If there are multiple deaths across a number of age-categories, the EB fitted age-standardised death rates increase will be larger than if it occurs in one age-group category because the prevalence of death increases with age (in most cases). The final graph (bottom-right) includes PHAs with COPD deaths of 4 or more, annually. These rates are calculated as above, with deaths in either one specific age group or across multiple age groups. The graph shows the structured nature, with EB fitted age-standardised death rates higher than for PHAs with lower number of deaths. In comparison, age-standardised rates were high for PHAs with one or two deaths indicating the impact of low populations on the calculation of the rate. This influence is now removed from the calculation. The higher EB fitted age-standardised death rates now reflect PHAs with higher number of deaths (four or more annually) rather than PHAs with a small number of deaths in a small population. Comparison of EB fitted age-standardised death rates and the age-standardised death rate show that while the absolute values of the EB fitted age-standardised death rates are lower than the age-standardised death rate there are similar relative magnitudes with high rates being high and low rates being low.

Figure 12 shows the effects on the rate calculations when cause of death prevalence is small. For comparison, Figure 13 shows the comparison of EB fitted age-standardised death rates to age-standardised death rates for the Premature – Persons death category. Once again, some of the high value age-standardised rates are driven by the interaction of a low number of deaths and small populations. This is apparent in the 11-30 Deaths per PHA category where there are age-standardised rates of around 750 deaths per 100,000 persons (top-right). The addition of a higher number of deaths (i.e. the 51 Deaths or more category) in other PHAs shows that the EB fitted age-standardised and age-standardised death rates are fairly comparative not moving substantially from the 1:1 line.

Comparison of the distribution of the EB fitted age standardised rates against the age-standardised death rates

In this section, we compare the population weighted 10 year average and standard deviations of the EB age-standardised and age-standardised death rates across all PHAs (Figure 14 and Figure 15). The figures show the reductions across all causes of death in both the 10 year averages and their corresponding standard deviations when the EB method is used. Of note, is the large standard deviations in some causes of death when the traditional age-standardisation method is implemented. This large amount of variation induced by this method in the calculation of the death rates will make the identification of PHAs with large death rates problematic.

For a final comparison of death rates, we compare the distributions of the annual EB-fitted age-standardised and the annual age-standardised death rate for the PHAs over the ten years using a both a scatter plot and histogram. Below are the examples for the COPD and Premature – Persons death categories. A comprehensive list by causes of death is available here

Figure 16 (left panel) shows the scatterplot of the rate distributions for COPD deaths by population while Figure 16 (right panel) shows the histogram of the rate distributions. Both plots show that the age-standardised death rates for COPD are positively skewed with long tail of outlying high rates. The average age-standardised death rate for COPD per PHA was 7.3 deaths per 100,000 persons over the ten years. The distribution of the EB fitted age-standardised death rates for COPD now shows a bimodal dynamic of the distribution with a group of PHAs with very low annual rates of death and a large group of PHAs with higher rates. The average EB fitted age-standardised death rate for COPD per PHA was estimated at 4.3 deaths per 100,000 persons.

For the Premature – Persons death category (Figure 17), the age-standardised mean death rate for Premature – Persons deaths was 205.1 deaths per 100,000, this compares to a lower average EB fitted age-standardised mean death rate of 191.7 death per 100,000 persons. The dynamics of the distribution of the EB fitted age standardised death rates shows a much more peaked and tighter distribution when compared to the distribution of age-standardised death rates which report a flatter distribution with a larger number of outliers. Of note, is the outlying PHA, top right of the left panel, which has a high death rate using the two methods. This suggest some degree of internal data consistency.

Some cautionary notes

We conducted an Empirical Bayes approach to estimate death rates at the PHA small area level across Australia. We undertook this to remove the influence of small numbers (both deaths and populations) on the calculation of annual PHA death rates. In general, we cannot easily distinguish between what is the true death rate of a PHA, what variation is caused by statistical issues and what is the degree of over smoothing or bias that is imposed by the EB methodology.

The major benefit of the EB methodology is that it provides a conservative estimate of the mortality rate, specifically reducing high outlier death rates which are high when a very small number of deaths occur randomly in a PHA and when age-specific populations are low. This is highlighted in the worked example in Table 2 and Table 3 and this will affect the robustness of EB method estimating the death rate. For example, for the cause of death categories which have large numbers of deaths such as the Premature – Persons death category (n=514,710 deaths), the difference between the age-standardised and the EB fitted age-standardised death rates are low. Nearly, 43% of the EB fitted death rates differ by ±10% and 76% of the values differed by ±20% when compared to the age-standardised death rates. However, for those causes of deaths with lower numbers of deaths, such as female breast cancer deaths (n=17,194 deaths), the difference between the EB fitted age-standardised death rates and the age-standardised death rate was high with over 70% of the values between 20-75% less than the age-standardised rate. These examples represent opposite ends of the spectrum with Premature – Persons deaths having a large number of deaths and large populations in comparison to the female breast cancer death category which have fewer deaths and is only calculated for the female populations (by age-group) in each PHAs. We therefore suggest that caution be used for those causes of death where low numbers of deaths occur.

Hot-spot analysis – Identifying the geographic variation and temporal persistence of death hot spots across Australia

We now have estimates of annual rates of death for each PHA which have been adjusted for known statistical issues. To identify the geographic and temporal persistence of high rates of death we undertook four steps:

• The first step involved deriving ratios between the annual EB fitted death rates by PHA by cause of death to their corresponding annual Australian average of the EB fitted death rates. For example, values of 0.5 or 1.5 meant a rate that was 50% lower or 50% greater than the Australian rate while values of 2 or 3 meant rates were two to three times greater than the Australian rate.
• The second step compared these ratios to a range of thresholds. This was undertaken to understand the sensitivity of the death rates within the PHA for each year and discern the disparities between PHA. A total of 20 thresholds were selected with the lowest threshold set to 50% of the Australian rate and the highest set to four times the Australian rate for each year. We recorded a value of one, when the ratio was higher than each threshold while a value of zero was recorded when it was lower over the ten years of analysis.
• In the third step we aggregated the ten years into a ten-digit code for each PHA by cause of death categories. This created unique codes of when the PHA was either above or below a chosen threshold for the ten years, as a whole. The codes were then classified into five categories; Cold, Cold-Warm, Warm, Warm-Hot and Hot based on a set of aggregation rules.
• This fourth step produced a dataset summarising the heat of cause of death categories by PHA and threshold. This information was used for input into the interactive web-based maps and heat map graphs

Hotspot differences for the Premature – Persons and Avoidable – Persons death categories by Australian State and Territory

There is a large variation in the number of hotspots identified across Australia. This variation differs by geographic location and by cause of death category. For the broad causes of death categories, the premature and potentially avoidable death categories, around 37% and 34% of all PHAs across Australia were classified as “Hot” i.e. those areas that had death rates consistently over the Australian average for the 10 years of data investigated (Figure 19 and Figure 20).

The distribution of PHAs within each state and territory that are classified as “Hot” and “Cold” by the Premature – Persons and Avoidable – Persons death categories for a range of thresholds are shown below. A comprehensive list for each cause of death is available here

Premature – Persons

The percentage of PHAs within each state or territory that are consistently, across all ten years, above (“Hot”) or below (“Cold”) the EB fitted Australian rate of deaths for the Premature – Persons death category are shown in Figure 19. The classifications represent the two extremes on the heat classification spectrum. The percentage of PHAs that were classified as “Hot” within a state or territory were by far the highest in the Northern Territory (NT), followed by a group of states; Tasmania (TAS), New South Wales (NSW) and Queensland (QLD). This classification represented over 70% of PHAs in the NT and between 28% and 48% of PHAs for the other states for the above the Australian Average threshold. As the thresholds increase the percentage of PHAs that are “Hot” decrease. At threshold 1.4 times the Australian average rate the percentage of PHAs classified as “Hot” reaches zero, except for the NT. The percentage of Population Health Areas that were classified as “Cold” within a state or territory were highest in Australian Capital Territory (ACT). This is followed by all states except the NT. This represented 75% of PHAs in the ACT, 50% of PHAs in Victoria and around 28% in TAS for the above Australian Average threshold. No PHAs in the NT were classified as “Cold” for this threshold. As the thresholds increase the percentage of PHAs that are “Cold” increase within all states and territories, reaching over 90% of PHAs by the 1.4 times the Australian average threshold. This occurs in all regions except for the NT.

If we take the example of Victoria (VIC) in Figure 19, we can see that 28% of PHAs are classified as “Hot” at the Australian average. At the opposite end of the classification spectrum, we can see that 50% of PHAs are classified as “Cold” at the Australian average. Therefore, the majority (78%) of PHAs sit in these two classifications with 15% or fewer PHAs distributed within the “Cold-Warm”, “Warm” and “Warm-Hot” heat categories.

Once again, it is important to mention that the EB methodology has shrunk the variation and therefore the statistical range of death rate estimates for the Premature-Persons death category. This is evident with only four thresholds being used before either zero or over 90% of PHAs being reached.

Avoidable – Persons

The percentage of PHAs within each state or territory that are consistently, across all ten years, above (“Hot”) or below (“Cold”) the selected thresholds for the Avoidable – Persons death category are shown in Figure 20. The percentage of Population Health Areas that were classified as “Hot” within a state or territory were by far the highest in the NT. This was followed QLD and to a lesser extent, NSW. For this death category which represents deaths that can be seen as avoidable, over 90% of PHAs in the NT are classified as “Hot”. These PHAs are always over the Australian Average threshold for the ten years of data. In comparison, the other states are between 25% and 42% of PHAs. As the thresholds increase the percentage of PHAs that are “Hot” decrease to zero, except for the NT. The percentage of PHAs that were classified as “Cold” within a state or territory were highest in the ACT followed by VIC. This represented nearly 80% of PHAs in the ACT and over 50% of PHAs in VIC. As the thresholds increase the percentage of PHAs that are “Cold” increase with all states and territories.

Hotspot differences for selected cause of death across Australia

Table 4 shows the number of PHAs across Australia that are classified as “Hot”, "Between Hot and Cold" and “Cold” by the Above Australian Average, 30% and 50% more than the Australian Average thresholds. There is some variation in the numbers of “Hot” and “Cold” PHAs by cause of death. As the thresholds are increased, the influence of the EB methodology is apparent with only a small number of PHAs classified as “Hot” under the 50% more than the Australian Average threshold. This is caused by the shrinkage in the statistical range of EB fitted death rates. However, there are some causes of death which have a notable number of PHAs classified as “Hot’ under this high threshold, for example, deaths from respiratory diseases and road traffic injuries.