Measuring excess mortality in Australia during the COVID-19 pandemic

Provisional deaths data for measuring changes in patterns of mortality during the COVID-19 pandemic and recovery period.

Released
25/11/2020
Source
Provisional Mortality Statistics, Jan - Aug 2020

Introduction

Since the emergence of COVID-19, deaths due to the virus have been closely monitored in Australia and around the world. During the COVID-19 pandemic many countries have experienced increased rates of mortality. These increases have been seen in part due to deaths attributed to COVID-19 deaths, but also by deaths to other causes. The concept of increased all-cause mortality is referred to as excess mortality, typically defined as the difference between the observed number of deaths in a specified time period and the expected numbers of deaths in that same time period. Measures of excess mortality can account for deaths due to COVID-19, any potentially misclassified or undiagnosed COVID-19 deaths and other mortality that may be indirectly related to the pandemic (e.g. relating to social isolation or changed access to health care). As such, excess mortality can provide a more complete measure of the overall impact of the pandemic.

To date, provisional mortality reports have focused on comparisons between weekly counts of 2020 deaths and 5-year historic averages (see graph below). These reports have provided a rapid snapshot of mortality in Australia, while also providing initial insights into any changes in mortality during the pandemic. However, these reports do not provide further context or measure the statistical significance of changes over time which may relate to natural weekly variation in numbers of deaths, changes in population size or structure, or cause specific trends in death rates.

This report provides further analysis of 2020 mortality data. The first part of the report focusses on age-standardised death rates that take account of population changes (including age structure and natural increase) when making historical comparisons. The second part focusses on whether Australia has experienced excess mortality during the COVID-19 pandemic, either at the all-causes level or among specific causes of death.

Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always have 7 days. Leap years result in slightly different dates being included in each week from week 9 on, with week 53 containing two days in a leap year but only one day in other years.
  4. Refer to explanatory notes on the Methodology page of this publication for more information regarding the data in this graph.
  5. Data for the number of COVID-19 infections has been sourced from the COVID-19 daily infections graph published on the Australian Government Department of Health website. Data extracted 1 November 2020.

Age-standardised death rates

Age-standardised death rates (SDRs) enable the comparison of deaths over time and between different population groups as they account for changes in both the size and age structure of the population. An additional table (1.2) has been added to the weekly dashboard with SDRs for doctor certified deaths in 2020 and for the comparison 5 year period. Confidence intervals are also provided to assist with interpretation.

The reported SDRs represent the rate of doctor certified deaths recorded per 100,000 people. SDRs for 2020 have been calculated using provisional doctor certified mortality data for the numerator and estimated resident population projections as the denominator. While they provide a robust measure for comparison across time periods they should not be taken as the official death rate for Australia. 

Key statistics

SDRs for doctor certified deaths in 2020 have been generally lower than the historical average recorded for the same weeks between 2015-2019. 

  • The exception to this is the week ending 31 March where a rate of 8.5 deaths per 100,000 people was recorded. This compares to a 5 year average of 8.3 deaths per 100,000 people. This week coincides with higher numbers of COVID-19 infections and stricter lockdown measures being implemented in Australia. 
  • During the winter months, the SDR is between 9-10 deaths per 100,000 people on average. During winter in 2020 the SDR has remained under 9.0. 
  • The SDR for deaths due to pneumonia was 0.2 per 100,000 people between the weeks ending 31 March and 14 April. This compares to a 5 year average SDR of 0.1 for the same period. 
  • SDRs for selected causes of death are included in Table 1.2 of the weekly dashboard in data downloads. 
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always have 7 days. Leap years result in slightly different dates being included in each week from week 9 on, with week 53 omitted.
  4. Refer to the Methodology page of this publication for more information regarding the data in this graph.

Measuring excess mortality in Australia

Across the world, health and statistical authorities have sought to measure excess mortality during the COVID-19 pandemic. There are many different methodologies that are designed for this purpose, with the suitability of particular methods often dependent on factors such as data quality and collection methods, or outcomes sought from analysis. 

In Australia, New South Wales (NSW) Health use methods developed for measuring excess mortality to monitor influenza infections and deaths. The methodology developed by NSW Health is based on a model developed by Serfling and the Centres for Disease Control and Prevention (CDC) in the United States. This model uses past patterns of disease mortality to predict expected numbers of deaths, as well as thresholds that help identify when temporary health hazards (i.e. the COVID-19 pandemic) have resulted in excess mortality. The ABS adopted aspects of this methodology and applied a cyclical linear regression with a robust estimation procedure to produce an expected number of deaths for 2020 for all cause and cause-specific mortality to identify significant changes in patterns of mortality over time. See the methodology section at the end of this article for detailed information about the model applied.

Interpreting results

Each section below presents two graphs – The first graph is a time series of doctor certified deaths from January 2015 to August 2020 that has the number of actual observations plotted against the expected number of deaths estimated from the regression. The upper and lower thresholds (1.96 standard errors) of the regression are also plotted. The second graph focusses only on 2020, allowing closer inspection of patterns of death during the COVID-19 pandemic. 

At any point in time, if no temporary health hazards influence the number of deaths (i.e. community transmission of influenza or COVID-19) then there is a 95% chance that the expected number of deaths would lie between the upper and lower bounds of the confidence intervals. When actual observations (counts of death) exceed the upper threshold or drop below the lower threshold this indicates a statistically significant change in the pattern of mortality. 

Different approaches can be used to calculate the number of excess deaths. Counts of excess can be taken as the difference between the average expected number of deaths and the actual observed number. This approach risks over-estimating the number of excess deaths as the expected count lies between the upper and lower bound. Counts of excess can also be taken as the difference between the upper threshold of the expected number of death and the actual observed number. This approach results in a lower number of excess deaths, but focusses more clearly on statistically significant changes in mortality. 

Identifying significant changes in the pattern of mortality during the COVID-19 pandemic compared to previous years is the aim of this report. Counts of deaths that are above the upper bound of the confidence interval (threshold) are considered to be “excess” and will be referred to as such in this report. Counts of excess deaths described in this report refer only to the number above the upper threshold.

A single week above threshold does not necessarily suggest excess mortality. Prolonged periods (2 or more weeks) where counts exceed thresholds suggest more strongly that the numbers of deaths are above or below normal. This should be considered when analysing the data in this report. 

All-cause mortality (doctor certified deaths)

  • Excess mortality at the all cause level for doctor certified deaths is observed in the week starting 30 March 2020. There were 48 deaths that exceeded the upper threshold of the expected observations. 
  • Although the week of 30 March 2020 is a single data point where deaths exceed the 95% threshold, this week coincides with high rates of COVID-19 infections and the implementation of stricter lock-down measures in Australia.
  • The second graph which focusses on mortality trends in 2020 shows counts of deaths both with and without COVID-19 deaths included. This highlights the effect that COVID-19 had on mortality at each point in time. 
  • There were 21 deaths due to COVID-19 certified by a doctor in the week beginning March 30. This week recorded the highest number of deaths due to COVID-19 during the first wave of infections in Australia. 
  • In the week beginning 6 April the number of deaths was still above expected projections even after excluding deaths due to COVID-19 (19 deaths). However, numbers of deaths do not exceed the upper threshold so no excess mortality is recorded for that week. 
  • Deaths are significantly lower than expected from the week beginning 1 June to mid-July 2020, dropping below lower thresholds. Winter months are typically associated with higher mortality. These decreases provide insights into how public health measures put in place to manage the COVID-19 pandemic have impacted on mortality. 
  • Mortality drops below lower thresholds in the last 2 weeks of August when deaths due to COVID-19 are excluded. 
  • There were 1,566 excess deaths recorded between July and September 2017. A severe influenza season largely accounted for these excess deaths.
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always start on a Monday, with each year following on from the previous one (eg week 52 has 28 December 2015 to 3 January 2016). In the main publication, week 1 is always 1-7 January. 
  4. Refer to explanatory notes in the Methodology section for this article for more information regarding the data in this graph.
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always start on a Monday, with each year following on from the previous one (eg week 52 has 28 December 2015 to 3 January 2016). In the main publication, week 1 is always 1-7 January. 
  4. Refer to explanatory notes in the Methodology section for this article for more information regarding the data in this graph.

Cause-specific mortality

While all cause mortality provides a more accurate picture of patterns of mortality during the pandemic it is important to look at cause-specific mortality to provide insights into how individual causes have changed. Specific causes may experience significant changes which can be masked at the all-cause level. The following section focusses on excess mortality analysis for selected causes of death certified by a doctor in 2020. Deaths are analysed by underlying cause of death only. 

Ischaemic heart disease (I20-I25)

  • There were 11 excess deaths due to ischaemic heart disease over the 2 weeks starting 10 and 17 February. These excess deaths occurred outside of the official COVID-19 pandemic period. 
  • The number and rate of deaths due to ischaemic heart disease has been declining in Australia over time. This has resulted in the regression producing a reduced number of expected deaths each year.
  • Some excess mortality was recorded for ischaemic heart disease in July and August of 2017. This aligns with a severe influenza season and excess mortality from other causes. 
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always start on a Monday, with each year following on from the previous one (eg week 52 has 28 December 2015 to 3 January 2016). In the main publication, week 1 is always 1-7 January. 
  4. Refer to explanatory notes in the Methodology section for this article for more information regarding the data in this graph.
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always start on a Monday, with each year following on from the previous one (eg week 52 has 28 December 2015 to 3 January 2016). In the main publication, week 1 is always 1-7 January. 
  4. Refer to explanatory notes in the Methodology section for this article for more information regarding the data in this graph.

Cerebrovascular diseases (I60-I69)

  • In 2020 cerebrovascular diseases exceeded the upper threshold at the week beginning 6 April accounting for 8 excess deaths and the week beginning 18 May accounting for 6 excess deaths. 
  • The weeks where cerebrovascular diseases exceeded the upper threshold are both single points in time and as such caution is advised when interpreting these results. 
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always start on a Monday, with each year following on from the previous one (eg week 52 has 28 December 2015 to 3 January 2016). In the main publication, week 1 is always 1-7 January. 
  4. Refer to explanatory notes in the Methodology section for this article for more information regarding the data in this graph.
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always start on a Monday, with each year following on from the previous one (eg week 52 has 28 December 2015 to 3 January 2016). In the main publication, week 1 is always 1-7 January. 
  4. Refer to explanatory notes in the Methodology section for this article for more information regarding the data in this graph.

Respiratory diseases (J00-J99)

  • Respiratory diseases as a group (including diseases such as influenza, pneumonia and chronic lower respiratory diseases) have not exceeded the upper threshold in 2020.
  • From the week starting 25 May there have been significantly lower than expected deaths due to respiratory diseases. This decrease has been sustained throughout the winter period in Australia. 
  • In 2017 there were 1,004 excess deaths due to respiratory diseases. The majority of these excess deaths were due to the severe influenza season. A severe influenza season in 2017 has been noted by Australian infectious disease surveillance systems. 
  • Deaths due to COVID-19 are not included in the analysis for respiratory diseases. For a weekly count of deaths due to COVID-19 that were certified by a doctor see Table 1.1 in the weekly dashboard in data downloads. 
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always start on a Monday, with each year following on from the previous one (eg week 52 has 28 December 2015 to 3 January 2016). In the main publication, week 1 is always 1-7 January. 
  4. Refer to explanatory notes in the Methodology section for this article for more information regarding the data in this graph.
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always start on a Monday, with each year following on from the previous one (eg week 52 has 28 December 2015 to 3 January 2016). In the main publication, week 1 is always 1-7 January. 
  4. Refer to explanatory notes in the Methodology section for this article for more information regarding the data in this graph.

Chronic lower respiratory conditions (J40-J47)

  • No periods of excess mortality have been recorded for chronic lower respiratory diseases in 2020. 
  • Deaths due to chronic lower respiratory conditions have been below the lower threshold of expectation for the majority of the winter months. 
  • There were 80 excess deaths from chronic lower respiratory conditions in 2017. Small numbers of excess deaths were recorded in 2015 and 2016. 
  • An increase in mortality from these conditions can be seen in winter months when common respiratory infections circulate. Infectious diseases can cause an acute infectious exacerbation of disorders such as chronic obstructive pulmonary disease and emphysema.
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always start on a Monday, with each year following on from the previous one (eg week 52 has 28 December 2015 to 3 January 2016). In the main publication, week 1 is always 1-7 January. 
  4. Refer to explanatory notes in the Methodology section for this article for more information regarding the data in this graph.
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always start on a Monday, with each year following on from the previous one (eg week 52 has 28 December 2015 to 3 January 2016). In the main publication, week 1 is always 1-7 January. 
  4. Refer to explanatory notes in the Methodology section for this article for more information regarding the data in this graph.

Influenza and pneumonia (J09-J18)

  • Though numbers of deaths were higher than historical averages and projected counts for influenza and pneumonia in the last week of March and start of April, numbers of deaths did not exceed the upper threshold of the expected range.
  • Deaths due to influenza and pneumonia have been significantly lower than expected since May 2020. The significant decrease in mortality has been sustained over the winter period. 
  • Public health measures put in place to prevent the spread of COVID-19 infections can also be effective in limiting the spread of other infectious agents including influenza. 
  • The influenza season resulted in excess mortality in 2015, 2016, 2017 and 2019. 
  • In 2017 there were 788 excess deaths due to influenza and pneumonia with excess deaths recorded between the weeks starting July 24 and October 9.
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always start on a Monday, with each year following on from the previous one (eg week 52 has 28 December 2015 to 3 January 2016). In the main publication, week 1 is always 1-7 January. 
  4. Refer to explanatory notes in the Methodology section for this article for more information regarding the data in this graph.
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always start on a Monday, with each year following on from the previous one (eg week 52 has 28 December 2015 to 3 January 2016). In the main publication, week 1 is always 1-7 January. 
  4. Refer to explanatory notes in the Methodology section for this article for more information regarding the data in this graph.

Pneumonia (J12-J18)

  • Pneumonia is commonly caused by seasonal viruses including influenza. Pneumonia is the most commonly certified consequence of COVID-19 leading to death. 
  • To remove the confounding effect of influenza on projected number of expected deaths pneumonia has been analysed separately with results detailed in the two graphs below. 
  • There have been 22 excess deaths due to pneumonia during the COVID-19 pandemic with observed counts exceeding the upper threshold for 3 consecutive weeks from the week beginning 23 March. 
  • While deaths due to pneumonia have exceeded the upper threshold in past years, the timing was between May and October when influenza activity is higher in Australia. 
  • From the week beginning 20 April deaths due to pneumonia have been lower than expected with multiple weeks dropping below the lower threshold. 
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always start on a Monday, with each year following on from the previous one (eg week 52 has 28 December 2015 to 3 January 2016). In the main publication, week 1 is always 1-7 January. 
  4. Refer to explanatory notes in the Methodology section for this article for more information regarding the data in this graph.
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always start on a Monday, with each year following on from the previous one (eg week 52 has 28 December 2015 to 3 January 2016). In the main publication, week 1 is always 1-7 January. 
  4. Refer to explanatory notes in the Methodology section for this article for more information regarding the data in this graph.

Cancer (C00-C97, D45, D46, D47.1, D47.3-D47.5)

  • No periods of excess mortality have been recorded for cancer in 2020. 
  • In the week beginning 27 April 2020 numbers of death dropped below the lower threshold. This is a single point in time and caution is advised when interpreting these results. 
  • Between 2015-2020 cancer mortality has generally fallen between expected ranges. Intermittent points where the number of deaths have exceeded the upper threshold or dropped below the lower threshold of expectation have not been prolonged. 
  • This analysis has covered all cancers grouped together. Results may differ for specific cancer types. 
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always start on a Monday, with each year following on from the previous one (eg week 52 has 28 December 2015 to 3 January 2016). In the main publication, week 1 is always 1-7 January. 
  4. Refer to explanatory notes in the Methodology section for this article for more information regarding the data in this graph.
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always start on a Monday, with each year following on from the previous one (eg week 52 has 28 December 2015 to 3 January 2016). In the main publication, week 1 is always 1-7 January. 
  4. Refer to explanatory notes in the Methodology section for this article for more information regarding the data in this graph.

Diabetes (E10-E14)

  • There were 38 excess deaths due to diabetes over 4 consecutive weeks between the weeks beginning 30 March and 20 April 2020. 
  • In the week beginning 10 August 2020 the number of deaths slightly exceeds the upper threshold (1 excess death). This is a single point in time and should be interpreted with caution. 
  • Some excess deaths from diabetes were recorded during the winter months in 2016, 2017 and 2019. 
  • People with diabetes can be susceptible to infections due to a compromised immune system. Mortality can be influenced by infectious disease activity in the community. 
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always start on a Monday, with each year following on from the previous one (eg week 52 has 28 December 2015 to 3 January 2016). In the main publication, week 1 is always 1-7 January. 
  4. Refer to explanatory notes in the Methodology section for this article for more information regarding the data in this graph.
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always start on a Monday, with each year following on from the previous one (eg week 52 has 28 December 2015 to 3 January 2016). In the main publication, week 1 is always 1-7 January. 
  4. Refer to explanatory notes in the Methodology section for this article for more information regarding the data in this graph.

Dementia (F01, F03, G30)

  • There were 10 excess deaths due to dementia recorded in the week beginning 23 March. This week coincides with stricter lockdown measures being implemented in Australia. 
  • Between the weeks beginning 22 June and 13 July the number of deaths due to dementia dropped below the lower threshold. 
  • There were excess deaths recorded due to dementia in 2017 during the winter months. This was likely related to the severe influenza season recorded in 2017. 
  • People who have dementia can be at higher risk of dying from acute respiratory infections including influenza and pneumonia. The level of activity of acute respiratory disease can affect the death rate of dementia.
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always start on a Monday, with each year following on from the previous one (eg week 52 has 28 December 2015 to 3 January 2016). In the main publication, week 1 is always 1-7 January. 
  4. Refer to explanatory notes in the Methodology section for this article for more information regarding the data in this graph.
Download
  1. This graph is compiled by the date the death occurred.
  2. This data is considered to be provisional and subject to change as additional data is received.
  3. Weeks in this graph always start on a Monday, with each year following on from the previous one (eg week 52 has 28 December 2015 to 3 January 2016). In the main publication, week 1 is always 1-7 January. 
  4. Refer to explanatory notes in the Methodology section for this article for more information regarding the data in this graph.

Discussion

This research focussed on identifying and applying a model that could provide insights into whether Australia experienced any excess mortality during the COVID-19 pandemic. Measuring excess mortality has proven informative internationally. It not only accounts for deaths from the pandemic disease (i.e. COVID-19), but also accounts for any potential misclassification of diseases, and deaths that may be related to the pandemic but not directly from the pandemic disease. The model was applied at both an 'all cause' level (i.e. all doctor certified deaths regardless of their cause) and to deaths from specific causes.

At the all cause level, a small number of excess deaths was recorded (48 deaths), with counts of deaths only exceeding thresholds in the week beginning 30 March. While the significance of potential excess deaths for a single point in time (e.g. a single week) should be considered carefully, the timing of this particular period of potential excess coincided with both the peak in first wave COVID-19 infections and the commencement of lock-down measures. It also followed a prolonged period where the number of deaths were above expectation, though not significantly. 

Analysis of cause-specific mortality is a key strength of this research. Understanding the impact of deaths from COVID-19 itself is important when seeking to understand patterns of mortality during the pandemic. However, patterns of mortality for other causes of death can follow different trajectories as a result of both the pandemic and the public health measures put in place to prevent spread of the virus. While some causes of death remained relatively stable during the pandemic (e.g. cancer) other causes have a changed pattern of death in 2020. 

Excess mortality was recorded for deaths due to pneumonia (22 deaths), diabetes (38 deaths) and dementia (10 deaths). While these are small numbers of excess deaths compared with other countries, increases in these conditions during the pandemic aligns with findings from many international studies. Information available to the ABS on these deaths is limited to that provided on the Medical Certificate of Cause of Death, so no further information on reasons for these excess deaths can be provided. However, international studies have cited potential misdiagnosis of COVID-19, changes in access to healthcare and social isolation as possible reasons.

Significantly lower than expected numbers of deaths were recorded from the end of May to mid-July 2020. Lower numbers of deaths from respiratory conditions including influenza, pneumonia and chronic respiratory conditions during that period were the main drivers of low numbers overall. Low numbers of influenza infections have also been recorded through influenza surveillance data. This period of lower than expected deaths provides some insight into the impact of measures put in place to fight the pandemic.

The model is also useful for looking at excess mortality in non-pandemic periods. Excess mortality has been experienced in Australia during past years, especially during the winter months. While this is largely linked to the severity of the influenza season, the model also demonstrated that excess deaths are seen in other causes of death including dementia, diabetes, ischaemic heart disease and chronic respiratory conditions when infectious agents have a high incidence in the community.

Limitations of data

There are a number of considerations that need to be taken into account when considering the utility of this research and conclusions that can be made.

The analysis has been conducted on provisional data for doctor certified deaths only. Each year just over 10% of deaths in Australia are certified by a coroner. These include reportable deaths including suicides, drug overdoses and assaults which can provide important insights into mental health and social well-being. This work will be completed again when a full dataset is available for analysis.

As the purpose was to provide insights into whether Australia has experienced excess mortality during the COVID-19 pandemic the model has only been applied at the national level for all causes and selected diseases. In light of the second wave of COVID-19 infections and associated mortality in Victoria this work will be replicated with a focus on excess mortality in states and territories. There may also be key differences between other demographics including age groups and sex. These areas will also benefit from closer analysis in future to understand how different populations in the community were affected during the COVID-19 pandemic.

There are many different methods for measuring excess mortality and countries must select the methods which best suit their needs. Applying an adjusted version of the Serfling model is suitable in the Australian setting for a number of reasons. The Serfling model has been previously applied to death registration data sourced from the Registry of Births Deaths and Marriages in NSW to track excess deaths from influenza. Death registration data is also used to compile the national mortality dataset. Additionally, the model applies a robust regression which adjusts for random outliers in the data. As the model was applied to a number of diseases with differing seasonal behaviours it was able to control for random variation when projecting expected numbers of death. Applying a different model may produce different results and this should be taken into account when comparing outcomes with other sources or international figures. Insights into the behaviour of specific diseases cannot be inferred from this model.

The actual calculation of the number of excess deaths produced by the model can differ. Excess deaths can be calculated by taking the expected number of deaths from the observed number. It can also be calculated by taking the the upper threshold of the expected number of deaths from the observed number. Given one of the key aims of this work was to determine the significance of any changes in mortality during the COVID-19 pandemic, the latter method has been used in this body of work. Methods will be reviewed when a full dataset for 2020 is available. 

Diseases and conditions have seasonal patterns based on organic elements of the disease but can also be dependent on social and health events. The model purposely applies a constant threshold to all diseases so the results can be observed in the context of COVID-19. The regression is not a prediction of the future number of deaths, but instead provides an expectation of mortality based on past patterns and numbers. Previous years that have had unusually high or low mortality from the norm may affect the final regression and its projection. As such some of the diseases will better fit the projection than others.  Individual models to address the unique qualities of diseases and conditions would need to be applied in order to make inferences on disease behaviour. This does not affect the perspective of the final results but should be kept in mind when observing and interpreting data presented throughout the report.   

Future directions

The overall aim of this research was to implement a framework that could provide a measure of excess mortality in Australia during the COVID-19 pandemic while considering key points such as natural variation in patterns of death, population changes and seasonality of diseases. The adjusted Serfling model was able to do this and proved to be suitable for identifying excess mortality outside of the pandemic period.

There are a number of key demographics that should be explored further including breakdown by age and sex and analysis of excess mortality in States and Territories.

The ABS will work with interested parties in research, policy and planning to identify sustainable ways of continuing this type of analysis into the future.

Acknowledgements

The ABS would like to thank Centre for Epidemiology and Evidence (CEE) and New South Wales Health acknowledging their openness and support in sharing information about the Influenza Surveillance System and associated methodologies. Special thanks goes to Dr David Muscatello from the University of New South Wales for providing his expertise and reviewing the adapted methodology applied by the ABS. The ABS would also like to thank the Australian Institute of Health and Welfare for peer reviewing this report. 

Methodology

The analysis of 2020 mortality data undertaken by the ABS is based on a model developed by Serfling (10) and later adapted by the US Center for Disease Control (CDC) and New South Wales Health (NSW Health). This section provides an overview of the how the model has been developed over time, key aspects of the model, and how the model has been adapted and applied by the ABS in this analysis.

Historical development of model

Historically this model has been used to estimate excess mortality caused by influenza.

  • As early as 1932, Collins determined that excess mortality during winter months in the United States of America (USA) was a consequence of epidemic influenza and therefore could be used as an indicator for the recognition of influenza outbreaks.²
  • In 1963, Serfling described a regression model, based on the seasonal pattern of pneumonia and influenza (P&I) deaths, to infer excess deaths due to influenza.
  • Since then, models based on Serfling's has been applied in a number of temperate countries (USA,³,⁴ France,⁴ Australia,⁴ and Italy⁵) to demonstrate that excess mortality occurring during winter months is associated with pandemic and seasonal epidemics of influenza.
  • The CDC took the approach further and established the 122 Cities Surveillance System to provide timely, prospective information on excess mortality due to influenza. ⁶,⁷
  • The New South Wales Influenza Surveillance Program runs from May to September each year, and uses an adaptation of the Serfling model to monitor for influenza epidemics. This system primarily monitors the proportion of influenza positive specimens among all respiratory specimens from the major public health laboratories, the proportion of emergency department visits diagnosed with influenza, and outbreaks of influenza reported to Public Health Units by residential care facilities.⁹ 

Aspects of the Serfling model and its implementations

The Serfling model uses historical data to predict current patterns of mortality. This then enables the identification of any deviation from that prediction that may signal an influenza epidemic. In developing this model, Serfling recognised that past influenza epidemics could overly influence the regression, limiting the ability of the model to identify any new epidemic. For this reason, Serfling manually excluded past epidemics when fitting the model. 

An important feature of this methodology is the identification of a threshold that can be used as the upper and/or lower limit for the data. Observations that lie between these upper and lower limits are considered to be normal. The chosen threshold will vary depending on the desired specificity of the analysis and the methods used for calculation. In statistics there are default or familiar thresholds used in measuring these boundaries such as 1.96 standard errors either side of the mid-point.

An upper threshold was chosen to define the upper limit of the expected number of deaths in the absence of an influenza epidemic. This upper threshold was calculated as the predicted number of deaths in a given week plus a constant multiple of the standard error of the time series (the differences between each value predicted by the model and the actual observed values, or the 'model residuals'). Serfling chose the constant multiple to be 1.64 standard errors, and considered 2 consecutive weeks above the threshold to indicate epidemic behaviour.¹⁰

The NSW Health model used for the Influenza Surveillance Program is based on the Serfling model. This variant of the model fits a cyclic linear regression to data for the previous 5 calendar years and then forecasts the current year's time series from data up to the end of the previous year. Receiver operating characteristic curves are used to identify the most useful balance between sensitivity (the true positive rate) and specificity (=1-false positive rate) when determining the threshold that helps identify the start of an Influenza epidemic.¹⁰

Data source

Refer to the methodology section of Provisional Mortality Statistics for a description of the scope of data used in this model.

ABS adaptation of Model

The ABS have applied key characteristics of the model used in the New South Wales Influenza Surveillance Program. The ABS model is also based on the application of a cyclical linear regression model to the time series of weekly numbers of deaths. The model has been applied to doctor certified deaths from all causes and to deaths from specific causes. Most diseases covered in both the Provisional Mortality Statistics reports and in this analysis show some cyclical pattern in numbers of deaths, indicating that the model should prove suitable for this broader application. 

The approach used by the ABS involves use of a 'robust' estimation procedure for fitting the model. Epidemics can cause increased mortality which leads to outliers in time series data. By down-weighting outliers it aims to fit the model to patterns of mortality that would have been expected to have occurred in the absence of an epidemic. This robust regression down-weights the influence of extreme observations (outliers) and is applied to a 5-year baseline time series. This baseline time series runs from 5 January 2015 to 5 January 2020 (weeks commencing on a Monday). The regression is then used to forecast expected numbers of deaths for the current year.

The cyclical regression model includes: a linear time term, t, with values 1, 2, 3, ... for each week of the time series, and the square of the time term, t2, to accommodate long-term linear and curvilinear changes in the background proportion of the cause of death arising from factors such as population growth or improved disease prevention or treatment.

Also included are annual seasonal harmonic variables to describe the cyclical seasonal background pattern. The harmonic variables are functions of the week number, t, and the periodicity in the same units – in this case, yearly (52.18 weeks). The 2 harmonic variables in this case are: sine(2π t/52.18) and cosine(2π t/52.18).

The final model was:

Expected(proportion) = A + Bt + Ct2 +D sine(2πt/52.18) + E cosine(2πt/52.18)

where A, B, C, D, and E are the coefficients calculated from the regression.

To evaluate the approach, the model was fitted using PROC ROBUSTREG in SAS Software with the simplest, default 'M estimation' method.¹² The PROC SCORE procedure was then used to forecast values for 2020.¹³ The standard error and threshold is derived from the stdi option in PROC ROBUSTREG which is run a second time with the 2020 results attached.

Threshold identification

Identifying the threshold at which to signal excess mortality will vary depending on the goals of the research, application of the method, seasonality of diseases of interest. Several national statistical organisations are using different models including Farrington surveillance algorithms, z -scores and prediction intervals to inform such thresholds.

While the ABS closely followed the NSW Health methodology for calculating the regression, a different method was chosen for calculating thresholds. This was because the ABS analysis covers deaths from several causes, rather than focusing specifically on influenza and pneumonia. While many of those diseases follow some cyclical pattern of mortality, they may not have a specific season in the same way as influenza and pneumonia. Supported systems developed by NSW Health to clarify the accuracy of results of the surveillance for influenza and pneumonia are also not available for analysis of deaths from other causes.

As a result, a method for calculating the threshold for this analysis was selected which had a broader application and was more easily interpreted across a range of conditions, namely 1.96 standard errors or a 95% confidence interval. Use of the 95% confidence interval does introduce the expectation that random weekly variation could lead to counts that exceed thresholds 5% of the time. As such, individual weeks that exceed thresholds should be interpreted with caution. However, any prolonged period of weeks, 2 or more, exceeding the expected range can suggest excess mortality has been signalled.

The SAS function PROC ROBUSTREG offers the STDI parameter to calculate the 'standard error of prediction' of the model.¹² Which is used to define 1 standard error. The standard error of prediction is a more logical choice for assigning the threshold of excess mortality than the root mean square error (standard error of the model residuals) because it incorporates not only the variance of the residuals but also the variance of the model parameter estimates. This provides an estimate of the expected variability of the observed values in the absence of an epidemic.

References

1. CDC. Excess Deaths Associated with COVID-19 https://www.cdc.gov/nchs/nvss/vsrr/covid19/excess_deaths.htm#techNotes

2. Collins SD. Excess mortality from causes other than influenza and pneumonia during influenza epidemics. Public Health Rep 1932;47:2159–2180.

3. Reichert T, Simonsen L, Sharma A, Pardo S, Fedson D, Miller M. Influenza and the winter increase in mortality in the United States, 1959–1999. Am J Epidemiol 2004;160:492–502.

4. Viboud C, Boelle P, Pakdaman K, Carrat F, Valleron A, Flahault A. Influenza epidemics in the United States, France, and Australia, 1972–1997. Emerg Infect Dis 2004;10:32–39.

5. Rizzo C, Bella A, Viboud C, Simonsen L, Miller M, Rota M, et al. Trends for influenza-related deaths during pandemic and epidemic seasons, Italy, 1969–2001. Emerg Infect Dis 2007;5:694–699.

6. Centers for Disease Control and Prevention. 122 Cities Mortality Reporting System, Manual of Procedures. Atlanta (Georgia, United States): US Department of Health and Human Services; 2004.

7. Simonsen L, Clarke MJ, Stroup DF, Williamson GD, Arden NH, Cox NJ. A method for timely assessment of influenza-associated mortality in the United States. Epidemiology 1997;8:90–395.

8. O'Brien K, Barr IG. Annual Report of the National Influenza Surveillance Scheme, 2006. Commun Dis Intell 2007;31:167–179.

9. NSW Health, Communicable Diseases Branch. New South Wales Influenza Surveillance Report. NSW Health  Available from: http://www.health.nsw.gov.au/PublicHealth/Infectious/ Accessed on 11 September 2008.

10. Serfling RE. Methods for current statistical analysis of excess pneumonia-influenza deaths. Public Health Rep 1963;78:494–506.

11. The ROBUSTREG procedure. SAS/STAT(R) 12.3 User's Guide. Cary (USA): SAS Institute, 2013. Available from: https://support.sas.com/documentation/onlinedoc/stat/123/rreg.pdf

12. Model Fit and Diagnostic Statistics. SAS/STAT(R) 9.2 User's Guide. Cary (USA): SAS Institute, 2008. Available from: http://support.sas.com/documentation/cdl/en/statug/59654/HTML/default/statug_reg_sect039.htm