Death Rates for COVID-19: Key Issues, Published Data and What it Tells Us

In this article, a group of life insurance actuaries interrogate published morbidity data from around the world. They identify what the key issues are, what actuaries should be looking out for when using the data and includes a list of useful data sources. 

Executive Summary 

  1. What’s Known

A number of territories across the world are publishing real-time data on COVID-19 deaths, including by age, gender and comorbidities. Rates of infection are also available. New York State in particular published deaths by age and comorbidity. This information, combined with data on general populations overall, provides evidence to consider some of the hypotheses about COVID-19.

  1. What’s New

Comorbidities

The most striking comorbidity relationship identified in the data is diabetics aged 20-39 that appear at a rate 10 times higher among COVID-19 fatalities than in the general population. This analysis is based on New York State Health data which indicated 26% of those aged 20-39 dying from COVID-19 were diabetic, compared with the level of diabetes in the general population age 20-39 of less than 3%.

This seems consistent with comorbidity risks between diabetes and COVID-19 which were also highlighted in a Guardian article noting NHS data indicating a quarter of those dying from COVID-19 in England were diabetic.

For significant illnesses other than diabetes and stroke, the analysis did not provide such clear evidence of association with COVID-19 fatality. However, one reason for this may be greater efforts to protect those with significant illness from infection in the first place (i.e. lower rates of infection amongst those with significant illness than the general population).

Gender

Excess of male COVID-19 mortality has been widely commented upon, and while there is some evidence of this particularly in earlier age groups, the data seems to suggest that it is primarily a reflection of the underlying excess of population male mortality.

Infections and Age-Based Mortality

The analysis for all nine territories considered support the view that fatality is significantly related to increased age. It is quite possible that differences between territories are largely due to relative levels of under-reporting. For example territories with high levels of positive rates in testing (a proxy for numbers the level of under-reporting of actual cases) generally have higher case fatality rates than those with lower levels of positive testing.

The analysis of four territories indicated lower rates of infection in ages 10-19 in their population (than their population overall) supporting the hypotheses that children do not seem to be as vulnerable to infection. This seems particularly clear when contrasting the infection for this age group with the next age band up 20-29 which actually have a higher level of infection (for Australia and Korea at least) than the average of the population.

  1. Key Assumptions / Data Sources

We support the call of many medical organisations and others studying the virus to increase the availability and granularity of data. To support further research we have published all of the websites used for the analysis in this article.

Introduction

Age and co-morbidities have been constantly mentioned by health professionals, administrators and politicians as significant drivers of fatality when infected by COVID-19.

This blog considers some of the publicly available data that may assist and inform views about potential relationships. It also briefly discusses some other risk factors that have been hypothesised but that we did not have time, or could not find the data, to explore.

The authors of this blog are Australian life insurance actuaries who are supporting the Actuaries Institute’s work on COVID-19, specifically on mortality. It is noted that work is being undertaken by the Institute on other aspects, such as morbidity, which may have potentially greater significance to life insurers in Australia.

We have reached out to actuaries overseas for feedback on the analysis presented in this blog and we welcome feedback from others (and not just actuaries!).

We would like to note that this blog seeks to provide information and perspectives that the authors consider may be informative, but none of it is perfect – for example in some cases we have had to compare data from different sources that may apply or record information differently – there may be better approaches – and available data will change, improve and increase over time.

We support the call of many medical organisations and others studying the virus to increase the availability and granularity of data.

Mortality by Age

Around the world there are now a number of sites that are publishing various data on COVID-19 deaths and infections. For those who are interested in a specific territory, a useful starting point is to search “<territory> health department”.

All these sites carry warnings about the accuracy of data, potential for change, etc. Nonetheless, we consider there is value in assembling this data and beginning to analyse it for signals – enabling further focus and potentially work to address shortcomings and identify with further confidence, relationships and information about the virus and its effects.

In relation to data quality/interpretation, there are a number of (sometimes known) biases and issues. Actuaries can adjust for these as they see fit – generally in this paper we use raw information and try to adjust as little as possible – leaving it for the reader to make their own adjustments as necessary. In some cases we mention or discuss known bias issues and how they might affect the results. However, we may not have identified all issues with the data (buyer beware!).

The best-known example is the both symptomatic and asymptomatic cases missing in the infection case numbers. These vary by country and over time. For example, in the US, with some states having up to a 40% positive test rate, there is likely to be a lot of (even symptomatic) cases not reported.  Australia has less than 1% positives, so would be expected to be missing less – although at the height of our case reporting/identification we suspect we still would have been missing cases.

Another example is deaths in the UK where, until about 1 May, deaths in nursing homes were not reported, and even now, UK authorities advise reporting may not be complete. In Australia, we understand reporting of deaths is reasonably comprehensive. The following countries/regions (hereafter referred to as territories) have published aged-based data for deaths and reported infections (or at least age based Case Fatality Rates (CFRs) i.e. death rates per reported infection). The data in this table is the basis for much of the analysis in this blog.

The territories are listed below from lowest to highest COVID-19 mortality, based on a measure that we call Standardised Equivalent Australian Case Fatality Rate (SEACFR). The SEACFR is the weighted average of a territory’s age-based CFRs, with the weights based on Australia’s age-based reported infection profile i.e. it is indicative of a territory’s overall population CFR if its age profile for infections was similar to Australia.

CFRs by age for each region are set out in the table and graphs below. The tables/graphs contain the same information with the following points noted.

  • The graphs show the median CFR which is calculated separately for each age-group. For the younger ages the median calculation does not include countries which have a zero CFR.
  • The second chart below shows ages 0-49. This was done purely because values/differences between countries are difficult to see on the chart that show all ages.
  • Texas had total deaths of 816 but only 372 had been classified by age. Similarly, it had 29,229 reported cases with only 7,480 classified by age. The age-based deaths and infections were separately grossed up to the known totals before calculating CFRs. We acknowledge that this is a significant approximation but we felt that it was worth including. Readers should be aware of this limitation and treat the Texas numbers with particular caution.

 

Critical in interpreting this information is understanding countries’ ability to identify and report infections (and to a lesser extent – e.g. the UK – deaths).

* Tests data from 4 May

As can be seen from the table, where testing is at some basic level, the emerging CFRs are strongly correlated to the discovery rate of positives in the testing, but clearly with other factors involved.

And although very, very broad-brush, it seems to us that allowing for these differences can potentially account for a portion of the very different apparent CFRs between territories.

For example, Connecticut is a high positive test territory at 28.6% and has a high CFR of 8.2% (SEACFR also high at 4.80%). We think that the high rate of positive testing may be one indicator of a higher level of “real” cases not identified and “missing” in the “reported” infected numbers. Therefore if/as testing increases, and all other things being equal, the CFRs of such territories may have potential to reduce somewhat.

It’s clear from the chart that the differences in aged-based rates between territories are not so extreme for the 80+ age bracket and are midway for the 70-79 age bracket. Our hypothesis regarding this is that all territories are testing their older populations more comprehensively resulting in less skew in these age categories than in the younger ages, resulting in more uniform CFRs.

Given the discussion above, one possible approach that might be worth actuaries considering is:

  • For the younger ages, up to say age 49, because of the very low numbers of deaths in the more credible territories, a credibility-based blend of the low-positive rate territories such as Australia, S. Korea and Texas, with a blend of the remaining available universe of results. Possibly with some transformation of the results to a standardised positives rate of, say, 2%.
  • For older ages, deprecate the rates from territories with higher positive rates and focus on the data-based rates in the more credible territories.

 

This suggested approach is based on the following observations:

  • Very broadly, for the 80+ age bracket, this approach is the same as any “averaging” approach and there are sufficient Australian 80+ deaths to give credibility, plus we know enough about infection and outcome at these ages to be able to make informed judgements about the data.
  • For 50~79, it is a matter of some judgement as the credible territory rates tend to look high on this basis. The key question in forming a judgement is whether younger age deaths in less-credible territories are adequately recorded, given that those territories are highly focused on older age deaths.

 

To illustrate what we believe are the potential distortions of incomplete testing regimes, the following chart presents index infection rates* (see the notes to the chart) for two “higher tests, lower positives” territories and two “lower tests, higher positives” territories.

Co-Morbidities

As noted in the introduction, the potential for other significant illnesses to be a driver of fatality when infected by COVID-19 has been frequently raised.

Detailed statistical analysis of relationships remains limited based on what we have been able to observe on publicly available data and is highly clouded by co-mingling with other factors.

The most relevant data we have found is New York State age-based co-morbidities of COVID-19 deaths. These rates amongst COVID-19 deaths are compared to age-based rates of co-morbidities in the general population:

The last table demonstrates the significant differences in population level co-morbidities by age between Australia and New York. Actuaries should keep these differences in mind when considering overseas co-morbidity data. However, as will shortly be shown, the differences are least for the co-morbidities that appear to be most significant.

The following table gives ratios or multipliers to the presumed base level of COVID-19 mortality from the key co-morbidities. This tends to show unexpected, and often low, co-morbidity outcomes, most notably for COPD, given COVID-19 is a respiratory disease. There appear to be, however, two highly significant co-morbidities for younger people, as shown in the table, diabetes and stroke.

We consider that further developing the analysis – based on more data sources – may assist decision making and risk-assessments as Australia and other countries look to adopt a risk-based approach to relaxing restrictions. For example, while COVID-19 mortality is generally very low at younger ages, special care may be needed to protect certain individuals (even if young), such as those with diabetes, (or other particular diseases) which may have dramatically higher vulnerability to risks associated with COVID-19 infection.

Actuaries should note that this analysis has been based on the known co-morbidities of the overall population, not the unknown actual co-morbidities of infected persons.

The results for diabetes and stroke are highlighted in the graphs below.

The following table and graphs show apparent expected CFR with co-morbidity:

While the data does not identify clear relationships with other co-morbidities – this is not to say they do not exist – just that they are not as clear from the data.

What seems likely is that some of the early commentary may be attributing too much significance to some co-morbidities as a COVID-19 risk factor. This may have occurred because early analysis of COVID-19 deaths was not properly standardised for age. That is, the analysis did not adjust for the fact that older people who have higher death rates from COVID-19 also have higher co-morbidities.

Another way of saying this is that, because COVID-19 CFRs increase with age, high levels of co-morbidities are expected among COVID-19 deaths even if they are not a driver of COVID-19 mortality risk. It is not clear that early commentary appropriately took this into account.

In this context, it is worth noting the recent Imperial College Report 17, which looked for mortality predictors among those admitted to hospital with COVID-19 symptoms.  This report found that co-morbidities were not a significant predictor of mortality, after taking into account gender, age and condition on presentation to the hospital.

However, we also accept that it is difficult to draw any firm conclusions from the available data and that there may be different views. For example, one hypotheses from anecdotal evidence is that the higher levels of northern Italian deaths may be related to higher co-morbidities, age adjusted, than other countries.

We particularly note the following issues that have the potential to impact the interpretation of the data:

  • Co-morbidities in infected population – as noted, we were unable to source data that indicates co-morbidities (by age) in the infected population. An example might be cancer which amongst COVID-19 deaths which, for ages above 40, shows as lower than amongst the population. However, this may be an outcome of lower infection rates, because of less contact with others – hence it can’t be concluded that those with cancer have lower fatality from COVID-19.
  • Definitions and data recording – there are issues of active treatment vs remission and prior severity levels of a condition, data recording issues, etc.
  • Limited data – unlike the age-based mortality analysis, we had co-morbid age based COVID-19 mortality data for only one region (New York). A larger data set would improve confidence in the results. However, we have seen a number of aggregated sets of co-morbidity data and/or comments on co-morbid outcomes which have not been inconsistent with the New York results.

 

Gender

Excess of male COVID-19 mortality has been widely commented upon, but the data seems to suggest that it is primarily a reflection of the underlying excess of population male mortality – but there are nuances.

The chart and tables below show the excess mortality of males (male mortality/female mortality-minus 1) for two countries that provided gender-based mortality and infection rates (Australia and the UK). We hope to do a similar analysis on USA data in due course.

It compares this with the excess mortality for non-COVID-19 mortality (using three data sources; FSC 2004-08 which is an Australian insured lives table, ALT 2015-17 which is the Australian population life table and UK NLT 2016-18 which is a UK population life table). It is noted that the mortality rates for 80+ were based on those for a 90 year-old while those for the other age bands are for specific ages 14, 24, etc. It is noted that adopting a different approach (such as a calculating a weighted average of mortality rates at each individual age within the age bands) has limited impact on the male to female mortality rate relativities.

This chart compares UK COVID-19 mortality with UK National Life Tables

For Australia, where young age deaths are far too small to draw any conclusion, the ratios suggest that the gender impact of COVID-19 is consistent with that for overall population causes of death.

For UK data, COVID-19 seems to have higher male excess mortality in younger ages (up to 39). For older ages (60+), there is no obvious difference between COVID-19 and overall population mortality. At this time, we do not have good insight into the unusual result for ages 40-49 and 50-59 in the UK data. For actuaries working from the Australian FSC mortality tables, it is necessary to remain aware of the differences between this and broader population mortality, arsing principally from positive selection effects.

COVID-19 Mortality: Other Key Issues and Hypotheses / Risk Factors

In addition to the potential risk factors considered in this paper there are a number of others that have been hypothesised that we consider worth being aware of but that we have not had the time or found available data to consider. For some of these data/studies may eventually become available while others may always remain a supposition.

Other Potential Risk Factors

Viral Load

It appears that COVID-19 spreads equally through the four transmission mechanisms by which other coronaviruses spread i.e. contact, droplets, faecal-oral and aerosolisation (note, we understand that is somewhat unclear whether aerolisation is a transmission mechanism). It follows then that certain locations, such as hospitals, nursing homes, cruise ships and possibly abattoirs will be prone to higher rates of infection due to both longer periods of exposure to infection and probably higher viral loads during exposure.

This conclusion may be important when considering COVID-19 effects on certain work / leisure places and for certain occupations. For example workload and viral load has been hypothesised as a factor in the deaths of Italian medical staff.

Speed of Diagnosis, Quality of Treatment

A possible interpretation of some of the better outcomes across the globe, and consistent with medical commentary, is that early diagnosis and early robust treatment of severe cases lead to significantly better mortality (and morbidity) outcomes. Actuaries need to consider this aspect of any situation for which they are considering the impact of COVID-19.

We note this can be impacted by medical systems becoming overrun. For example under an adverse scenario a country/region which has “good” experience may have potential to deteriorate if infections increase dramatically. In Australia his may be relevant for actuaries considering adverse scenarios.

It seems that the Australian government’s policy (at least early on) was primarily focussed on trying to avoid this occurring – for example, draft analysis from the Doherty Institute included analysis of COIVD-19 infection scenarios and whether ICU, general wards, emergency departments, GP clinics etc were overrun.

Level of Vaccination and Level of Influenza/Pneumonia

A critical complication of COVID-19 in the severely ill is pneumonia. Medical authorities recommend getting vaccinated against pneumococcal disease to reduce bacterial complications if a patient contracts COVID-19. This in no way protects against COVID-19 or its complications, but it eliminates one other battle.

At this time there is no vaccine, no candidate vaccines and no effective medicinal treatments for COVID-19. There are clinical trials being undertaken for both antivirals (i.e. treatments which help ameliorate the effects of a virus) and positive treatments such as inhaled Heparin (an intravenous anticoagulant widely used in hospitals which might ameliorate COVID-19’s effect on the lungs).

Seasonal Factors

In the Southern Hemisphere there are concerns that COVID-19 will “become” much more aggressive in our winter. Influenza is a highly infectious disease, mildly affected by temperature and so becomes more aggressive in cooler times and climates – but the major seasonal influence is behavioural – more time indoors, closer proximity etc. This fact is highlighted by the plunge in seasonal influenza in Australia to date in our lockdown.

No one knows yet if COVID-19 is responsive to temperature or humidity. However, we can observe different infection and death rates in different countries across the globe.

The wide disparity of response to both the emerging outbreak in a locality and the emerged outbreak, both by government-mandated actions and by citizen-initiated actions makes it near impossible to draw meaningful conclusions from the available data.

However, it is our impression that COVID-19 is likely to be only mildly seasonal in terms of mortality impact. Its morbidity impact may be much larger. We make that impression from two observations:

  • While significant, in the context of overall mortality, seasonal influenza is only a minor increment to a nation’s overall month-by-month mortality.
  • The discernible variation of COVID-19 infectiousness, severity of morbidity and mortality in localities with very different climates, both temperature and humidity might be real, and material, but is clearly not bigger than the effects of adequate healthcare and broadscale infection control. Our best illustration of this is S.Korea which we have earlier used as a near analogue to Australia, yet S.Korea has a climate radically different to Australia’s, particularly during March and April – which would be a chilly winter for any Australian (March 0-10, April 7-17).

 

Ethnicity

Many of the territories which publish useful data also give ethnic breakdowns, sometimes in up to eight categories. This is not surprising, because these territories have significant ethnic mixes and ethnicity is often an important issue for health and social outcomes. Our working hypothesis is, for Australia, ethnicity is not material. As an example, we have seen data that suggests poorer outcomes for Asians in one territory and then the opposite outcome in another territory.

There are some clear ethnic patterns for disadvantaged groups, but the most likely issue is that it is their disadvantage, not their ethnicity that is relevant.

For any Actuary who needs to consider ethnicity effects of COVID-19, on either mortality or morbidity, we would strongly suggest they first consider occupational – both general class and specific industry – and then other socio-economic effects ahead of any consideration of ethnicity effects.

Different Strains?

The most up-to-date information indicates that, unlike influenza viruses, SARS-CoV-19 is more or less stable. Multiple “alleles” have emerged, but their behaviour and pathogenicity does not appear to be materially different from the progenitor.

A general behaviour of new viruses is to be more pathological at emergence and then slowly reduce.

Developing Countries

As noted above, one key aspect of improved mortality (and morbidity) outcomes at this time is the effectiveness of medical interventions. Any actuary considering the impact of COVID-19 outside Australia would need to adjust their assumptions to reflect their understanding of the health systems ability to respond, both in unstressed situations, and perhaps also in stressed / overloaded situations.

Delays to Medical Treatment Because of COVID-19 Fears, Availability, etc

Significant concerns are currently being raised by oncologists and cardiologists about what they consider to be a significant increase in failure to seek treatment following symptoms (for example the number of people presenting for examination has dramatically reduced). This includes patients with existing treatment plans, failing to follow them, attend/schedule appointments etc. There is concern this will lead to a “bulge” in deaths in later periods as conditions worsen because of the delay in receiving treatment.

This issue is simply noted here for readers who are interested in mortality impacts overall to be aware of.

Other Studies

This is the Imperial College study of hospital admissions, suggesting that co-morbidities were not a significant predictor of mortality, after taking into account gender, age and condition on presentation to the hospital here.

A list of publicly available sources is set out below:

Co-morbidities by age of COVID-19 deaths:

 

Co-morbidities by population statistics

 

 

Infection rates and CFR:

 

Cases, tests per population:

 

Life tables:

 

Population demographics:

CPD: Actuaries Institute Members can claim two CPD points for every hour of reading articles on Actuaries Digital.

Comments

Image of Aaron
Aaron says

15 May 2020

thanks for this fantastic and comprehensive quantification of risk. I was actually trying to do a far less comprehensive analysis of my own risk the other night (because why not).

Of course you start with a base rate, and then work on multipliers or dividers for risk and protective factors. One thing I noticed in your analysis that I hadn’t considered and really appreciated is your starting point of considering CFR by country rather than as a static factor that’s is common to the disease. Thanks for highlighting this main effect in the data.

The two things worth considering that may add (or hopefully subtract) from your model)

1/ orthogonality of risk factors. The multiplication beta weights assume that risks do not overlap. For instance someone with heart disease and hypertension may have an interaction effect where the sum of factors is greater than either on their own (eg: smoking and asbestos exposure on cancer risk), or where the risk of one is virtually factored into the risk of the other. I think some bigger data will be needed to tease this apart

2/ I havent gotten to the bottom of this yet in my reading, but the blood group hypothesis seems to be robust as a finding, even though most medical people are pretty clear there is no reason why it should be

Would love to see more of this work as it evolves. Can we have a “how likely to die of COVID-19 are you?” Calculator please :-)


Comment on the article (Be kind)

Your comment will be revised by the site if needed.