The CMI’s approach to the use of 2020/2021 data
This blog sets out the CMI’s intentions regarding the use of 2020 and 2021 data in forthcoming investigations, in particular regarding pensioner, annuitant and life assurance mortality. We comment also on the similar but somewhat different question of income protection experience investigations and the use of population data for mortality projections.
The conclusions of this note follow discussion within the CMI in the COVID-19 workstream, as well as amongst the Chairs of the various experience investigation Committees, and subsequent discussion in the Executive Committee. We have also had regard to recent conversations on the issues with key Subscribers, as well as responses to the recent CMI Subscriber survey.
Unlike our other blogs, this does not describe a recently published CMI Working Paper. The high-level ‘direction of travel’ outlined here did not seem to warrant a Working Paper in its own right. Readers interested in more of the background to the issues discussed below should refer to “Considerations relating to COVID-19 for mortality and morbidity assumptions”, produced by the COVID-19 Working Party (Chaired by Steve Bale) and released in October 2020.
The year 2020 was an extraordinary (and we hope unique) year for mortality because of the pandemic. In the UK there were approximately 73,000 excess deaths above that expected based on mortality in 2019 (the CMI’s excess deaths measure). We are now most of the way through 2021, and deaths from COVID-19 are still a significant number (67,000 to 15 October 2021, based on ONS information on deaths with COVID-19 listed on the death certificate). However, the mortality experience of 2021 is unusual owing to factors beyond ‘just’ the COVID-19 deaths: in particular, we have increased other-cause (non-COVID) deaths arising from delayed diagnoses and treatment due to the lockdowns, and a reduction in deaths in respect of those people who died from COVID-19 in 2020 who might otherwise have been expected to die in 2021 (the ‘forward displacement’ effect).
This poses a challenging question to the CMI: to what extent can mortality data from 2020 and 2021 be of use? Exactly the same question faces actuaries working in life insurance and pensions.
The CMI’s use of experience data
The experience investigations we carry out in the CMI fall into broadly two types:
- Actual versus Expected, where we assess how the experience of a year or group of years compares with what would be expected based on the most appropriate tables.
The CMI will carry on doing this type of analysis on 2020 and 2021 data. This will help subscribers to benchmark how their own experience compares with the data provided by Subscribers to the CMI. Also, in doing this type of analysis, it may be useful to look at (for instance) the observed 2020 experience against what we might have expected based on prior years adjusted using a mortality improvement assumption. This corresponds to ‘option 3’ in the data adjustment section of Working Paper 139.
- Development of new mortality tables, from analysis of the probability of death at any age followed by smoothing across the age range (‘graduation’, primarily to remove noise) and extension to younger and older ages (where the CMI’s dataset lacks credibility).
This work aims to derive mortality tables that are predictive of future experience. Clearly, deriving tables based partially on unadjusted 2020 and/or 2021 CMI data is unlikely to be predictive. However, we have not found a satisfactory way to adjust 2020 or 2021 CMI data for this purpose, as we discuss below. Therefore, as a general tenet, the CMI is not intending to develop new mortality tables using data from 2020 or 2021. This work would not be a useful application of our Subscribers’ fees and tables produced this way are likely to be misrepresentative.
Possible approaches to adjusting the 2020 and 2021 experience
We have spent considerable time trying to answer the fundamental question of whether we can remove the pandemic’s effect from the 2020 and 2021 mortality data. In other words, what would mortality have been absent the pandemic? (This is clearly a different question from ‘what will mortality be post-pandemic?’, and we comment on that further below.)
We have considered two approaches – a 'bottom up' approach using data on deaths directly attributable to COVID-19, and more of a 'top-down' approach looking at ‘excess deaths’ (deaths above those expected, and hence likely attributable to the pandemic).
Bottom-up approach – adjusting the data using COVID-19 mortality
A 'bottom-up approach' could work using ONS data on deaths with COVID-19 listed on the death certificate, UKHSA (formerly PHE) data on deaths within 28 days of a positive COVID-19 test, or in some cases possibly information provided by CMI data contributors on the cause of death/claim; perhaps these could be combined in some way to enhance accuracy. Although this sounds like a good start, there are several areas of difficulty:
- Testing for COVID-19 in the first wave of the pandemic was not as extensive and as established as during the second wave (and beyond) of the pandemic and so some COVID-19 deaths are likely to have been “missed” and assigned as other causes of death. These missed COVID-19 deaths would need to be estimated.
- We would need to calculate from high-level public domain data an appropriate COVID-19 mortality age curve.
- For the CMI’s work on life insurance and pension experience, where the socio-economic profile differs from that of the general population, we would then need to allow for how COVID-19 affects these different ‘insured’ lives. Although we have started to receive basic socio-economic markers (Index of Multiple Deprivation and “NUTS 1” region for a subset of the main datasets), we do not have enough risk factor information that would enable us to make an appropriate adjustment with confidence.
- We would need to calculate an ‘amounts weighted’ equivalent of the above (because our analyses generally provide results weighted by the amount of insurance or pension benefit per life).
Each of these steps involves substantial subjectivity and potential room for error; the combination of these steps would likely lead to results that would be of little use, other than perhaps indicating a reasonable range.
A particular concern with the above approach is that, while it looks just about plausible in dealing with 2020 (albeit with a large degree of uncertainty), the approach would be of no use in 2021 because the other elements that make 2021 an abnormal year (for instance, forward displacement and delayed diagnoses) would not be appropriately allowed for.
If the CMI were to adopt a method to allow for abnormal mortality, we would want this to be effective in both years (and possibly even 2022). The degree of uncertainty inherent in the above approach, and its inapplicability to 2021, therefore make us reluctant to pursue it at this stage.
Top-down approach – adjusting the data using excess mortality
A 'top-down approach' would seek to define deaths caused by the pandemic as the difference between actual deaths, and those that would otherwise have been expected: this difference is the ‘excess’. This has been a very useful approach for quantifying the pandemic’s overall mortality impact for the purpose of the CMI’s regular .
However, this approach is not so well-suited to the question here of adjusting 2020 and 2021 data, and then going on to conduct subsequent analyses of that adjusted data. The reason is that the ‘actual less expected’ method is sensitive to what we define ‘expected mortality’ to be. This sensitivity has not been a material issue for our pandemic mortality monitoring, where the ‘pandemic deaths’ calculated do not vary greatly on slightly different assumptions of expected deaths.
For an attempted analysis of ‘non-pandemic deaths’, this issue is significant. In simple terms, we would be quantifying ‘non-pandemic deaths’ as (Actual deaths less Excess deaths), where Excess deaths are themselves defined as (Actual deaths less expected deaths). In a circular fashion, we, therefore, end up calculating non-pandemic deaths as expected deaths.
This means we are not bringing into our analysis any information on actual 2020 or 2021 mortality: we have simply brought in a prior expectation through the back door. For this reason, the top-down approach is of no use in adjusting data to arrive at an idea of what 2020 (or 2021) mortality has been ‘absent the pandemic’.
What does all this mean for the next few years, and when will we start to report our findings on post-pandemic mortality? Much depends on the experience of 2022 and 2023. If they are comparable to the years preceding the pandemic, it may be feasible to produce tables based on the years (for instance) 2018, 2019, 2022 and 2023. However, this may not be a viable option if post-pandemic mortality (i.e. from 2022 onwards) shows a ‘step-change’ difference from pre-pandemic mortality, or there is some continued mortality displacement.
It may be that the first consecutive four-year period of broadly similar mortality that we are able to use for developing tables is the period 2022-2025, in which case the underlying work would not be done until 2027 at the earliest. The individual investigation Committees will of course be analysing their data for 2022 and onwards to understand the various changes emerging and will report to Subscribers on their findings and any proposed changes to their analytical approach.
The above comments have dealt largely with experience investigations of pensions and life insurance mortality, but we should note how the issues may differ in other contexts.
For the Mortality Projections Model, there are similar concerns about the unrepresentative nature of 2020 and 2021. In addition, the Model does not react well to abnormally negative mortality improvements. The Mortality Projections Committee has therefore needed to modify the method used in CMI_2020 and CMI_2021. For CMI_2020, “weights” were introduced into the Model, with nil weight being the default option for 2020 (in other words, the Model excludes the data for 2020 during its fitting process), and full weight placed on data for other years. The current intention is to take a similar approach for CMI_2021, placing nil weight on data for both 2020 and 2021.
The Income Protection Committee is currently in the process of collecting data for 2020. At this stage, it is unclear what impact the pandemic will have had on income protection experience – experience may have increased due to claims made as a result of COVID-19, or reduced as a result of furlough and working from home. The Committee recently surveyed Subscribers to understand, qualitatively, what impact the pandemic has had on claims. This analysis will be published in an upcoming working paper.
We hope this has been useful in outlining the CMI’s direction of travel concerning the use of 2020 and 2021 data. We are in a very strange period of mortality experience but we will do the most we can for our Subscribers with the available data – just as we have done throughout the pandemic. Through the ‘Actual versus Expected’ work described above, we should have some view of what post-pandemic mortality is by 2024.
If you’d like to provide any feedback on this blog, or on the CMI’s COVID-19 activities more generally, please email your comments to .