What does the latest CMI Mortality Projections Model tell us?

By Andrew Gaches for

COVID-19 Actuaries Response Group – Learn. Share. Educate. Influence.

Summary

The CMI’s latest mortality projections model CMI_2020, and accompanying working paper, show that users adopting the default parameterisations may observe around a 4 week decrease in projected life expectancy for a 65 year old man (compared to last year’s model), and around a 1 week decrease for a 65 year old woman.  It should be noted that these impacts do not include any Covid-19 related effects (as they give no weight to the highly unusual 2020 mortality data, and do not allow for any change in the longer-term outlook).

In order to incorporate the likely impact of Covid-19 on future mortality evolution, users of the model will need to form their own view on the longer term impact of Covid-19, and use the parameters available in the model to capture those impacts.

Impact on life expectancy

The CMI recently published its latest annual mortality projections model (CMI_2020).  These models are widely used by actuaries to express how mortality could evolve in the future.  Alongside the latest version of the model, the CMI has released a working paper covering, amongst other things, the impact of moving from the previous version of the model (CMI_2019) to CMI_2020.

The working paper illustrates that adopting CMI_2020 in place of CMI_2019 (and assuming default parameters are used in each case) could result in projected life expectancies from age 65 falling by around 4 weeks for men and 1 week for women.  Before we focus too much on those statistics, we need to consider what they really mean.

What does the 4 week decrease in life expectancy for men represent?

First, it’s worth being clear that the decrease in illustrated life expectancy is nothing to do with Covid-19.  Covid-19 had a massive impact on mortality in 2020, to the extent that reflecting 2020 data in CMI_2020 in the usual way would fundamentally distort the projections, and so the CMI_2020 model (using default parameters) places no weight on 2020 data.

Why then is there any change in life expectancy in the new model?  On the face of it, if there’s no new data (being used by the model) why would the outputs change?

A key driver of this impact is the concept of the transition from short term improvements (typically low, reflecting the low improvements observed in the latter half of the 2010s) to long term improvements (typically higher).  As usual, the latest mortality projections model effectively pushes out the long term by 1 year, so delaying that transition to higher rates of improvement by a year; and it’s this shift that is a driver of the decrease in life expectancy projected by CMI_2020.

So one way to conceptualise the default parameterisation of CMI_2020 (vs the previous version of the model) is (broadly!) “make no allowance for the exceptional mortality seen in 2020, but shift out the move to (typically) higher long term improvements by 1 year”.

Using the flexibility of the CMI_2020 model

Above, it’s only the default parameterisation of CMI_2020 that is discussed.  And “default” does not mean “appropriate” for any given user; nor is it intended to be in any way “recommended” by the CMI.  The real value of the mortality projections model is that it provides a common framework within which users can express their views, through the use of the various parameters available in the model.

The latest version of the model introduces additional flexibility, specifically to address the issue of the exceptional 2020 experience being unrepresentative of underlying trends.  In particular, it provides a mechanism to specify the weighting given to data in each year enabling users to specify, for example, anything from 0% weight to 2020 data (ie the CMI_2020 default) to 100% weight (ie treating 2020 as a “normal” year).

Building on these flexibilities, there are several ways in which users who currently use the CMI_2019 model could evolve their trend assumptions including:

• Retain CMI_2019 model (and existing parameterisation). This approach effectively reflects a view that 2020 data tells us nothing new about how mortality will evolve, and that the user’s view on future trends is therefore unchanged.  It may seem very odd to conclude that the many impacts of Covid-19 have nil net impact on mortality trends, hence this approach may be more of a holding assumption, adopted while the longer term impacts of Covid-19 are being assessed.
• Adopt CMI_2020 model (with existing parameterisation, and no weight to 2020 data). As noted above, this approach could be consistent with a view that the 2020 data is not helpful in updating views on mortality trends, but that the 2020 experience has (in broad terms) pushed out the transition to higher long term improvements by a year.
• Adopt i) CMI_2020 model with some weight to 2020 data; or ii) CMI_2020 model with some other amendment (eg use of A parameter); or iii) retain existing CMI model but amend in some way (eg using the A parameter). These are all more sophisticated updates to trend assumptions (and are a non-exhaustive list of possible approaches) which may seek to reflect a user’s view of the short-medium term future impact of Covid-19.  These may be combined with change in long term improvement rate, if the user views long term prospects for mortality improvements to have changed.  The key challenge in implementing approaches of this kind is the user will need to form a view on the impact of Covid-19 on future mortality trends outside the mortality projection model, and then set the parameters of the model to (broadly) reflect those views.

As noted above, there are a wide range of possible approaches to updating mortality projection assumptions, ranging from the simplistic “no amendment as we cannot yet form a robust view” to much more sophisticated updates, reflecting a more comprehensive review (and quantification) of the ways in which Covid-19 is likely to change the evolution of mortality rates.

These Covid-19 related drivers of mortality were classified and outlined in our earlier bulletin on the longer-term mortality and morbidity impacts of the pandemic.  In practice we may expect many users will have adopted approaches akin to the former over the course of 2020, but will move to more sophisticated assessments of the longer term impacts during 2021.

4 March 2021