Andrew JG Cairns, David Blake, Kevin Dowd, Guy Coughlan, Owen Jones and Jeffrey Rowney share their findings on the mortality experience analysis for the Universities Superannuation Scheme
The Universities Superannuation Scheme (USS) is the UK’s largest funded pension scheme and has a highly educated and homogeneous membership. We analysed members’ mortality experience and compared it with English mortality subdivided into deprivation centiles using the Index of Multiple Deprivation (IMD) and with other developed countries. Here, we review our key findings and lessons, and consider wider implications.
Male and female USS pensioners’ mortality rates were found to be significantly lower than those of the least deprived (and longest-lived) national IMD deciles. Figure 1 shows the development of age-standardised mortality rates of males aged 65–89 for selected IMD centiles (the 100th centile being the least deprived) during 2001–16 and for USS primary males (male principal scheme members) during 2005–16. We can see that the USS male mortality rates are lower than the average of the 1% least deprived areas in England. The implication is that the USS has a different base mortality table from the English national population and from even the least deprived IMD centile. This finding should be helpful in valuing USS pension liabilities.
Figure 1: Age-standardised mortality rates (ASMRs) of males aged 65–89 for selected English IMD centiles (2001–16) and for USS primary males (2005–16).
The USS males and females’ mortality rate improvements were found not to be significantly different from corresponding IMD-10 (the least deprived IMD decile) improvement rates. This is useful for forecasting, since IMD-10 has a greater volume of data (implying less small-population noise) and a longer run of data.
International mortality trends also provide a useful comparison. We find that USS mortality rates are significantly lower than those of other countries’ populations. As shown in Figure 2, the spread between countries is much narrower than the spread between the centiles in one country. There is a broadly similar downward trend, although we can see levelling off in some countries in recent years.
Figure 2: Age-standardised mortality rates (for males aged 65–89) for six developed countries (1980–2016) and for English IMD centiles (2001–16).
We found that neither an individual’s IMD decile nor the region in which they live has any explanatory power. The only statistically significant co-variate was pension amount, but this was a much weaker predictor of low mortality than is normally the case for pension schemes. The implication is that the homogeneity of the USS occupational group (university academics and senior administrators) overrides other potential co-variates as a predictor of mortality.
Lessons for the USS
Based on these findings, the USS drew the following lessons:
- The USS should use its own base mortality tables for both males and females.
- The short-term future mortality rate improvement for USS pensioners should be the same as for IMD-10 (using the full experience of IMD-10 from 2001 onwards) and this should be different from (higher than) the short-term mortality rate improvement for the English national population.
- The principle of coherence suggests that the long-term improvement assumption should be compatible with related populations including IMD-10, the English national population and other similar populations internationally. Note that, while IMD-10 might have different short-term improvement assumptions from other IMD deciles, long-term improvement assumptions should be consistent (all equal). That means all IMD deciles should have the same long-term improvement rate as the English national population, which, in turn, needs to be consistent with that of other developed countries.
General mortality analysis framework
In the process of examining USS mortality, we developed a general framework for analysing the mortality experience of a large portfolio of lives, such as those in any large pension scheme or annuity book. The objective is to provide a firm evidence base on which to support the setting of future mortality assumptions for the portfolio as a whole or subgroup-by-subgroup.
The framework uses a stochastic mortality model ‒ in this case the CBDX model ‒ to smooth the mortality data. This could also be used for forecasting future mortality rates.
It also uses a wide variety of graphical diagnostics. These should be seen as an essential part of the process for several reasons:
- The graphics help to identify distinctive characteristics of the portfolio, such as swings in mean age at retirement and concentrations of early retirements in specific years.
- Knowledge of these characteristics can be used to formulate hypotheses about the data, or to group the data in sensible ways. For example, we were able to relate increases in retirement or concentrations of early retirements to specific legislative measures or government policy changes.
- Graphical diagnostics also help to identify errors in the data (such as birth date errors).
The framework is designed to work with a variety of larger benchmark datasets that can be helpful, again using graphical diagnostics, for determining the base mortality table and forecasting future mortality trends in the short and long term. A key requirement is to find a stable relationship between the mortality rates of the portfolio of lives of interest, and those of one or more of these benchmark datasets. We were able to find a stable relationship between the mortality data on USS pensioners and that of English mortality for one of the IMD deciles. The principal benefit of this is that data for IMD deciles, particularly IMD-10, are more comprehensive than the USS in having more years of observations and a significantly larger number of lives, resulting in substantially lower small-population noise. This can be exploited to improve predictions about future USS mortality rates.
Of particular importance is the ability to map a member of the dataset to a particular socioeconomic group or region whose mortality can also be modelled independently. We were able to do this with the USS dataset by mapping a member’s postcode to a Lower Layer Super Output Area (a geographic hierarchy designed to improve the reporting of small area statistics in England and Wales) and thence to an IMD decile. This index is specific to England, but there are other available geodemographic classification techniques based on the notion of ‘linking people to places’, such as:
- The US’s Neighborhood Socioeconomic Status Index
- Geodemographic profiling – this includes commercial organisations that collate multi-source information to produce sociodemographic measures (such as Experian’s Mosaic consumer classification system)
- Customised socioeconomic mortality indices (such as the Longevity Index for England).
Finally, it is important to recognise that different datasets will have their own idiosyncratic features, which need to be teased out using the graphical tools of the general framework and then exploited to help set the base mortality table and improve mortality forecasts. For example, our discovery of USS members’ strong homogeneity and high longevity can be exploited to improve the robustness of future mortality rate projections. Other datasets are likely to have a more heterogeneous membership, but if the socioeconomic composition can be reliably determined, it may still be possible to get a good fix on the most appropriate base mortality table and to generate reliable forecasts of short and long-term future mortality trends.
A useful framework
The USS was found to have significantly lower mortality rates than even IMD-10 (the least deprived of the English deciles), but with similar mortality improvement rates to that decile from 2005–16. We found that other potential co-variates derived from an individual’s postcode (geographical region and the IMD associated with their local area) typically had no explanatory power, or much weaker explanatory power than is normally the case. This lack of dependence is an important conclusion of the USS-specific analysis, and contrasts with others that consider the mortality of more heterogeneous scheme memberships. Although the key findings are likely to be particular to the USS, we believe our analytical framework will be useful for other large pension schemes and life annuity providers.
Professor Andrew JG Cairns is a professor in the Department of Actuarial Mathematics and Statistics at Heriot-Watt University
Professor David Blake is a professor of pension economics at Bayes Business School and director of the Pensions Institute
Professor Kevin Dowd is a professor of finance and economics at Durham University
Guy Coughlan, Owen Jones and Jeffrey Rowney are employees of the Universities Superannuation Scheme
Cairns AJG, Blake D, Dowd K, Coughlan GD, Jones O and Rowney J. A general framework for analysing the mortality experience of a large portfolio of lives: with an application to the UK universities superannuation scheme. European Actuarial Journal 2022; 12, 381-415.
Cairns AJG, Wen J and Kleinow T. Drivers of mortality: risk factors and inequality. Heriot-Watt University 2021.
Dowd K, Cairns AJG and Blake D. CBDX: a workhorse mortality model from the Cairns–Blake–Dowd family. Annals of Actuarial Science 2020; 14, 445-60.
Hyndman R, Booth H and Yasmeen F. Coherent mortality forecasting: the product-ratio method with functional time series models, Demography 2013; 50, 261-83.
Li N and Lee R. Coherent mortality forecasts for a group of populations: an extension of the Lee-Carter method. Demography 2005; 42, 575-94.
Ministry of Housing, Communities and Local Government. National statistics: English indices of deprivation 2019.
Richards SJ. Applying survival models to pensioner mortality data. British Actuarial Journal 2008; 14, 257-303.
Image credit | iStock