International Mortality and Longevity Symposium 2016

• Current key developments in biology of ageing
• Current key developments in the intervention of the ageing process
• Implications for population morbidity, mortality and longevity

Professor Thomas Kirkwood, CBE, Associate Dean for Ageing, Newcastle University

A1: Coherent Mortality Projections for the Netherlands Taking into Account Mortality Delay and Smoking

Aims: Estimates of future mortality often prove inaccurate as conventional extrapolative mortality projection methods do not capture the impact of smoking nor the mortality delay: the shift in the age-at-death distribution towards older ages. Also, mortality projection models are often applied to single populations/countries, which over time leads to divergence in the forecasted mortality, whereas this is not a likely outcome.
Recent analyses of mortality trends reveal a transition from mortality compression (a changing shape of the age-at death distribution with more deaths occurring around the modal age at death) towards mortality delay. Mortality projections including this shift have been applied before, albeit very rarely, and only for single populations/countries. Also these mortality projections did not account for the effect of smoking, whereas smoking is known to greatly affect future mortality trends and levels. Recently developed coherent mortality projections have not been applied much, let alone in combination with smoking and mortality delay. 
In this paper we will estimate future life expectancy for the Netherlands by simultaneously taking into account the effect of smoking, developments in delay and compression of mortality and the mortality experience of the opposite sex and in other countries.
Methods: Based on lung-cancer and all-cause mortality data for the Netherlands from 1950 to 2012, we (1) compare the increases in the modal age at death for all-cause mortality versus non-smoking-related mortality using our CoDe mortality model, (2) project non-smoking-related mortality by extrapolating the obtained delay (and compression) parameter values up until 2050, both individually and coherently, and (3) project all-cause mortality by combining this projection of non-smoking-related mortality with an earlier projection of smoking-related mortality. 
Smoking-related mortality will be estimated by means of an adapted and simplified version of the indirect Peto-Lopez method, for both the Netherlands and additional European countries as input for the coherent mortality projection.
Results: Compared to all-cause mortality, for non-smoking-related mortality increases in the modal age at death – indicating mortality delay - are much more linear, and more similar for men and women. 
Simple extrapolation of the parameters of the CoDe model for nonsmokers, either solely the modal age at death or all parameters, resulted in a higher e40 in 2050 compared to a Lee-Carter projection. The effect of taking into account the compression parameters resulted in slightly lower e40 for men, and a slightly higher e40 for women. 
Adding to the CoDe extrapolations for nonsmokers the extrapolation of mortality for smokers resulted as well in higher e40 than the direct Lee-Carter extrapolation for all-cause mortality.  
We expect that adding the experience of the opposite sex and the experience of including additional countries will be smaller for non-smoking-related mortality than for all-cause mortality. 
Conclusions: For the Netherlands, the necessity of taking into account smoking when performing projections based on mortality delay clearly showed. Our coherent mortality projection for the Netherlands which took into account both mortality delay and smoking resulted in higher life expectancy values in 2050 than a conventional (=Lee-Carter) projection, and more deaths at higher ages. 
Joint work by Fanny Janssen, Population Research Centre, University of Groningen and Joop de Beer, Netherlands Interdisciplinary Demographic Institute. 

Speaker: Fanny Janssen, Population Research Centre, University of Groningen

A2: Extreme Scenarios for Pandemic Risk

It is almost a century since the last catastrophic influenza pandemic.  The most recent influenza pandemic was also H1N1, and occurred in 2009.  Because this was mild, there is a psychological outcome bias towards people downgrading pandemic risk as a cause of mass excess mortality.  This perception may be strengthened by advances in public health, medical technology, and medical care which have all had a mitigating impact on influenza mortality.     In order to understand the scope of pandemic risk, the spectrum of possible future scenarios needs to be constructed based on the current human population and infectious disease environment.  This structured approach to dynamic modelling is more insightful for risk management than a purely statistical method. Intrinsic to this approach is the maximal incorporation of medical and public health knowledge into the modelling process.  This extends to the inclusion of hypothetical scenarios that have no historical precedent, but yet are scientifically plausible.    Of particular actuarial interest for life insurers are the extreme pandemic scenarios.  How costly could one realistically be?  A benchmark for a pandemic of insurance catastrophe proportions is the 1918 pandemic. The high case fatality rate of 2.5% is much higher than for the other pandemics of the 20th century in 1957 and 1968, which were of the order of 0.1%.   Far more people died of the influenza pandemic than in the Great War itself.  Significantly for risk assessment, the two global disasters were causally connected.  The influenza was most likely brought to the western front by a cohort of 100,000 Chinese labourers despatched there by the Chinese government seeking favour from the western powers.    In general, standards of public health today are far higher than in 1918, and actuaries may ponder the pathways by which a pandemic disaster of 1918 proportions could re-emerge. The nexus between political conflict and a global pandemic provides one clear route to disaster.  If an epidemic were to emerge in one of the numerous developing regions in a state of political unrest, civil strife or anarchy, the absence of disease surveillance and fragile public health system could well allow the contagion to become established there and then spread abroad to other continents via refugees with little constraint.   Already, there has been an upsurge in communicable diseases in Lebanese refugee camps.    In exploring the future prospects for extreme pandemics, special consideration will be given to the mitigating effect of veterinary surveillance in developing countries.  The stark deficiencies in veterinary surveillance in poorer countries were tragically exposed during the 2014 Ebola crisis.  Of interest to life reinsurers in particular is the cost-effectiveness of an initiative to improve veterinary surveillance in key countries where a new pathogen might emerge.   A reduction in capital charge might be justified by the pandemic risk reduction associated with a systematic programme of veterinary surveillance.   

Speaker: Gordon Woo, RMS

A3: Forecasting Mortality by Cause of Death

Aims: The aim of this research is to develop methods for forecasting mortality, disaggregated by cause of death. Cause-specific forecasts would help analysts identify the causes of current trends and assess their future. Policy makers might use the forecasts to target areas for future reductions in mortality. In addition, the ultimate cause of death may be a good guide to the period of ill health that precedes most deaths and help to indicate its cost to society. At the moment, very few institutions try to forecast mortality by cause of death because of the technical difficulties. One of the reasons it is more difficult is that even though the average age at death is generally rising, each person must die, so if the forecast suggests that a cause will become less important in the future , one or more of the other causes has to absorb extra deaths and not necessarily at the same age. A second problem arises because we would like the sum of the forecasts for each cause to match the existing forecasts that do not separate the causes of death. Previous research has shown that life expectancy forecasts for individual cause groups, when combined, are more pessimistic than the same methodology applied to total mortality. This project undertakes basic research to explore new statistical methods that are explicitly designed to address these problems. Methods Rather than forecast mortality rates, we forecast the cause-specific distributions of life table deaths using methods from Compositional Data Analysis (CoDa). These methods are designed for data constrained to a constant sum, in this case the radix of the life table in each year. This ensures that reducing deaths from any cause or in any age-group is automatically compensated by an increase in deaths elsewhere. Thus compensation is explicitly built into the model. The forecasting model follows the Lee-Carter structure, translated into CoDa form and extended to multiple-decrement life tables. Results The CoDa methodology applied to all-cause mortality was compared with the Lee-Carter method across 10 advanced economies. The CoDa method out-performed Lee-Carter in about half of the countries. Generally, CoDa forecasts of life expectancy are more optimistic than Lee-Carter. Applying the CoDa method to cause-specific data for Japan and France showed that rank 2 approximations capture more than 90% of the variance in survival, even when modelling seven causes of death. The model suggests that the share of deaths attributable to Cancer is currently at a peak for French women and will decline in the future. The aggregate of the cause-specific forecasts are remarkably similar to the CoDa results for all-causes combined. Using this methodology, there is no evidence that cause-specific analysis leads to more pessimistic forecasts.
Conclusions: This paper presents an application of cause-specific mortality forecasting, explicitly addressing the necessity for coherence between the causes. The structure of the model is a simple translation of the Lee-Carter approach, but applied to the age- and cause-specific distributions of life table deaths. When recombined, the cause-specific forecasts closely approximate the all-cause forecast.

Speaker: Jim Oeppen, Max Planck Odense Center on the Biodemography of Aging

A4: Use of Routinely Collected Primary Care Data to Model Longevity and Longevity Improvement

Aims: Longevity and morbidity risks are of essential importance to the actuarial community. We believe that to be able to establish the drivers of their changes, and to predict how they may change over time and how this would affect life expectancy, actuarial researchers need to engage in statistical modelling of mortality experience using large scale population-based individual level data collected over the long term. Big Actuarial Data such as the CMI data are then required to translate the results to the reference population of relevance to the actuarial community. We describe our new Research Programme for 2016-2020 recently funded by IFoA and provide two case studies: on the use of statins and on longevity after myocardial infarction (MI). The aims of our research are to identify and quantify the key factors affecting mortality and longevity, such as lifestyle choices, medical conditions and/or interventions; modelling of temporal changes in the factors affecting morbidity and mortality; evaluation of plausible scenarios in mortality trends of insureds due to medical advances or lifestyle changes; and development of software tools to forecast longevity risk.
Methods: To determine the main factors affecting longevity and dynamics of their changes, we are using the subset of the THIN primary care database comprising 3.4 million patients born before 1960. For a target condition such as MI we design a population-based prospective cohort study using an appropriate extract of the primary care data. We use a case-control design with cases matched with several (3 to 5) controls from the same GP practice. This provides balanced and comparable cohorts of cases and controls and simplifies the study of comparatively rare conditions without loss of efficiency.  After the main factors affecting longevity and dynamics of their changes are established, we need to adjust the results for basis risk using the CMI data
Results: We provide a case-study on longevity improvement due to the widening of the prescription of statins, the only known longevity-improving drug in general use. We considered four prospective cohorts: people aged 60, 65, 70, and 75. Sample sizes were 118,700; 199,574; 247,149; and 194,085, respectively. Our preliminary research demonstrated that patients who are prescribed statins by the age of 65, have no survival benefits in low risk group, a  benefit of 11% in the moderate risk group, and of 14% to 18% in the high risk group. In the course of the project we shall combine these results with a novel model for the uptake of statins over time and we will develop an adjustment for the basis risk based on the CMI data to provide a plausible scenario of temporal changes in longevity due to statins. 
Conclusions: The use of Big Data enables evidence-based actuarial research into longevity improvement. We intend to develop an R package incorporating our models and providing analytical and graphical means to forecast longevity of a general UK population, and also of a population of a user defined composition under a number of scenarios for changes in disease incidence, health behaviors and treatments.  

Speakers: Lisanne Gitsels and Elena Kulinskaya, University of East Anglia

A5: Medical Innovation vs. Risk Factors: A Future Perspective on Breast and Lung Cancer

Improvements in cancer survival are generally reliant upon medical advancements in terms of both treatments and diagnostics, as well reductions in risk, particularly individual lifestyle risk behaviours. 
An overview of the current and future trends in breast cancer and lung cancer incidence and mortality as a consequence of these drivers will be presented. Evidence from empirical research will be discussed and results will be stratified by gender, age, and socioeconomic status where possible, including older ages, with a view to understanding future impact.

Speaker: Nicola Oliver, Medical Intelligence (UK) Ltd

Cancelled workshop - A6: An exploration of the Use of 'Leaders' in Mortality Forecasting

• Data and analyses to understand relationships between medical information and mortality.
• Case study to show features of causal models.
• The use of causal models.

Speakers:

Matthew Edwards, Senior Consultant, Willis Towers Watson

Joseph Lu, Longevity Science Director, Legal & General

Chris Martin, Clinical Modelling Consultant, Legal & General

The Actuarial Research Centre (ARC) is the Institute and Faculty of Actuaries’ network of actuarial researchers around the world.  The ARC supports actuarial researchers around the world in the delivery of cutting-edge research programmes that aim to address some of the significant challenges in actuarial science.

Recognising the role that IFoA members can make in advancing actuarial science, the PhD workshop sessions at the International Mortality and Longevity Symposium 2016 provide a platform for industry practitioners to engage with research students funded through the ARC. 

Delegates will be presented with students’ current research and can assist in offering advice for their future studies, ensuring the remainder of their research projects are both relevant and will have significant industry impact.

Chair: Andrew Cairns, ARC Director

• ARC overview (ARC Directors and research programmes)
• Two PhD student presentations with Q&A

Student 1: Vasil Enchev, Heriot Watt - Optimal design of structured products

Student 2: Liang Chen, Heriot Watt - Longevity risk modelling

Paul Johnson, Director, Institute for Fiscal Studies

• Developments in trends in healthy life expectancy
• Implications for health and care providers

Professor Carol Jagger, AXA Professor of Epidemiology of Ageing, Newcastle University

B1: Inferences for Maximum Country Life Expectancy using Provincial Data

In this paper we introduce a new approach to modelling and projecting (maximum) life expectancy for a region using only data from subregions within this larger region by applying principles from the statistical theory of Extreme Values.
The most popular mortality forecasting models, the Lee Carter Model (Lee and Carter, 1992) and its numerous extensions and variants e.g Li et al. (2004); Renshaw and Haberman (2003); Hyndman et al. (2007) fit trends to age-standardized (log) death rates. However, there is a strong argument for using life expectancy in forecasting. White (2002) found that linear trends in life expectancy give a better empirical fit to the experience of individual countries than linear trends in age-standardized (log) death rates in his study of 21 developed countries. Among those who have forecast life expectancy are Alho and Spencer (2005); Andreev and Vaupel (2006); Lee (2006); Torri and Vaupel (2012). Extreme value theory has previously applied to the study of global maximum life expectancy by Medford (2015) and this paper builds upon ideas presented therein.
From the Canadian Human Mortality Database, high quality life expectancy data is available for the various Canadian provinces over the period 1921- 2011. We take the maximum life expectancy from among the various provinces over each of the years for which the data is available and fit a time varying Generalized Extreme Value (GEV) Distribution.
One advantage of our approach is that we are able to make probabilistic statements about future maximum Canadian life expectancy. Secondly, we are able to project maximum life expectancy for Canada overall by using only information about life expectancy at the provincial level. This method could be applied in other situations where regional-only data is available but one would like to have an idea about life expectancy at a supra-regional or more aggregated level.

Speaker: Anthony Medford, Max Planck Centre for the Biodemography of Aging University of Southern Denmark

B2: Explaining the Female Longevity Puzzle

The observed increase in life expectancy for developed countries does not follow a general pattern, as periods with gender-specific divergence between countries have emerged. In particular in the Scandinavian countries we have observed a decrease in the mortality rates for Norwegian and Swedish females, but a stable or in some cases even increasing mortality rate for Danish females in the time period 1980-1995. The idea of this paper is to understand the complexity of female longevity improvements in Scandinavia by using detailed register data for Denmark and Norway and explain why this puzzle has emerged. We partition every individual in the +50 population at each age and year into ten socio-economic groups based on an affluence measure derived from the individual’s income and wealth. We show that a very clear ranking of life expectancy exists as it is regularly increasing from the lowest affluence group to the highest affluence group for both genders. In particular we are able to identify differences in life expectancy of up to 7-8 years for 50 year old males and 5-6 years for 50 year old females in the mid-1980s. We analyze the female longevity puzzle and identified which specific socio-economic groups have been driving the "stagnation" in life expectancy in Scandinavia. We find that it is the lower middle to middle affluence groups that are responsible for the slowdown in mortality improvements for the Danish women. All subgroups for Norwegian women exhibit positive improvements over the period. Interestingly, we see that the two groups with lowest affluence measure for Danish women actually see large improvements in life expectancy for the oldest ages, i.e. 80-95 year olds. This corresponds well with the findings that the decline in longevity for Danish women cannot solely be explained by higher smoking prevalence for the lower socio-economic groups. Further, we estimate the mortality behavior of the ten socio-economic groups using three well-known stochastic mortality models (the Lee-Carter model, the Age-Period-Cohort model and the Lee-Li model) and investigate their forecast performance. Overall, we find that the model fit only decrease slightly by having smaller populations. Also we clearly show the importance of using available information for the subgroups as it considerably reduces the forecast errors. Moreover, we find that the combination of total population data with group specific data improve the forecasting performance substantially in the Lee and Li model. This is especially the case when investigating longer forecast horizons as it avoids the unrealistic cases where the lowest affluence groups improve so much that they show higher life expectancy than higher income groups.

Speakers: Malene Kallestrup-Lamb and Carsten P.T.Rosenskjold, Department of Economics, Aarhus University

B3: CMI Update on Longevity Modelling and High Age Mortality

The Continuous Mortality Investigation (CMI) carries out research into mortality and morbidity experience, providing outputs that are widely used by UK life insurance companies and pension funds. We will provide an update on our research particularly focusing on recent mortality in the general and annuitant populations, including high age mortality, as well as the current developments on the Mortality Projection Model. 

Speakers: Steve Bale, Legal & General and Tim Gordon, Aon Hewitt

B4: Stochastic Mortality Forecasting with Smoothing and Overdispersion

We propose a unified mortality forecasting framework, building on existing approaches, but developed to address limitations of those models. For example, where the stochastic model does not adequately fit the data, this can lead to estimates and forecasts which over-fit and are insufficiently robust. We account for Poisson overdispersion with a negative binomial error structure, which mitigates over-fitting. Another feature is the facility to coherently impose smoothness in parameter series which vary over age, cohort, and time. Such constraints are integrated into the modelling process, so that there is a natural feedback, whereby the smoothing of parameter series can appropriately impact other estimates, rather than being performed in a post hoc fashion. Generalised additive models (GAMs) are a flexible class of semiparametric statistical models which allow parametric functions and unstructured (but smooth) functions of explanatory variables to appear in the model simultaneously. We demonstrate the potential of GAMs for mortality modelling and forecasting. In particular, GAMs allow us to differentially smooth components, such as cohorts, more aggressively in areas of sparse data for the component concerned. While GAMs can provide a reasonable fit for the ages where there is adequate data, estimation and extrapolation of mortality rates using a GAM at higher ages is problematic due to high variation in crude rates. At these ages, parametric models can give a more robust fit, enabling a borrowing of strength across age groups. Our forecasting methodology is based on a smooth transition between a GAM at lower ages and a fully parametric model at higher ages. Finally, our framework is fully probabilistic, and provides a coherent description of forecast uncertainty.

Speaker: Jon Forster, University of Southampton

B5: Introduction of China Life Insurance Mortality Table 2010-2013

Aims: To introduce what is the current mortality experiences in China.  
Methods: Dealing with 340 million policies with 1.85 milion claims requires hard work from 25 volunteers and a good IT solution. In developing this table, the team compared models from popular mortality development methods.   
Results: The third insurance mortality table is built for valuation purpose and related researches were carried out.

Speakers: Yao Zhang, Director, Actuarial Division, Life Insurance Department, CIRC(China Insurance Regulatory Commission) and Chu Zhang, Experience Study Manager, FSA, New China Life

Two short presentations followed by panel discussion:
•    How publicly held data can help tackle unprecedented challenges posed by the UK’s ageing population.
•    How private data and the internet of things can change analyses and trends of morbidity, mortality and longevity.

Speakers:
Dame Karen Dunnell, Chair of Longevity Science Panel
Professor Elena Kulinskaya, University of East Anglia

C1: On Bayesian Two-Population Mortality Models for the Assessment of Basis Risk in Longevity Hedges

Background: Index-based hedges have the potential for providing pension funds and annuity provider with an effective solution for the management of their longevity risk. However, the lack of an appropriate and applicable approach for quantifying the basis risk emerging from the potential mismatch between the mortality in the index reference portfolio and the pension fund/annuity book being hedged is one of the main obstacles that prevents many pension schemes and insurers from progressing in their consideration of index-based solutions. Two-population stochastic mortality models allow the joint modelling and projection of the mortality of the reference and target populations providing a possible approach to overcome this obstacle. However, often the portfolio experience data will be sparse, posing a challenge for the accurate calibration and projection of two-population models using standard estimation techniques. The aim of this presentation is to illustrate how a Bayesian estimation framework can be used to deal with some of the data limitation arising in the application of two-population stochastic mortality model for the assessment of basis risk in longevity hedges.
Methods: We formulate under a Bayesian framework several two-population extensions of the well-known Lee-Carter and CBD models and implement such an approach using the R package rjags. We then contrast the performance of the Bayesian estimation approach with that of the more traditional maximum likelihood estimation discussed in Haberman et al (2014). We first assess the performance of the models based on their goodness-of-fit and on the plausibility of their projections. We then evaluate model performance using illustrative hedge-effectiveness evaluation exercises and pay particular attention to the impact that different volumes of data may have on the assessment of basis risk. We use England and Wales population data for the reference population. We also generate synthetic portfolio data using a combination of ONS data for England split by deprivation quintile and data on the typical socio-economic composition of pension schemes.
Results: The sampling noise present in mortality datasets with limited exposures implies significant parameter estimation error in the calibration of two-population mortality models and, if not appropriately handled, can lead to a misestimation of longevity basis risk. A Bayesian estimation approach –which accounts naturally for parameter uncertainty– results in a more accurate assessment of the hedge effectiveness of index-based solutions for pension schemes with a small population.
Conclusion: A Bayesian approach to the fitting of two-population mortality models offers an alternative for the assessment of basis risk when exposures sizes are modest, as is the case of most pension scheme datasets. Furthermore, our hedge effectiveness results show that index-based hedges have the potential to provide an effective and flexible solution to mitigate longevity risk.
References: Haberman, S., Kaishev, V. K., Millossovich, P., Villegas, A. M., Baxter, S., Gaches, A., Gunnlaugsson, S., and Sison, M. (2014). Longevity Basis Risk: A methodology for assessing basis risk. Institute and Faculty of Actuaries Sessional Research Paper.

Speaker: Andrés M. Villegas, ARC Centre of Excellence in Population Ageing Research (CEPAR), UNSW Australia

C2: When is a Cohort Not a Cohort? Spurious Parameters in Stochastic Longevity Models

Aims: Stochastic models are widely used to assess the risk of pensioners and annuitants living longer than expected. In particular, many UK insurers have implemented stochastic longevity models for Solvency II; and the Prudential Regulation Authority (PRA) considers a range of stochastic models when determining the quantitative indicators that it uses to assess insurers’ internal models. UK mortality data shows very strong cohort features – rates of mortality improvement that differ according to year of birth – so it is natural to include cohort parameters in models to reflect cohort effects. However the fitted cohort parameters of some stochastic models do not reflect patterns of historical mortality improvements, and projections of future improvements can be misleading as a result. Our research demonstrates this concern and shows how the problem arises for particular model structures.
Methods: We fit a range of models, from the Lee-Carter, age-period-cohort (APC) and Cairns-Blake-Dowd families, to international historical data from the Human Mortality Database and to synthetic data. The synthetic data reflects the key features of historical data, but is constructed to have certain known properties – such as a lack of a cohort effect. Working with synthetic data allows us to test how features of the data are reflected in fitted parameters, and also makes it possible to analyse the parameters algebraically.
Results: We fit the APC model to synthetic data that has no cohort effect, but find that the fitted cohort parameters are non-zero and have a distinctive parabolic shape, even after allowing for identifiability constraints. Algebraic analysis shows that the cohort parameters are being used to model age rather than cohort patterns, and the size of the cohort parameters is directly related to the slope of mortality improvements by age. The age and period parameters also have a quadratic shape. When we fit the same model to historical data from a range of countries we find that the fitted cohort parameters are dominated by the same parabolic shape as for the synthetic data, rather than reflecting genuine cohort effects. Projecting these fitted parameters using the standard approach in the existing literature would give misleading results – the time series typically used do not take account of the quadratic shape of the parameters. We repeat the analysis for a range of models, and find that there are similar concerns for the cohort extension of Lee-Carter, and for some members of the Cairns-Blake-Dowd family.
Conclusion: It is helpful to test models by fitting them to synthetic data with known simple features. If model parameters do not reflect features of the synthetic data then the models should be treated with caution – it is difficult to make a sensible projection of parameters which do not have a clear interpretation. Only a few of the models that we have tested can model a cohort effect without having spurious parameters. This makes an assessment of model risk difficult.

Speaker: Jon Palin, Barnett Waddingham

C3: Semi-Parametric Extensions of the Cairns-Blake-Dowd Model: A One-Dimensional  Kernel Smoothing Approach

Aims: Briefly outline what is already known about the subject and what the research aims to address.
In this paper, we propose a time-varying coefficient mortality model aiming to combine good characteristics of existing models with efficient model calibration methods.
Methods: Describe what research was done and how it was carried out. This may include specific methodology or models used and/or a description of the participants (e.g. adults aged 50+ in England). 
Nonparametric kernel smoothing techniques have been applied in the literature of mortality modeling and based on the findings from Li et al.'s study (2015), such techniques can significantly improve the forecasting performance of mortality models. Thus, in this study we follow the same path and adopt a kernel smoothing approach along the time dimension. Since we follow the model structure of the Cairns-Blake-Dowd (CBD) model, the time-varying coefficient model we propose can be seen as semi-parametric extensions of the CBD model and it gives specific model design according to different countries' mortality experience. 
Results: Detail the findings of the research. 
The fitting and forecasting results from empirical studies have shown superior performances of the model over a selection of well-known mortality models in the current literature.
Conclusions: Should summarise the most important outcome(s) of the research.
The proposed model maintains and further enhances the strengths of existing models, while avoids the issue of overdispersion in death data. We conduct an empirical study based on male mortality data from ten developed countries. The fitting and forecasting results show that the proposed model fits historical mortality data very well and provides more accurate mortality predictions compared to a selection of well-known mortality models. Since local linear kernel smoothing technique is used to estimate and forecast the time-varying coefficient model, this has in turn made the semi-parametric approach more attractive compared to the time-series type of forecasting methods in the literature.

Speaker: Colin O'Hare and Han Li, Monash University

C4: Socioeconomic Differentials in Multimorbidity and Health Expectancy Using Electronic Health Records - Methodological Challenges

Background and Aims: The socioeconomic gradient in life expectancy is well established, but little is known about the impact of variations in disease patterns and trajectories underlying lifespan inequalities. Multimorbidity is the co-occurrence of two or more chronic diseases in an individual; and most people aged over 75 are multimorbid. For our analysis, we selected 30 major chronic diseases, including all those relevant to the UK insurance industry. The objectives of our research are twofold: (1) To quantify the impact of differences in the age of onset and type of disease/s on survival: i.e. do disadvantaged groups acquire more, or more lethal combinations of, diseases; or do they simply become ill at younger ages? (2) Furthermore, while mortality from major diseases in isolation is well-studied, less investigated is the mortality risk from interactions between diseases (e.g. are the risks of having 2 or more diseases additive, multiplicative, or does one more lethal disease trump the others?).

Data: CPRD (Clinical Practice Research Datalink) database.  Our cohort consists of 1.3 million English patients aged 45+, registered in 225 practices, followed up from 2001 to 2009 using linked electronic health records which span primary care consultations, hospital admissions and ONS death registration.

Methods: We use two analytic methods. Initially, both approaches have been applied to a disease-count based analysis. In future, we will extend these methods to analyse specific disease combinations.

(1) Survival models: We use Cox proportional hazard models  to assess the differences in hazard ratios, survival probabilities and modelled life expectancies by deprivation, smoker status and health status at baseline, separately for each sex. Health status is categorised into patients with no chronic disease (of 30 diseases), 1 disease, 2 diseases, or 3+ diseases.

(2) Multi-state models (MSM): We use continuous-time Markov models to partition total life expectancies into health expectancies, and to quantify the ages of disease onset, disease progression and death. No recoveries from these chronic diseases are assumed: i.e. it has a progressive structure. We calculate health expectancies based on transitions between 5 health states – no diseases, 1 disease, 2 diseases, 3+ diseases and death, for each deprivation, sex and smoker status sub-group. Life expectancies and years spent with 0, 1, 2, 3+ diseases are derived for any age from 65 onwards.

Results and conclusions: We will discuss the methodological challenges of using the CPRD database and present provisional results from our disease-count based models. We will also assess the strengths and limitations of our analytical approaches.  As populations age, we need to move beyond the single-disease analysis framework. Results from this project will inform underwriting philosophy and the calibration of medical longevity models.

Speakers: Dr Madhavi Bajekal and Mei Sum Chan, University College London

C5: Parameter Risk in Time-Series Mortality Forecasts

Aims: We aim to quantify the impact of parameter uncertainty and volatility on mortality projections, show how the overall prediction error can be decomposed into those two components, and evaluate the impact of parameter uncertainty, model risk and volatility on solvency capital requirements 
Methods: We apply a bootstrap procedure to the parameter estimation problem for ARIMA time series models to project the period effect in a Lee-Carter model for England & Wales mortality rates. 
Results: We find that the best fitting ARIMA model leads to more uncertainty and less robust mortality projections than a simpler ARIMA model which provides a less good fit. There is a trade-off between goodness of fit and robustness of models. This has implications for solvency-capital requirements. 
Conclusions: The paper shows that a model for mortality forecasting cannot be selected on closeness of fit alone. The counter-example in the paper - that of the best-fitting ARIMA model producing less stable forecasts - shows the importance of acknowledging parameter uncertainty in solvency work.

Speakers: Torsten Kleinow, Heriot-Watt University and Stephen Richards, Longevitas Ltd.

• Scientific evidence that we age at different pace
• Big Data opportunity to better understand the pace of ageing
• Opportunities for the insurance sector

Professor Jay Olshansky, Chief Scientist and Co-Founder, Lapetus Solutions

The Actuarial Research Centre (ARC) is the Institute and Faculty of Actuaries’ network of actuarial researchers around the world. The ARC supports actuarial researchers around the world in the delivery of cutting-edge research programmes that aim to address some of the significant challenges in actuarial science.

Recognising the role that IFoA members can make in advancing actuarial science, the PhD workshop sessions at the International Mortality and Longevity Symposium 2016 provide a platform for industry practitioners to engage with research students funded through the ARC. 

Delegates will be presented with students’ current research and can assist in offering advice for their future studies, ensuring the remainder of their research projects are both relevant and will have significant industry impact.

Chair: Andrew Cairns, ARC Director

• Short ARC overview 
• Two PhD student presentations with Q&A

Student 1: Raj Bahl, University of Edinburgh - Mortality linked derivatives and their pricing

Student 2: Ruhao Wu, University of Leicester - Coherent mortality forecasting: the multilevel functional principal component analysis approach

The Conference Dinner will be held in the Founder's Building -  widely recognised as one of the most spectacular university buildings in the world. It is an example of Gothic Revival architecture in the United Kingdom and houses the Picture Gallery, containing a collection of over 70 pieces of Victorian era art given to the college at the time of its founding by Thomas Holloway. 

D1: Forecasting Socio-Economic Differences in the Mortality of Danish Males

Aims: To analyse the mortality of Danish males subdivided into 10 socio-economic groups; what can we infer about the mortality dynamics of different sub-groups that might be mimicked in other countries.
Methods: The Statistics Denmark database contains accurate and very detailed information about all Danish residents since 1981. We exploit this richresource to subivide the males population into 10 equal-sized groups based on a measure of affluence based on individual income and wealth. We then build a new multi-population model that models jointly the mortality of each of the 10 groups in a parsimonious way covering ages 55 to 94 and years 1985 to 2012.
Results: The affluence metric is shown to provide consistent sub-group rankings based on mortality rates across all ages and over all years in a way that improves significantly on previous studies that have focused on life expectancy. The gap between the most and least affluent is confirmed to be widest at young ages and has widened over time. The modelling framework allows us to draw statistically stronger results than would be possible using crude mortality data alone, enabling us to model the larger number of smaller sub-groups than has been previously possible without losing the essential character of the underlying data. The model produces bio-demographically reasonable forecasts of mortality that preserve the group rankings at all ages. The model is also shown to satisfy reasonableness criteria related to the term structure of correlation across ages and through time through consideration of both future mortality rates and survival rates.
Conclusions: We have developed a detailed mortality database for Danish males divided into 10 socio-economics groups (deciles) that will be made publicly available for use by researchers. The 10 groups can be combined in various ways to simulate mortality datasets for other populations that are broader than single deciles but also potentially much smaller. Alongside this we have developed a general multi-population mortality model that helps researchers to extract the maximum signal from the relatively noisy dataset. The new dataset and accompanying model can be used to investigate a variety of important risk measurement and risk management problems including the assessment of basis risk in longevity hedges and economic capital assessment.

Speaker: Andrew Cairns, Heriot-Watt University

D2: Evidence and Implications of Socioeconomic Differences in Mortality

Aims: This paper provides evidence on the differences in life expectancy around retirement age across different socioeconomic groups in selected OECD countries based on measures of education, income and occupation. The implications of these differences in mortality for pensions and annuity markets and for policy are discussed.    
Methods: Based on existing research    
Results and Conclusions: Evidence shows that there are significant differences in life expectancy across socioeconomic groups as measured by education, income and occupation, and that there are also differences in the gradient of improvements in mortality and life expectancy across socioeconomic groups. Differences in life expectancy present a challenge for pension funds and annuity providers to manage the longevity risk they face, first with respect to establishing appropriate mortality assumptions and secondly to effectively mitigate their exposure to the risk. These differences also present an opportunity for pensions and insurers to expand their markets and diversify their longevity risk by adapting product offerings to different segments of society. Policy makers could help to facilitate the measurement and management of the longevity risk exposure of pension funds and annuity providers by making accurate and timely mortality data available by socioeconomic groups. Policy makers should encourage and facilitate product innovation to meet the various needs of different market segments, though they should also ensure that the risks arising from these products are managed appropriately.  Policymakers must be aware of this fragmentation to ensure that the rules with respect to how the overall access to funds earmarked for retirement is governed do not penalise lower socioeconomic groups, as policies defined “on average” may be regressive.   

Speaker: Jessica Mosher, OECD

D3: Has now moved to C4

D4: Behavioural and Psychological Drivers of Mortality

When assessing mortality risk, insurers have historically focused most of their attention on traditional biometric risk factors.  This session looks at the impact of behavioural and other psychological drivers of mortality – including how data underlying US credit scores can be remodelled to give a very strong predictor of mortality – and considers the future possibilities of incorporating new evidence into underwriting insurance.

Speakers: Peter Banthorpe and Chris Falkous, RGA

D5: Does Money Buy you Longevity?

Aim: Better understanding of longevity differentials as apply to higher earners who dominate pension scheme / annuity liabilities
Methods: Using Club Vita data of over 2.5m pensioners (1 in 7 of UK population) we explore the quantum of mortality differentials for higher income individuals and the importance of choosing the most appropriate wealth metric using a mix of GLM analysis and case studies. We explore the shape of mortality and life expectancy with rising affluence. Emerging patterns of improvement amongst different wealth groups will be explored, identifying the quantum of difference and financial impact. This will provide an update for most recent developments compared to our previous research which covered trends up to 2010. Implications for financial management of longevity risk will be explored including how to structure ’top-slicing’ in light of our findings.
Results & Conclusions:

• Choice of wealth metric important to discern meaningful longevity insights (e.g. salary better than pension)
• Wealth and lifestyle have similar impact on spread of longevity outcomes (which we quantify)
• Longevity ‘flattens’ with increasing wealth – there is a point beyond which there is limited benefit of any additional income on life expectancy
• We will explore this shape (including confidence we can have in it) and the underlying causal dynamics •
• Life expectancy has (recently) been converging – the lower income parts of society have been ‘catching up’
• We will present the latest evidence on these socio-economic trends, consistency with other results, and explore what the underlying pathways may be driving these differences
• We will explore how exposure to higher affluence individuals varies between industrial sectors and different organisations
• We explore the impact on longevity risk management techniques of these findings - including illustrating how ‘top-slicing’ can be applied in ways which mitigate selection risk on the residual lives. 

Speaker: Steven Baxter, Hymans Robertson LLP

• How economy affects longevity?
• How would longevity or ageing population affect the UK economy?
• How would longevity or ageing population affect the UK’s health and social care cost?
• Watch out for these emerging global demographic trends

Dr Amlan Roy, Senior Research Associate, Financial Markets Group, London School of Economics

Final thoughts on the next big things in longevity and morbidity research

• Comments on issues raised at this Symposium.
• What new developments can we expect to see in the next 5 years? 
• What new developments should we begin now?

Speakers:

Joseph Lu, Longevity Science Director, Legal & General

Professor Jay Olshansky, Chief Scientist and Co-Founder, Lapetus Solutions