TY - GEN
T1 - New Approaches in Mortality Modelling and Forecasting
AU - Basellini, Ugofilippo
PY - 2020/2/24
Y1 - 2020/2/24
N2 - Mortality modelling and forecasting are deeply rooted in demographic and actuarial sciences.
Models to describe mortality patterns over age and time have long been used and developed since
John Graunt (1662) introduced one of the first models of mortality, the life table. Forecasts of
mortality have also been produced for many years: the first examples trace back to the beginning
of the twentieth century, when English actuaries started to measure the financial burden of
unanticipated longevity improvements on insurance and pension providers’ reserves.
Today, the study of human mortality still occupies a central role in demographic and actuarial analyses. Most of the attention received by this area of research has been stimulated by two
pressing challenges faced by modern societies: population ageing and longevity risk. According
to the latest World Population Prospects, virtually every country of the world is experiencing
growth in the number and proportion of older persons, resulting from continuous mortality and
fertility declines (United Nations, 2019). Furthermore, the demographic transition has been impacting both public and private pension systems, whose retirement liabilities lie between $60
and $80 trillions in developed economies due to unexpected mortality improvements (Michaelson and Mulholland, 2014). Funding public policies and retirement products for the elderly
becomes increasingly difficult as working-age populations shrink and dependency ratios increase
worldwide.
The enormous size of unexpected public and private retirement liabilities is the result of
overly conservative forecasts of mortality during most recent decades. Despite the great advances in the field of mortality forecasting, including the shift from deterministic to stochastic
approaches, currently and widely used methods have repeatedly failed to anticipate the sustained
rate of mortality improvements observed in many low-mortality countries. The need for novel
models that can predict longevity improvements more accurately than established methodologies is evident and timely. Therefore, this dissertation aims to bring new insights to the analysis
and forecasting of human mortality by introducing novel statistical methods that offer different
perspectives on mortality developments.
This dissertation comprises six chapters, five of which are studies that have been devised to
address this goal. Each study takes the form of a research manuscript, which has been published
or submitted to scientific journals; furthermore, routines for reproducing the results presented in the thesis have been made publicly available. The first chapter introduces the basic notions and
measures employed in the study of human mortality, reviews the main contributions in the history
of modelling and forecasting mortality, and provides a short overview of the five studies developed
in the thesis. In Chapter 2, we illustrate a general framework for modelling adult mortality
that reconciles the well-known laws of mortality into a single flexible family. Re-parameterizing
mortality models in terms of the proposed location–scale family has two important advantages:
the model’s parameters have a direct demographic interpretation, and their estimation is more
precise due to their lower correlation.
From the third to the fifth chapters, the attention is shifted from mortality rates to age-atdeath distributions as an alternative, yet informative (and neglected), function for modelling and
forecasting human mortality. Chapter 3 proposes a relational approach to model and forecast
adult mortality by transforming the age-axis of a standard distribution of deaths. The proposed
Segmented Transformation Age-at-death Distributions (STAD) model successfully captures mortality developments over age and time, and its forecasts are more accurate and optimistic than
those obtained with the seminal Lee-Carter (LC) model (Lee and Carter, 1992) and its extensions. The STAD model is further employed and generalized in the following two chapters. In
Chapter 4, the methodology is extended to the entire age-range. The age-pattern of mortality
is first smoothly decomposed into three independent components that operate upon childhood,
middle and old ages (as originally proposed by Thiele, 1871). The three components are then
modelled and forecast with specialized versions of the STAD model. The resulting forecasts are
shown to be more accurate and optimistic than those of traditional and well-established models.
Chapter 5 presents a generalization and application of the STAD methodology for modelling and
forecasting cohort mortality data. Models developed to forecast cohort data are very scarce in
the literature, and our proposed approach allows us to precisely complete the mortality experience of partially observed cohorts. Finally, Chapter 6 proposes a new extension of the influential
LC model that overcomes some of its known drawbacks. Working in a penalized composite link
framework, we simultaneously smooth and decompose the mortality pattern into three independent components, which are modelled, estimated and forecast within an LC smooth framework.
Fitted and forecast mortality profiles do not show the jaggedness typically displayed by the LC
model; furthermore, mortality rates can vary more flexibly across age and time, as they result
from a combination of three component-specific schedules of mortality changes.
AB - Mortality modelling and forecasting are deeply rooted in demographic and actuarial sciences.
Models to describe mortality patterns over age and time have long been used and developed since
John Graunt (1662) introduced one of the first models of mortality, the life table. Forecasts of
mortality have also been produced for many years: the first examples trace back to the beginning
of the twentieth century, when English actuaries started to measure the financial burden of
unanticipated longevity improvements on insurance and pension providers’ reserves.
Today, the study of human mortality still occupies a central role in demographic and actuarial analyses. Most of the attention received by this area of research has been stimulated by two
pressing challenges faced by modern societies: population ageing and longevity risk. According
to the latest World Population Prospects, virtually every country of the world is experiencing
growth in the number and proportion of older persons, resulting from continuous mortality and
fertility declines (United Nations, 2019). Furthermore, the demographic transition has been impacting both public and private pension systems, whose retirement liabilities lie between $60
and $80 trillions in developed economies due to unexpected mortality improvements (Michaelson and Mulholland, 2014). Funding public policies and retirement products for the elderly
becomes increasingly difficult as working-age populations shrink and dependency ratios increase
worldwide.
The enormous size of unexpected public and private retirement liabilities is the result of
overly conservative forecasts of mortality during most recent decades. Despite the great advances in the field of mortality forecasting, including the shift from deterministic to stochastic
approaches, currently and widely used methods have repeatedly failed to anticipate the sustained
rate of mortality improvements observed in many low-mortality countries. The need for novel
models that can predict longevity improvements more accurately than established methodologies is evident and timely. Therefore, this dissertation aims to bring new insights to the analysis
and forecasting of human mortality by introducing novel statistical methods that offer different
perspectives on mortality developments.
This dissertation comprises six chapters, five of which are studies that have been devised to
address this goal. Each study takes the form of a research manuscript, which has been published
or submitted to scientific journals; furthermore, routines for reproducing the results presented in the thesis have been made publicly available. The first chapter introduces the basic notions and
measures employed in the study of human mortality, reviews the main contributions in the history
of modelling and forecasting mortality, and provides a short overview of the five studies developed
in the thesis. In Chapter 2, we illustrate a general framework for modelling adult mortality
that reconciles the well-known laws of mortality into a single flexible family. Re-parameterizing
mortality models in terms of the proposed location–scale family has two important advantages:
the model’s parameters have a direct demographic interpretation, and their estimation is more
precise due to their lower correlation.
From the third to the fifth chapters, the attention is shifted from mortality rates to age-atdeath distributions as an alternative, yet informative (and neglected), function for modelling and
forecasting human mortality. Chapter 3 proposes a relational approach to model and forecast
adult mortality by transforming the age-axis of a standard distribution of deaths. The proposed
Segmented Transformation Age-at-death Distributions (STAD) model successfully captures mortality developments over age and time, and its forecasts are more accurate and optimistic than
those obtained with the seminal Lee-Carter (LC) model (Lee and Carter, 1992) and its extensions. The STAD model is further employed and generalized in the following two chapters. In
Chapter 4, the methodology is extended to the entire age-range. The age-pattern of mortality
is first smoothly decomposed into three independent components that operate upon childhood,
middle and old ages (as originally proposed by Thiele, 1871). The three components are then
modelled and forecast with specialized versions of the STAD model. The resulting forecasts are
shown to be more accurate and optimistic than those of traditional and well-established models.
Chapter 5 presents a generalization and application of the STAD methodology for modelling and
forecasting cohort mortality data. Models developed to forecast cohort data are very scarce in
the literature, and our proposed approach allows us to precisely complete the mortality experience of partially observed cohorts. Finally, Chapter 6 proposes a new extension of the influential
LC model that overcomes some of its known drawbacks. Working in a penalized composite link
framework, we simultaneously smooth and decompose the mortality pattern into three independent components, which are modelled, estimated and forecast within an LC smooth framework.
Fitted and forecast mortality profiles do not show the jaggedness typically displayed by the LC
model; furthermore, mortality rates can vary more flexibly across age and time, as they result
from a combination of three component-specific schedules of mortality changes.
M3 - Ph.D. thesis
PB - Syddansk Universitet. Det Samfundsvidenskabelige Fakultet
CY - Odense
ER -