Interpreting mortality trends in the presence of heterogeneity ......Modeling heterogeneous...
Transcript of Interpreting mortality trends in the presence of heterogeneity ......Modeling heterogeneous...
Interpreting mortality trends in the presence of
heterogeneity: A population dynamics approach
Sarah Kaakaı
Laboratoire Manceau de Mathematiques, Le Mans Universite
Based on joint works with S. Arnold, N. El Karoui, H. Labit-Hardy and
A. Villegas
Longevity Heterogeneity and Pension Design workshop, 28 january 2020
Outline
1 Introduction
2 How can a cause-of-death reduction be compensated for by the population
heterogeneity? A dynamic approach.
3 Consistence of mortality forecast
1
Socioeconomic gradient in mortality
§ Research on the relationship between socioeconomic status (SES) and
mortality is longstanding (Villerme (1830), General Register Office (1851))
ñ broad consensus on the strong correlation between SES and
mortality.
§ New trends observed in the past decades: increasing of socioeconomic
gaps in health and mortality.
US female life expectancy gaps at birth between less and more educated
women: 7.7 years in 1990, 10.3 years in 2008 (Olshansky et al. (2012)).
Gap in male life expectancy at 65 between higher managerial and routine
occupations (England Wales): 2.4 years 1982-1986, 3.9 years 2007-2011,
ONS).
Widening gaps: socioeconomic subgroups experience rather different
mortality than national mortality rates.
2
Taking heterogeneity into account
§ SES inequities declared key public issues by the World Health
Organization in its last report on ageing and health.
§ Not taking into account heterogeneity can lead to:
Increased inequalities due public health reforms (Alai et al. (2017)) or
“unfair” redistribution properties of pensions systems.
Errors in funding of annuity and pension obligation (Meyricke and
Sherris (2013), Villegas and Haberman (2014)).
§ Better understanding of heterogeneity allows for a better understanding
of basis risk (Longevity basis risk report (2014))
3
Modeling heterogeneous mortality rates
§ Growing literature in the joint modeling and forecasting of the
mortality of socioeconomic subgroups Bensusan (2010), Jarner and
Kryger (2011), Villegas and Haberman (2014), Cairns et al. (2016) ...
§ Remaining questions:
Interpreting targets set by institutions (Department of Health, WHO)
(Alai et al. (2017)).
Consistency of sub-national and national estimates/forecasts (Shang
and Hyndman (2017), Shang and Haberman (2017)).
§ Standard tool for modelling and forecasting longevity: mortality rates.
§ Approach: take into account all population data rather that just
mortality data.
4
Modeling heterogeneous mortality rates
§ Growing literature in the joint modeling and forecasting of the
mortality of socioeconomic subgroups Bensusan (2010), Jarner and
Kryger (2011), Villegas and Haberman (2014), Cairns et al. (2016) ...
§ Remaining questions:
Interpreting targets set by institutions (Department of Health, WHO)
(Alai et al. (2017)).
Consistency of sub-national and national estimates/forecasts (Shang
and Hyndman (2017), Shang and Haberman (2017)).
§ Standard tool for modelling and forecasting longevity: mortality rates.
§ Approach: take into account all population data rather that just
mortality data.
4
What can we learn from population dynamic?
Population divided in p risk classes:
§ One year central death rate in the global population for individuals age
x during calendar year t:
µxt “Dxt
Ext“
pÿ
j“1
D jxt
Ext
The relative exposure W jxt “
E jxt
Extis linked to the proportion of
individual of age x in risk class j .
Evolution of these quantities are determined by the population
dynamics and can vary a lot depending on the age x and time t.
How changes in the socioeconomic composition of the population affect
aggregated indicators? Could we miss a cause-of-death reduction in
presence of heterogeneity?
5
What can we learn from population dynamic?
Population divided in p risk classes:
§ One year central death rate in the global population for individuals age
x during calendar year t:
µxt “Dxt
Ext“
pÿ
j“1
D jxt
Ext“
pÿ
j“1
D jxtE
jxt
E jxtExt
The relative exposure W jxt “
E jxt
Extis linked to the proportion of
individual of age x in risk class j .
Evolution of these quantities are determined by the population
dynamics and can vary a lot depending on the age x and time t.
How changes in the socioeconomic composition of the population affect
aggregated indicators? Could we miss a cause-of-death reduction in
presence of heterogeneity?
5
What can we learn from population dynamic?
Population divided in p risk classes:
§ One year central death rate in the global population for individuals age
x during calendar year t:
µxt “Dxt
Ext“
pÿ
j“1
D jxt
Ext“
pÿ
j“1
µjxt
E jxt
Ext
The relative exposure W jxt “
E jxt
Extis linked to the proportion of
individual of age x in risk class j .
Evolution of these quantities are determined by the population
dynamics and can vary a lot depending on the age x and time t.
How changes in the socioeconomic composition of the population affect
aggregated indicators? Could we miss a cause-of-death reduction in
presence of heterogeneity?
5
Outline
1 Introduction
2 How can a cause-of-death reduction be compensated for by the population
heterogeneity? A dynamic approach.
3 Consistence of mortality forecast
6
Data
§ Two datasets:
1981-2007: Department of Applied Health Research, UCL.
2001-2015: Office for National Statistics, UK .
§ English cause-specific number of deaths and mid-year population
estimates per socioeconomic circumstances, age and gender.
Socioeconomic circumstances are measured by the Index of multiple
deprivation (IMD), based on the postcode of individuals.
§ Small areas (LSOA) are ranked based on seven broad criteria: income,
employment, health, education, barriers to housing and services, living
environment and crime.
§ This ranking permits to divide the population in 5 quintiles with about
same number of individuals in each quintile.
7
Life expectancy
Figure: Evolution of period life expectancy at age 65
● ●●
● ● ●● ● ●
● ●●
●● ● ● ●
●●
●
● ● ●
●●
●●
●
● ●
● ● ●●
●
● ● ●
●●
●● ● ●
● ●●
●
● ● ●●
● ●
●
● ●●
●●
●●
●
● ●
● ● ●
●●
● ●●
●●
●● ● ●
● ●●
●
● ● ● ● ●●
●
● ● ●
● ●
●● ●
●●
●● ●
●
●
● ● ●●
●
●● ● ●
● ●●
●
●●
● ● ●●
●
● ● ●
● ●
●● ●
●●
●● ●
●
●
● ●●
●●
●● ●
●
●●
●
●
● ● ●● ●
●
●
● ●●
●●
● ●●
●● ●
● ●
●
●
● ●●
●●
●● ● ●
● ●●
●
● ●●
● ● ●
●
● ● ●
●●
●● ●
●●
● ● ●
●
●
Females
1980 1985 1990 1995 2000 2005 2010 2015
12
14
16
18
20
22
Year
Per
iode
life
exp
ecta
ncy
IMD Quintiles●
●
●
●
●
●
1
2
3
4
5
6
(a) Females
●●
●
●●
●● ●
●
● ●●
●
●●
●●
●●
●
●●
●
●
●
●●
●
●●
● ●●
●●
● ● ●
● ●●
● ●●
●●
●●
● ●●
● ●●
●
● ●●
●●
●●
●
●●
● ● ●● ●
●●
●
●●
●
● ● ●
● ●
●●
●●
●
●●
●
●
●●
●
●●
●●
●
●●
● ●●
●●
● ● ●●
●●
● ● ●●
●●
●
● ●●
●●
●
●
● ●●
●●
●
●●
●●
● ● ●
●●
● ● ●
●●
●
● ● ●● ●
● ●
●●
●●
●●
●● ●
●
●
●●
● ●
●
●● ● ●
●●
● ● ●
●●
●
● ● ●●
●●
●
● ●
●●
●●
●
●●
●
●
●
●●
●
●●
● ● ●●
●
Males
1980 1985 1990 1995 2000 2005 2010 2015
12
14
16
18
20
22
Year
Per
iode
life
exp
ecta
ncy
IMD Quintiles●
●
●
●
●
●
1
2
3
4
5
6
(b) Males
§ Increase in gap: 2.2 ñ 4.2 years (females), 2.9 ñ 3.9 (males).
8
Mortality by deprivation
Figure: Males average annual mortality improvement rates by age
−2.5%
−2%
−1.5%
−1%
−0.5%
0%
0.5%
1%
1.5%
2%
2.5%
3%
3.5%
25 30 35 40 45 50 55 60 65 70 75 80
Age
Ave
rage
d an
nual
impr
ovem
ent r
ates
IMD Quintiles1
2
3
4
5
6
(a) 1981-1995
−2.5%
−2%
−1.5%
−1%
−0.5%
0%
0.5%
1%
1.5%
2%
2.5%
3%
3.5%
25 30 35 40 45 50 55 60 65 70 75 80
Age
Ave
rage
d an
nual
impr
ovem
ent r
ates
IMD Quintiles1
2
3
4
5
6
(b) 2001-2015
§ 2001-2015: widening of mortality improvement rates gap above 60.
9
Population composition, 2015
Age-pyramids by IMD quintile, 2015
0
5
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80
85
90
95
100+
20152010
20052000
19951990
19851980
1975
19701965
19601955
19501945
19401935
19301925
19201915
100 80 60 40 20 0 20 40 60 80 100Number of individuals (in thousands)
Age
Year of birth
Type Males Females
(a) Most deprived quintile ($)
Median age: 33y
0
5
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80
85
90
95
100+
20152010
20052000
19951990
19851980
1975
19701965
19601955
19501945
19401935
19301925
19201915
100 80 60 40 20 0 20 40 60 80 100Number of individuals (in thousands)
Age
Year of birth
Type Males Females
(b) Least deprived quintile ($$$$$)
Median age: 44.2y
§ Baby-boom cohort less deprived than younger/older cohorts.10
Changes in the population composition
Figure: Compostion of males age classes in years 1981, 1990, 2005, 2015.
18%
20%21%20%
21%
19%
21%21%20%20%
22%22%21%
18%
17%
23%23%
21%
17%
14%
Age 65−74 in 1981 (Cohorts 1907−1916)
Age 65−74 in 1990 (Cohorts 1916−1925)
Age 65−74 in 2005 (Cohorts 1931−1940)
Age 65−74 in 2015 (Cohorts 1941−1950)
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
0
10
20
Pro
port
ion
of p
opul
atio
n
IMD Quintiles 1 2 3 4 5
(a) Age class 65-74
17%
19%
20%22%22%
17%
19%
20%
22%22%
16%
18%
20%
23%22%
14%
17%
20%
25%24%
Age 25−34 in 1981 (Cohorts 1947−1956)
Age 25−34 in 1990 (Cohorts 1956−1965)
Age 25−34 in 2005 (Cohorts 1971−1980)
Age 25−34 in 2015 (Cohorts 1981−1990)
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
0
10
20
Pro
port
ion
of p
opul
atio
n
IMD Quintiles 1 2 3 4 5
(b) Age class 25-34
§ Decrease of deprivation over time for older age classes.
§ Increase of deprivation for younger age classes.
11
Heterogeneous population dynamics
§ Simple age-structured population dynamics framework to illustrate
different impacts of heterogeneity on the aggregated mortality.
§ Deterministic evolution of each subgroup is described by a
McKendrick (1926) -Von Foerster (1959).
§ Equation for each gender ε “ m or f and subgroup:
Aging law:
pBa ` Btqgεj pa, tq “ ´µ
εj pa, tqg
εj pa, tq
Birth law:
g εj p0, tq “
ż a:
0
pεg fj pa, tqbj pa, tqda
Initial Pyramid:
g εj pa, 0q
12
Aggregated population
§ Aggregated population:
g εpa, tq “
řpj“1 g
εj pa, tq
Aging law: pBa ` Btqgεpa, tq “ ´dε
pa, tqg εpa, tq
§ Aggregated death rate:
Weighted sum of the subpopulations death rates:
dεpa, tq “
ÿ
j
w εj pa, tqµ
εj pa, tq, w ε
j pa, tq “g εj pa, tq
g εpa, tq(1)
d depends non-linearly on the population inputs: g 0j , µj , and bj .
§ Even with time-independent rates µεj pa, �Atq
ñ the aggregate death rate dεpa, tq depends on time, due to changes
of composition of the heterogeneous population.
13
Numerical results
§ Two applications
1 Impact of the age-pyramid heterogeneity.ë Compare order of magnitude of mortality changes induced by
compositional changes to constant mortality improvements.
2 Cause specific mortality reduction vs “reverse” cohort effect.
ë Compensation of cause-specific mortality reduction due to adverse
compositional changes in some cohorts.
§ We consider a synthetic population composed of the most and least
deprived IMD quintile (illustrative purpose).
14
Three demographic scenarios
A Scenario A: Population evolution with time-invariant mortality
Compositional changes isolated ñ death rates in each subpopulation
do not depend on time:
dεpa, tq “ µε1paqwε1 pa, tq ` µ
ε5paqw
ε5 pa, tq.
B Scenario B: Population evolution with mortality improvement
Constant annual mortality improvement rates of r “ 0.5%:
dεpa, tq “ µε1paqp1´ rqtw ε1 pa, tq ` µ
ε5paqp1´ rqtw ε
5 pa, tq.
C Scenario C: Mortality improvements without composition changes
dεpa, tq “ µε1paqp1´ rqtw ε1 paq ` µ
ε5paqp1´ rqtw ε
5 paq.
Mortality rates and initial age-pyramid fitted to the data for year 1981 and
2015.
15
Mortality improvement rates (males)
Figure: Average annual mortality improvement rates over years 0-30
● ● ● ● ●
● ● ● ● ●● ● ● ● ●
●
●
●●
●
●
● ● ● ● ●
● ● ● ● ●● ● ● ● ●
●
●●
●●
●
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
−0.5%
−0.4%
−0.3%
−0.2%
−0.1%
0%
0.1%
0.2%
0.3%
0.4%
0.5%
0.6%
0.7%
0.8%
0.9%
65 70 75 80 85
Age
Ave
rage
d an
nual
impr
ovem
ent r
ates
Scenario ● ● ●A B C
(a) 1981 Inputs
●
●
●●
●
●
●
●
●
● ●●
●●
● ● ●● ●
●●
●
●
●●
●
●
●
●
●
● ●●
●●
● ● ●● ●
●●
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
−0.5%
−0.4%
−0.3%
−0.2%
−0.1%
0%
0.1%
0.2%
0.3%
0.4%
0.5%
0.6%
0.7%
0.8%
0.9%
65 70 75 80 85
Age
Ave
rage
d an
nual
impr
ovem
ent r
ates
Scenario ● ● ●A B C
(b) 2015 Inputs
16
Summary
§ 1981 initial population: positive contribution of changes in the
composition of the 65+ age class.
§ 2015 initial population: negative contribution of composition changes
ñ might offset future mortality improvement rates.
§ Order of magnitude of age-pyramid heterogeneity impact in annual
mortality improvement rates of 0.2%- 0.5%.
17
Composition changes and compensation effect
Example of scenario illustrating impact of changes of demographic rates:
§ Cause specific reduction of mortality vs “reverse” cohort effect (adverse
composition changes quantified by changes in birth patterns).
§ Baseline (“neutral”) scenario:
Death rates: time-invariant death rates in each population (year 2015).
Birth rates same birth rates in each population (England, 2015).
Initial pyramid our aim is to limit influence of initial pyramids
evolution: (((((((
((hhhhhhhhhSame initial pyramid ñ Stable pyramids.
18
Scenario 1a: cause of death reduction
Figure: Reduction of mortality rates from cardiovascular disease (CVD)
Males
0 10 20 30 40 50 60 70 80 90 10053.4
53.6
53.8
54
54.2
54.4
54.6
54.8
55
55.2
Time
Life
exp
ecta
ncy
at a
ge 2
5
Baseline scenario
Scenario 1a cause 2 (0%,−10%)
Scenario 1a cause 2 (0%,−20%)
Scenario 1a cause 2 (0%,−30%)
(a) Male period life expectancy at 25
Males
0 10 20 30 40 50 60 70 80 90 10053.4
53.6
53.8
54
54.2
54.4
54.6
54.8
55
55.2
Time
Life
exp
ecta
ncy
at a
ge 2
5
Baseline scenario
Scenario 1a cause 2 (0%,−10%)
Scenario 1a cause 2 (0%,−20%)
Scenario 1a cause 2 (0%,−30%)
(b) Male cohort life expectancy at 25
§ Reduction of CoD mortality from CVD (10, 20 and 30%) over a period
of 30 years, starting at t “ 40.19
Scenario 1b: “Reverse” Cohort effect
Males
0 10 20 30 40 50 60 70 80 90 10053.4
53.6
53.8
54
54.2
54.4
54.6
54.8
55
55.2
Time
Life
exp
ecta
ncy
at a
ge 2
5
Baseline scenario
Scenario 1b (20%,0%)
Scenario 1b (40%,0%)
Scenario 1b (60%,0%)
(a) Male period life expectancy at 25
Males
0 10 20 30 40 50 60 70 80 90 10053.4
53.6
53.8
54
54.2
54.4
54.6
54.8
55
55.2
TimeLi
fe e
xpec
tanc
y at
age
25
Baseline scenario
Scenario 1b (20%,0%)
Scenario 1b (40%,0%)
Scenario 1b (60%,0%)
(b) Male cohort life expectancy at 25
§ Reverse cohort effect: increase of birth rates in most deprived subgroup
over period r0, 20s.
§ Õ 60% ñ cohorts composed of 63% of most deprived subgroup.
20
Scenario 2: combined CoD reduction and cohort effect
Males
0 10 20 30 40 50 60 70 80 90 10053.4
53.6
53.8
54
54.2
54.4
54.6
54.8
55
55.2
Time
Life
exp
ecta
ncy
at a
ge 2
5
Baseline scenario
Scenario 1a cause 2 (0%,−20%)
Scenario 2 cause 2 (20%,−20%)
Scenario 2 cause 2 (40%,−20%)
Scenario 2 cause 2 (60%,−20%)
Figure: Male period life expectancy at 25
When the population heterogeneity is not taken into account, cause-of-death
mortality reduction could be compensated for and/or misinterpreted depending on
the population composition evolution.
21
Outline
1 Introduction
2 How can a cause-of-death reduction be compensated for by the population
heterogeneity? A dynamic approach.
3 Consistence of mortality forecast
22
The consistency issue
§ Population divided in p risk classes.
Example: national population divided in socioeconomic or geographical
subgroups...
§ Many forecasts are made simultaneously at the national level and at
the subgroup level.
§ Issue Are forecasts consistent?
(Aggregation relation) µxt “
pÿ
j“1
µjxtW
jxt ?
§ Problem when comparing policies, question robustness of forecasts.
23
Existing approaches
§ Relative approach (e.g.Jarner and Kryger (2011), Cairns et al. (2011),
Villegas and Haberman (2014), Li et al. (2015))
Method: First forecast of the national population and then mortality
differentials µjxt ´ µxt .
Advantages: national mortality data available for longer period, more
precise age-groups, estimations less volatile.
Drawback: No consistency!
§ Bottom-up approach:
Method: First forecast subnational mortality and then pose
µxt “řp
j“1 µjxtW
jxt .
Advantages: consistent forecasts.
Drawbacks: poor/missing data quality, small datasets.
§ Optimal combination : Combines the advantages of the two
approaches.
24
Existing approaches
§ Relative approach (e.g.Jarner and Kryger (2011), Cairns et al. (2011),
Villegas and Haberman (2014), Li et al. (2015))
Method: First forecast of the national population and then mortality
differentials µjxt ´ µxt .
Advantages: national mortality data available for longer period, more
precise age-groups, estimations less volatile.
Drawback: No consistency!
§ Bottom-up approach:
Method: First forecast subnational mortality and then pose
µxt “řp
j“1 µjxtW
jxt .
Advantages: consistent forecasts.
Drawbacks: poor/missing data quality, small datasets.
§ Optimal combination : Combines the advantages of the two
approaches.
24
Existing approaches
§ Relative approach (e.g.Jarner and Kryger (2011), Cairns et al. (2011),
Villegas and Haberman (2014), Li et al. (2015))
Method: First forecast of the national population and then mortality
differentials µjxt ´ µxt .
Advantages: national mortality data available for longer period, more
precise age-groups, estimations less volatile.
Drawback: No consistency!
§ Bottom-up approach:
Method: First forecast subnational mortality and then pose
µxt “řp
j“1 µjxtW
jxt .
Advantages: consistent forecasts.
Drawbacks: poor/missing data quality, small datasets.
§ Optimal combination : Combines the advantages of the two
approaches.
24
Optimal combination
§ Forecast reconciliation method used in economics, operation research
and recently for mortality rates (Shang and Hyndman (2017), Shang and
Haberman (2017)).
§ Vector notations: βnpxq “
¨
˚
˚
˝
µ1n
...
µpn
˛
‹
‹
‚
pxq, Ytpxq “
˜
µt
βt
¸
pxq.
§ Consistency equation:
Yn “Wnβn, Wn “
¨
˚
˚
˚
˚
˚
˝
W 1n W 2
n ¨ ¨ ¨ W pn
1 0 ¨ ¨ ¨ 0...
. . . 0
0 ¨ ¨ ¨ 1
˛
‹
‹
‹
‹
‹
‚
.
25
Optimal combination
Method:
1 First forecast mortality rates pYnqn“t..T using any method:
Yn ‰Wnβn.
2 Revise the subpopulations forecasts βn “tpµ1
n, ¨ ¨ ¨ , µpnq so that the
forecast Yn “Wnβn obtained by bottom-up is as close as possible as
the initial forecast Y .
Yn “Wnβn ` εn.
3 Replace Y by the consistent forecast Yn “Wnβn.
How can the relative exposures W jnpxq “
E jnpxq
Enpxqbe forecasted?
26
Optimal combination
Method:
1 First forecast mortality rates pYnqn“t..T using any method:
Yn ‰Wnβn.
2 Revise the subpopulations forecasts βn “tpµ1
n, ¨ ¨ ¨ , µpnq so that the
forecast Yn “Wnβn obtained by bottom-up is as close as possible as
the initial forecast Y .
Yn “Wnβn ` εn.
3 Replace Y by the consistent forecast Yn “Wnβn.
How can the relative exposures W jnpxq “
E jnpxq
Enpxqbe forecasted?
26
Population composition forecast
Proposed methods
§ Forecast Exposure to Risk at each age x pE jnpxqqn“t..T using
independent time series.
§ Constant Exposure to risk E jt`hpxq “ E j
t px ´ hq.
Reseach program
§ Population dynamics model provides a natural framework for the
projection of relative exposures, using all the age-pyramid information.
§ Adapt the bottom-up and optimal combination method in this context.
§ Challenge:
Yn “Wnppβsqs“t...nqβn,
27
Population composition forecast
Proposed methods
§ Forecast Exposure to Risk at each age x pE jnpxqqn“t..T using
independent time series.
§ Constant Exposure to risk E jt`hpxq “ E j
t px ´ hq.
Reseach program
§ Population dynamics model provides a natural framework for the
projection of relative exposures, using all the age-pyramid information.
§ Adapt the bottom-up and optimal combination method in this context.
§ Challenge:
Yn “Wnppβsqs“t...nqβn,
27
Conclusion and perspectives
Population dynamics framework ñ study of interactions between
changes in socioeconomic composition and mortality indicators:
§ Evidence from data: future evolution of the population socioeconomic
composition could have a different impact on mortality than 30y ago.
§ Cause-of-death mortality changes can be compensated and/or
misinterpreted in the presence of heterogeneity.
Perspectives
§ Take into account internal migrations: ñ non-linear mortality models.
§ Stochastic modelling
ë R package IBMPopSim in preparation,(with D. Giorgi and V. Lemaire).
§ Consistency of mortality forecasts under different demographic
scenarios.
28
Thank you for you attention!
29
Initial age-pyramids
0
5
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80
85
90
95
100+
19811976
19711966
19611956
19511946
1941
19361931
19261921
19161911
19061901
18961891
18861881
150 100 50 0 50 100 150Number of individuals (thousands)
Age
Year of birth
Type Males Females IMD Quintile 1 5
(a) 1981 Inputs
0
5
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80
85
90
95
100+
20152010
20052000
19951990
19851980
1975
19701965
19601955
19501945
19401935
19301925
19201915
150 100 50 0 50 100 150Number of individuals (thousands)
Age
Year of birth
Type Males Females IMD Quintile 1 5
(b) 2015 Inputs
30
Period life expectancy
Figure: Evolution of male life expectancy at age 65
65 60 55 50 45 40 35 30 25
13.0
13.1
13.2
13.3
13.4
13.5
13.6
13.7
13.8
13.9
14.0
14.1
14.2
14.3
14.4
14.5
14.6
14.7
0 5 10 15 20 25 30 35 40
Age at initial time
Time
Per
iod
life
expe
ctan
cy
Scenario A B C
(a) 1981 Inputs
65 60 55 50 45 40 35 30 25
17.9
18.0
18.1
18.2
18.3
18.4
18.5
18.6
18.7
18.8
18.9
19.0
19.1
19.2
19.3
19.4
19.5
19.6
19.7
19.8
19.9
0 5 10 15 20 25 30 35 40
Age at initial time
Time
Per
iod
life
expe
ctan
cy
Scenario A B C
(b) 2015 Inputs
§ Life expectancy at t “ 30: surviving individuals initially 35+31
Birth-Death-Swap processes
Pathwise construction of Birth-Death-Swap systems leading to an averaging result in presence of two
timescales, with N. El Karoui
§ Goal: To study the random evolution of an heterogeneous population
including:
A time-varying random environment.
Model changes in the population’s composition induced by interacting
individuals changing characteristics.
§ Main contributions:
General mathematical framework and tools to study such processes.
Study of the aggregated “macro” dynamic produced by such models.
§ Averaging result: aggregated mortality rates are approximated by
“averaged” rates depending non-trivially on the number of individuals
in the population.
32