Skip to contents

The package calculates health expectancies, and in particular disability-free life expectancies, with the Sullivan method. It also smooths prevalences of disability observed by age bracket, and projects prevalences and health expectancies. This vignette walks through a typical workflow, step by step:

  1. calculating life expectancies with and without disability with CompleteDFLEtable();
  2. converting prevalences observed by age bracket into prevalences by single year of age with ApproxPrevalenceTable();
  3. projecting prevalences under simple hypotheses on life expectancies with prevalenceForecast();
  4. projecting prevalences with models estimated on past observations, with fitPrevalenceModel(), predict() and projectPrevalence().

The methods of step 4 are described in detail in the vignette vignette("projection-methods", package = "healthexpectancies"). The last section lists the datasets included in the package.

1. Health expectancies with the Sullivan method

Indicators

Three life expectancies at age xx are used throughout the package:

Indicator Meaning Variable
total life expectancy expected number of years lived ex
disability-free life expectancy (DFLE) expected number of years lived without disability (espérance de vie sans incapacité, EVSI, in French) DFLEx
disability life expectancy (DLE) expected number of years lived with disability (espérance de vie en incapacité, EVI, in French) DLEx

They add up: ex = DFLEx + DLEx. The shares of DFLE and DLE in total life expectancy are pctDFLEx and pctDLEx (in percent).

The Sullivan method combines a life table with the prevalence of disability at each age (pix, a proportion between 0 and 1): the person-years lived at each age are split into years lived with and without disability according to the prevalence at this age, and then summed over the following ages.

CompleteDFLEtable()

CompleteDFLEtable() takes a data frame with at least the variables age, qx (probability of death between age xx and x+1x+1) or mx (central death rate), and pix. It returns the same data frame, completed with all the variables of the life table and the health expectancies. If pix is missing, only total life expectancy is calculated.

The package includes the first example of the Sullivan manual (sullivan), which CompleteDFLEtable() reproduces:

res <- CompleteDFLEtable(sullivan[, c("year", "age", "mx", "qx", "pix")])

# life expectancies at birth, at 65 and at 85
res[res$age %in% c(0, 65, 85), c("age", "ex", "DFLEx", "DLEx", "pctDFLEx")]
#>    age        ex     DFLEx      DLEx pctDFLEx
#> 1    0 81.415174 66.573158 14.842015 81.76996
#> 66  65 19.865979 12.295134  7.570845 61.89040
#> 86  85  5.371792  2.616062  2.755729 48.70000

The meaning of each variable of the output is given in the dataset description_sullivan:

description_sullivan
#>      heading                                                    description
#> 1       year                                                           Year
#> 2        age                                                            Age
#> 3         Px                                            Mid-year population
#> 4         Dx                                                     No. deaths
#> 5         mx                                             central death rate
#> 6         qx                               conditional probability of death
#> 7         lx                                     numbers surviving to age x
#> 8         Lx                                    person years lived at age x
#> 9         Tx                            total number of years lived  from x
#> 10        ex                                          total life expectancy
#> 11       pix                                     proportion with disability
#> 12      DFLx                          person years lived without disability
#> 13      DFTx                total years lived without disability from age x
#> 14     DFLEx                                disability-free life expectancy
#> 15  pctDFLEx                       proportion of life spent disability-free
#> 16       DLx                             person years lived with disability
#> 17      DLEx                                with disability life expectancy
#> 18   pctDLEx                       proportion of life spent with disability
#> 19   MeanDAx average conjonctural age of person years lived with disability
#> 20 MedianDAx  median conjonctural age of person years lived with disability
#> 21  ModalDAx   modal conjonctural age of person years lived with disability

Several series at once

If the input data frame contains the variables year and/or sex, life expectancies are calculated separately for each year and sex. Other grouping variables (for instance a region or a social category) can be given with the argument categories. This is used in the next section, with two years.

2. From age brackets to single years of age

Prevalences of disability are usually observed by age bracket only (in surveys, sample sizes are too small to measure them at each age), whereas the Sullivan method needs prevalences by single year of age. ApproxPrevalenceTable() converts prevalences by bracket into smoothed prevalences by single year of age. By default, it minimises the second differences of the logit of prevalences, under the constraint that the average prevalence of each bracket, weighted by the population, is equal to the observed one.

The example uses the prevalence of the APA (allocation personnalisée d’autonomie, the French benefit for older people with long-term care needs), observed by 5-year age bracket from 60 to 85 and over (FRAPAprevalence).

# Prevalences of the APA by age bracket, 2010 to 2024 (proportions)
apa <- FRAPAprevalence[FRAPAprevalence$decomp == 6, c("year", "age", "prevalence")]
head(apa)
#>   year age  prevalence
#> 1 2010  60 0.005174792
#> 2 2010  65 0.014897804
#> 3 2010  70 0.031050453
#> 4 2010  75 0.069224164
#> 5 2010  80 0.149172768
#> 6 2010  85 0.363697393

# Population by single year of age (both sexes), used as weights.
# The population is measured on 1 January: the population on 1 January of year Y
# is the population at the end of year Y-1, like the prevalences of the APA.
pop <- FRInseePopulation[FRInseePopulation$geo == "france" &
                           FRInseePopulation$age0101 %in% 60:99, ]
pop$year <- pop$year - 1
pop <- aggregate(popx ~ year + age0101, data = pop, FUN = sum)
names(pop) <- c("year", "age", "popx")
pix <- ApproxPrevalenceTable(apa,
                             agecuts     = seq(60, 85, 5),   # lower bounds of the brackets
                             agemin      = 60,
                             agemax      = 99,
                             weights.tab = pop,
                             categories  = "year")

# one row per year and single year of age, with the observed prevalence
# of the bracket (prevalence.raw) and the smoothed prevalence (prevalence.approx)
head(pix)
#>   year age agebracket prevalence.raw prevalence.approx
#> 1 2010  60    [60,65)    0.005174792       0.002991059
#> 2 2010  61    [60,65)    0.005174792       0.003846648
#> 3 2010  62    [60,65)    0.005174792       0.004941532
#> 4 2010  63    [60,65)    0.005174792       0.006328417
#> 5 2010  64    [60,65)    0.005174792       0.008052231
#> 6 2010  65    [65,70)    0.014897804       0.010127897
p19 <- pix[pix$year == 2019, ]

plot(p19$age, p19$prevalence.raw,
     type = "s", lwd = 2,
     xlab = "Age", ylab = "Prevalence of the APA")
lines(p19$age, p19$prevalence.approx, lwd = 2, col = "#0072B2")
legend("topleft", bty = "n", lwd = 2, col = c("black", "#0072B2"),
       legend = c("Observed, by age bracket", "Smoothed, by single year of age"))
Prevalence of the APA in 2019: observed by age bracket (steps) and smoothed by single year of age (line).

Prevalence of the APA in 2019: observed by age bracket (steps) and smoothed by single year of age (line).

Health expectancies of the APA

To calculate life expectancies, the prevalences are merged with mortality rates. FRInseeMortalityForecast2026 contains the mortality rates observed until 2022 and projected by Insee afterwards. The life table stops at age 104, where the probability of death is set to 1, and prevalences after age 99 are set equal to the prevalence at 99.

# keep year, age and smoothed prevalence
pix <- data.frame(year = as.numeric(pix$year),
                  age  = pix$age,
                  pix  = pix$prevalence.approx)

# extension to ages 100-104 (same prevalence as at age 99)
last <- pix[pix$age == 99, ]
pix  <- rbind(pix, do.call(rbind, lapply(100:104, function(a) transform(last, age = a))))

# mortality rates, both sexes, ages 60 to 104
mort <- FRInseeMortalityForecast2026
mort <- mort[mort$sex == "all" &
               mort$geo %in% c("France excluding Mayotte", "France") &
               mort$def.age == "current age (approx)" &
               mort$age %in% 60:104,
             c("year", "age", "qx")]
mort$qx[mort$age == 104] <- 1

# life expectancies in 2010 and 2019
tab <- merge(pix[pix$year %in% c(2010, 2019), ], mort, by = c("year", "age"))
le  <- CompleteDFLEtable(tab)

le[le$age %in% c(60, 75, 85), c("year", "age", "ex", "DFLEx", "DLEx", "pctDLEx")]
#>    year age        ex     DFLEx     DLEx   pctDLEx
#> 6  2010  60 24.803463 22.323771 2.479692  9.997363
#> 21 2010  75 13.088135 10.382902 2.705233 20.669351
#> 31 2010  85  6.867716  4.146822 2.720894 39.618619
#> 51 2019  60 25.624908 23.162815 2.462093  9.608203
#> 66 2019  75 13.848563 11.169316 2.679247 19.346750
#> 76 2019  85  7.255699  4.543023 2.712676 37.386828

Here, the “disability” is being a beneficiary of the APA: DLEx is the expected number of years spent as a beneficiary after age xx. Between 2010 and 2019, total life expectancy at 60 increased, while the expected number of years spent as a beneficiary remained stable: the share of life expectancy at 60 spent as a beneficiary decreased.

3. Projections under simple hypotheses on life expectancies

prevalenceForecast() projects prevalences from a reference year, given projected mortality rates and a hypothesis on the evolution of health expectancies:

  • "cstPreval": prevalences by age remain constant;
  • "cstDFLE": DFLE remains constant at each age, so that all the gains in life expectancy are spent with disability;
  • "cstDLE": DLE remains constant at each age, so that all the gains in life expectancy are spent without disability;
  • "cstPctDFLE" or "cstPctDLE": the share of DFLE (or of DLE) in total life expectancy remains constant at each age (the two hypotheses are equivalent).

With the hypotheses on life expectancies, prevalences are obtained by inverting the Sullivan method. At the youngest ages, where prevalences are very low, this inversion can require negative prevalences: the argument on_unfeasible = "clip" bounds them to the interval 0 to 1 (with a warning), and recalculates life expectancies accordingly.

# reference year 2019, projections for 2030 and 2040
tabref  <- tab[tab$year == 2019, ]
tabproj <- mort[mort$year %in% c(2030, 2040), ]

hypotheses <- c("cstPreval", "cstDFLE", "cstDLE", "cstPctDLE")

fc <- do.call(rbind, lapply(hypotheses, function(h) {
  f <- prevalenceForecast(tabref, tabproj, hypo = h, on_unfeasible = "clip")
  f <- f[f$age == 60 & f$year == 2040, c("year", "ex", "DFLEx", "DLEx")]
  data.frame(hypothesis = h, f)
}))

fc
#>      hypothesis year       ex    DFLEx     DLEx
#> 461   cstPreval 2040 27.06672 24.18326 2.883453
#> 4611    cstDFLE 2040 27.06672 23.16282 3.903902
#> 4612     cstDLE 2040 27.06672 24.59989 2.466832
#> 4613  cstPctDLE 2040 27.06672 24.46018 2.606536

The four hypotheses give the same total life expectancy at 60 in 2040, since mortality is the same, but different numbers of years spent as a beneficiary of the APA.

4. Projections with models estimated on past observations

Instead of setting a hypothesis directly, fitPrevalenceModel() estimates a model of the evolution of prevalences on past years, and predict() uses it to project prevalences. Four families of methods are available:

  • method = "trend": extrapolation of the trend of (the logit of) prevalences, at each age;
  • method = "mortality": link between prevalences and mortality rates (possibly with an age shift);
  • method = "target": hypotheses on life expectancies (constant value, linear trend, or change of a given percentage at a given horizon), estimated on past years;
  • method = "shift": shift of the curve of prevalences towards older ages at a constant speed.

projectPrevalence() estimates the model, projects prevalences and calculates life expectancies in one step. The example below extrapolates the trend of the logit of prevalences at each age, estimated on 2010-2019 (the years of the Covid-19 epidemic are left out):

The projected prevalences are used to calculate life expectancies, so the model must cover all the ages of the life table: here, it is estimated on ages 60 to 104 (prevalences after age 99 being equal to the prevalence at 99). Since the projection goes far beyond the estimation period, predict() warns that the trend is extrapolated over a long horizon (the warning is not displayed here).

# estimation table: prevalences by single year of age (60 to 104), 2010 to 2019
obs <- pix[pix$year %in% 2010:2019, ]

# projection years: mortality rates up to the last age of the life table
newdata <- mort[mort$year %in% c(2030, 2040), ]

proj <- projectPrevalence(obs, newdata = newdata, method = "trend", fit_years = 2010:2019)

proj[proj$age == 60, c("year", "ex", "DFLEx", "DLEx", "pctDLEx")]
#>    year       ex    DFLEx     DLEx  pctDLEx
#> 1  2030 26.31772 23.98168 2.336044 8.876315
#> 46 2040 27.06672 24.83155 2.235168 8.257996

The estimated model is kept as an attribute of the result, and its results are available with the usual methods (coef(), fitted(), residuals(), summary(), as.data.frame()):

model <- attr(proj, "model")
model
#> Model of prevalences by age, method "trend" (scale: logit)
#>   Years used for the estimation: 2010 to 2019 (10 years)
#>   Series: 1
#>   Ages: 60 to 104
#>   Estimator: ols
#> Use coef(), as.data.frame(, what = ) and predict() to retrieve results and project prevalences.

The assumptions of each method, their limits and a comparison on the example of the APA are given in vignette("projection-methods", package = "healthexpectancies").

5. Interactive example

The package includes a Shiny app, which illustrates projections of prevalences under the hypotheses of prevalenceForecast(). It requires the packages shiny, shinydashboard, plotly and tidyverse, which are not needed to install the package itself:

6. Datasets included in the package

All the datasets, except sullivan and description_sullivan, concern France. Their full description is available in their help pages (for instance ?FRAPAprevalence).

Dataset Content
sullivan, description_sullivan first example of the Sullivan manual, and description of the variables of the output of CompleteDFLEtable()
Prevalences
FRAPAprevalence prevalence of the APA by age bracket, all sexes, 2010-2024 (DREES, Aide sociale survey)
FRDreesAPA share of beneficiaries of the APA by sex and age bracket, at home and in institutions, 2016-2024 (DREES)
FRDreesAPA2017 share of beneficiaries of the APA by sex and age bracket, 2017 (DREES)
FRDreesVQSsurvey2014, FRDreesVQSsurvey2021 prevalences of disabilities by sex and age bracket, from the Vie quotidienne et santé (VQS) surveys of 2014 and 2021 (DREES)
FRGaliEUSilc prevalence of general activity limitations (GALI) by sex and age bracket (SRCV survey, French part of EU-SILC)
FRDreesEHPA share of people living in institutions for older people by sex and age (DREES, EHPA survey)
Population
FRInseePopulation population by sex and age on 1 January (Insee annual estimates)
FRInseePopulationForecast2016, FRInseePopulationForecast2021, FRInseePopulationForecast2026 population projections of Insee (central scenario)
Mortality
FRInseeMortalityrates, FRInseeMortalityrates_t69 observed mortality rates by sex and age (Insee)
FRInseeMortalityForecast2016, FRInseeMortalityForecast2021, FRInseeMortalityForecast2026 observed and projected mortality rates by sex and age, from the population projections of Insee (central scenario)