Skip to content

vkontis/maple

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

39 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

maple

Welcome to maple: model averaged projections of life expectancy.

maple is an R package for fitting statistical models on demographic data to produce projections of death rates and life expectancy. It contains the code used in V. Kontis, J.E. Bennett, C.D. Mathers, G. Li, K. Foreman, M. Ezzati, Future life expectancy in 35 industrialised countries: projections with a Bayesian model ensemble, Lancet (2017), doi: 10.1016/S0140-6736(16)32381-9.

Prerequisites

The code uses the INLA algorithm to fit most of the models, implemented in the R package INLA. More information on how to download and install it can be found on the INLA website. Note that you may need to install package sp first by running install.packages("sp").

Installation

The easiest way to install maple is using the devtools package:

devtools::install_github("vkontis/maple")

If you do not have devtools, you can install it from CRAN, by typing install.packages("devtools").

Alternatively, clone or download the repository to a local directory and install it by running install.packages(<local_dir>, repos = NULL, type = "source").

Running the models

Getting some data

You will need some data on death counts and population. These need to be formatted into matrices, with one row for each age group and one column for each year of data. For now, maple only handles 5-year age groups 0-4, 5-9, ..., 80-84 and 85+ years, so it expects matrices with 18 rows. The column names of these matrices must be set to the years of data. The package comes with example data, stored in maple.deaths and maple.population:

> data(maple.deaths)
> maple.deaths[1:5, 49:53]
         2008       2009       2010       2011       2012
0   1717.9084  1669.4182  1613.4061  1587.3184  1516.6746
5    144.5131   127.9170   128.8043   112.0376   129.0808
10   148.1691   161.2160   146.1746   127.1724   131.4167
15   509.2995   485.2724   435.0701   368.0076   351.2908
20   729.2682   666.7927   649.6085   604.0687   564.8684

There is another matrix, maple.ax, which can optionally be provided to the model fitting methods, and holds the number of years lived on average in each age group, by those who die in that age group (denoted nax in demography). We will not worry about this for now, more info can be found in the help file ?maple_plt.

Choosing which models to fit

The models contained in the package can be obtained by typing models <- maple_models(). This will return a list of models, all or a subset of which can be used in the model fitting methods. More details on the models can be found in the published paper and its appendix. To choose a subset of models, run models.to.run <- maple_models()[c("RW1AGE", "LC_1PC")].

Running individual models

1) Running a single model
fit <- maple_fit_model(models.to.run$LC_1PC,
                       deaths = maple.deaths,
                       population = maple.population,
                       forecast.horizon = 10)

This will fit a Lee-Carter model to the data and return an object containing information about the fit, including the estimated model parameters and the estimated death rates

>fit$rates[1:5, 49:53]
           2008         2009         2010         2011         2012
0  9.465266e-04 8.793542e-04 8.290460e-04 7.423325e-04 7.234237e-04
5  8.532998e-05 7.960949e-05 7.530884e-05 6.786028e-05 6.622960e-05
10 1.052803e-04 1.000591e-04 9.606762e-05 8.900666e-05 8.743379e-05
15 2.839088e-04 2.739918e-04 2.663051e-04 2.524669e-04 2.493400e-04
20 3.881237e-04 3.787097e-04 3.713404e-04 3.579050e-04 3.548378e-04
2) Running an ensemble of models and calculating posterior samples
fits <- maple_fit_ensemble(deaths = maple.deaths,
                           population = maple.population,
                           forecast.horizon = 10,
                           models = models.to.run)

This will fit the models currently listed in models.to.run to calculate some posterior samples for age-specific death rates, life expectancy and probability of dying, and statistical summaries from these samples. Samples for each individual model are stored in fits$samples. For example, fits$samples$RW1AGE is a list of length 1000 (the default number of samples) each element of which holds a life table draw, for example the 12th element:

> head(fits$samples$RW1AGE[[12]])
  age year           mx          qx       ex
1   0 1960 0.0050361105 0.024622544 73.29968
2   5 1960 0.0003788656 0.001892535 70.13744
3  10 1960 0.0002911676 0.001454779 65.26569
4  15 1960 0.0005034493 0.002514267 60.35713
5  20 1960 0.0006476626 0.003233193 55.50260
6  25 1960 0.0006784103 0.003386501 50.67434

Statistical summaries calculated from these samples are stored as a data frame under fits$sample.summaries, including the mean, median, standard deviation and 95% credible interval:

> head(fits$sample.summaries)
   model year age metric         mean       median           sd           lb           ub
1 RW1AGE 1960   0   rate 0.0049698784 0.0049693898 4.805665e-05 0.0048756943 0.0050636170
2 RW1AGE 1960   5   rate 0.0003692761 0.0003690454 9.588001e-06 0.0003510777 0.0003873282
3 RW1AGE 1960  10   rate 0.0002847784 0.0002848100 7.571953e-06 0.0002696875 0.0002997572
4 RW1AGE 1960  15   rate 0.0005147145 0.0005144855 1.246367e-05 0.0004909577 0.0005385888
5 RW1AGE 1960  20   rate 0.0006396528 0.0006395553 1.376471e-05 0.0006151523 0.0006686139
6 RW1AGE 1960  25   rate 0.0006925368 0.0006929623 1.591788e-05 0.0006625961 0.0007242749

The values in the age column 0, 5, ..., 80, 85 correspond to age groups 0-4, 5-9, ..., 80-84, 85+. The metrics are:

  • "rate": The death rate.
  • "ex": Life expectancy at the beginning of the age group in the age column; e.g. when age == 0 this is life expectancy at birth, and when age == 65 life expectancy at age 65 years.
  • "qx": Probability of dying by the end of the age group in the age column; e.g. when age == 65 (which corresponds to the age group 65-69 years), this is the probability of dying between birth and age 70 years (this probability is always 1 for the open ended age group 85+).

The individual model fits are stored under fits$model.fits.

Forecasting and model averaging

bma <- maple(deaths = maple.deaths, population = maple.population,
              forecast.horizon = 10, holdout = 13, models = models.to.run,
              num.draws = 5000)

This will fit all selected models twice: once to calculate model weights and once to produce the projections. The first run uses all but the last 13 years of data (specified in the holdout argument) to estimate model parameters and project them forward. The projection error over these 13 years is used to calculate model weights. The second run uses all the data to fit the models and produce forecasts for the next 10 years (specified in the forecast.horizon argument). The weights from the first run are then used to combine the models into a model average.

Due to the multimodality and the usually large variance of the posterior model average, a large number of samples (at least 1,000 and preferably 5,000 as specified in the num.draws argument) is recommended.

Model weights are stored under bma$model.weights:

> bma$model.weights
  RW1AGE    LC_1PC
0.5320914 0.4679086

Model average posterior samples for death rates, life expectancy and probability of dying are stored under bma$sample.summaries, while those of individual models are stored under bma$individual.model.sample.summaries.

Questions

Please feel free to open an issue with any questions you may have.

License

The maple code is available under GPLv3. INLA uses external libraries which come with different software licenses (more info on the INLA website).

About

maple: Model averaged projections of life expectancy

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages