Runs (or continues running) MCMCs for simulating the total fertility rate of all countries of the world (phase II), using a Bayesian hierarchical model.
run.tfr.mcmc(nr.chains = 3, iter = 62000,
output.dir = file.path(getwd(), "bayesTFR.output"),
thin = 1, replace.output = FALSE,
start.year = 1950, present.year = 2015, wpp.year = 2017,
my.tfr.file = NULL, my.locations.file = NULL, buffer.size = 100,
U.c.low = 5.5, U.up = 8.8, U.width = 3,
mean.eps.tau0 = -0.25, sd.eps.tau0 = 0.4, nu.tau0 = 2,
Triangle_c4.low = 1, Triangle_c4.up = 2.5,
Triangle_c4.trans.width = 2,
Triangle4.0 = 0.3, delta4.0 = 0.8, nu4 = 2,
S.low = 3.5, S.up = 6.5, S.width = 0.5,
a.low = 0, a.up = 0.2, a.width = 0.02,
b.low = a.low, b.up = a.up, b.width = 0.05,
sigma0.low = 0.01, sigma0.up = 0.6, sigma0.width = 0.1,
sigma0.min = 0.001,
const.low = 0.8, const.up = 2, const.width = 0.3,
d.low = 0.05, d.up = 0.5, d.trans.width = 1,
chi0 = -1.5, psi0 = 0.6, nu.psi0 = 2,
alpha0.p = c(-1, 0.5, 1.5), delta0 = 1, nu.delta0 = 2,
dl.p1 = 9, dl.p2 = 9,
S.ini = NULL, a.ini = NULL, b.ini = NULL, sigma0.ini = NULL,
Triangle_c4.ini = NULL, const.ini = NULL, gamma.ini = 1,
proposal_cov_gammas = NULL,
seed = NULL, parallel = FALSE, nr.nodes = nr.chains,
save.all.parameters = FALSE, compression.type = 'None',
auto.conf = list(max.loops=5, iter=62000, iter.incr=10000,
nr.chains=3, thin=80, burnin=2000),
verbose = FALSE, verbose.iter = 10, …)
continue.tfr.mcmc(iter, chain.ids=NULL,
output.dir=file.path(getwd(), "bayesTFR.output"),
parallel = FALSE, nr.nodes = NULL, auto.conf = NULL,
verbose=FALSE, verbose.iter = 10, …)
Number of MCMC chains to run.
Number of iterations to run in each chain. In addition to a single value, it can have the value ‘auto’ in which case the function runs for the number of iterations given in the auto.conf
list (see below), then checks if the MCMCs converged (using the auto.conf
settings). If it did not converge, the procedure is repeated until convergence is reached or the number of repetition exceeded auto.conf$max.loops
.
Directory which the simulation output should be written into.
Thinning interval between consecutive observations to be stored on disk.
If TRUE
, existing outputs in output.dir
will be replaced by results of this simulation.
Start year for using historical data.
End year for using historical data.
Year for which WPP data is used. The functions loads a package called wpp\(x\) where \(x\) is the wpp.year
and uses the tfr*
datasets.
File name containing user-specified TFR time series for one or more countries. See Details below.
File name containing user-specified locations. See Details below.
Buffer size (in number of iterations) for keeping data in the memory. The smaller the buffer.size
the more often will the process access the hard disk and thus, the slower the run. On the other hand, the smaller the buffer.size
the less data will be lost in case of failure.
Lower and upper bound of the parameter \(U_c\), the start level of the fertility transition. The lower bound is set for each country as the maximum of U.c.low
and the minimum of historical TFR for that country.
Width for slice sampling used when updating the \(U_c\) parameter.
Mean and standard deviation of the prior distribution of \(m_{\tau}\) which is the mean of the distortion terms \(\epsilon_{c,\tau}\) in start periods \(\tau_c\).
The shape parameter of the prior Gamma distribution of \(1/s_{\tau}^2\) is nu.tau0
/2. \(s_{\tau}\) is standard deviation of the distortion terms in start periods \(\tau_c\).
Lower and upper bound of the \(\Delta_{c4}\) parameter.
Width for slice sampling used when updating the logit-transformed \(\Delta_{c4}\) parameter.
Mean and standard deviation of the prior distribution of the \(\Delta_4\) parameter which is the hierarchical mean of the logit-transformed \(\Delta_{c4}\).
The shape parameter of the prior Gamma distribution of \(1/\delta_4^2\) is nu4
/2. \(\delta_4\) is standard deviation of the logit-transformed \(\Delta_{c4}\).
Lower and upper bound of the uniform prior distribution of the \(S\) parameter which is the TFR at which the distortion has maximum variance.
Width for slice sampling used when updating the \(S\) parameter.
Lower and upper bound of the uniform prior distribution of the \(a\) parameter which is a coefficient for linear decrease of the TFR for TFR larger than \(S\).
Width for slice sampling used when updating the \(a\) parameter.
Lower and upper bound of the uniform prior distribution of the \(b\) parameter which is a coefficient for linear decrease of the TFR for TFR smaller than \(S\).
Width for slice sampling used when updating the \(b\) parameter.
Lower and upper bound of the uniform prior distribution of the \(\sigma_0\) parameter. \(\sigma_0^2\) is the maximum variance of the distortions at TFR equals \(S\).
Width for slice sampling used when updating the \(\sigma_0\) parameter.
Minimum value that \(\sigma_0\) can take.
Lower and upper bound of the uniform prior distribution of the \(c\) parameter which is the multiplier of standard deviation of the distortions before 1975.
Width for slice sampling used when updating the \(c\) parameter.
Lower and upper bound of the parameter \(d_c\), the maximum annual decrement for country \(c\). (Note that in Alkema et al. this parameter is a five-years decrement.)
Width for slice sampling used when updating the logit-transformed \(d_c\) parameter.
Prior mean and standard deviation of the \(\chi\) parameter which is the hierarchical mean of logit-transformed maximum decline parameter \(d_c\).
The shape parameter of the prior Gamma distribution of \(1/\psi^2\) is nu.psi0
/2. \(\psi\) is the standard devation of logit-transformed maximum decline parameter \(d_c\).
Vector of prior means of the \(\alpha_i\) parameters, \(i=1,2,3\). \(\alpha_i\) is the hierarchical mean of \(\gamma_{ci}\).
Prior standard deviation of the \(\alpha_i\) parameters. It is a single value, i.e. the same standard deviation is used for all \(i\).
The shape parameter of the prior Gamma distribution of \(1/\delta_i^2\) is nu.delta0
/2. \(\delta_i\) is the standard deviation of \(\gamma_{ci}\).
Values of the parameters \(p_1\) and \(p_2\) of the double logistic function.
Initial value for the \(S\) parameter. It can be a single value or an array of the size nr.chains
. By default, if nr.chains
is 1, it is the middle point of the interval [S.low, S.up
]. Otherwise, it is equally spaced distributed between S.low
and S.up
.
Initial value for the \(a\) parameter. It can be a single value or an array of the size nr.chains
. By default, if nr.chains
is 1, it is the middle point of the interval [a.low, a.up
]. Otherwise, it is equally spaced distributed between a.low
and a.up
.
Initial value for the \(b\) parameter. It can be a single value or an array of the size nr.chains
. By default, if nr.chains
is 1, it is the middle point of the interval [b.low, b.up
]. Otherwise, it is equally spaced distributed between b.low
and b.up
.
Initial value for the \(\sigma_0\) parameter. It can be a single value or an array of the size nr.chains
. By default, if nr.chains
is 1, it is the middle point of the interval [sigma0.low, sigma0.up
]. Otherwise, it is equally spaced distributed between sigma0.low
and sigma0.up
.
Initial value for the \(\Delta_{c4}\) parameter. It can be a single value or an array of the size nr.chains
. By default, if nr.chains
is 1, it is the middle point of the interval [Triangle_c4.low, Triangle_c4.up
]. Otherwise, it is equally spaced distributed between Triangle_c4.low
and Triangle_c4.up
.
Initial value for the \(c\) parameter. It can be a single value or an array of the size nr.chains
. By default, if nr.chains
is 1, it is the middle point of the interval [const.low, const.up
]. Otherwise, it is equally spaced distributed between const.low
and const.up
.
Initial value for the \(\gamma_c\) parameter. The same value is used for all countries.
Proposal for the gamma covariance matrices for each country. It should be a list with two values: values
and country_codes
. The structure corresponds to the object returned by the function get.cov.gammas
. By default the covariance matrices are obtained using data(proposal_cov_gammas_cii)
. This argument overwrite the defaults for countries contained the argument.
Seed of the random number generator. If NULL
no seed is set. It can be used to generate reproducible results.
Logical determining if the simulation should run multiple chains in parallel. If it is TRUE
, the package snowFT is required. In such a case further arguments can be passed, see description of ….
Relevant only if parallel
is TRUE
. It gives the number of nodes for running the simulation in parallel. By default it equals to the number of chains.
If TRUE
, the distortion terms \(\epsilon_{c,t}\) for all \(t\) are stored on disk, otherwise not.
One of ‘None’, ‘gz’, ‘xz’, ‘bz’, determining type of a compression of the MCMC files. Important: Do not use this option for a long MCMC simulation as this tends to cause very long run times due to slow reading!
List containing a configuration for an ‘automatic’ run (see description of argument iter
). Item iter
gives the number of iterations in the first chunk of the MCMC simulation; item iter.incr
gives the number of iterations in the following chunks; nr.chains
gives the number of chains in all chunks of the MCMC simulation; items thin
and burnin
are used in the convergence diagnostics following each chunk; max.loops
controls the maximum number of chunks. All items must be integer values. This argument is only used if the function argument iter
is set to ‘auto’.
Logical switching log messages on and off.
Integer determining how often (in number of iterations) log messages are outputted during the estimation.
Additional parameters to be passed to the function performParallel
, if parallel
is TRUE
. For example cltype
which is ‘SOCK’ by default but can be set to ‘MPI’.
Array of chain identifiers that should be resumed. If it is NULL
, all existing chains in output.dir
are resumed.
An object of class bayesTFR.mcmc.set
which is a list with two components:
An object of class bayesTFR.mcmc.meta
.
A list of objects of class bayesTFR.mcmc
, one for each MCMC.
The function run.tfr.mcmc
creates an object of class bayesTFR.mcmc.meta
and stores it in output.dir
. It launches nr.chains
MCMCs, either sequentially or in parallel. Parameter traces of each chain are stored as (possibly compressed) ASCII files in a subdirectory of output.dir
, called mc
x where x is the identifier of that chain. There is one file per parameter, named after the parameter with the suffix “.txt”, possibly followed by a compression suffix if compression.type
is given. Country-specific parameters (\(U, d, \gamma\)) have the suffix _c
y where y is the country code. In addition to the trace files, each mc
x directory contains the object bayesTFR.mcmc
in binary format. All chain-specific files are written into disk after the first, last and each buffer.size
-th iteration.
Using the function continue.tfr.mcmc
one can continue simulating an existing MCMCs by iter
iterations for either all or selected chains.
The function loads observed data (further denoted as WPP dataset) from the tfr
and tfr_supplemental
datasets in a wpp\(x\) package where \(x\) is the wpp.year
. It is then merged with the include
dataset that corresponds to the same wpp.year
. The argument my.tfr.file
can be used to overwrite those default data. Such a file can include a subset of countries contained in the WPP dataset, as well as a set of new countries. In the former case,
the function replaces the corresponding country data from the WPP dataset by values in this file. Only columns are replaced that match column names of the WPP dataset, and in addition, columns ‘last.observed’ and ‘include_code’ are used, if present. Countries are merged with WPP using the column ‘country_code’. In addition, in order the countries to be included in the simulation, in both cases (whether they are included in the WPP dataset or not), they must be contained in the table of locations (UNlocations
). In addition, their corresponding include_code
must be set to 2. If the column ‘include_code’ is present in my.tfr.file
, its value overwrites the default include code, unless is -1.
The default UN table of locations mentioned above can be overwritten/extended by using a file passed as the my.locations.file
argument. Such a file must have the same structure as the UNlocations
dataset. Entries in this file will overwrite corresponding entries in UNlocations
matched by the column ‘country_code’. If there is no such entry in the default dataset, it will be appended. This option of appending new locations is especially useful in cases when my.tfr.file
contains new countries/regions that are not included in UNlocations
. In such a case, one must provide a my.locations.file
with a definition of those countries/regions.
For simulation of the hyperparameters of the Bayesian hierarchical model, all countries are used that are included in the WPP dataset, possibly complemented by the my.tfr.file
, that have include_code
equal to 2. The hyperparameters are used to simulate country-specific parameters, which is done for all countries with include_code
equal 1 or 2. The following values of include_code
in my.tfr.file
are recognized: -1 (do not overwrite the default include code), 0 (ignore), 1 (include in prediction but not estimation), 2 (include in both, estimation and prediction). Thus, the set of countries included in the estimation and prediction can be fully user-specific.
Optionally, my.tfr.file
can contain a column called last.observed
containing the year of the last observation for each country. In such a case, the code would ignore any data after that time point. Furthermore, the function tfr.predict
fills in the missing values using the median of the BHM procedure (stored in tfr_matrix_reconstructed
of the bayesTFR.prediction
object). For last.observed
values that are below a middle year of a time interval \([t_i, t_{i+1}]\) (computed as \(t_i+3\)) the last valid data point is the time interval \([t_{i-1}, t_i]\), whereas for values larger equal a middle year, the data point in \([t_i, t_{i+1}]\) is valid.
The package contains a dataset called my_tfr_template
(in extdata
directory) which is a template for user-specified my.tfr.file
.
L. Alkema, A. E. Raftery, P. Gerland, S. J. Clark, F. Pelletier, Buettner, T., Heilig, G.K. (2011). Probabilistic Projections of the Total Fertility Rate for All Countries. Demography, Vol. 48, 815-839.
# NOT RUN {
m <- run.tfr.mcmc(nr.chains=1, iter=5, verbose=TRUE)
summary(m)
m <- continue.tfr.mcmc(iter=5, verbose=TRUE)
summary(m)
# }
Run the code above in your browser using DataLab