rbi.helpers is collection of helper functions to use with rbi, an R interface to LibBi, a library for Bayesian Inference.
This vignette builds on the rbi vignette, applying the higher-level functions contained in rbi.helpers to the same model introduced there. For the lower-level functions to run LibBi and read the results, please refer to the documentation and vignette that comes with rbi.
The easiest way to install the latest stable version of
rbi.helpers is via CRAN. The package name is
rbi.helpers
(all lowercase):
Alternatively, the current development version can be installed using
the remotes
package
Most functions in the rbi.helpers package require a working installation of LibBi. Please see the rbi vignette for how to get one via homebrew or linuxbrew.
These steps are reproduced from the rbi vignette, where there is more information on the individual steps
In the rbi
vignette, a stochastic
SIR model is fitted to simulated data from the same model using
particle Markov-chain Monte Carlo with 32 particles. Given a model and
data, how do we know how many particles we need? This question does not
have a simple answer, as the “optimal” number of particles may depend on
the state of the Markov chain. A possible rule-of-thumb is to choose the
number of particles such that the variance of the log-likelihood near
the mode is approximately one. This suggests a strategy
by which first and approximate location of the mode or mean of the
posterior distribution is obtained in a trial run, before the numer of
particles is adjusted by monitoring the variance of the log-likelihood
while keeping the parameters fixed. rbi.helpers
implements the second part of this strategy (adjusting the number of
particles at a given location in parameter space) with the
adapt_particles
method. For the first part (finding the
mode), a crude method is to take a fixed number of samples from the
prior distribution and choose the one that maximises the posterior. In
rbi, this can be achieved with
bi_prior <- sample(
proposal = "prior", sir_model, nsamples = 1000, end_time = 16 * 7,
nparticles = 4, obs = sir_data, seed = 1234
)
This runs particle MCMC with the prior distribution as proposal
distribution (because proposal = "prior"
), in this case
with 4 particles. In other words, when sampling from the posterior the
proposals will be drawn independently from the prior distribution. Note
that we set a seed to make the results reproducible. It is worth trying
the commands with a different seed and seeing the difference to the
results obtained below. The location in parameters of the sampler at the
end of the 1000 iterations will give an approximation of the mode of the
posterior distribution. This can then be used to adjust the number of
particles using
This will take the last sample of the output file contained in the
libbi
object bi_prior
, and use it to adjust
the number of particles by starting with 1 particle (or a given
min
) and doubling it until the variance of the
loglikelihood crosses 1. The number of particles is then saved in the
adapted
object:
Having adjusted the number of particles, the second important
information to give the posterior sampler is the proposal distribution.
This can, again, be obtained using a sequence of trial runs, whereby the
proposal distribution is sequentially adjusted from previous samples to
be proportional to the empirical covariance of the posterior samples.
The way this is implemented in the adapt_proposal
function
in rbi.helpers is that the proposal distribution is
adjusted to come from a multivariate normal taking into account the
covariance of samples obtained so far, until the acceptance rate lies
between the given minimum and maximumad. For example, to adjust the
proposal distribution for an acceptance rate between 0.05 and 0.4, we
can run:
The covariance matrices for parameters and initial conditions are
stored in the input file contained in the libbi
object
adapted
To compute the Deviance
Information Criterion (DIC), use DIC
:
This samples from the posterior distribution using the adapted number of particles and proposal distribution and then calculates the DIC.
LibBi uses real-valued times. To convert these to time or date
objects for use in R, use the numeric_to_time
function:
The function time_to_numeric
does the converse,
converting R times or dates into numeric values for use by LibBi:
With the pipe operator available since R version 4.1, it is possible to construct inference chains more elegantly. For example, to get posterior samples from adapted Metropolis-Hastings including sampled observations, we could have written