Learn R Programming

nimble (version 1.2.1)

waic: Using WAIC

Description

Details of the WAIC measure for comparing models. NIMBLE implements an online WAIC algorithm, computed during the course of the MCMC iterations.

Arguments

<code>controlWAIC</code> list

The controlWAIC argument is a list that controls the behavior of the WAIC algorithm and is passed to either configureMCMC or (if not using configureMCMC) buildMCMC. One can supply any of the following optional components:

online: Logical value indicating whether to calculate WAIC during the course of the MCMC. Default is TRUE and setting to FALSE is primarily for backwards compatibility to allow use of the old calculateWAIC method that calculates WAIC from monitored values after the MCMC finishes.

dataGroups: Optional list specifying grouping of data nodes, one element per group, with each list element containing the node names for the data nodes in that group. If provided, the predictive density values computed will be the joint density values, one joint density per group. Defaults to one data node per 'group'. See details.

marginalizeNodes: Optional set of nodes (presumably latent nodes) over which to marginalize to compute marginal WAIC (i.e., WAIC based on a marginal likelihood), rather than the default conditional WAIC (i.e., WAIC conditioning on all parent nodes of the data nodes). See details.

niterMarginal: Number of Monte Carlo iterations to use when marginalizing (default is 1000).

convergenceSet: Optional vector of numbers between 0 and 1 that specify a set of shorter Monte Carlo simulations for marginal WAIC calculation as fractions of the full (niterMarginal) Monte Carlo simulation. If not provided, NIMBLE will use 0.25, 0.50, and 0.75. NIMBLE will report the WAIC, lppd, and pWAIC that would have been obtained for these smaller Monte Carlo simulations, allowing assessment of the number of Monte Carlo samples needed for stable calculation of WAIC.

thin: Logical value for specifying whether to do WAIC calculations only on thinned samples (default is FALSE). Likely only useful for reducing computation when using marginal WAIC.

nburnin_extra: Additional number of pre-thinning MCMC iterations to discard before calculating online WAIC. This number is discarded in addition to the usual MCMC burnin, nburnin. The purpose of this option is to allow a user to retain some samples for inspection without having those samples used for online WAIC calculation (default = 0).

Extracting WAIC

The calculated WAIC and related quantities can be obtained in various ways depending on how the MCMC is run. If using nimbleMCMC and setting WAIC = TRUE, see the WAIC component of the output list. If using runMCMC and setting WAIC = TRUE, either see the WAIC component of the output list or use the getWAIC method of the MCMC object (in the latter case WAIC = TRUE is not required). If using the run method of the MCMC object, use the getWAIC method of the MCMC object.

The output of running WAIC (unless one sets online = FALSE) is a list containing the following components:

WAIC: The computed WAIC, on the deviance scale. Smaller values are better when comparing WAIC for two models.

lppd: The log predictive density component of WAIC.

pWAIC: The pWAIC estimate of the effective number of parameters, computed using the pWAIC2 method of Gelman et al. (2014).

To get further information, one can use the getWAICdetails method of the MCMC object. The result of running getWAICdetails is a list containing the following components:

marginal: Logical value indicating whether marginal (TRUE) or conditional (FALSE) WAIC was calculated.

niterMarginal: Number of Monte Carlo iterations used in computing marginal likelihoods if using marginal WAIC.

thin: Whether WAIC was calculated based only on thinned samples.

online: Whether WAIC was calculated during MCMC sampling.

nburnin_extra: Number of additional iterations discarded as burnin, in addition to original MCMC burnin.

WAIC_partialMC, lppd_partialMC, pWAIC_partialMC: The computed marginal WAIC, lppd, and pWAIC based on fewer Monte Carlo simulations, for use in assessing the sensitivity of the WAIC calculation to the number of Monte Carlo iterations.

niterMarginal_partialMC: Number of Monte Carlo iterations used for the values in WAIC_partialMC, lppd_partialMC, pWAIC_partialMC.

WAIC_elements, lppd_elements, pWAIC_elements: Vectors of individual WAIC, lppd, and pWAIC values, one element per data node (or group of nodes in the case of specifying dataGroups). Of use in computing the standard error of the difference in WAIC between two models, following Vehtari et al. (2017).

Online WAIC

As of version 0.12.0, NIMBLE provides enhanced WAIC functionality, with user control over whether to use conditional or marginal versions of WAIC and whether to group data nodes. In addition, users are no longer required to carefully choose MCMC monitors. WAIC by default is now calculated in an online manner (updating the required summary statistics at each MCMC iteration), using all post-burnin samples. The WAIC (Watanabe, 2010) is calculated from Equations 5, 12, and 13 in Gelman et al. (2014) (i.e., using 'pWAIC2').

Note that there is not a unique value of WAIC for a model. By default, WAIC is calculated conditional on the parent nodes of the data nodes, and the density values used are the individual density values of the data nodes. However, by modifying the marginalizeNodes and dataGroups elements of the control list, users can request a marginal WAIC (using a marginal likelihood that integrates over user-specified latent nodes) and/or a WAIC based on grouping observations (e.g., all observations in a cluster) to use joint density values. See the MCMC Chapter of the NIMBLE User Manual for more details.

For more detail on the use of different predictive distributions, see Section 2.5 from Gelman et al. (2014) or Ariyo et al. (2019).

Note that based on a limited set of simulation experiments in Hug and Paciorek (2021) our tentative recommendation is that users only use marginal WAIC if also using grouping.

Author

Joshua Hug and Christopher Paciorek

Details

To obtain WAIC, set WAIC = TRUE in nimbleMCMC. If using a more customized workflow, set enableWAIC = TRUE in configureMCMC or (if skipping configureMCMC) in buildMCMC, followed by setting WAIC = TRUE in runMCMC, if using runMCMC to manage sample generation.

By default, NIMBLE calculates WAIC using an online algorithm that updates required summary statistics at each post-burnin iteration of the MCMC.

One can also use calculateWAIC to run an offline version of the WAIC algorithm after all MCMC sampling has been done. This allows calculation of WAIC from a matrix (or dataframe) of posterior samples and also retains compatibility with WAIC in versions of NIMBLE before 0.12.0. However, the offline algorithm is less flexible than the online algorithm and only provides conditional WAIC without the ability to group data points. See help(calculateWAIC) for details.

References

Watanabe, S. (2010). Asymptotic equivalence of Bayes cross validation and widely applicable information criterion in singular learning theory. Journal of Machine Learning Research 11: 3571-3594.

Gelman, A., Hwang, J. and Vehtari, A. (2014). Understanding predictive information criteria for Bayesian models. Statistics and Computing 24(6): 997-1016.

Ariyo, O., Quintero, A., Munoz, J., Verbeke, G. and Lesaffre, E. (2019). Bayesian model selection in linear mixed models for longitudinal data. Journal of Applied Statistics 47: 890-913.

Vehtari, A., Gelman, A. and Gabry, J. (2017). Practical Bayesian model evaluation using leave-one-out cross-validation and WAIC. Statistics and Computing 27: 1413-1432.

Hug, J.E. and Paciorek, C.J. (2021). A numerically stable online implementation and exploration of WAIC through variations of the predictive density, using NIMBLE. arXiv e-print <arXiv:2106.13359>.

See Also

calculateWAIC configureMCMC buildMCMC runMCMC nimbleMCMC

Examples

Run this code
code <- nimbleCode({
  for(j in 1:J) {
    for(i in 1:n) 
      y[j, i] ~ dnorm(mu[j], sd = sigma)
    mu[j] ~ dnorm(mu0, sd = tau)
  }
  sigma ~ dunif(0, 10)
  tau ~ dunif(0, 10)
})
J <- 5
n <- 10
groups <- paste0('y[', 1:J, ', 1:', n, ']') 
y <- matrix(rnorm(J*n), J, n)
Rmodel <- nimbleModel(code, constants = list(J = J, n = n), data = list(y = y),
                      inits = list(tau = 1, sigma = 1))

## Various versions of WAIC available via online calculation.
## Conditional WAIC without data grouping:
conf <- configureMCMC(Rmodel, enableWAIC = TRUE)
## Conditional WAIC with data grouping
conf <- configureMCMC(Rmodel, enableWAIC = TRUE, controlWAIC = list(dataGroups = groups))
## Marginal WAIC with data grouping:
conf <- configureMCMC(Rmodel, enableWAIC = TRUE, controlWAIC =
            list(dataGroups = groups, marginalizeNodes = 'mu'))
if (FALSE) {
Rmcmc <- buildMCMC(conf)
Cmodel <- compileNimble(Rmodel)
Cmcmc <- compileNimble(Rmcmc, project = Rmodel)
output <- runMCMC(Cmcmc, niter = 1000, WAIC = TRUE)
output$WAIC              # direct access
## Alternatively call via the `getWAIC` method; this doesn't require setting
## `waic=TRUE` in `runMCMC`
Cmcmc$getWAIC()          
Cmcmc$getWAICdetails()
}

Run the code above in your browser using DataLab