Skip to contents

Uses results from MBNMA JAGS models to calculate pD via the plugin method (Spiegelhalter et al. 2002) . Can only be used for models with known standard errors or covariance matrices (typically univariate).

Usage

pDcalc(
  obs1,
  obs2,
  fups = NULL,
  narm,
  NS,
  theta.result,
  resdev.result,
  likelihood = "normal",
  type = "time"
)

Arguments

obs1

A matrix (study x arm) or array (study x arm x time point) containing observed data for y (normal likelihood) or r (binomial or Poisson likelihood) in each arm of each study. This will be the same array used as data for the JAGS model.

obs2

A matrix (study x arm) or array (study x arm x time point) containing observed data for se (normal likelihood), N (binomial likelihood) or E (Poisson likelihood) in each arm of each study. This will be the same array used as data for the JAGS model.

fups

A numeric vector of length equal to the number of studies, containing the number of follow-up mean responses reported in each study. Required for time-course MBNMA models (if type="time")

narm

A numeric vector of length equal to the number of studies, containing the number of arms in each study.

NS

A single number equal to the number of studies in the dataset.

theta.result

A matrix (study x arm) or array (study x arm x time point) containing the posterior mean predicted means/probabilities/rate in each arm of each study. This will be estimated by the JAGS model.

resdev.result

A matrix (study x arm) or array (study x arm x time point) containing the posterior mean residual deviance contributions in each arm of each study. This will be estimated by the JAGS model.

likelihood

A character object of any of the following likelihoods:

  • univariate

  • binomial (does not work with time-course MBNMA models)

  • multivar.normal (does not work with time-course MBNMA models)

type

The type of MBNMA model fitted. Can be either "time" or "dose"

Value

A numeric value for the effective number of parameters, pD, calculated via the plugin method

Details

Method for calculating pD via the plugin method proposed by (Spiegelhalter et al. 2002) . Standard errors / covariance matrices must be assumed to be known. To obtain values for theta.result and resdev.result these parameters must be monitored when running the JAGS model.

For non-linear time-course MBNMA models residual deviance contributions may be skewed, which can lead to non-sensical results when calculating pD via the plugin method. Alternative approaches are to use pV (pv) as an approximation (Plummer 2008) or pD calculated by Kullback–Leibler divergence (pd.kl) or using an optimism adjustment (popt) (Plummer 2008) .

References

TO ADD pV REF

Examples

# \donttest{
# Using the alogliptin dataset
network <- mb.network(alog_pcfb)
#> Reference treatment is `placebo`
#> Studies reporting change from baseline automatically identified from the data

# Run Emax model saving predicted means and residual deviance contributions
emax <- mb.run(network, fun=temax(),
  parameters.to.save=c("theta", "resdev"), intercept=FALSE)
#> 'et50' parameters must take positive values.
#>  Default half-normal prior restricts posterior to positive values.
#> Compiling model graph
#>    Resolving undeclared variables
#>    Allocating nodes
#> Graph information:
#>    Observed stochastic nodes: 233
#>    Unobserved stochastic nodes: 38
#>    Total graph size: 4166
#> 
#> Initializing model
#> 

# Get matrices of observed data
jagsdat <- getjagsdata(network$data.ab)

# Plugin estimation of pD is problematic with non-linear models as it often leads to
#negative values, hence use of pV, pd.kl and popt as other measures for the effective
#number of parameters
pDcalc(obs1=jagsdat$y, obs2=jagsdat$se,
  fups=jagsdat$fups, narm=jagsdat$narm, NS=jagsdat$NS,
  theta.result = emax$BUGSoutput$mean$theta,
  resdev.result = emax$BUGSoutput$mean$resdev
  )
#> [1] -9738.838
# }