This is just a short demonstration on running a nimble model “the long way.” It lets you check your initial values before compiling to avoid the Nimble equivalent of JAGS’s “Node inconsistent with parents” or other errors like that.
This assumes you have already created appropriate data, constant, and initial value objects and functions.
Add: A more comprehensive version.
#this won't actually run.
# first bundle your stuff together
my.nimmod <- nimble::nimbleModel(code = model.code,
constants = nim.const,
data = nim.dat,
inits = inits())
my.nimmod$calculate() # this should come back as a number (the loglikelihood)
# if nodes have appropriate initial values. You may also
# need to give initial values for latent states.
my.nimmod$info()#this will show you any errors with nodes. It will tell you what parameters need initial values/
test.config<-nimble::configureMCMC(my.nimmod) # configure step
#now you can alter the type of sampler on specific parameters if needed.
# block samplers are slower, but produce a larger ESS for things like the different parameters in a single model (intercept, regression coefficients)
# test.config$replaceSamplers(target = c('phi','p'),
# type= 'RW',targetByNode=TRUE)
#
# test.config$replaceSamplers(target = c('mean.FID','beta.FID'),
# type= 'AF_slice')
test.config
#call to see about the samplers and what parameters
test.build<-nimble::buildMCMC(test.config)
comp.mod<-nimble::compileNimble(my.nimmod) #first compile step
comp.mcmc<-nimble::compileNimble(test.build, project=comp.mod) # second compile step
# run
samples<-nimble::runMCMC(comp.mcmc,
niter=10000,
nburnin = 1000,
nchains = 3)
MCMCvis::MCMCplot(samples,pdf=FALSE)