kmbayes_combine_lowmem {bkmrhat}

Combine multiple BKMR chains in lower memory settings

Description

Combine multiple chains comprising BKMR fits at different starting values. This function writes some results to disk, rather than trying to process fully within memory which, in some cases, will result in avoiding "out of memory" errors that can happen with kmbayes_combine.

Usage

kmbayes_combine_lowmem(
  fitkm.list,
  burnin = NULL,
  excludeburnin = FALSE,
  reorder = TRUE
)

comb_bkmrfits_lowmem(
  fitkm.list,
  burnin = NULL,
  excludeburnin = FALSE,
  reorder = TRUE
)

Arguments

fitkm.list

output from kmbayes_parallel

burnin

(numeric, or default=NULL) add in custom burnin (number of burnin iterations per chain). If NULL, then default to half of the chain

excludeburnin

(logical, default=FALSE) should burnin iterations be excluded from the final chains? Note that all bkmr package functions automatically exclude burnin from calculations.

reorder

(logical, default=TRUE) ensures that the first half of the combined chain contains only the first half of each individual chain - this allows unaltered use of standard functions from bkmr package, which automatically trims the first half of the iterations. This can be used for posterior summaries, but certain diagnostics may not work well (autocorrelation, effective sample size) so the diagnostics should be done on the individual chains #' @param ... arguments to as.mcmc.bkmrfit

Details

Chains are not combined fully sequentially (see "reorder")

Value

a bkmrplusfit object, which inherits from bkmrfit (from the kmbayes function) with multiple chains combined into a single object and additional parameters given by chain and iters, which index the specific chains and iterations for each posterior sample in the bkmrplusfit object

Examples

# following example from https://jenfb.github.io/bkmr/overview.html
set.seed(111)
library(bkmr)
dat <- bkmr::SimData(n = 50, M = 4)
y <- dat$y
Z <- dat$Z
X <- dat$X
set.seed(111)

future::plan(strategy = future::multisession, workers=2)
# run 4 parallel Markov chains (low iterations used for illustration)
fitkm.list <- kmbayes_parallel(nchains=2, y = y, Z = Z, X = X, iter = 500,
  verbose = FALSE, varsel = TRUE)
# use bkmr defaults for burnin, but keep them
bigkm = kmbayes_combine_lowmem(fitkm.list, excludeburnin=FALSE)
ests = ExtractEsts(bigkm) # defaults to keeping second half of samples
ExtractPIPs(bigkm)
pred.resp.univar <- PredictorResponseUnivar(fit = bigkm)
risks.overall <- OverallRiskSummaries(fit = bigkm, y = y, Z = Z, X = X,
  qs = seq(0.25, 0.75, by = 0.05), q.fixed = 0.5, method = "exact")

# additional objects that are not in a standard bkmrfit object:
summary(bigkm$iters) # note that this reflects how fits are re-ordered to reflect burnin
table(bigkm$chain)


closeAllConnections()