gibbsDPP: Dirichlet Process Prior Gibbs Sampler

In BushmanLab/cellTypeCompositions: Cell Type Compositions Gibbs Sampler

Dirichlet Process Prior Gibbs Sampler

Description

Auxiliary Gibbs Sampler for Dirichlet Process Prior

Usage

gibbsDPP(
  wtab,
  om,
  eta = NULL,
  etaN = NULL,
  dataToEta = NULL,
  lambda = NULL,
  lambdaN = NULL,
  dataToLambda = NULL,
  etaM = 0L,
  auxEtaM = 5L,
  alphaEta = 100,
  lambdaM = 0L,
  auxLambdaM = 5L,
  alphaLambda = 1,
  verbose = 0L,
  etaCols = 500L,
  lambdaSize = 10L,
  dprior = 1,
  lambdaShape = 1,
  lambdaRate = 0.01,
  ijvals = 0L,
  ...
)

gibbsScan(
  wtab,
  om,
  eta = NULL,
  etaN = NULL,
  dataToEta = NULL,
  etaM = 0L,
  alphaEta = 1,
  auxEtaM = 10L,
  lambda = NULL,
  lambdaN = NULL,
  dataToLambda = NULL,
  lambdaM = 0L,
  alphaLambda = 1,
  auxLambdaM = 10L,
  nkeep = 1L,
  nthin = 1L,
  nburn = 0L,
  etaCols = 50L,
  lambdaSize = 20L,
  ijvals = 0L,
  abEta = c(1e-04, 1e-04),
  abLambda = c(0.01, 0.01),
  verbose = FALSE,
  dprior = 1,
  lambdaShape = 1,
  lambdaRate = 0.01,
  keep = TRUE,
  niter.tune = 10L,
  ...
)

## S3 method for class 'ctScan'
update(object, elt = length(object), ...)

keep.default(pass, log.posterior, i, nkeep)

Arguments

wtab	the result of uniTab
om	the object returned by omega(paramList)
eta	(optional) initialization and workspace for the sampled parameters
etaN	(optional) initialization and workspace for the counts of sampled parameter vectors
dataToEta	index from the original data to the value of eta sampled
lambda	(optional) initialization and workspace for the sampled abundance parameters
lambdaN	(optional) initialization and workspace for the counts of sampled parameters
dataToLambda	index from the original data to the value of lambda sampled
etaM	number of samples provided in the intialization
auxEtaM	how many auxiliary paramters to draw
alphaEta	how much weight to place on new samples
lambdaM	how many auxiliary paramters to draw
auxLambdaM	how much space to allot for new samples. Only one new sample will be drawn if greater than zero
alphaLambda	how much weight to place on new samples
verbose	whether to report intermediate actions
etaCols	how large to make the workspace (too small a value will result in termination with an error)
lambdaSize	dprior prior value for draws of eta
dprior	symmetric Dirichlet prior value
lambdaShape	base gamma distribution prior shape
lambdaRate	base distribution prior rate
ijvals	0L by default, but for (not yet implemented) splitting and merging a value of 2 is needed. Users should not usually change the default.
...	parameters to revise using update
nkeep	the maximum number of iterations
nthin	how many iterations per saved value
nburn	discard this many iterations before saving any values
abEta	if NULL treat alphaEta as a fixed value. Otherwise the first two elements are the shape and rate parameters of Gamma prior for alphaEta.
abLambda	akin to abEta
keep	either TRUE to apply keep.default to the result in each iteration or a function called with pass, the intermediate results, and the other arguments of keep.default
niter.tune	how many cycles of tuning for each gibbs scan
object	a “ctScan” object
elt	integer to select a sample to use as the starting point
pass	a list of parameter values following an iteration, viz. the values in elements of the default return value of gibbsScan.
log.posterior	the loglikelihood evaluated at the current values of pass
i	the iteration number

Details

The auxiliary Gibbs sampler for a Dirichlet Process prior is implemented. The usual call is to gibbsScan, but update can be used to continue a chain (with no arguments) possibly changing the parameters.

gibbsDPP only does an initial Gibbs scan with no following tuning of parameters or subsequent iteration. It may be helpful for low-level access to the scanning routine, but rarely is used.

gibbsScan is the usual entry point. It conducts the Gibbs scan for eta and lambda and subseqeunt tuning alone the lines of Algorithm 8 in Neal, 2000, but the scan for lambda uses the integrated posterior from the base distribution as in Algorithm 1 of Neal, 2000 to decide whether to add a new value.

The method for update allows iterations under gibbsScan(..., keep = TRUE) to resume. If another value is used for keep some care is needed to mimic the return value as described next.

keep.default is used internally under the default (gibbsScan(..., keep = TRUE)) to return values from each iteration. Users can provide their own function to alter the collection of values from each iterate. The function should have the same number of arguments as keep.default and may pass those arguments to keep.default when i == nkeep to allow update to resume iteration.

Value

list which by default has elements eta, etaN, etaM, lambda, lambdaN, lambdaM, and dataToEta and dataToLambda which are updates to the correspondingly named inputs for gibbsDPP and a list of lists of such elements for gibbsScan. The element logLik is the loglikelihood for lambda and eta conditioning on the values of all other parameters.

Author(s)

Charles Berry

References

Radford M. Neal (2000) Markov Chain Sampling Methods for Dirichlet Process Mixture Models, Journal of Computational and Graphical Statistics, 9:2, 249-265, DOI: 10.1080/10618600.2000.10474879

Examples

tab <- diag(10, nrow = 3)[ rep(1:3,2), ] +
  rbinom(18,2,0.5)
uop <- prop.table(diag(3)+0.05,1)
gibbsScan(uniTab(tab), uop)

BushmanLab/cellTypeCompositions documentation built on June 14, 2022, 9:29 p.m.