flowReMix_control: Auxiliary for Controlling flowReMix Fitting.

In RGLab/flowReMix: Generalized Linear Modeling of FCM Data With Clustered Coefficients

Description

Auxiliary function for flowReMix fitting. Can be used to generate an appropriate object for modifying the fitting process of the stochastic EM algorithm used by flowReMix.

Usage

flowReMix_control(updateLag = 10, randomAssignProb = 1e-08, nsamp = 50,
  lastSample = NULL, initMHcoef = 2.5, nPosteriors = 3,
  maxDispersion = 10^3, minDispersion = 10^7, isingInit = -4.59512,
  keepEach = 5, centerCovariance = FALSE, intSampSize = 100,
  initMethod = "robust", ncores = NULL, preAssignCoefs = 1,
  markovChainEM = TRUE, seed = 100, prior = 0, isingWprior = TRUE,
  zeroPosteriorProbs = FALSE, clusterType = c("AUTO", "FORK", "SOCK"),
  isingStabilityReps = 200, randStabilityReps = 0, stabilityGamma = 0.9,
  stabilityAND = TRUE, learningRate = 0.6, keepWeightPercent = 0.9,
  sampleNew = FALSE, subsetDiscardThreshold = 0, threads = NULL)

Arguments

updateLag	number of iterations before the algorithm is assumed to converge, at which time the parameter estimates will be aggregated.
randomAssignProb	an optional parameter for adding noise to the cluster assignments generated by the Gibbs sampler. Should only be changed if ising model estimates are unstable.
nsamp	number of Gibbs/componentwise MH cycles to perform for each subject. Must be larger than keepEach.
lastSample	how many samples to keep from the final iteration.
initMHcoef	the initial value for the shrinkage/inflation to perform on the estimated covariance in the componentwise MH sampler. The initial value does not matter much as this parameter self-tunes as the algorithm runs and usually converges to a good value within a few iterations.
nPosteriors	number of posterior samples to take per subject. If left as NULL then a value will be determined within flowReMix according to a preset formula.
maxDispersion	the maximum overdispersion level allowed. The lower the value of the variable the more overdispersion is allowed. Must be larger than 0.
minDispersion	the minimum overdispersion allowed. The larger the value of the variable the less overdispersion is allowed.
isingInit	initialize the Ising model with this value.
keepEach	one out of how many Gibbs/MH samples to keep. This is used to reduce the dependence between posterior samples.
centerCovariance	whether to center random effect estimates before computing the covariance estimate or not.
intSampSize	number of importance samples to take when performing univariate numerical integration in the Gibbs sampler.
initMethod	the method used to initialize the regression coefficients. Options are either "sparse" for cv.glmnet or "binom" for glm. If left as NULL then it will be determined according to the regression_method specified for the flowReMix call.
ncores	The number of cpu cores to use to fit the model in parallel.
preAssignCoefs	coefficients to multiply the posterior probabilities. 0 is a hard assignment and observations that are designated non-responders based on pu > ps will have posterior probabilities of 0. > 0 is a soft assignment, and a prior will be placed on the prior probability of non-response in the Ising model.
markovChainEM	logical use the mcEM algorithm to fit the model. Default TRUE.
seed	numeric a random seed for reproducible initialization. Default 100.
prior	numeric value, a prior for response and non-response in the Ising model used constrain non-responders (e.g. when pu>ps).
isingWprior	logical fit the Ising model with a prior on the baseline response using the parameter in prior. Default TRUE.
zeroPosteriorProbs	boolean. TRUE will zero out posterior response probabilities where pu>ps, equivalent to a one-sided test. The full set of responses will still be used to estimate the Ising model. Default FALSE. Can be used together with the prior argument on the Ising model.
clusterType	character type of cluster. AUTO, FORK, SOCK. Default AUTO. Can be changed if the default doesn't work.
sampleNew	logical should the stability selection draw new samples. Default FALSE.

Value

An object of type flowReMix_control.

RGLab/flowReMix documentation built on May 8, 2019, 5:55 a.m.