BMNPseudo: Pseudo-likelihood inference in L1-penalized Binary Markov Model

Description

Functions to calculate approximate parameter estimates for L1-penalized Binary Markov Models.

Usage

BMNPseudo(X, rhoVec, Delta=NULL, ThetaStart=NULL, 
maxError=1e-5, verbose=FALSE, maxIter=100, penalize.diag=FALSE, 
stepSize=1, performLineSearch=FALSE)
BMNPseudo.single(X, rho, Delta=NULL, ThetaStart=NULL, 
maxError=1e-5, maxIter=100, penalize.diag=FALSE, 
stepSize=1, performLineSearch=FALSE)

Arguments

Input data matrix consisting of 0-1 entries. Has n rows and p columns.

rho

Value of the penalty parameter; If a non-negative p-by-p matrix is given, it is used as the penalty structure.

rhoVec

Gives all values of rho for which the solution should be calculated.

Delta

Adjustmeant to the gradient.

ThetaStart

Starting value for Theta, has to be a p-by-p matrix.

maxError

convergence threshold for the algorithm.

verbose

Print status messages.

maxIter

Maximum number of iteratios to run.

penalize.diag

Should the diagonal be penalized?

stepSize

Stepsize of the algorithm; should be 1 or less.

performLineSearch

If TRUE, a line search is performed; takes longer but is guarateed to converge.

Value

rho: Vector of non-negative penalty parameters sorted in decreasing order.
ThetaList: A list of Theta pxp matrices, corresponding to the penalty parameters in rho.
success: A logical vector of the same length as rho. True, if the function succeeded for the corresponding value in rho.
penalize.diag: Logical. Indicates if the diagonal was penalized (same as input value penalize.diag.

Details

The function BMNPseudo fits an approximate penalized pairwise binary Markov model to the data provided as matrix X for each of the elements in the penalty parameter vector rhoVec (note that rhoVec will be sorted in increasing order). Internally, the function BMNExact.single is called for each entry in rhoVec and the results are collected as described below.

Examples

Run this code

library(BMN)
Theta = matrix(numeric(25), ncol=5);
Theta[1,1]=0.5; Theta[2,2]=0.5; Theta[3,3]=0; Theta[4,4]= -0.5; Theta[5,5]= 0.5;
Theta[1,2]=Theta[2,1]=1; Theta[1,4]=Theta[4,1]=1; Theta[2,3]=Theta[3,2]= -1;
numSamples=1000; burnIn=100; skip=1;

simData = BMNSamples(Theta, numSamples, burnIn, skip)

rhoVec = c(0.01, 0.02, 0.03)

pseudoPath = BMNPseudo(simData, rhoVec)
pseudoSingle = BMNPseudo.single(simData, 0.02)