analyse.models: Summarize MCMC sample of Bayesian logic regression models

Description

For an MCMC sample of Bayesian logic regression models obtained with fblr the distribution of the model size and the most common logic predictors with up to three binaries are reported.

Usage

analyse.models(file, size.freq = TRUE, moco = c(20, 10), int.freq = TRUE, 
                kmax = 10, int.level = 2, bin.names = NULL)

Arguments

file

character string naming file where MCMC output of fblr is stored.

size.freq

determines whether distribution of model size is reported as frequencies (default) or proportions.

moco

a vector of length 2 or 3 that determines how many of the most common main effects, two-factor interactions and (possibly) three-factor interactions are reported.

int.freq

determines whether the number (default) or the proportion of models containing a specific interaction is reported.

kmax

the maximum number of allowed logic predictors used in fblr.

int.level

the maximum number of allowed binaries in a logic predictor used in fblr.

bin.names

character vector of names for the binary variables. If no names are supplied, binaries are referred to with their indices.

Value

size

table of model sizes.

ones

table of the moco[1] most common single-binary predictors.

twos

table of the moco[2] most common two-binaries predictors.

threes

table of the moco[3] most common three-binaries predictors.

Details

The logic regression models visited during the MCMC run are stored by fblr in the rows of a matrix in the following fashion: Position 1 contains the number of logic predictors in the model. The next kmax * (int.level + 1) positions contain the predictors, each predictor being coded as c(number of binaries in predictor, indices of binaries), where negative indices denote the complement of a variable. It follow the log-likelihood of the model, the value of the precision of the regression parameters and the kmax+1 regression parameters. Zeros indicate empty entries. analyse.models extracts some of the most interesting information, namely which logic predictors occur most often in the visited models, from the sample. The complement of a binary is indicated with a minus sign preceding its name.

Examples

Run this code

# NOT RUN {
# Use fblr on some simulated SNP data
snp <- matrix(rbinom(500*20,2,0.3),ncol=20)
bin <- snp2bin(snp)
int <- apply(bin,1,function(x) (x[1] == 1 & x[3] == 0)*1)
case.prob <- exp(-0.5+log(5)*int)/(1+exp(-0.5+log(5)*int))
y <- rbinom(nrow(snp),1,prob=case.prob)
fblr(y, bin, niter=1000, nburn=0)

analyse.models("fblr_mcmc.txt")

# with SNP names
name.snp <- LETTERS[1:20]
name.bin <- paste(rep(name.snp,each=2), c("_d","_r"),sep="")

analyse.models("fblr_mcmc.txt", bin.names=name.bin)
  
# }