Functions for computing density and producing random samples from the LKJ onion method correlation matrix distribution.
dlkjcorr( x , eta=1 , log=TRUE )
rlkjcorr( n , K , eta=1 )Matrix to compute probability density for
Parameter controlling shape of distribution
Dimension of correlation matrix
If TRUE, returns log-probability instead of probability
Number of random matrices to sample
The LKJ correlation matrix distribution is based upon work by Lewandowski, Kurowicka, and Joe. When the parameter eta is equal to 1, it defines a flat distribution of correlation matrices. When eta > 1, the distribution is instead concentrated towards to identity matrix. When eta < 1, the distribution is more concentrated towards extreme correlations at -1 or +1.
It can be easier to understand this distribution if we recognize that the individual correlations within the matrix follow a beta distribution defined on -1 to +1. Thus eta resembles theta in the beta parameterization with a mean p and scale (sample size) theta.
The function rlkjcorr returns an 3-dimensional array in which the first dimension indexes matrices. In the event that n=1, it returns instead a single matrix.
Lewandowski, Kurowicka, and Joe. 2009. Generating random correlation matrices based on vines and extended onion method. Journal of Multivariate Analysis. 100:1989-2001.
Stan Modeling Language User's Guide and Reference Manual, v2.6.2
# NOT RUN {
R <- rlkjcorr(n=1,K=2,eta=4)
dlkjcorr(R,4)
# plot density of correlation
R <- rlkjcorr(1e4,K=2,eta=4)
dens( R[,1,2] )
# visualize 3x3 matrix
R <- rlkjcorr(1e3,K=3,eta=2)
plot( R[,1,2] , R[,1,3] , col=col.alpha("black",0.2) , pch=16 )
# }
Run the code above in your browser using DataLab