Jan Smyčka, Petr Keil

This is an R package that enables to perform Bayesian Principal Components Analysis, which we call bPCA.

For easy installation directly from this repository just type this in R:

library(devtools)

install_github("bPCA", username="petrkeil")

library(bPCA)

The idea is to fit a MultiVariate Normal (MVN) distribution to set of continuous variables. The fitting is done using MCMC sampler in JAGS. The means and covariances (parameters of the MVN) are monitored during the MCMC sampling and stored as MCMC chains, and subsequently subjected to various summary procedures.

The potential advantages over classical PCA are:

• Prior information about associations between variables can be provided.
• Stability of the PCA can be assessed, especially when only small sample sizes are available.
• The posterior distributions for the PCA scores and loadings can be extracted for further use, e. g. anywhere where propagation of uncertainty is of interest.

The project was initiated by Jan Smyčka who approached Petr Keil with the idea of performing principal components analysis (PCA) in Bayesian setting. After some discussions Jan put together the first JAGS code and essentially the core function of the project. Petr then build the rest of the package, and together they wrote the documentation.