ptycho: Bayesian Variable Selection with Hierarchical Priors

ptycho is an R package for Bayesian variable selection for linear regression models using hierarchical priors. There is a prior that combines information across responses and one that combines information across covariates, as well as a standard spike and slab prior for comparison. An MCMC samples from the marginal posterior distribution for the 0-1 variables indicating if each covariate belongs to the model for each response.

• Manual
• Package at CRAN
• Input data: To demonstrate that the priors increase power when they capture structural characteristics of the data, we simulated data from priors used by ptycho and applied each of the priors to each of the data sets. Each of the following files is 19MB and contains one of these data sets in the form of an R object output by the function createData and described in the package manual. Each has 100 replicates, and each replicate has 5000 observations, 50 covariates, and 5 responses.
  • Standard spike and slab prior. For each response, all covariates have the same probability of being in the model.
  • Share information across responses. Each covariate has the same probability of being in the model for all five responses.
  • Share information across covariates. Covariates are put into 10 groups, and the covariates in a group all have the same probability of being in the model for a response.
• Manuscript: Stell and Sabatti (2016) Genetic variant selection: learning across traits and sites. This paper also used:
  • Actual genotype and phenotype data from Service et al. (2014) Re-sequencing expands our understanding of the phenotypic impact of variants at GWAS loci, PLoS Genetics, doi:10.1371/journal.pgen.1004147.
  • R code for processing ptycho output
  • R code for computing posterior marginals exactly and processing the output