Purpose

fbseqDebug is a companion package for debugging the fbseq and fbseqCUDA packages, available on GitHub. You will need both of these packages installed, and fbseqCUDA requires CUDA. See the fbseqCUDA package vignette for more details.

Checking full conditionals

The full_conditionals function uses both real maize data and simulated data to check each MCMC parameter against its true full conditional distribution and output relevant plots. For the histograms, MCMC samples are plotted with the true full conditional density overlayed as a line. Traceplots are also output.

Checking approximate credible intervals

The fbseq package estimates credible intervals using posterior means and posterior variances of parameters. This approach saves hundreds of megabytes of memory relative to calculations involving quantiles of Monte Carlo samples. However, the required distributional assumptions only approximately hold, so these credible intervals are only approximately correct. To that end, the credible_debug function checks the memory-efficient credible intervals with conventional quantile-intensive ones to make sure the two match well enough. Multiple plots are produced for that purpose.

Running example MCMCs

The paschold_mcmc and simulated_mcmc functions run fbseq on real and simulated data, respectively, and produce output to assess how well the program is working.



wlandau/fbseqDebug documentation built on May 4, 2019, 8:44 a.m.