First release.
compare_pt() compares two samples of continuous multivariate observations
with the hidden Markov Pólya tree of Awaya and Ma (2024)
doi:10.1080/01621459.2022.2105223. It returns the posterior probability
that the two distributions differ, a representative partition tree, and local
effect sizes for the divisions of that tree. Posterior computation combines
sequential Monte Carlo over partitions with recursive message passing. The
two samples are supplied either as two matrices or as one data set with a
grouping variable.
pt_prior() sets the model: the depth of the tree, the number of candidate
split points, the state transition probabilities, and the spike-and-slab
prior on split locations of Awaya and Ma (2024, Appendix A), which is used by
default. pt_control() sets the computational budget: the number of
particles, the resampling threshold and power, and the number of posterior
draws used for the effect sizes.
Variables are put on a common scale before the space is divided.
transform = "rank" uses pooled midranks and is the default;
transform = "range" uses the linear scaling of the paper and is the default
when support is supplied. Thresholds and region conditions are reported in
the original units.
regions() lists the divisions of the representative tree with their
probability of a difference and their signed and absolute log odds ratios,
ordered by either quantity. region_rows() and region_data() return the
observations of a region.
plot() draws the representative tree, the regions with the largest
differences on scatter plots, the regions of each tree depth shaded by their
probability of a difference (two variables only), and the effective sample
size of the sequential Monte Carlo. summary() and diagnostics() report
the settings and the computation.
vignette("ptcompare") works through a simulated comparison in two variables
and a comparison of 30 measurements of cell nuclei.