NEWS.md
mcmc_areas()
tries to use less vertical blank space. (#218, #230)
Fix bug in color_scheme_view()
minimal theme (#213).
Fix error in mcmc_acf()
for certain input types. (#244, #245, @hhau)
New plotting functions ppc_dens_overlay_grouped()
and ppc_ecdf_overlay_grouped()
for plotting density and cumulative distributions of the posterior predictive distribution (versus observed data) by group. (#212)
New plotting function ppc_km_overlay()
for outcome variables that are
right-censored. Empirical CCDF estimates of yrep
are compared with the Kaplan-Meier estimate of y
. (#233, #234, @fweber144)
ppc_loo_pit_overlay()
now uses a boundary correction for an improved kernel density estimation. The new argument boundary_correction
defaults to TRUE but can be set to FALSE to recover the old version of the plot. (#171, #235,
@ecoronado92)
CmdStanMCMC objects (from CmdStanR) can now be used with extractor functions nuts_params()
, log_posterior()
, rhat()
, and neff_ratio()
. (#227)
On the y axis, ppc_loo_pit_qq(..., compare = "normal")
now plots standard normal quantiles calculated from the PIT values (instead of the standardized PIT values). (#240, #243, @fweber144)
mcmc_rank_overlay()
gains argument facet_args
. (#221, @hhau)
For mcmc_intervals()
the sizeof the points and interval lines can be set with
mcmc_intervals(…, outer_size, inner_size, point_size)`. (#215, #228, #229)
Compatibility with dplyr 1.0.0 (#219)
Release requested by CRAN to fix errors at https://cran.r-project.org/web/checks/check_results_bayesplot.html due to matrices also inheriting from “array” in R 4.0.
(GitHub issue/PR numbers in parentheses)
The pars
argument of all MCMC plotting functions now supports tidy variable selection. See ?"tidy-params"
for details and examples. (#161, #183, #188)
Two new plots have been added for inspecting the distribution of ranks. Rank histograms were introduced by the Stan team’s new paper on MCMC diagnostics. (#178, #179)
mcmc_rank_hist()
: A traditional traceplot (mcmc_trace()
) visualizes how sampled values the MCMC chains mix over the course of sampling. A rank histogram (mcmc_rank_hist()
) visualizes how the ranks of values from the chains mix together. An ideal plot would show the ranks mixing or overlapping in a uniform distribution.
mcmc_rank_overlay()
: Instead of drawing each chain’s histogram in a separate panel, this plot draws the top edge of the chains’ histograms in a single panel.
Added mcmc_trace_data()
, which returns the data used for plotting the trace plots and rank histograms. (Advances #97)
ColorBrewer palettes are now available as color schemes via color_scheme_set()
. For example, color_scheme_set("brewer-Spectral")
will use the Spectral palette. (#177, #190)
MCMC plots now also accept objects with an as.array
method as input (e.g., stanfit objects). (#175, #184)
mcmc_trace()
gains an argument iter1
which can be used to label the traceplot starting from the first iteration after warmup. (#14, #155, @mcol)
mcmc_areas()
gains an argument area_method
which controls how to draw the density curves. The default "equal area"
constrains the heights so that the curves have the same area. As a result, a narrow interval will appear as a spike of density, while a wide, uncertain interval is spread thin over the x axis. Alternatively "equal height"
will set the maximum height on each curve to the same value. This works well when the intervals are about the same width. Otherwise, that wide, uncertain interval will dominate the visual space compared to a narrow, less uncertain interval. A compromise between the two is "scaled height"
which scales the curves from "equal height"
using height * sqrt(height)
. (#163, #169)
mcmc_areas()
correctly plots density curves where the point estimate does not include the highest point of the density curve. (#168, #169, @jtimonen)
mcmc_areas_ridges()
draws the vertical line at x = 0 over the curves so that it is always visible.
mcmc_intervals()
and mcmc_areas()
raise a warning if prob_outer
is ever less than prob
. It sorts these two values into the correct order. (#138)
MCMC parameter names are now always converted to factors prior to plotting. We use factors so that the order of parameters in a plot matches the order of the parameters in the original MCMC data. This change fixes a case where factor-conversion failed. (#162, #165, @wwiecek)
The examples in ?ppc_loo_pit_overlay()
now work as expected. (#166, #167)
Added "viridisD"
as an alternative name for "viridis"
to the supported colors.
Added "viridisE"
(the cividis version of viridis) to the supported colors.
ppc_bars()
and ppc_bars_grouped()
now allow negative integers as input. (#172, @jeffpollock9)
(GitHub issue/PR numbers in parentheses)
Loading bayesplot no longer overrides the ggplot theme! Rather, it sets a theme specific for bayesplot. Some packages using bayesplot may still override the default ggplot theme (e.g., rstanarm does but only until next release), but simply loading bayesplot itself will not. There are new functions for controlling the ggplot theme for bayesplot that work like their ggplot2 counterparts but only affect plots made using bayesplot. Thanks to Malcolm Barrett. (#117, #149).
The Visual MCMC Diagnostics vignette has been reorganized and has a lot of useful new content thanks to Martin Modrák. (#144, #153)
The LOO predictive checks now require loo version >= 2.0.0
. (#139)
Histogram plots gain a breaks
argument that can be used as an alternative to binwidth
. (#148)
mcmc_pairs()
now has an argument grid_args
to provide a way of passing optional arguments to gridExtra::arrangeGrob()
. This can be used to add a title to the plot, for example. (#143)
ppc_ecdf_overlay()
gains an argument discrete
, which is FALSE
by default, but can be used to make the Geom more appropriate for discrete data. (#145)
PPC intervals plots and LOO predictive checks now draw both an outer and an inner probability interval, which can be controlled through the new argument prob_outer
and the already existing prob
. This is consistent with what is produced by mcmc_intervals()
. (#152, #154, @mcol)
(GitHub issue/PR numbers in parentheses)
New package documentation website: https://mc-stan.org/bayesplot/
Two new plots that visualize posterior density using ridgelines. These work well when parameters have similar values and similar densities, as in hierarchical models. (#104)
mcmc_dens_chains()
draws the kernel density of each sampling chain.mcmc_areas_ridges()
draws the kernel density combined across chains._data()
function to return the data plotted by each function.mcmc_intervals()
and mcmc_areas()
have been rewritten. (#103)
mcmc_areas()
now uses geoms from the ggridges package to draw density curves.Added mcmc_intervals_data()
and mcmc_areas_data()
that return data plotted by mcmc_intervals()
and mcmc_areas()
. (Advances #97)
New ppc_data()
function returns the data plotted by many of the PPC plotting functions. (Advances #97)
Added ppc_loo_pit_overlay()
function for a better LOO PIT predictive check. (#123)
Started using vdiffr to add visual unit tests to the existing PPC unit tests. (#137)
(GitHub issue/PR numbers in parentheses)
New plotting function mcmc_parcoord()
for parallel coordinates plots of MCMC draws (optionally including HMC/NUTS diagnostic information). (#108)
mcmc_scatter
gains an np
argument for specifying NUTS parameters, which allows highlighting divergences in the plot. (#112)
New functions with names ending with suffix _data
don’t make the plots, they just return the data prepared for plotting (more of these to come in future releases):
ppc_stat_grouped()
, ppc_stat_freqpoly_grouped()
gain a facet_args
argument for controlling ggplot2 faceting (many of the mcmc_
functions already have this).
The divergences
argument to mcmc_trace()
has been deprecated in favor of np
(NUTS parameters) to match the other functions that have an np
argument.
Fixed an issue where duplicated rhat values would break mcmc_rhat()
(#105).
(GitHub issue/PR numbers in parentheses)
bayesplot::theme_default()
is now set as the default ggplot2 plotting theme when bayesplot is loaded, which makes changing the default theme using ggplot2::theme_set()
possible. Thanks to @gavinsimpson. (#87)
mcmc_hist()
and mcmc_hist_by_chain()
now take a freq
argument that defaults to TRUE
(behavior is like freq
argument to R’s hist
function).
Using a ts
object for y
in PPC plots no longer results in an error. Thanks to @helske. (#94)
mcmc_intervals()
doesn’t use round lineends anymore as they slightly exaggerate the width of the intervals. Thanks to @tjmahr. (#96)
A lot of new stuff in this release. (GitHub issue/PR numbers in parentheses)
Avoid error in some cases when divergences
is specified in call to mcmc_trace()
but there are not actually any divergent transitions.
The merge_chains
argument to mcmc_nuts_energy()
now defaults to FALSE
.
For mcmc_*()
functions, transformations are recycled if transformations
argument is specified as a single function rather than a named list. Thanks to @tklebel. (#64)
For ppc_violin_grouped()
there is now the option of showing y
as a violin, points, or both. Thanks to @silberzwiebel. (#74)
color_scheme_get()
now has an optional argument i
for selecting only a subset of the colors.
New color schemes: darkgray, orange, viridis, viridisA, viridisB, viridisC. The viridis schemes are better than the other schemes for trace plots (the colors are very distinct from each other).
mcmc_pairs()
, which is essentially a ggplot2+grid implementation of rstan’s pairs.stanfit()
method. (#67)
mcmc_hex()
, which is similar to mcmc_scatter()
but using geom_hex()
instead of geom_point()
. This can be used to avoid overplotting. (#67)
overlay_function()
convenience function. Example usage: add a Gaussian (or any distribution) density curve to a plot made with mcmc_hist()
.
mcmc_recover_scatter()
and mcmc_recover_hist()
, which are similar to mcmc_recover_intervals()
and compare estimates to “true” values used to simulate data. (#81, #83)
New PPC category Discrete with functions:
ppc_rootogram()
for use with models for count data. Thanks to
@paul-buerkner. (#28)ppc_bars()
, ppc_bars_grouped()
for use with models for ordinal, categorical and multinomial data. Thanks to @silberzwiebel. (#73)New PPC category LOO (thanks to suggestions from @avehtari) with functions:
ppc_loo_pit()
for assessing the calibration of marginal predictions. (#72)ppc_loo_intervals()
, ppc_loo_ribbon()
for plotting intervals of the LOO predictive distribution. (#72)(GitHub issue/PR numbers in parentheses)
Images in vignettes should now render properly using png
device. Thanks to TJ Mahr. (#51)
xaxis_title(FALSE)
and yaxis_title(FALSE)
now set axis titles to NULL
rather than changing theme elements to element_blank()
. This makes it easier to add axis titles to plots that don’t have them by default. Thanks to Bill Harris. (#53)
Add argument divergences
to mcmc_trace()
function. For models fit using HMC/NUTS this can be used to display divergences as a rug at the bottom of the trace plot. (#42)
The stat
argument for all ppc_stat_*()
functions now accepts a function instead of only the name of a function. (#31)
ppc_error_hist_grouped()
for plotting predictive errors by level of a grouping variable. (#40)
mcmc_recover_intervals)(
for comparing MCMC estimates to “true” parameter values used to simulate the data. (#56)
bayesplot_grid()
for juxtaposing plots and enforcing shared axis limits. (#59)