New sample size specific diagnostic threshold for Pareto
k. The pre-2022 version of the PSIS paper recommended
diagnostic thresholds of k < 0.5 "good",
0.5 <= k < 0.7 "ok",
0.7 <= k < 1 "bad", k>=1 "very bad".
The 2022 revision of the PSIS paper now recommends
k < min(1 - 1/log10(S), 0.7) "good",
min(1 - 1/log10(S), 0.7) <= k < 1 "bad",
k > 1 "very bad", where S is the sample
size. There is now one fewer diagnostic threshold ("ok" has
been removed), and the most important threshold now depends on the
sample size S. With sample sizes 100,
320, 1000, 2200,
10000 the sample size specific part
1 - 1/log10(S) corresponds to thresholds of
0.5, 0.6, 0.67, 0.7,
0.75. Even if the sample size grows, the bias in the PSIS
estimate dominates if 0.7 <= k < 1, and thus the
diagnostic threshold for good is capped at 0.7 (if
k > 1, the mean does not exist and bias is not a valid
measure). The new recommended thresholds are based on more careful
bias-variance analysis of PSIS based on truncated Pareto sums theory.
For those who use the Stan default 4000 posterior draws, the
0.7 threshold will be roughly the same, but there will be
fewer warnings as there will be no diagnostic message for
0.5 <= k < 0.7. Those who use smaller sample sizes
may see diagnostic messages with a threshold less than 0.7,
and they can simply increase the sample size to about 2200
to get the threshold to 0.7.
No more warnings if the r_eff argument is
not provided, and the default is now r_eff = 1.
The summary print output showing MCSE and ESS now shows diagnostic
information on the range of r_eff. The change was made to
reduce unnecessary warnings. The use of r_eff does not
change the expected value of elpd_loo, p_loo,
and Pareto k, and is needed only to estimate MCSE and ESS.
Thus it is better to show the diagnostic information about
r_eff only when MCSE and ESS values are shown.
k Inf if it is NA by @topipa in #224E_loo() when type is variance by @jgabry in #226E_loo() now allows type="sd" by @jgabry in #226pointwise() convenience function for extracting
pointwise estimates by @jgabry in #241k threshold by @avehtari in #235mcse_elpd using log-normal approximation by
@avehtari in
#246n_eff/ESS if
k > k_threshold by @avehtari in #248E_loo() Pareto-k diagnostics by @avehtari in #247loo_subsample.R by @avehtari in #238New loo_predictive_metric() function for computing
estimates of leave-one-out predictive metrics: mean absolute error, mean
squared error and root mean squared error for continuous predictions,
and accuracy and balanced accuracy for binary classification. (#202,
@LeeviLindgren)
New functions crps(), scrps(),
loo_crps(), and loo_scrps() for computing the
(scaled) continuously ranked probability score. (#203, @LeeviLindgren)
New vignette “Mixture IS leave-one-out cross-validation for high-dimensional Bayesian models.” This is a demonstration of the mixture estimators proposed by Silva and Zanella (2022). (#210)
loo_model_weights() to make them consistent with
loo_compare(). (#217)New Frequently Asked Questions page on the package website. (#143)
Speed improvement from simplifying the normalization when fitting the generalized Pareto distribution. (#187, @sethaxen)
Added parallel likelihood computation to speedup
loo_subsample() when using posterior approximations. (#171,
@kdubovikov)
Switch unit tests from Travis to GitHub Actions. (#164)
save_psis = TRUE (#166, @fweber144).Fixed a bug in relative_eff.function() that caused
an error on Windows when using multiple cores. (#152)
Fixed a potential numerical issue in
loo_moment_match() with split=TRUE.
(#153)
Fixed potential integer overflow with
loo_moment_match(). (#155, @ecmerkle)
Fixed relative_eff() when used with a
posterior::draws_array. (#161, @rok-cesnovar)
elpd() (and methods for matrices
and arrays) for computing expected log predictive density of new data or
log predictive density of observed data. A new vignette demonstrates
using this function when doing K-fold CV with rstan. (#159, @bnicenboim)loo_moment_match() that prevented
... arguments from being used correctly. (#149)Added Topi Paananen and Paul Bürkner as coauthors.
New function loo_moment_match() (and new vignette),
which can be used to update a loo object when Pareto k
estimates are large. (#130)
The log weights provided by the importance sampling functions
psis(), tis(), and sis() no
longer have the largest log ratio subtracted from them when returned to
the user. This should be less confusing for anyone using the
weights() method to make an importance sampler. (#112,
#146)
MCSE calculation is now deterministic (#116, #147)
Added Mans Magnusson as a coauthor.
New functions loo_subsample() and
loo_approximate_posterior() (and new vignette) for doing
PSIS-LOO with large data. (#113)
Added support for standard importance sampling and truncated
importance sampling (functions sis() and
tis()). (#125)
compare() now throws a deprecation warning
suggesting loo_compare(). (#93)
A smaller threshold is used when checking the uniqueness of tail values. (#124)
For WAIC, warnings are only thrown when running
waic() and not when printing a waic object.
(#117, @mcol)
Use markdown syntax in roxygen documentation wherever possible. (#108)
New function loo_compare() for model comparison that
will eventually replace the existing compare() function.
(#93)
New vignette on LOO for non-factorizable joint Gaussian models. (#75)
New vignette on “leave-future-out” cross-validation for time series models. (#90)
New glossary page (use help("loo-glossary")) with
definitions of key terms. (#81)
New se_diff column in model comparison results.
(#78)
Improved stability of psis() when
log_ratios are very small. (#74)
Allow r_eff=NA to suppress warning when specifying
r_eff is not applicable (i.e., draws not from MCMC).
(#72)
Update effective sample size calculations to match RStan’s version. (#85)
Naming of k-fold helper functions now matches scikit-learn. (#96)
This is a major release with many changes. Whenever possible we have opted to deprecate rather than remove old functionality, but it is possible that old code that accesses elements inside loo objects by position rather than name may error.
New package documentation website http://mc-stan.org/loo/ with vignettes, function reference, news.
Updated existing vignette and added two new vignettes demonstrating how to use the package.
New function psis() replaces psislw()
(now deprecated). This version implements the improvements to the PSIS
algorithm described in the latest version of
https://arxiv.org/abs/1507.02646. Additional diagnostic information is
now also provided, including PSIS effective sample sizes.
New weights() method for extracting smoothed weights
from a psis object. Arguments log and
normalize control whether the weights are returned on the
log scale and whether they are normalized.
Updated the interface for the loo() methods to
integrate nicely with the new PSIS algorithm. Methods for log-likelihood
arrays, matrices, and functions are provided. Several arguments have
changed, particularly for the loo.function method. The
documentation at help("loo") has been updated to describe
the new behavior.
The structure of the objects returned by the loo()
function has also changed slightly, as described in the
Value section at
help("loo", package = "loo").
New function loo_model_weights() computes weights
for model averaging as described in https://arxiv.org/abs/1704.02030.
Implemented methods include stacking of predictive distributions,
pseudo-BMA weighting or pseudo-BMA+ weighting with the Bayesian
bootstrap.
Setting options(loo.cores=...) is now deprecated in
favor of options(mc.cores=...). For now, if both the
loo.cores and mc.cores options have been set,
preference will be given to loo.cores until it is removed
in a future release. (thanks to @cfhammill)
New functions example_loglik_array() and
example_loglik_matrix() that provide objects to use in
examples and tests.
When comparing more than two models with compare(),
the first column of the output is now the elpd difference
from the model in the first row.
New helper functions for splitting observations for K-fold CV:
kfold_split_random(), kfold_split_balanced(),
kfold_split_stratified(). Additional helper functions for
implementing K-fold CV will be included in future releases.
E_loo() function for computing weighted
expectations (means, variances, quantiles).pareto_k_table() and pareto_k_ids()
convenience functions for quickly identifying problematic
observations(-Inf, 0.5],
(0.5, 0.7], (0.7, 1], (1, Inf)
(didn’t used to include 0.7)psislw() instead of
print.looprint.loo() shows a table of pareto k estimates (if any
k > 0.7)compare() to allow loo objects to be
provided in a list rather than in '...'extract_log_lik()compare(). In previous
versions of loo model weights were also reported by
compare(). We have removed the weights because they were
based only on the point estimate of the elpd values ignoring the
uncertainty. We are currently working on something similar to these
weights that also accounts for uncertainty, which will be included in
future versions of loo.This update makes it easier for other package authors using
loo to write tests that involve running the
loo function. It also includes minor bug fixes and
additional unit tests. Highlights:
cores=1.psislw function is called in an
interactive session.This update provides several important improvements, most notably an alternative method for specifying the pointwise log-likelihood that reduces memory usage and allows for loo to be used with larger datasets. This update also makes it easier to to incorporate loo’s functionality into other packages.
matrix and function
methods for both loo() and waic(). The matrix
method provide the same functionality as in previous versions of
loo (taking a log-likelihood matrix as the input). The
function method allows the user to provide a function for computing the
log-likelihood from the data and posterior draws (which are also
provided by the user). The function method is less memory intensive and
should make it possible to use loo for models fit to
larger amounts of data than before.plot and print methods.
plot also provides label_points argument,
which, if TRUE, will label any Pareto k points
greater than 1/2 by the index number of the corresponding observation.
The plot method also now warns about
Inf/NA/NaN values of
k that are not shown in the plot.compare now returns model weights and accepts more than
two inputs.options(loo.cores = NUMBER).loo_and_waic function in favor of separate
functions loo and waicloo_and_waic_diff. Use compare
instead.