Stats
PSIS.PSISResult
PosteriorStats.AbstractELPDResult
PosteriorStats.AbstractModelWeightsMethod
PosteriorStats.BootstrappedPseudoBMA
PosteriorStats.ModelComparisonResult
PosteriorStats.PSISLOOResult
PosteriorStats.PseudoBMA
PosteriorStats.Stacking
PosteriorStats.SummaryStats
PosteriorStats.WAICResult
PSIS.PSISPlots.paretoshapeplot
PSIS.ess_is
PSIS.psis
PSIS.psis!
PosteriorStats.compare
PosteriorStats.default_diagnostics
PosteriorStats.default_stats
PosteriorStats.default_summary_stats
PosteriorStats.elpd_estimates
PosteriorStats.hdi
PosteriorStats.hdi!
PosteriorStats.information_criterion
PosteriorStats.loo
PosteriorStats.loo_pit
PosteriorStats.model_weights
PosteriorStats.r2_score
PosteriorStats.smooth_data
PosteriorStats.summarize
PosteriorStats.waic
StatsBase.summarystats
Summary statistics
PosteriorStats.SummaryStats
— Typestruct SummaryStats{D, V<:(AbstractVector)}
A container for a column table of values computed by summarize
.
This object implements the Tables and TableTraits column table interfaces. It has a custom show
method.
SummaryStats
behaves like an OrderedDict
of columns, where the columns can be accessed using either Symbol
s or a 1-based integer index.
name::String
: The name of the collection of summary statistics, used as the table title in display.data::Any
: The summary statistics for each parameter. It must implement the Tables interface.parameter_names::AbstractVector
: Names of the parameters
SummaryStats([name::String,] data[, parameter_names])
SummaryStats(data[, parameter_names]; name::String="SummaryStats")
Construct a SummaryStats
from tabular data
with optional stats name
and param_names
.
data
must not contain a column :parameter
, as this is reserved for the parameter names, which are always in the first column.
PosteriorStats.default_summary_stats
— Functiondefault_summary_stats(focus=Statistics.mean; kwargs...)
Combinatiton of default_stats
and default_diagnostics
to be used with summarize
.
PosteriorStats.default_stats
— Functiondefault_stats(focus=Statistics.mean; prob_interval=0.94, kwargs...)
Default statistics to be computed with summarize
.
The value of focus
determines the statistics to be returned:
Statistics.mean
:mean
,std
,hdi_3%
,hdi_97%
Statistics.median
:median
,mad
,eti_3%
,eti_97%
If prob_interval
is set to a different value than the default, then different HDI and ETI statistics are computed accordingly. hdi
refers to the highest-density interval, while eti
refers to the equal-tailed interval (i.e. the credible interval computed from symmetric quantiles).
See also: hdi
PosteriorStats.default_diagnostics
— Functiondefault_diagnostics(focus=Statistics.mean; kwargs...)
Default diagnostics to be computed with summarize
.
The value of focus
determines the diagnostics to be returned:
Statistics.mean
:mcse_mean
,mcse_std
,ess_tail
,ess_bulk
,rhat
Statistics.median
:mcse_median
,ess_tail
,ess_bulk
,rhat
PosteriorStats.summarize
— Functionsummarize(data, stats_funs...; name="SummaryStats", [var_names]) -> SummaryStats
Compute the summary statistics in stats_funs
on each param in data
.
stats_funs
is a collection of functions that reduces a matrix with shape (draws, chains)
to a scalar or a collection of scalars. Alternatively, an item in stats_funs
may be a Pair
of the form name => fun
specifying the name to be used for the statistic or of the form (name1, ...) => fun
when the function returns a collection. When the function returns a collection, the names in this latter format must be provided.
If no stats functions are provided, then those specified in default_summary_stats
are computed.
var_names
specifies the names of the parameters in data
. If not provided, the names are inferred from data
.
To support computing summary statistics from a custom object, overload this method specifying the type of data
.
See also SummaryStats
, default_summary_stats
, default_stats
, default_diagnostics
.
Examples
Compute mean
, std
and the Monte Carlo standard error (MCSE) of the mean estimate:
julia> using Statistics, StatsBase
julia> x = randn(1000, 4, 3) .+ reshape(0:10:20, 1, 1, :);
julia> summarize(x, mean, std, :mcse_mean => sem; name="Mean/Std")
Mean/Std
mean std mcse_mean
1 0.0003 0.990 0.016
2 10.02 0.988 0.016
3 19.98 0.988 0.016
Avoid recomputing the mean by using mean_and_std
, and provide parameter names:
julia> summarize(x, (:mean, :std) => mean_and_std, mad; var_names=[:a, :b, :c])
SummaryStats
mean std mad
a 0.000305 0.990 0.978
b 10.0 0.988 0.995
c 20.0 0.988 0.979
Note that when an estimator and its MCSE are both computed, the MCSE is used to determine the number of significant digits that will be displayed.
julia> summarize(x; var_names=[:a, :b, :c])
SummaryStats
mean std hdi_3% hdi_97% mcse_mean mcse_std ess_tail ess_bulk r ⋯
a 0.0003 0.99 -1.92 1.78 0.016 0.012 3567 3663 1 ⋯
b 10.02 0.99 8.17 11.9 0.016 0.011 3841 3906 1 ⋯
c 19.98 0.99 18.1 21.9 0.016 0.012 3892 3749 1 ⋯
1 column omitted
Compute just the statistics with an 89% HDI on all parameters, and provide the parameter names:
julia> summarize(x, default_stats(; prob_interval=0.89)...; var_names=[:a, :b, :c])
SummaryStats
mean std hdi_5.5% hdi_94.5%
a 0.000305 0.990 -1.63 1.52
b 10.0 0.988 8.53 11.6
c 20.0 0.988 18.5 21.6
Compute the summary stats focusing on Statistics.median
:
julia> summarize(x, default_summary_stats(median)...; var_names=[:a, :b, :c])
SummaryStats
median mad eti_3% eti_97% mcse_median ess_tail ess_median rhat
a 0.004 0.978 -1.83 1.89 0.020 3567 3336 1.00
b 10.02 0.995 8.17 11.9 0.023 3841 3787 1.00
c 19.99 0.979 18.1 21.9 0.020 3892 3829 1.00
StatsBase.summarystats
— Functionsummarystats(data::InferenceData; group=:posterior, kwargs...) -> SummaryStats
summarystats(data::Dataset; kwargs...) -> SummaryStats
Compute default summary statistics for the data using summarize
.
General statistics
PosteriorStats.hdi
— Functionhdi(samples::AbstractArray{<:Real}; prob=0.94) -> (; lower, upper)
Estimate the unimodal highest density interval (HDI) of samples
for the probability prob
.
The HDI is the minimum width Bayesian credible interval (BCI). That is, it is the smallest possible interval containing (100*prob)
% of the probability mass.[Hyndman1996]
samples
is an array of shape (draws[, chains[, params...]])
. If multiple parameters are present, then lower
and upper
are arrays with the shape (params...,)
, computed separately for each marginal.
This implementation uses the algorithm of [ChenShao1999].
Any default value of prob
is arbitrary. The default value of prob=0.94
instead of a more common default like prob=0.95
is chosen to reminder the user of this arbitrariness.
Examples
Here we calculate the 83% HDI for a normal random variable:
julia> x = randn(2_000);
julia> hdi(x; prob=0.83) |> pairs
pairs(::NamedTuple) with 2 entries:
:lower => -1.38266
:upper => 1.25982
We can also calculate the HDI for a 3-dimensional array of samples:
julia> x = randn(1_000, 1, 1) .+ reshape(0:5:10, 1, 1, :);
julia> hdi(x) |> pairs
pairs(::NamedTuple) with 2 entries:
:lower => [-1.9674, 3.0326, 8.0326]
:upper => [1.90028, 6.90028, 11.9003]
hdi(data::InferenceData; kwargs...) -> Dataset
hdi(data::Dataset; kwargs...) -> Dataset
Calculate the highest density interval (HDI) for each parameter in the data.
PosteriorStats.hdi!
— Functionhdi!(samples::AbstractArray{<:Real}; prob=0.94) -> (; lower, upper)
A version of hdi
that sorts samples
in-place while computing the HDI.
PosteriorStats.r2_score
— Functionr2_score(y_true::AbstractVector, y_pred::AbstractArray) -> (; r2, r2_std)
$R²$ for linear Bayesian regression models.[GelmanGoodrich2019]
Arguments
y_true
: Observed data of lengthnoutputs
y_pred
: Predicted data with size(ndraws[, nchains], noutputs)
Examples
julia> using ArviZExampleData
julia> idata = load_example_data("regression1d");
julia> y_true = idata.observed_data.y;
julia> y_pred = PermutedDimsArray(idata.posterior_predictive.y, (:draw, :chain, :y_dim_0));
julia> r2_score(y_true, y_pred) |> pairs
pairs(::NamedTuple) with 2 entries:
:r2 => 0.683197
:r2_std => 0.0368838
r2_score(idata::InferenceData; y_name, y_pred_name) -> (; r2, r2_std)
Compute $R²$ from idata
, automatically formatting the predictions to the correct shape.
Keywords
y_name
: Name of observed data variable inidata.observed_data
. If not provided, then the only observed data variable is used.y_pred_name
: Name of posterior predictive variable inidata.posterior_predictive
. If not provided, theny_name
is used.
Examples
julia> using ArviZExampleData, PosteriorStats
julia> idata = load_example_data("regression10d");
julia> r2_score(idata) |> pairs
pairs(::NamedTuple) with 2 entries:
:r2 => 0.998385
:r2_std => 0.000100621
Pareto-smoothed importance sampling
PSIS.PSISResult
— TypePSISResult
Result of Pareto-smoothed importance sampling (PSIS) using psis
.
Properties
log_weights
: un-normalized Pareto-smoothed log weightsweights
: normalized Pareto-smoothed weights (allocates a copy)pareto_shape
: Pareto $k=ξ$ shape parameternparams
: number of parameters inlog_weights
ndraws
: number of draws inlog_weights
nchains
: number of chains inlog_weights
reff
: the ratio of the effective sample size of the unsmoothed importance ratios and the actual sample size.ess
: estimated effective sample size of estimate of mean using smoothed importance samples (seeess_is
)tail_length
: length of the upper tail oflog_weights
that was smoothedtail_dist
: the generalized Pareto distribution that was fit to the tail oflog_weights
. Note that the tail weights are scaled to have a maximum of 1, sotail_dist * exp(maximum(log_ratios))
is the corresponding fit directly to the tail oflog_ratios
.normalized::Bool
:indicates whetherlog_weights
are log-normalized along the sample dimensions.
Diagnostic
The pareto_shape
parameter $k=ξ$ of the generalized Pareto distribution tail_dist
can be used to diagnose reliability and convergence of estimates using the importance weights [VehtariSimpson2021].
- if $k < \frac{1}{3}$, importance sampling is stable, and importance sampling (IS) and PSIS both are reliable.
- if $k ≤ \frac{1}{2}$, then the importance ratio distributon has finite variance, and the central limit theorem holds. As $k$ approaches the upper bound, IS becomes less reliable, while PSIS still works well but with a higher RMSE.
- if $\frac{1}{2} < k ≤ 0.7$, then the variance is infinite, and IS can behave quite poorly. However, PSIS works well in this regime.
- if $0.7 < k ≤ 1$, then it quickly becomes impractical to collect enough importance weights to reliably compute estimates, and importance sampling is not recommended.
- if $k > 1$, then neither the variance nor the mean of the raw importance ratios exists. The convergence rate is close to zero, and bias can be large with practical sample sizes.
See PSISPlots.paretoshapeplot
for a diagnostic plot.
PSIS.ess_is
— Functioness_is(weights; reff=1)
Estimate effective sample size (ESS) for importance sampling over the sample dimensions.
Given normalized weights $w_{1:n}$, the ESS is estimated using the L2-norm of the weights:
\[\mathrm{ESS}(w_{1:n}) = \frac{r_{\mathrm{eff}}}{\sum_{i=1}^n w_i^2}\]
where $r_{\mathrm{eff}}$ is the relative efficiency of the log_weights
.
ess_is(result::PSISResult; bad_shape_nan=true)
Estimate ESS for Pareto-smoothed importance sampling.
ESS estimates for Pareto shape values $k > 0.7$, which are unreliable and misleadingly high, are set to NaN
. To avoid this, set bad_shape_nan=false
.
PSIS.PSISPlots.paretoshapeplot
— Functionparetoshapeplot(values; showlines=false, ...)
paretoshapeplot!(values; showlines=false, kwargs...)
Plot shape parameters of fitted Pareto tail distributions for diagnosing convergence.
values
may be either a vector of Pareto shape parameters or a PSIS.PSISResult
.
If showlines==true
, horizontal lines indicating relevant Pareto shape thresholds are drawn. See PSIS.PSISResult
for an explanation of the thresholds.
All remaining kwargs
are forwarded to the plotting function.
See psis
, PSISResult
.
Examples
using PSIS, Distributions, Plots
proposal = Normal()
target = TDist(7)
x = rand(proposal, 1_000, 100)
log_ratios = logpdf.(target, x) .- logpdf.(proposal, x)
result = psis(log_ratios)
paretoshapeplot(result)
We can also plot the Pareto shape parameters directly:
paretoshapeplot(result.pareto_shape)
We can also use plot
directly:
plot(result.pareto_shape; showlines=true)
PSIS.psis
— Functionpsis(log_ratios, reff = 1.0; kwargs...) -> PSISResult
psis!(log_ratios, reff = 1.0; kwargs...) -> PSISResult
Compute Pareto smoothed importance sampling (PSIS) log weights [VehtariSimpson2021].
While psis
computes smoothed log weights out-of-place, psis!
smooths them in-place.
Arguments
log_ratios
: an array of logarithms of importance ratios, with size(draws, [chains, [parameters...]])
, wherechains>1
would be used when chains are generated using Markov chain Monte Carlo.reff::Union{Real,AbstractArray}
: the ratio(s) of effective sample size oflog_ratios
and the actual sample sizereff = ess/(draws * chains)
, used to account for autocorrelation, e.g. due to Markov chain Monte Carlo. If an array, it must have the size(parameters...,)
to matchlog_ratios
.
Keywords
warn=true
: Iftrue
, warning messages are deliverednormalize=true
: Iftrue
, the log-weights will be log-normalized so thatexp.(log_weights)
sums to 1 along the sample dimensions.
Returns
result
: aPSISResult
object containing the results of the Pareto-smoothing.
A warning is raised if the Pareto shape parameter $k ≥ 0.7$. See PSISResult
for details and PSISPlots.paretoshapeplot
for a diagnostic plot.
PSIS.psis!
— Functionpsis(log_ratios, reff = 1.0; kwargs...) -> PSISResult
psis!(log_ratios, reff = 1.0; kwargs...) -> PSISResult
Compute Pareto smoothed importance sampling (PSIS) log weights [VehtariSimpson2021].
While psis
computes smoothed log weights out-of-place, psis!
smooths them in-place.
Arguments
log_ratios
: an array of logarithms of importance ratios, with size(draws, [chains, [parameters...]])
, wherechains>1
would be used when chains are generated using Markov chain Monte Carlo.reff::Union{Real,AbstractArray}
: the ratio(s) of effective sample size oflog_ratios
and the actual sample sizereff = ess/(draws * chains)
, used to account for autocorrelation, e.g. due to Markov chain Monte Carlo. If an array, it must have the size(parameters...,)
to matchlog_ratios
.
Keywords
warn=true
: Iftrue
, warning messages are deliverednormalize=true
: Iftrue
, the log-weights will be log-normalized so thatexp.(log_weights)
sums to 1 along the sample dimensions.
Returns
result
: aPSISResult
object containing the results of the Pareto-smoothing.
A warning is raised if the Pareto shape parameter $k ≥ 0.7$. See PSISResult
for details and PSISPlots.paretoshapeplot
for a diagnostic plot.
LOO and WAIC
PosteriorStats.AbstractELPDResult
— Typeabstract type AbstractELPDResult
An abstract type representing the result of an ELPD computation.
Every subtype stores estimates of both the expected log predictive density (elpd
) and the effective number of parameters p
, as well as standard errors and pointwise estimates of each, from which other relevant estimates can be computed.
Subtypes implement the following functions:
PosteriorStats.PSISLOOResult
— TypeResults of Pareto-smoothed importance sampling leave-one-out cross-validation (PSIS-LOO).
See also: loo
, AbstractELPDResult
estimates
: Estimates of the expected log pointwise predictive density (ELPD) and effective number of parameters (p)pointwise
: Pointwise estimatespsis_result
: Pareto-smoothed importance sampling (PSIS) results
PosteriorStats.WAICResult
— TypeResults of computing the widely applicable information criterion (WAIC).
See also: waic
, AbstractELPDResult
estimates
: Estimates of the expected log pointwise predictive density (ELPD) and effective number of parameters (p)pointwise
: Pointwise estimates
PosteriorStats.elpd_estimates
— Functionelpd_estimates(result::AbstractELPDResult; pointwise=false) -> (; elpd, elpd_mcse, lpd)
Return the (E)LPD estimates from the result
.
PosteriorStats.information_criterion
— Functioninformation_criterion(elpd, scale::Symbol)
Compute the information criterion for the given scale
from the elpd
estimate.
scale
must be one of (:deviance, :log, :negative_log)
.
information_criterion(result::AbstractELPDResult, scale::Symbol; pointwise=false)
Compute information criterion for the given scale
from the existing ELPD result
.
scale
must be one of (:deviance, :log, :negative_log)
.
If pointwise=true
, then pointwise estimates are returned.
PosteriorStats.loo
— Functionloo(log_likelihood; reff=nothing, kwargs...) -> PSISLOOResult{<:NamedTuple,<:NamedTuple}
Compute the Pareto-smoothed importance sampling leave-one-out cross-validation (PSIS-LOO). [Vehtari2017][LOOFAQ]
log_likelihood
must be an array of log-likelihood values with shape (chains, draws[, params...])
.
Keywords
reff::Union{Real,AbstractArray{<:Real}}
: The relative effective sample size(s) of the likelihood values. If an array, it must have the same data dimensions as the corresponding log-likelihood variable. If not provided, then this is estimated usingMCMCDiagnosticTools.ess
.kwargs
: Remaining keywords are forwarded to [PSIS.psis
].
See also: PSISLOOResult
, waic
Examples
Manually compute $R_\mathrm{eff}$ and calculate PSIS-LOO of a model:
julia> using ArviZExampleData, MCMCDiagnosticTools
julia> idata = load_example_data("centered_eight");
julia> log_like = PermutedDimsArray(idata.log_likelihood.obs, (:draw, :chain, :school));
julia> reff = ess(log_like; kind=:basic, split_chains=1, relative=true);
julia> loo(log_like; reff)
PSISLOOResult with estimates
elpd elpd_mcse p p_mcse
-31 1.4 0.9 0.34
and PSISResult with 500 draws, 4 chains, and 8 parameters
Pareto shape (k) diagnostic values:
Count Min. ESS
(-Inf, 0.5] good 7 (87.5%) 151
(0.5, 0.7] okay 1 (12.5%) 446
loo(data::Dataset; [var_name::Symbol,] kwargs...) -> PSISLOOResult{<:NamedTuple,<:Dataset}
loo(data::InferenceData; [var_name::Symbol,] kwargs...) -> PSISLOOResult{<:NamedTuple,<:Dataset}
Compute PSIS-LOO from log-likelihood values in data
.
If more than one log-likelihood variable is present, then var_name
must be provided.
Examples
Calculate PSIS-LOO of a model:
julia> using ArviZExampleData, PosteriorStats
julia> idata = load_example_data("centered_eight");
julia> loo(idata)
PSISLOOResult with estimates
elpd elpd_mcse p p_mcse
-31 1.4 0.9 0.34
and PSISResult with 500 draws, 4 chains, and 8 parameters
Pareto shape (k) diagnostic values:
Count Min. ESS
(-Inf, 0.5] good 6 (75.0%) 135
(0.5, 0.7] okay 2 (25.0%) 421
PosteriorStats.waic
— Functionwaic(log_likelihood::AbstractArray) -> WAICResult{<:NamedTuple,<:NamedTuple}
Compute the widely applicable information criterion (WAIC).[Watanabe2010][Vehtari2017][LOOFAQ]
log_likelihood
must be an array of log-likelihood values with shape (chains, draws[, params...])
.
See also: WAICResult
, loo
Examples
Calculate WAIC of a model:
julia> using ArviZExampleData
julia> idata = load_example_data("centered_eight");
julia> log_like = PermutedDimsArray(idata.log_likelihood.obs, (:draw, :chain, :school));
julia> waic(log_like)
WAICResult with estimates
elpd elpd_mcse p p_mcse
-31 1.4 0.9 0.33
waic(data::Dataset; [var_name::Symbol]) -> WAICResult{<:NamedTuple,<:Dataset}
waic(data::InferenceData; [var_name::Symbol]) -> WAICResult{<:NamedTuple,<:Dataset}
Compute WAIC from log-likelihood values in data
.
If more than one log-likelihood variable is present, then var_name
must be provided.
Examples
Calculate WAIC of a model:
julia> using ArviZExampleData, PosteriorStats
julia> idata = load_example_data("centered_eight");
julia> waic(idata)
WAICResult with estimates
elpd elpd_mcse p p_mcse
-31 1.4 0.9 0.33
Model comparison
PosteriorStats.ModelComparisonResult
— TypeModelComparisonResult
Result of model comparison using ELPD.
This struct implements the Tables and TableTraits interfaces.
Each field returns a collection of the corresponding entry for each model:
name
: Names of the models, if provided.rank
: Ranks of the models (ordered by decreasing ELPD)elpd_diff
: ELPD of a model subtracted from the largest ELPD of any modelelpd_diff_mcse
: Monte Carlo standard error of the ELPD differenceweight
: Model weights computed withweights_method
elpd_result
:AbstactELPDResult
s for each model, which can be used to access useful stats like ELPD estimates, pointwise estimates, and Pareto shape values for PSIS-LOOweights_method
: Method used to compute model weights withmodel_weights
PosteriorStats.compare
— Functioncompare(models; kwargs...) -> ModelComparisonResult
Compare models based on their expected log pointwise predictive density (ELPD).
The ELPD is estimated either by Pareto smoothed importance sampling leave-one-out cross-validation (LOO) or using the widely applicable information criterion (WAIC). We recommend loo. Read more theory here - in a paper by some of the leading authorities on model comparison dx.doi.org/10.1111/1467-9868.00353
Arguments
models
: aTuple
,NamedTuple
, orAbstractVector
whose values are eitherAbstractELPDResult
entries or any argument toelpd_method
.
Keywords
weights_method::AbstractModelWeightsMethod=Stacking()
: the method to be used to weight the models. Seemodel_weights
for detailselpd_method=loo
: a method that computes anAbstractELPDResult
from an argument inmodels
.sort::Bool=true
: Whether to sort models by decreasing ELPD.
Returns
ModelComparisonResult
: A container for the model comparison results. The fields contain a similar collection tomodels
.
Examples
Compare the centered and non centered models of the eight school problem using the defaults: loo
and Stacking
weights. A custom myloo
method formates the inputs as expected by loo
.
julia> using ArviZExampleData
julia> models = (
centered=load_example_data("centered_eight"),
non_centered=load_example_data("non_centered_eight"),
);
julia> function myloo(idata)
log_like = PermutedDimsArray(idata.log_likelihood.obs, (2, 3, 1))
return loo(log_like)
end;
julia> mc = compare(models; elpd_method=myloo)
┌ Warning: 1 parameters had Pareto shape values 0.7 < k ≤ 1. Resulting importance sampling estimates are likely to be unstable.
└ @ PSIS ~/.julia/packages/PSIS/...
ModelComparisonResult with Stacking weights
rank elpd elpd_mcse elpd_diff elpd_diff_mcse weight p ⋯
non_centered 1 -31 1.4 0 0.0 1.0 0.9 ⋯
centered 2 -31 1.4 0.06 0.067 0.0 0.9 ⋯
1 column omitted
julia> mc.weight |> pairs
pairs(::NamedTuple) with 2 entries:
:non_centered => 1.0
:centered => 5.34175e-19
Compare the same models from pre-computed PSIS-LOO results and computing BootstrappedPseudoBMA
weights:
julia> elpd_results = mc.elpd_result;
julia> compare(elpd_results; weights_method=BootstrappedPseudoBMA())
ModelComparisonResult with BootstrappedPseudoBMA weights
rank elpd elpd_mcse elpd_diff elpd_diff_mcse weight p ⋯
non_centered 1 -31 1.4 0 0.0 0.52 0.9 ⋯
centered 2 -31 1.4 0.06 0.067 0.48 0.9 ⋯
1 column omitted
PosteriorStats.model_weights
— Functionmodel_weights(elpd_results; method=Stacking())
model_weights(method::AbstractModelWeightsMethod, elpd_results)
Compute weights for each model in elpd_results
using method
.
elpd_results
is a Tuple
, NamedTuple
, or AbstractVector
with AbstractELPDResult
entries. The weights are returned in the same type of collection.
Stacking
is the recommended approach, as it performs well even when the true data generating process is not included among the candidate models. See [YaoVehtari2018] for details.
See also: AbstractModelWeightsMethod
, compare
Examples
Compute Stacking
weights for two models:
julia> using ArviZExampleData
julia> models = (
centered=load_example_data("centered_eight"),
non_centered=load_example_data("non_centered_eight"),
);
julia> elpd_results = map(models) do idata
log_like = PermutedDimsArray(idata.log_likelihood.obs, (2, 3, 1))
return loo(log_like)
end;
┌ Warning: 1 parameters had Pareto shape values 0.7 < k ≤ 1. Resulting importance sampling estimates are likely to be unstable.
└ @ PSIS ~/.julia/packages/PSIS/...
julia> model_weights(elpd_results; method=Stacking()) |> pairs
pairs(::NamedTuple) with 2 entries:
:centered => 5.34175e-19
:non_centered => 1.0
Now we compute BootstrappedPseudoBMA
weights for the same models:
julia> model_weights(elpd_results; method=BootstrappedPseudoBMA()) |> pairs
pairs(::NamedTuple) with 2 entries:
:centered => 0.483723
:non_centered => 0.516277
The following model weighting methods are available
PosteriorStats.AbstractModelWeightsMethod
— Typeabstract type AbstractModelWeightsMethod
An abstract type representing methods for computing model weights.
Subtypes implement model_weights
(method, elpd_results)
.
PosteriorStats.BootstrappedPseudoBMA
— Typestruct BootstrappedPseudoBMA{R<:Random.AbstractRNG, T<:Real} <: AbstractModelWeightsMethod
Model weighting method using pseudo Bayesian Model Averaging using Akaike-type weighting with the Bayesian bootstrap (pseudo-BMA+)[YaoVehtari2018].
The Bayesian bootstrap stabilizes the model weights.
BootstrappedPseudoBMA(; rng=Random.default_rng(), samples=1_000, alpha=1)
BootstrappedPseudoBMA(rng, samples, alpha)
Construct the method.
rng::Random.AbstractRNG
: The random number generator to use for the Bayesian bootstrapsamples::Int64
: The number of samples to draw for bootstrappingalpha::Real
: The shape parameter in the Dirichlet distribution used for the Bayesian bootstrap. The default (1) corresponds to a uniform distribution on the simplex.
See also: Stacking
PosteriorStats.PseudoBMA
— Typestruct PseudoBMA <: AbstractModelWeightsMethod
Model weighting method using pseudo Bayesian Model Averaging (pseudo-BMA) and Akaike-type weighting.
PseudoBMA(; regularize=false)
PseudoBMA(regularize)
Construct the method with optional regularization of the weights using the standard error of the ELPD estimate.
This approach is not recommended, as it produces unstable weight estimates. It is recommended to instead use BootstrappedPseudoBMA
to stabilize the weights or Stacking
. For details, see [YaoVehtari2018].
See also: Stacking
PosteriorStats.Stacking
— Typestruct Stacking{O<:Optim.AbstractOptimizer} <: AbstractModelWeightsMethod
Model weighting using stacking of predictive distributions[YaoVehtari2018].
Stacking(; optimizer=Optim.LBFGS(), options=Optim.Options()
Stacking(optimizer[, options])
Construct the method, optionally customizing the optimization.
optimizer::Optim.AbstractOptimizer
: The optimizer to use for the optimization of the weights. The optimizer must support projected gradient optimization via amanifold
field.options::Optim.Options
: The Optim options to use for the optimization of the weights.
See also: BootstrappedPseudoBMA
Predictive checks
PosteriorStats.loo_pit
— Functionloo_pit(y, y_pred, log_weights; kwargs...) -> Union{Real,AbstractArray}
Compute leave-one-out probability integral transform (LOO-PIT) checks.
Arguments
y
: array of observations with shape(params...,)
y_pred
: array of posterior predictive samples with shape(draws, chains, params...)
.log_weights
: array of normalized log LOO importance weights with shape(draws, chains, params...)
.
Keywords
is_discrete
: If not provided, then it is set totrue
iff elements ofy
andy_pred
are all integer-valued. Iftrue
, then data are smoothed usingsmooth_data
to make them non-discrete before estimating LOO-PIT values.kwargs
: Remaining keywords are forwarded tosmooth_data
if data is discrete.
Returns
pitvals
: LOO-PIT values with same size asy
. Ify
is a scalar, thenpitvals
is a scalar.
LOO-PIT is a marginal posterior predictive check. If $y_{-i}$ is the array $y$ of observations with the $i$th observation left out, and $y_i^*$ is a posterior prediction of the $i$th observation, then the LOO-PIT value for the $i$th observation is defined as
\[P(y_i^* \le y_i \mid y_{-i}) = \int_{-\infty}^{y_i} p(y_i^* \mid y_{-i}) \mathrm{d} y_i^*\]
The LOO posterior predictions and the corresponding observations should have similar distributions, so if conditional predictive distributions are well-calibrated, then all LOO-PIT values should be approximately uniformly distributed on $[0, 1]$.[Gabry2019]
Examples
Calculate LOO-PIT values using as test quantity the observed values themselves.
julia> using ArviZExampleData
julia> idata = load_example_data("centered_eight");
julia> y = idata.observed_data.obs;
julia> y_pred = PermutedDimsArray(idata.posterior_predictive.obs, (:draw, :chain, :school));
julia> log_like = PermutedDimsArray(idata.log_likelihood.obs, (:draw, :chain, :school));
julia> log_weights = loo(log_like).psis_result.log_weights;
julia> loo_pit(y, y_pred, log_weights)
╭───────────────────────────────╮
│ 8-element DimArray{Float64,1} │
├───────────────────────────────┴──────────────────────────────────────── dims ┐
↓ school Categorical{String} [Choate, Deerfield, …, St. Paul's, Mt. Hermon] Unordered
└──────────────────────────────────────────────────────────────────────────────┘
"Choate" 0.943511
"Deerfield" 0.63797
"Phillips Andover" 0.316697
"Phillips Exeter" 0.582252
"Hotchkiss" 0.295321
"Lawrenceville" 0.403318
"St. Paul's" 0.902508
"Mt. Hermon" 0.655275
Calculate LOO-PIT values using as test quantity the square of the difference between each observation and mu
.
julia> using Statistics
julia> mu = idata.posterior.mu;
julia> T = y .- median(mu);
julia> T_pred = y_pred .- mu;
julia> loo_pit(T .^ 2, T_pred .^ 2, log_weights)
╭───────────────────────────────╮
│ 8-element DimArray{Float64,1} │
├───────────────────────────────┴──────────────────────────────────────── dims ┐
↓ school Categorical{String} [Choate, Deerfield, …, St. Paul's, Mt. Hermon] Unordered
└──────────────────────────────────────────────────────────────────────────────┘
"Choate" 0.873577
"Deerfield" 0.243686
"Phillips Andover" 0.357563
"Phillips Exeter" 0.149908
"Hotchkiss" 0.435094
"Lawrenceville" 0.220627
"St. Paul's" 0.775086
"Mt. Hermon" 0.296706
loo_pit(idata::InferenceData, log_weights; kwargs...) -> DimArray
Compute LOO-PIT values using existing normalized log LOO importance weights.
Keywords
y_name
: Name of observed data variable inidata.observed_data
. If not provided, then the only observed data variable is used.y_pred_name
: Name of posterior predictive variable inidata.posterior_predictive
. If not provided, theny_name
is used.kwargs
: Remaining keywords are forwarded to the base method ofloo_pit
.
Examples
Calculate LOO-PIT values using already computed log weights.
julia> using ArviZExampleData, PosteriorStats
julia> idata = load_example_data("centered_eight");
julia> loo_result = loo(idata; var_name=:obs);
julia> loo_pit(idata, loo_result.psis_result.log_weights; y_name=:obs)
╭───────────────────────────────────────────╮
│ 8-element DimArray{Float64,1} loo_pit_obs │
├───────────────────────────────────────────┴──────────────────────────── dims ┐
↓ school Categorical{String} [Choate, Deerfield, …, St. Paul's, Mt. Hermon] Unordered
└──────────────────────────────────────────────────────────────────────────────┘
"Choate" 0.943511
"Deerfield" 0.63797
"Phillips Andover" 0.316697
"Phillips Exeter" 0.582252
"Hotchkiss" 0.295321
"Lawrenceville" 0.403318
"St. Paul's" 0.902508
"Mt. Hermon" 0.655275
loo_pit(idata::InferenceData; kwargs...) -> DimArray
Compute LOO-PIT from groups in idata
using PSIS-LOO.
Keywords
y_name
: Name of observed data variable inidata.observed_data
. If not provided, then the only observed data variable is used.y_pred_name
: Name of posterior predictive variable inidata.posterior_predictive
. If not provided, theny_name
is used.log_likelihood_name
: Name of log-likelihood variable inidata.log_likelihood
. If not provided, theny_name
is used ifidata
has alog_likelihood
group, otherwise the only variable is used.reff::Union{Real,AbstractArray{<:Real}}
: The relative effective sample size(s) of the likelihood values. If an array, it must have the same data dimensions as the corresponding log-likelihood variable. If not provided, then this is estimated usingess
.kwargs
: Remaining keywords are forwarded to the base method ofloo_pit
.
Examples
Calculate LOO-PIT values using as test quantity the observed values themselves.
julia> using ArviZExampleData, PosteriorStats
julia> idata = load_example_data("centered_eight");
julia> loo_pit(idata; y_name=:obs)
╭───────────────────────────────────────────╮
│ 8-element DimArray{Float64,1} loo_pit_obs │
├───────────────────────────────────────────┴──────────────────────────── dims ┐
↓ school Categorical{String} [Choate, Deerfield, …, St. Paul's, Mt. Hermon] Unordered
└──────────────────────────────────────────────────────────────────────────────┘
"Choate" 0.943511
"Deerfield" 0.63797
"Phillips Andover" 0.316697
"Phillips Exeter" 0.582252
"Hotchkiss" 0.295321
"Lawrenceville" 0.403318
"St. Paul's" 0.902508
"Mt. Hermon" 0.655275
Utilities
PosteriorStats.smooth_data
— Functionsmooth_data(y; dims=:, interp_method=CubicSpline, offset_frac=0.01)
Smooth y
along dims
using interp_method
.
interp_method
is a 2-argument callabale that takes the arguments y
and x
and returns a DataInterpolations.jl interpolation method, defaulting to a cubic spline interpolator.
offset_frac
is the fraction of the length of y
to use as an offset when interpolating.
- Hyndman1996Rob J. Hyndman (1996) Computing and Graphing Highest Density Regions, Amer. Stat., 50(2): 120-6. DOI: 10.1080/00031305.1996.10474359jstor.
- ChenShao1999Ming-Hui Chen & Qi-Man Shao (1999) Monte Carlo Estimation of Bayesian Credible and HPD Intervals, J Comput. Graph. Stat., 8:1, 69-92. DOI: 10.1080/10618600.1999.10474802jstor.
- GelmanGoodrich2019Andrew Gelman, Ben Goodrich, Jonah Gabry & Aki Vehtari (2019) R-squared for Bayesian Regression Models, The American Statistician, 73:3, 307-9, DOI: 10.1080/00031305.2018.1549100.
- VehtariSimpson2021Vehtari A, Simpson D, Gelman A, Yao Y, Gabry J. (2021). Pareto smoothed importance sampling. arXiv:1507.02646v7 [stat.CO]
- VehtariSimpson2021Vehtari A, Simpson D, Gelman A, Yao Y, Gabry J. (2021). Pareto smoothed importance sampling. arXiv:1507.02646v7 [stat.CO]
- VehtariSimpson2021Vehtari A, Simpson D, Gelman A, Yao Y, Gabry J. (2021). Pareto smoothed importance sampling. arXiv:1507.02646v7 [stat.CO]
- Vehtari2017Vehtari, A., Gelman, A. & Gabry, J. Practical Bayesian model evaluation using leave-one-out cross-validation and WAIC. Stat Comput 27, 1413–1432 (2017). doi: 10.1007/s11222-016-9696-4 arXiv: 1507.04544
- LOOFAQAki Vehtari. Cross-validation FAQ. https://mc-stan.org/loo/articles/online-only/faq.html
- Watanabe2010Watanabe, S. Asymptotic Equivalence of Bayes Cross Validation and Widely Applicable Information Criterion in Singular Learning Theory. 11(116):3571−3594, 2010. https://jmlr.csail.mit.edu/papers/v11/watanabe10a.html
- Vehtari2017Vehtari, A., Gelman, A. & Gabry, J. Practical Bayesian model evaluation using leave-one-out cross-validation and WAIC. Stat Comput 27, 1413–1432 (2017). doi: 10.1007/s11222-016-9696-4 arXiv: 1507.04544
- LOOFAQAki Vehtari. Cross-validation FAQ. https://mc-stan.org/loo/articles/online-only/faq.html
- YaoVehtari2018Yuling Yao, Aki Vehtari, Daniel Simpson, and Andrew Gelman. Using Stacking to Average Bayesian Predictive Distributions. 2018. Bayesian Analysis. 13, 3, 917–1007. doi: 10.1214/17-BA1091 arXiv: 1704.02030
- YaoVehtari2018Yuling Yao, Aki Vehtari, Daniel Simpson, and Andrew Gelman. Using Stacking to Average Bayesian Predictive Distributions. 2018. Bayesian Analysis. 13, 3, 917–1007. doi: 10.1214/17-BA1091 arXiv: 1704.02030
- YaoVehtari2018Yuling Yao, Aki Vehtari, Daniel Simpson, and Andrew Gelman. Using Stacking to Average Bayesian Predictive Distributions. 2018. Bayesian Analysis. 13, 3, 917–1007. doi: 10.1214/17-BA1091 arXiv: 1704.02030
- YaoVehtari2018Yuling Yao, Aki Vehtari, Daniel Simpson, and Andrew Gelman. Using Stacking to Average Bayesian Predictive Distributions. 2018. Bayesian Analysis. 13, 3, 917–1007. doi: 10.1214/17-BA1091 arXiv: 1704.02030
- Gabry2019Gabry, J., Simpson, D., Vehtari, A., Betancourt, M. & Gelman, A. Visualization in Bayesian Workflow. J. R. Stat. Soc. Ser. A Stat. Soc. 182, 389–402 (2019). doi: 10.1111/rssa.12378 arXiv: 1709.01449