This function returns one or more probability intervals of posterior samples.
p.interval(obj, HPD=TRUE, MM=TRUE, prob=0.95, plot=FALSE, PDF=FALSE, …)
This can be either a vector or matrix of posterior samples,
or an object of class demonoid
, iterquad
,
laplace
, pmc
, or vb
. If it is an object of
class demonoid
, then it will use only stationary posterior
samples and monitored target distributions (automatically discarding
the burn-in; if stationarity does not exist, then it will use all
samples).
Logical. This argument defaults to TRUE
, in which
case one or more Highest Posterior Density (HPD) intervals is
returned. When FALSE
, one or more quantile-based probability
intervals is returned.
Logical. This argument defaults to TRUE
, in which
case each column vector is checked for multimodality, and if found,
the multimodal form of a Highest Posterior Density (HPD) interval is
additionally estimated, even when HPD=FALSE
.
This is a numeric scalar in the interval (0,1) giving the target probability interval, and defaults to 0.95, representing a 95% probability interval. A 95% probability interval, for example, is an interval that contains 95% of a posterior probability distribution.
Logical. When plot=TRUE
, each kernel density is
plotted and shaded gray, and the area under the curve within the
probability interval is shaded black. If the kernel density is
considered to be multimodal, then up to three intervals are shaded
black. A vertical, red, dotted line is added at zero. The
plot
argument defaults to FALSE
.
Logical. When PDF=TRUE
, and only when
plot=TRUE
, plots are saved as a .pdf file in the working
directory.
Additional arguments are unused.
A matrix is returned with rows corresponding to the parameters (or
deviance or monitored variables), and columns "Lower"
and
"Upper"
. The elements of the matrix are the unimodal
probability intervals. The attribute "Probability"
is the
user-selected probability width. If MM=TRUE
and multimodal
posterior distributions are found, then multimodal HPD intervals are
printed to the screen in a character string.
A probability interval, also called a credible interval or Bayesian
confidence interval, is an interval in the domain of a posterior
probability distribution. When generalized to multivariate forms, it
is called a probability region (or credible region), though some
sources refer to a probability region (or credible region) as the
area within the probability interval. Bivariate probability regions
may be plotted with the joint.pr.plot
function.
The p.interval
function may return different probability
intervals: a quantile-based probability interval, a unimodal
Highest Posterior Density (HPD) interval, and multimodal HPD
intervals. Another type of probability interval is the Lowest
Posterior Loss (LPL) interval, which is calculated with the
LPL.interval
function.
The quantile-based probability interval is used most commonly,
possibly because it is simple, the fastest to calculate, invariant
under transformation, and more closely resembles the frequentist
confidence interval. The lower and upper bounds of the
quantile-based probability interval are calculated with the
quantile
function. A 95% quantile-based probability interval
reports the values of the posterior probability distribution that
indicate the 2.5% and 97.5% quantiles, which contain the central
95% of the distribution. The quantile-based probability interval is
centered around the median and has equal-sized tails.
The HPD (highest posterior density) interval is identical to the quantile-based probability interval when the posterior probability distribution is unimodal and symmetric. Otherwise, the HPD interval is the smallest interval, because it is estimated as the interval that contains the highest posterior density. Unlike the quantile-based probability interval, the HPD interval could be one-tailed or two-tailed, whichever is more appropriate. However, unlike the quantile-based interval, the HPD interval is not invariant to reparameterization (Bernardo, 2005).
The unimodal HPD interval is estimated from the empirical CDF of the sample for each parameter (or deviance or monitored variable) as the shortest interval for which the difference in the ECDF values of the end-points is the user-specified probability width. This assumes the distribution is not severely multimodal.
As an example, imagine an exponential posterior distribution. A quantile-based probability interval would report the highest density region near zero to be outside of its interval. In contrast, the unimodal HPD interval is recommended for such skewed posterior distributions.
When MM=TRUE
, the is.multimodal
function is
applied to each column vector after the unimodal interval (either
quantile-based or HPD) is estimated. If multimodality is found, then
multimodal HPD intervals are estimated with kernel density and
printed to the screen as a character string. The original unimodal
intervals are returned in the output matrix, because the matrix is
constrained to have a uniform number of columns per row, and because
multimodal HPD intervals may be disjoint.
Disjoint multimodal HPD intervals have multiple intervals for one posterior probability distribution. An example may be when there is a bimodal, Gaussian distribution with means -10 and 10, variances of 1 and 1, and a 95% probability interval is specified. In this case, there is not enough density between these two distant modes to have only one probability interval.
The user should also consider LPL.interval
, since it is
invariant to reparameterization like the quantile-based probability
interval, but could be one- or two-tailed, whichever is more
appropriate, like the HPD interval. A comparison of the quantile-based
probability interval, HPD interval, and LPL interval is available
here: https://web.archive.org/web/20150214090353/http://www.bayesian-inference.com/credible.
Bernardo, J.M. (2005). "Intrinsic Credible Regions: An Objective Bayesian Approach to Interval Estimation". Sociedad de Estadistica e Investigacion Operativa, 14(2), p. 317--384.
is.multimodal
,
IterativeQuadrature
,
joint.pr.plot
,
LaplaceApproximation
,
LaplacesDemon
,
LPL.interval
,
PMC
, and
VariationalBayes
.
# NOT RUN {
##First, update the model with the LaplacesDemon function.
##Then
#p.interval(Fit, HPD=TRUE, MM=TRUE, prob=0.95)
# }
Run the code above in your browser using DataLab