HDInterval (version 0.2.4)

hdi: Highest (Posterior) Density Interval

Description

Calculate the highest density interval (HDI) for a probability distribution for a given probability mass. This is often applied to a Bayesian posterior distribution and is then termed "highest posterior density interval", but can be applied to any distribution, including priors.

The function is an S3 generic, with methods for a range of input objects.

Usage

hdi(object, credMass = 0.95, ...)
# S3 method for default
hdi(object, credMass = 0.95, ...)
# S3 method for function
hdi(object, credMass = 0.95, tol, ...)
# S3 method for matrix
hdi(object, credMass = 0.95, ...)
# S3 method for data.frame
hdi(object, credMass = 0.95, ...)
# S3 method for list
hdi(object, credMass = 0.95, ...)
# S3 method for density
hdi(object, credMass = 0.95, allowSplit=FALSE, ...)
# S3 method for mcmc
hdi(object, credMass = 0.95, ...)  
# S3 method for mcmc.list
hdi(object, credMass = 0.95, ...)  
# S3 method for mcarray
hdi(object, credMass = 0.95, ...)  
# S3 method for bugs
hdi(object, credMass = 0.95, ...)  
# S3 method for jagsUI
hdi(object, credMass = 0.95, ...)  
# S3 method for rjags
hdi(object, credMass = 0.95, ...)  
# S3 method for runjags
hdi(object, credMass = 0.95, ...)  

Value

a vector of length 2 or a 2-row matrix with the lower and upper limits of the HDI, with an attribute "credMass".

The density method with allowSplit=TRUE produces a matrix with a row for each component of a discontinuous HDI and columns for begin and end. It has an additional attribute "height" giving the probability density at the limits of the HDI.

Arguments

object

an object specifying the target distribution; see Details.

credMass

a scalar [0, 1] specifying the mass within the credible interval.

tol

the desired accuracy; see optimize; default is 1e-8.

allowSplit

only available for objects of class density; if FALSE and the proper HDI is discontinuous, a single credible interval is returned, but this is not HDI; see Value.

...

named parameters to be passed to other methods; see Examples.

Author

Mike Meredith and John Kruschke. Code for hdi.function based on hpd by Greg Snow, corrected by John Kruschke.

Details

The HDI is the interval which contains the required mass such that all points within the interval have a higher probability density than points outside the interval.

In contrast, a symmetric density interval defined by (eg.) the 10% and 90% quantiles may include values with lower probability than those excluded.

For a distribution that is not severely multimodal, the HDI is the narrowest interval containing the specified mass, and the hdi function actually returns the narrowest interval. This does not always work properly for multimodel densities, where the HDI may be discontinuous (the horizontal black line in the Figure below). The single interval returned by hdi (the blue line) may incorrectly include values between the modes with low probability density. The density method with allowSplit = TRUE gives separate limits for discontinuous HDIs.

The default method expects a vector representing draws from the target distribution, such as is produced by an MCMC process. Missing values are silently ignored; if the vector has no non-missing values, NAs are returned.

The matrix and data.frame methods expect an object with vectors of the above type for each parameter in columns. The result is a matrix with parameters in columns and rows with the upper and lower limits of the HDI.

The list method is a wrapper for lapply(object, hdi, ...). It is intended for output from rjags::jags.samples which returns a list of mcarray objects.

The mcmc.list, mcmc and mcarray methods expect objects of the respective types as defined in package coda.

The packages R2winbugs and R2openbugs produce bugs objects; R2jags produces rjags; jagsUI produces jagsUI; runjags produces runjags. For the rjags package, rjags::coda.samples produces a mcmc.list object and rjags::jags.samples produces a list of mcarray objects.

None of the above use interpolation: the values returned correspond to specific values in the data object, and will be conservative (ie, too wide rather than too narrow). Results thus depend on the random draws, and will be unstable if few values are provided. For a 95% HDI, 10,000 independent draws are recommended; a smaller number will be adequate for a 80% HDI, many more for a 99% HDI.

The function method requires the name for the inverse cumulative density function (ICDF) of the distribution; standard R functions for this have a q- prefix, eg. qbeta. Arguments required by the ICDF must be specified by their (abbreviated) names; see the examples.

References

Kruschke, J. K. 2011. Doing Bayesian data analysis: a tutorial with R and BUGS. Elsevier, Amsterdam, section 3.3.5.

Examples

# for a vector:
tst <- rgamma(1e5, 2.5, 2)
hdi(tst)
hdi(tst, credMass=0.8)
# For comparison, the symmetrical 80% CrI:
quantile(tst, c(0.1,0.9))

# for a density:
dens <- density(tst)
hdi(dens, credMass=0.8)

# Now a data frame:
tst <- data.frame(mu = rnorm(1e4, 4, 1), sigma = rlnorm(1e4))
hdi(tst, 0.8)
apply(tst, 2, quantile, c(0.1,0.9))
tst$txt <- LETTERS[1:25]
hdi(tst, 0.8)

# For a function:
hdi(qgamma, 0.8, shape=2.5, rate=2)
# and the symmetrical 80% CrI:
qgamma(c(0.1, 0.9), 2.5, 2)

# A multimodal distribution:
set.seed(2020)
tst2 <- c(rnorm(500, 30, 3), rnorm(200, 50, 1.5), 60 + rexp(200, 0.1))
hist(tst2, breaks = 100, freq=FALSE)
(hdiMC <- hdi(tst2))
segments(hdiMC[1], 0, hdiMC[2], 0, lwd=4, col='red', lend='butt')
# This is a valid 95% CrI, but not a Highest Density Interval

dens2 <- density(tst2, bw="SJ") # not the default for 'bw', see ?density
lines(dens2, lwd=2, col='blue')
(hdiD <- hdi(dens2, allowSplit=TRUE))
(ht <- attr(hdiD, "height"))
segments(hdiD[, 1], ht, hdiD[, 2], ht, lwd=4, col='blue', lend='butt')
# This is the correct 95% HDI.