PeriodCPT.bern: ~~function to evaluate within period changepoint events for...

In taylors2/PeriodCPT: Within Period Changepoint Detection for Periodic Data

~~function to evaluate within period changepoint events for binary data~~

Description

Evaluation of within period changepoint events for Bernoulli distributed observations.

Usage

PeriodCPT.bern(data, periodlength = NULL, minseglen = 1, Mprior = c("pois", "unif"),
                           Mhyp = 1, spread = 1, param.a = 1, param.b = 1,
                           inits = NULL, n.iter = 1e6, n.chains = 1, n.burn = 0,
                           cachesize = 50, quiet = FALSE, ...)

Arguments

data	A vector or ts object containing the data within which you which to find within period changepoint events.
periodlength	Positive integer specifying the period length. If NULL, then information is obtained from ts object specified by data.
minseglen	Positive integer giving the minumum segment length (no. of observations between within-period changes), set at 1 for default.
Mprior	Option specifying the prior for the number of within period changepoint events.
Mhyp	Numeric vector giving the hyper-parameters for the number of within period chagepoint prior, see details.
spread	Positive number specifying the prior degree of spread around the period for the changepoint positions. By default, 1 indicates random uniform spread.
param.a, param.b	Positive numbers giving the shape hyper-parameters for the beta prior distribution on the within period segment parameter. Both set to 1 by default, giving sepecifying a uniform distribution.
inits	Initial values for the chains. See details
n.iter	Positive integer giving the number of iterations for the MCMC.
n.chains	Positive integer giving the number of chains to evaluate.
n.burn	Positive integer giving the initial number iterations to discard as the burn-in period.
cachesize	Positive integer giving the cache size to minimise re-evaluation of repeated scenarios within the MCMC algorithm. Set to 50 by default.
quiet	Logical flag indicating whether a progress bar should be printed
...	Additional inputs, see details.

Details

This function is used to find with period changepoint events for selected sampling distribution and statistic. A changepoint is denoted as the (!)LAST OBSERVATION OF THE CURRENT SEGMENT/REGIME(!)

Segment parameters: This model assumes that, for a given segment, data~Bern(theta) with prior distribution theta~beta(prior.a, prior.b).

Within period changepoint parameters: The prior for the number of within period changepoint events is either:

pois

Poisson distribution truncated to be between 1 and the maximum number permitted within the period given the minimum segment length constraint. In this case, Mhyp should contain a single positive number that denotes the lambda hyper-parameter.

unif

Discrete uniform distribution on the integers from 1 to the maximum number permitted within the period given the minimum segment length constraint. The argument Mhyp is ignored.

Note: the scenario where the number of within period changepoints is equal to 1 defines the case where there are no within-period changepoint events and so model reduces to just the sampling distribution. Given the total number of changepoint events, the prior for their positions is given in terms of their excess segment length beyond the minimum condition: delta[i] = pcpt[i] - pcpt[i-1] - minseglen. The prior on the vector of excess segment lengths is a Dirichlet-multinomial distribution with shape vector given by spread (same size for all components). If spread<1 then there is a prior preference that the changepoints cluster, spread>1 indicates a prior preference for repellent events, else spread=1 specifies a uniform distribution. The excess segment lengths are mapped back to the within period changepoint positions with the aid of a anchoring event that follows a uniform distribution over the within peroid positions. If necessary, events are permuted such that all are defined within the interval [1, periodlength].

The chain inital values, init, can be specified in one of the following ways:

NULL

Default, positions are generated from the prior for each chain.

"ends"

This option specifies inital values for two chains (adjusts n.chains() accordingly) where the first denotes the trivial case of no within period changepoint events, and the other denots a packed case by fitting in as many events as possible with respect to the minimum segment length.

vector

A vector of integers giving the inital value for one (and only one) requested chain.

list

A list where each item specifies the vector of initial values for each chain.

function

The initalisation of a given chain is determined/generated by a provided function. Input arguments for this function can be specified via ... of the main PeriodCPT funciton. The input names pcpt.object and chain.index can be specified to respectively obtain information held within the internally generated pcpt class object or to specify a function that depends on a specific chain via an integer index.

Note: For codeing ease, the trivial case of no changepoint events is specified by c(1).

Value

An object of S4 class "pcpt" is returned.

Author(s)

Simon Taylor

See Also

pcpt, PeriodCPT

Examples

  set.seed(1)
  data = ts( rbinom(120, size = 1, prob = rep(c(0.2, 0.8), each=6)), freq = 12)
  ans  = PeriodCPT.bern(data, niter = 1e3)

taylors2/PeriodCPT documentation built on June 28, 2024, 12:32 p.m.