est.changepoints function

est.changepoints

This function estimates multiple change points using marginal likelihood method proposed by Du, Kao and Kou (2015), which we would denoted as DKK2015 afterward.

est.changepoints(data.x, model, prior, max.segs, logH, logMD)

Arguments

• data.x: Observed data in vector or matrix form. When the data is in matrix form, each column should represent a single observation.

• model: The specified distributional assumption. Currently we have implemented two arguments: "normal" (data follows one dimensional Normal distribution with unknown mean and variance) and "poisson" (data follows Poisson distribution with unknown intensity). A third argument "user" is also accepted, given that the prior and the log marginal likelihood function are specified in the parameter prior and logMD.

• prior: The prespecified prior parameters in consistent with the form used in logMD.

  For the proposed priors in DKK2015, use the corresponding prior function provided.

• max.segs: (Opt.) The maximum number of segments allowed, which is the value M in DKK2015. Must be a positive integer greater then 1. If missing, the function would process using the algorihtm by Jackson et al.(2005).

• logH: (Opt.) A Boolean algebra determine whether to report the log H matrix in DKK2015. Default is FALSE.

• logMD: (Opt.) The log marginal likelihood function (which is the log of D function in DKK2015). The function must be in the form of logMD(data.x, prior).

Returns

If logH is FALSE, the function returns the set of estimated change-points by the index of the data, where each index is the end point of a segment. If the result is no change-points, the function returns NULL.

If logH is TRUE, then the function returns a list with three components: changePTs is the set of estimated change-points, log.H is the log value for the H matrix used in the algorithm, where $log.H(m,i) = log H(x1, x2, ..., xi | m)$, and max.j

records the $j$ that maximizes the marginal likelihood in each step. See the manual in data folder for more details.

Details

See Manual.pdf in "data" folder.

Examples

library(StepSignalMargiLike)

n <- 5
data.x <- rnorm(n, 1, 1)
data.x <- c(data.x, rnorm(n, 10,1))
data.x <- c(data.x, rnorm(n, 2,1))
data.x <- c(data.x, rnorm(n, 10,1))
data.x <- c(data.x, rnorm(n, 1,1))

prior <- prior.norm.A(data.x)
max.segs <- 10

est.changepoints(data.x=data.x, model="normal", prior=prior)
est.changepoints(data.x=data.x, model="normal", prior=prior, max.segs=max.segs)
est.changepoints(data.x=data.x, model="normal", prior=prior, max.segs=max.segs,logH=TRUE)

References

Chao Du, Chu-Lan Michael Kao and S. C. Kou (2015), "Stepwise Signal Extraction via Marginal Likelihood". Forthcoming in Journal of American Statistical Association.