bcov {eglhmm}

Bootstrap covariance.

Description

Creates an estimate of the covariance matrix of the parameter estimates for an extended generalised linear hidden Markov model via parametric bootstrapping.

Usage

   bcov(object, nsim = 50, itmax = 500, verbose = TRUE)

Arguments

object

An object of class eglhmm as produced by eglhmm().

nsim

The number of data sets to simulate, from which to estimate parameters. From each data set a vector of parameters is estimated; the estimated covariance matrix is the empirical covariance matrix of these nsim vectors.

itmax

The maximum number of iterations to be used in attempting to achieve convergence when fitting models to the simulated data sets. Note that if convergence is not achieved, the simulated data set being used is discarded (i.e. it “doesn't count”) and a replacement data set is simulated.

verbose

Logical scalar. Should a “progress report” be printed out at each step of the fitting procedure?

Value

A list with components:

C_hat

The parametric bootstrap estimate of the covariance matrix of the parameter estimates.

nc.count

A count of the total number of times that the algorithm failed to converge during the bootstrapping procedure.

an.count

A count of the “anomalies” that occurred, i.e. the number of times that there was a decrease in the log likelihood. Present only if the method used in fitting the models is "em".

Remarks

Although this documentation refers to “extended generalised linear models”, the only such models currently (13/02/2024) available are the Gaussian model with the identity link, the Poisson model, with the log link, the Binomial model with the logit link, the Dbd (discretised beta distribution model), and the Multinom model. The latter two are generalised linear models only in the “extended” sense. Other models may be added at a future date.

When eglhmm() is called by bcov() the argument checkDecrLL is set equal to FALSE. This has an effect only when the method used in fitting the models is "em". In this case a decrease in the log likelihood is treated as meaning that the algorithm has converged. Setting checkDecrLL equal to FALSE is done so as to decrease the number of discarded data sets and thereby speed up the rate at which the iterations proceed.

Author(s)

Rolf Turner [e-mail]

References

See the help for eglhmm() for references.

See Also

fitted.eglhmm() reglhmm() reglhmm.default() reglhmm.eglhmm()

Examples

    ## Not run:  # Takes too long.
        fitP <- eglhmm(y~locn+depth,data=SydColCount,distr="P",cells=c("locn","depth"),
                     K=2,contr="sum",verb=TRUE,itmax=300)
	cvrP <- bcov(fitP)
        fitD <- eglhmm(y~locn+depth,data=SydColCount,distr="D",cells=c("locn","depth"),
                     K=2,nbot=0,ntop=11,contr="sum",verb=TRUE)
	cvrD <- bcov(fitD)
        fitM <- eglhmm(y~locn+depth,data=SydColDisc,distr="M",cells=c("locn","depth"),
                     K=2,contr="sum",verb=TRUE)
	cvrM <- bcov(fitM)
    
## End(Not run)