eglhmm (version 0.1-3)

reorder.eglhmm: Reorder the states of a fitted eglhmm model

Description

Reorder the states of a fitted eglhmm model so that the state effects are in decreasing order.

Usage

# S3 method for eglhmm
reorder(x, ...)

Value

An object of class c("eglhmm","reordered") which is identical to the argument x in most respects. The components which (may) differ are:

• tpm

• ispd

• phi

• theta

• Hessian

• gradient

• mean and sd, or lambda or p, or alpha and beta (depending on which distribution is being used)

• fy

The entries of these components will have the same numerical values as before but, given that the ordering of the states has actually changed, will have different orderings, corresponding to the new ordering of the states.

Note that the attribute

preSpecSigma of the component theta, may differ from what it was in x.

The returned value will also have a component neworder

which is an integer vector providing the indices of the reordering of the states. It also currently (format(Sys.Date(),"%d/%m/%Y")) has a component newlog.like. This should (if there is any justice in the world --- but there isn't!) have the same value as the component log.like. Once I am confident that everything is working as it should, the newlog.like

component will be removed.

Arguments

x

An object of class "eglhmm" as returned by the function eglhmm().

...

Not used.

Author

Rolf Turner

Details

The states of a fitted hidden Markov model are usually in a rather arbitrary order, which can sometimes make it difficult to compare different fits. This function reorders the states so that the state corresponding to the “largest state effect” comes first (and so on down the line). What is meant by “largest state effect” depends on whether the distribution used in the model is “Dbd”. If the distribution is not “Dbd”, then what is meant is simply the largest of those entries of phi which correspond to state. (The vector phi is the vector of coefficients of the linear predictor in the model. Note that, since the formula for the model is constructed as y~0+state+..., the “state” coefficients are unconstrained and there are as many of them as there are states.)

If the distribution in question is “Dbd” then things are a bit more complicated. We calculate the theoretical expected values for “Dbd”s with parameters \(\alpha =\) alpha[k] and \(\beta =\) beta[k] where alpha[k] and beta[k] are the parameter values corresponding to the kth state. The states are then ordered according to the decreasing order of these expected values. These expected values are the expected values of the emissions given that all predictors other than the state predictors are zero.

See Also

eglhmm()

Examples

    loc4 <- c("LngRf","BondiE","BondiOff","MlbrOff")
    SCC4 <- SydColCount[SydColCount$locn %in% loc4,]
    SCC4$locn <- factor(SCC4$locn) # Get rid of unused levels.
    rownames(SCC4) <- 1:nrow(SCC4)
    fit  <- eglhmm(y~locn+depth,data=SCC4,cells=c("locn","depth"),
                  K=2,distr="P",verb=TRUE)
    ofit <- reorder(fit)