M-step for models from class MixARgen

M-step for models from class MixARgen. This function is for use by other functions.

Usage

mixgenMstep(y, tau, model, index, fix = NULL, comp_sigma = FALSE,
            method = "BBsolve", maxit = 100, trace = FALSE,
            lessverbose = TRUE, ...)

Arguments

y

time series, a numeric vector.

tau

conditional probabilities, an object of class "MixComp".

model

the current model, an object from a subclass of class "MixAR".

index

indices of observations for which to compute residuals, a vector of positive integers, see 'Details'.

method

optimisation or equation solving method for package BB

...

arguments to pass on to optimisation functions, not thought over yet. Do not use until this notice is removed.

comp_sigma

If TRUE optimise the scale parameters using univariate optimisation. (note: does not work with argument 'fix' yet.)

fix

specify parameters to be held fixed during optimisation, see 'Details'.

maxit

maximal number of iterations for BB optimisers and solvers. Meant mainly for testing.

trace

if TRUE, BB optimisers and solvers will print information about their proceedings. Meant mainly for testing.

lessverbose

if TRUE, print a dot instead of more verbose information.

Details

mixgenMstep is an implementation of the M-step of the EM algorithm for mixture autoregressive models specified by objects of class "MixARgen". The function was build and modified incrementally with the main goal of providing flexibility. Speed will be addressed later.

By default optimisation is done with respect to all parameters. Argument fix may be a list with elements "prob", "shift", "scale" and "arcoef". These elements should be logical vectors containing TRUE in the positions of the fixed parameters. Elements with no fixed parameters may be omitted. (Currently the "prob" element is ignored, i.e. it is not possible to fix any of the component probabilities.)

If fix = "shift" the shift parameters are kept fixed. This is equivalent to fix = list(shift = rep(TRUE,g)).

The parameters (if any) of the distributions of the error components are estimated by default. Currently the above method cannot be used to fix some of them. This can be achieved however by modifying the distribution part of the model since that incorporates information about the parameters and whether they are fixed or not.

See also

fit_mixAR and mixARgenemFixedPoint which are meant to be called by users.