mssm: Get Multivariate State Space Model Functions
In mssm: Multivariate State Space Models

Description Usage Arguments Value See Also Examples

Returns an object with a function that can be used to run a particle filter, a function to perform parameter estimation using a Laplace approximation, and a function to perform smoothing of particle weights.

1 2	mssm(fixed, family, data, random, weights, offsets, ti, control = mssm_control())

`fixed`	`formula` with outcome variable on the left hand side and covariates with fixed effects on the right hand side.
`family`	family for the observed outcome given the state variables and covariates.
`data`	`data.frame` or environment containing the variables in `fixed` and `random`.
`random`	`formula` for covariates with a random effect. Left hand side is ignored.
`weights`	optional prior weights.
`offsets`	optional a priori known component in the linear predictor.
`ti`	integer vector with time indices matching with each observation of `fixed` and `random`.
`control`	list with arguments passed to `mssm_control`.

An object of class mssmFunc with the following elements

`pf_filter`	function to perform particle filtering. See mssm-pf.
`Laplace`	function to perform parameter estimation with a Laplace approximation. See mssm-Laplace.
`smoother`	function to compute smoothing weights for an `mssm` object returned by the `pf_filter` function. See mssm-smoother.
`terms_fixed`	`terms.object` for the covariates with fixed effects.
`terms_random`	`terms.object` for the covariates with random effects.
`y`	vector with outcomes.
`X`	covariates with fixed effects.
`Z`	covariates with random effects.
`ti`	time indices for each observation.
`weights`	prior weights for each observation.
`offsets`	a priori known component in the linear predictor for each observation.
`call`	the matched call.
`family`	character describing the conditional distribution of the outcomes.

The README of the package contains examples of how to use this function. See https://github.com/boennecd/mssm.

if(require(Ecdat)){
  # load data and fit glm to get starting values
  . <- print
  data("Gasoline", package = "Ecdat")
  glm_fit <- glm(lgaspcar ~ factor(country) + lincomep + lrpmg + lcarpcap,
                 Gamma("log"), Gasoline)

  # get object to perform estimation
  library(mssm)
  ll_func <- mssm(
  fixed = formula(glm_fit), random = ~ 1, family = Gamma("log"),
    data = Gasoline, ti = year, control = mssm_control(
      N_part = 1000L, n_threads = 1L))
  .(ll_func)

  # fit model with time-varying intercept with Laplace approximation
  disp <- summary(glm_fit)$dispersion
  laplace <- ll_func$Laplace(
    cfix = coef(glm_fit), disp = disp, F. = diag(.5, 1), Q = diag(1))
  .(laplace)

  # compare w/ glm
  .(logLik(laplace))
  .(logLik(glm_fit))
  .(rbind(laplace = laplace$cfix, glm = coef(glm_fit)))

  # run particle filter
  pf <- ll_func$pf_filter(
    cfix = laplace$cfix, disp = laplace$disp, F. = laplace$F., Q = laplace$Q)
  .(pf)

  # compare approximate log-likelihoods
  .(logLik(pf))
  .(logLik(laplace))

  # predicted values from filtering (does not appear random...)
  plot(pf)

  # plot predicted values from smoothing distribution
  pf <- ll_func$smoother(pf)
  plot(pf, which_weights = "smooth")
}