ctsem (version 3.0.4)

Kalman: Kalman

Description

Takes list containing ctsem subject matrices, as well as long form data object, and calculates predicted and updated latent states, likelihoods, and predicted observations using the Kalman filter.

Usage

Kalman(kpars, datalong, manifestNames, latentNames,
  imputeMissings = FALSE, TDpredNames = NULL, continuoustime = TRUE,
  idcol = "id", timecol = "time", derrind = "all",
  optimize = FALSE, ukf = FALSE, plotoptim = FALSE)

Arguments

kpars

list object containing DRIFT,T0VAR,DIFFUSION,CINT,T0MEANS,TDPREDEFFECT, MANIFESTMEANS, LAMBDA, and MANIFESTVAR matrices, with list elements named accordingly. Such a list is returned by ctStanContinuousPars.

datalong

long format data object as used by ctStanFit, but must contain only a single subjects' data and does not need an id column.

manifestNames

String vector of names of manifest varifables to use from datalong.

latentNames

String vector of names of latent variables.

imputeMissings

Logical. If TRUE, randomly generate any missing observations of manifest variables according to model.

TDpredNames

If model contains time dependent predictors, string vector of their names in the data.

continuoustime

Logical, whether to use a continuous time Kalman filter or discrete time. Refers only to latent states, observations are always at discrete time points.

idcol

Character string giving name of subject identification column in data.

timecol

name of time column in datalong. Note that time column must be an ascending sequence of numeric values from row 1 to row n. Ignored if continuoustime=FALSE.

derrind

vector of integers denoting which latent variables are involved in covariance calcs.

optimize

Set to TRUE when using for optimization.

ukf

set to TRUE to use the unscented Kalman filter, only necessary for fitting non-linear models, currently only for optimizing.

plotoptim

set to TRUE to plot / print optimization steps.

Value

When optimize=TRUE, returns log likelihood. Else, returns a list containing matrix objects etaprior, etaupd, etasmooth, y, yprior, yupd, ysmooth, prederror, time, loglik, with values for each time point in each row. eta refers to latent states and y to manifest indicators - y itself is thus just the input data. Covariance matrices etapriorcov, etaupdcov, etasmoothcov, ypriorcov, yupdcov, ysmoothcov, are returned in a row * column * time array.

Examples

# NOT RUN {
### ctstantestfit is a dummy ctStanFit object with 2 manifest indicators,
###  4 latents, and 1 time dependent predictor.

### get parameter matrices
kpars <- ctStanContinuousPars(ctstantestfit)

#construct dummy data
datalong <- cbind(0:9, 1, matrix(rnorm(20,2,1),ncol=2))
datalong[c(1:3,9:10),3:4]<-NA #missing data to pre/fore cast
colnames(datalong) <- c('time', 'id', paste0('Y',1:2))
print(datalong)

#obtain Kalman filtered estimates
kout <- Kalman(kpars=kpars, datalong=datalong,
  manifestNames=paste0('Y',1:nrow(kpars$MANIFESTMEANS)),
  latentNames=paste0('eta',1:nrow(kpars$DRIFT)))

#print and plot smoothed estimates (conditional on all states) of indicators.
print(kout$ysmooth)
matplot(kout$time,kout$ysmooth,type='l')
matplot(kout$time,datalong[,3:4],type='p',add=TRUE,pch=1)
# }