ctsem (version 2.8.2)

ctFit: Fit a ctsem object

Description

This function fits continuous time SEM models specified via ctModel to a dataset containing one or more subjects.

Usage

ctFit(dat, ctmodelobj, dataform = "wide", objective = "auto",
  stationary = c("T0TRAITEFFECT", "T0TIPREDEFFECT"),
  optimizer = "CSOLNP", retryattempts = 15, iterationSummary = FALSE,
  carefulFit = TRUE, carefulFitWeight = 100, showInits = FALSE,
  asymptotes = FALSE, meanIntervals = FALSE,
  plotOptimization = FALSE, crossEffectNegStarts = TRUE, fit = TRUE,
  nofit = FALSE, discreteTime = FALSE, verbose = 0,
  useOptimizer = TRUE, omxStartValues = NULL,
  transformedParams = TRUE, datawide = NA)

Arguments

dat

the data you wish to fit a ctsem model to, in either wide format (one individual per row), or long format (one time point of one individual per row). See details.

ctmodelobj

the ctsem model object you wish to use, specified via the ctModel function.

dataform

either "wide" or "long" depending on which input format you wish to use for the data. See details and or vignette.

objective

'auto' selects either 'Kalman', if fitting to single subject data, or 'mxRAM' for multiple subjects. For single subject data, 'Kalman' uses the mxExpectationStateSpace function from OpenMx to implement the Kalman filter. For more than one subject, 'mxRAM' specifies a wide format SEM with a row of data per subject. 'cov' may be specified, in which case the 'meanIntervals' argument is set to TRUE, and the covariance matrix of the supplied data is calculated and fit instead of the raw data. This is much faster but only a rough approximation, unless there are no individual differences in time interval and no missing data. 'Kalman' may be specified for multiple subjects, however as no trait matrices are used by the Kalman filter one must consider how average level differences between subjects are accounted for. See ctMultigroupFit for the possibility to apply the Kalman filter over multiple subjects)

stationary

Character vector of T0 matrix names in which to constrain any free parameters to stationarity. Defaults to c('T0TRAITEFFECT','T0TIPREDEFFECT'), constraining only between person effects to stationarity. Use NULL for no constraints, or 'all' to constrain all T0 matrices.

optimizer

character string, defaults to the open-source 'CSOLNP' optimizer that is distributed in all versions of OpenMx. However, 'NPSOL' may sometimes perform better for these problems, though requires that you have installed OpenMx via the OpenMx web site, by running: source('http://openmx.psyc.virginia.edu/getOpenMx.R')

retryattempts

Number of times to retry the start value randomisation and fit procedure, if non-convergance or uncertain fits occur.

iterationSummary

if TRUE, outputs limited fit details after every fit attempt.

carefulFit

if TRUE, first fits the specified model with a penalised likelihood function to force MANIFESTVAR, DRIFT, TRAITVAR, MANIFESTTRAITVAR parameters to remain close to 0, then fits the specified model normally, using these estimates as starting values. Can help to ensure optimization begins at sensible, non-exteme values, though results in any user specified start values being ignored for the final fit (though they are still used for initial fit).

carefulFitWeight

Positive numeric. Sets the weight for the penalisation (or prior) applied by the carefulFit algorithm. Generally unnecessary to adjust, may be helpful to try a selection of values (perhaps between 0 and 1000) when optimization is problematic.

showInits

if TRUE, prints the list of starting values for free parameters. These are the 'raw' values used by OpenMx, and reflect the log (var / cov matrices) or -log(DRIFT matrices) transformations used in ctsem. These are saved in the fit object under fitobject$omxStartValues.

asymptotes

when TRUE, optimizes over asymptotic parameter matrices instead of continuous time parameter matrices. Can be faster for optimization and in some cases makes reliable convergance easier. Will result in equivalent models when continuous time input matrices (DRIFT, DIFFUSION, CINT) are free, but fixing the values of any such matrices will result in large differences - a value of 0 in a cell of the normal continuous time DIFFUSION matrix does not necessarily result in a value of 0 for the asymptotic DIFFUSION matrix, for instance.

meanIntervals

Use average time intervals for each column for calculation (both faster and inaccurate to the extent that intervals vary across individuals).

plotOptimization

If TRUE, uses checkpointing for OpenMx function mxRun, set to checkpoint every iteration, output checkpoint file to working directory, then creates a plot for each parameter's values over iterations.

crossEffectNegStarts

Logical. If TRUE (default) free DRIFT matrix cross effect parameters have starting values set to small negative values (e.g. -.05), if FALSE, the start values are 0. The TRUE setting is useful for easy initialisation of higher order models, while the FALSE setting is useful when one has already estimated a model without cross effects, and wishes to begin optimization from those values by using the omxStartValues switch. are re-transformed into regular continuous time parameter matrices, and may be interpreted as normal.

fit

if TRUE, output only openmx model without fitting

nofit

Deprecated. If FALSE, output only openmx model without fitting

discreteTime

Estimate a discrete time model - ignores timing information, parameter estimates will correspond to those of classical vector autoregression models, OpenMx fit object will be directly output, thus ctsem summary and plot functionality will be unavailable. Time dependent predictor type also becomes irrelevant.

verbose

Integer between 0 and 3. Sets mxComputeGradientDescent messaging level, defaults to 0.

useOptimizer

Logical. Defaults to TRUE. Passes argument to mxRun, useful for using custom optimizers or fitting to specified parameters.

omxStartValues

A named vector containing the raw (potentially log transformed) OpenMx starting values for free parameters, as captured by OpenMx function omxGetParameters(ctmodelobj$mxobj). These values will take precedence over any starting values already specified using ctModel.

transformedParams

Logical indicating whether or not to log transform certain parameters internally to allow unconstrained estimation over entire 'sensible' range for parameters. When TRUE (default) raw OpenMx parameters (only reported if verbose=TRUE argument used for summary function) will reflect these transformations and may be harder to interpret, but summary matrices are reported as normal.

datawide

included for compatibility with scripts written for earlier versions of ctsem. Do not use this argument, instead use the dat argument, and the dataform argument now specifies whether the data is in wide or long format.

Details

For full discussion of how to structure the data and use this function, see the vignette using: vignette('ctsem'), or the data examples data("longexample") ; longexample for long and data("datastructure") ; datastructure for wide. If using long format, the subject id column must be numeric and grouped by ascending time within subject, and named 'id'. The time column must also be numeric, and representing absolute time (e.g., since beginning of study, *not* time intervals), and called 'time'. Models are specified using the ctModel function. For help regarding the summary function, see summary.ctsemFit, and for the plot function, plot.ctsemFit. Multigroup models may be specified using ctMultigroupFit. Confidence intervals for any matrices and or parameters may be estimated using ctCI. Difficulties during estimation can sometimes be alleviated using ctRefineTo instead of ctFit -- this uses a multistep fit procedure.

Examples

# NOT RUN {
## Examples set to 'dontrun' because they take longer than 5s.
# }
# NOT RUN {
mfrowOld<-par()$mfrow
par(mfrow=c(2, 3))

### example from Driver, Oud, Voelkle (2017), 
### simulated happiness and leisure time with unobserved heterogeneity.
data(ctExample1)
traitmodel <- ctModel(n.manifest=2, n.latent=2, Tpoints=6, LAMBDA=diag(2), 
  manifestNames=c('LeisureTime', 'Happiness'), 
  latentNames=c('LeisureTime', 'Happiness'), TRAITVAR="auto")
traitfit <- ctFit(dat=ctExample1, ctmodelobj=traitmodel)
summary(traitfit)
plot(traitfit, wait=FALSE)


###Example from Voelkle, Oud, Davidov, and Schmidt (2012) - anomia and authoritarianism.  
data(AnomAuth) 
AnomAuthmodel <- ctModel(LAMBDA = matrix(c(1, 0, 0, 1), nrow = 2, ncol = 2), 
Tpoints = 5, n.latent = 2, n.manifest = 2, MANIFESTVAR=diag(0, 2), TRAITVAR = NULL) 
AnomAuthfit <- ctFit(AnomAuth, AnomAuthmodel)
summary(AnomAuthfit)


### Single subject time series - using Kalman filter (OpenMx statespace expectation)
data('ctExample3')
model <- ctModel(n.latent = 1, n.manifest = 3, Tpoints = 100, 
  LAMBDA = matrix(c(1, 'lambda2', 'lambda3'), nrow = 3, ncol = 1), 
  CINT= matrix('cint'),
  MANIFESTMEANS = matrix(c(0, 'manifestmean2', 'manifestmean3'), nrow = 3, 
    ncol = 1))
fit <- ctFit(dat = ctExample3, ctmodelobj = model, objective = 'Kalman', 
  stationary = c('T0VAR'))


###Oscillating model from Voelkle & Oud (2013). 
data("Oscillating")

inits <- c(-39, -.3, 1.01, 10.01, .1, 10.01, 0.05, .9, 0)
names(inits) <- c("crosseffect","autoeffect", "diffusion",
  "T0var11", "T0var21", "T0var22","m1", "m2", 'manifestmean')

oscillatingm <- ctModel(n.latent = 2, n.manifest = 1, Tpoints = 11,
  MANIFESTVAR = matrix(c(0), nrow = 1, ncol = 1),
  LAMBDA = matrix(c(1, 0), nrow = 1, ncol = 2),
  T0MEANS = matrix(c('m1', 'm2'), nrow = 2, ncol = 1),
  T0VAR = matrix(c("T0var11", "T0var21", 0, "T0var22"), nrow = 2, ncol = 2),
  DRIFT = matrix(c(0, "crosseffect", 1, "autoeffect"), nrow = 2, ncol = 2),
  CINT = matrix(0, ncol = 1, nrow = 2),
  MANIFESTMEANS = matrix('manifestmean', nrow = 1, ncol = 1),
  DIFFUSION = matrix(c(0, 0, 0, "diffusion"), nrow = 2, ncol = 2),
  startValues=inits)

oscillatingf <- ctFit(Oscillating, oscillatingm, carefulFit = FALSE)
# }