See the NEWS file for recent updates, and below for quick start!

• Quick start: https://github.com/cdriveraus/ctsem/
• Manual: https://github.com/cdriveraus/ctsem/raw/master/vignettes/hierarchicalmanual.pdf
• Tutorial: https://osf.io/preprints/psyarxiv/4q9ex_v2
• Conceptual / Individual differences: https://www.researchgate.net/publication/324093594_Hierarchical_Bayesian_Continuous_Time_Dynamic_Modeling
• Interventions: https://www.researchgate.net/publication/328221807_Understanding_the_Time_Course_of_Interventions_with_Continuous_Time_Dynamic_Models
• Blog: https://cdriver.netlify.app/
• Discussion: https://github.com/cdriveraus/ctsem/discussions

ctsem allows for easy specification and fitting of a range of continuous and discrete time dynamic models, including multiple indicators (dynamic factor analysis), multiple, potentially higher order processes, and time dependent (varying within subject) and time independent (not varying within subject) covariates. Classic longitudinal models like latent growth curves and latent change score models are also possible. Version 1 of ctsem provided SEM based functionality by linking to the OpenMx software, allowing mixed effects models (random means but fixed regression and variance parameters) for multiple subjects. For version 2 of the R package ctsem, we include a hierarchical specification and fitting routine that uses the Stan probabilistic programming language, via the rstan package in R. This allows for all parameters of the dynamic model to individually vary, using an estimated population mean and variance, and any time independent covariate effects, as a prior. Version 3 allows for state dependencies in the parameter specification (i.e. time varying parameters).

The current manual is at https://cran.r-project.org/package=ctsem/vignettes/hierarchicalmanual.pdf. The original ctsem is documented in a JSS publication (Driver, Voelkle, Oud, 2017), and in R vignette form at https://cran.r-project.org/package=ctsemOMX/vignettes/ctsemOMX.pdf, however these OpenMx based functions have been split off into a sub package, ctsemOMX. For most use cases the newer formulation (with Kalman filtering coded in Stan) is faster, more robust, and more flexible, and both default to maximum likelihood. For cases with many subjects, few time points, and no individual differences in timing, ctsemOMX may be faster.

For questions (or to see past answers) please use https://github.com/cdriveraus/ctsem/discussions

For some tutorials and another quick start, see . The very quick start is below.

To cite ctsem please use the citation(“ctsem”) command in R.

Function name update

From version 3.11.0 (June 2026), ctsem documentation and examples use shorter function names that avoid Stan-specific wording in the main user-facing API. Existing code using the older names should continue to work because the old names remain as compatibility aliases.

Use ctFit() as the main fitting function; ctStanFit() is now an alias. The model argument to ctFit() is now called model; the older ctstanmodel argument is deprecated but still accepted for existing scripts. For model specification, the usual modern workflow is:

model <- ctModel(type = "ct", ...)
fit <- ctFit(data, model)

or ctModel(type = "dt", ...) for discrete time models. ctModel(type = "omx") creates an old matrix-list object retained primarily for data generation and legacy workflows; these objects are not fitted directly by the current ctsem package. To adapt such an object to the modern fit-ready format, use:

model <- ctModelConvertOMX(omxmodel)

ctStanModel() remains an alias for ctModelConvertOMX(), but new material should use ctModelConvertOMX() when discussing conversion from old OpenMx-style model objects.

Other common name updates are:

• ctStanGenerate() -> ctGenerateFromPriors()
• ctStanGenerateFromFit() -> ctGenerateFromFit()
• ctStanKalman() -> ctKalmanArray()
• ctStanPlotPost() -> ctPlotPosterior()
• ctStanPostPredict() -> ctPostPredict()
• ctStanSubjectPars() -> ctSubjectPars()
• ctStanTIpredeffects() -> ctTIpredEffects()
• ctStanFitUpdate() -> ctFitUpdate()
• ctStanDiscretePars() -> ctDiscretePars()
• ctStanDiscreteParsPlot() -> ctDiscreteParsPlot()
• ctStanContinuousPars() -> ctSummaryMatrices()
• ctStanParnames() -> ctRawParnames()

Modern model objects can also be edited in matrix form via the pars-backed model$matrices view, for example model$matrices$DRIFT[1, 2] <- "cross". See the ctsem GitHub repository for current details and examples: https://github.com/cdriveraus/ctsem.

To install the github version, first install rstan and Rtools, then from a fresh R session:

remotes::install_github('cdriveraus/ctsem', INSTALL_opts = "--no-multiarch", dependencies = c("Depends", "Imports"))

Or just use the CRAN version, but rstan compiler setup is needed separately for some models:

install.packages('ctsem')

Troubleshooting Rstan / Rtools install for Windows:

Ensure recent version of R and Rtools is installed. If the installctsem.R code has never been run before, be sure to run that (see above).

Place this line in ~/.R/makevars.win , and if there are other lines, delete them:

CXX17FLAGS += -mtune=native -Wno-ignored-attributes -Wno-deprecated-declarations

For compile issues, check if you can use rstan, check forum posts on

In case of compile errors like g++ not found, ensure the devtools package is installed:

install.packages('devtools')

Quick start – univariate panel data with covariate effects on parameters

#’ The basic long data structure. Diet, (our covariate) is a categorical variable so needs dummy / ‘one hot’ encoding.

head(ChickWeight) 

#’ Setup dummy coding

library(data.table)
library(mltools)
chickdata <- one_hot(as.data.table(ChickWeight),cols = 'Diet')

#’ Scaling of continuous variables makes for easier estimation and more sensible default priors (if used). Time intervals can also benefit

chickdata$weight <- scale(chickdata$weight) 
head(chickdata) #now we have the four diet categories

#’ Setup continuous time model – in this case we are estimating a regular first order autoregressive

library(ctsem)

m <- ctModel(
  LAMBDA=diag(1), #Factor loading matrix of latent processes on measurements, fixed to 1
  type = 'ct', #Could specify 'dt' here for discrete time.
  tipredDefault = FALSE, #limit covariate effects on parameters to those explicitly specified
  manifestNames='weight', #Observed measurements of the latent processes
  latentNames='Lweight', #Names here simply make parameters and plots more interpretable
  TIpredNames = paste0('Diet_',2:4), #Covariates, in this case one category needs to be baseline...
  DRIFT='a11 | param', #normally self feedback (diagonal drift terms) are restricted to negative
  MANIFESTMEANS=0, #For identification CINT is normally zero with this freely estimated
  CINT='cint ||||Diet_2,Diet_3,Diet_4', #diet covariates specified in 5th 'slot' (four '|' separators)
  time='Time',
  id='Chick')

#’ View model in pdf/ latex form

ctModelLatex(m)

#’ Fit model to data – here using priors because Hessian problems are reported otherwise

f <- ctFit(chickdata,m,priors=TRUE)

#’ Summarise fit, view covariate effects – Diets 3 and 4 seem most obviously successful

s=summary(f)

print(s$tipreds )

#’ Predictions conditional on all earlier data

ctKalman(f,plot=TRUE,subjects=2:4,kalmanvec=c('yprior','ysmooth')) 

#’ Predictions conditional only on covariates, showing 1 chick from each diet

ctKalman(f,plot=T, 
  subjects=as.numeric(chickdata$Chick[!duplicated(ChickWeight$Diet)]),
  removeObs = T,polygonalpha=0)

#’ Plot temporal regression coefficients conditional on time interval – increases in this case!

ctDiscretePars(f,plot=T)

#’ Other useful functions:

#’ Compare two fits: ctChisqTest()

#’ Fit and summarise / plot a list of models: ctFitMultiModel()

#’ Add samples to fit to increase estimate precision: ctAddSamples()

#’ Return dynamic system parameters in matrix forms: ctSummaryMatrices()

#’ Compute cross validation statistics: ctLOO()

#’ Plot time independent predictor (covariate effects on parameters): ctTIpredEffects()

#’ Generate data from a specified model of fixed parameters: ctGenerate()

#’ Generate data from a specified model of fixed and free parameters / priors: ctGenerateFromPriors()

#’ Generate data from a fitted model: ctGenerateFromFit()

#’ Get samples from the fitted object: ctExtract()

#’ In samples, pop_DRIFT refers to the population drift matrix, subj_DRIFT refers to the subject matrix. Subject matrices only computed for max likelihood / posterior mode by default, and found in the $stanfit$transformedparsfull object.