BayesPGMM: Fit Bayesian Piecewise Growth Mixture Model

In lockEF/BayesianPGMM: Bayesian Piecewise Growth Mixture Modeling

Fit Bayesian Piecewise Growth Mixture Model

Description

Estimates a Bayesian piecewise growth mixture model with linear segments, with a given latent number of classes and a latent number of possible change points in each class. See [1] for methodological details.

Usage

BayesPGMM(X,Y,n_clust=2,max_cp=2,cp_prior='binomial',binom_prob=0.5, 
          alpha=1, scale_prior='uniform', iters_BurnIn=20000,
          iters_sampling=30000,Thin=15,SaveChains=FALSE)

Arguments

X	An array of dimensions N x M, where the ij'th value gives the j'th time point for subject i. The default priors assume the first time point is x=0.
Y	An array of dimension N x M, where the ij'th value gives the outcome for subject i at the j'th time point. Missing measurements should have the value NA.
n_clust	Number of possible latent classes. Default is 2. For one latent class, use BayesPGM(...).
max_cp	Maximum number of changepoints in each latent class.
cp_prior	Prior for the number of changepoints in each class, with options 'binomial' or 'uniform'. Default is 'binomial'.
binom_prob	Probability for binomial prior, if specified. Default is 0.5.
alpha	Concentration parameter for Dirichlet clustering prior, Dirichlet(alpha,alpha,...,alpha). Default is 1
scale_prior	Prior for the scale parameter for the hierarchical random effects. The default is 'uniform', a scaled uniform prior; the option 'hc' is a scaled half-cauchy prior. See [1] for more details.
iters_BurnIn	(optional) Number of Gibbs sampling iterations to run for the burn in. Default is 20000.
iters_sampling	(optional) Number of Gibbs sampling iterations to run for posterior sampling. Default is 30000.
Thin	(optional) Will save every k'th posterior sample, where k=Thin. Default is 15.
SaveChains	(optional) If TRUE, raw MCMC samples from rjags will be returned.

Details

For more information on the priors implemented in this package, see [1]. For more control over the priors and aspects of posterior computation, the source functions can be modified directly (enter BayesPGMM in the R console to view source code).

Value

An object of class S3 class PGMM, with attributes

Gelman.msrf	Gelman multivariate scale reduction factor to assess convergence, with the potential scale reduction factor (psrf) for each parameter
ClusProb	Marginal probability of each class.
Clust	Vector of length N, giving the class allocation for each subject (given by the class with highest posterior probability).
y.mean	NxM array giving the fitted value at each time point for each subject.
error.sd	The residual standard deviation.
error.sd.CI	95% CI for the residual standard deviation.
DIC	Deviance information criterion for fitted model.
Save	If SaveChains=TRUE in input, raw MCMC samples from rjags
C	List with class specific results; C[[i]] gives parameter estimates that are specific to class i.
C[[i]]$K_prob	Posterior probabilities for the number of changepoints in class i.

C[[i]]$K[[k+1]] gives parameters for the k changepoint model for class i:

C[[i]]$K[[k+1]]$b.mean	Vector giving the coefficient means [intercept,slope,change at cp1,change at cp2,...] for the k-changepoint model
C[[i]]$K[[k+1]]$b.sd	Vector giving the coefficient standard deviations for the k-changepoint model
C[[i]]$K[[k+1]]$b.mean.CI	95% CI for the coefficient means for the k-changepoint model
C[[i]]$K[[k+1]]$b.sd.CI	95% CI for the coefficient standard deviations for the k-changepoint model
C[[i]]$K[[k+1]]$cp.mean	Vector giving mean location of each changepoint in the k-changepoint model
C[[i]]$K[[k+1]]$cp.sd	Vector giving standard deviation of each changepoint in the k-changepoint model

Author(s)

Eric F. Lock

References

[1] EF Lock, N Kohli & M Bose (2018). Detecting multiple random changepoints in Bayesian piecewise growth mixture models. Psychometrika, 83 (3): 733-750.

Examples

data(SimData)  ##load simple simulated dataset
plotPGMM(X,Y) ##Plot the data
Results <- BayesPGMM(X,Y) ##Fit PGMM (can take about 5 minutes)
summary(Results) ##summarize results
plotPGMM(X,Y,Results) ##Plot results

lockEF/BayesianPGMM documentation built on Sept. 30, 2022, 6:53 a.m.