Print diagnostic plots

Description

Generates diagnostic plots for assessing the convergence and autocorrelation behavior of pan's and jomo's MCMC algorithms.

Usage

## S3 method for class 'mitml'
plot(x, print=c("beta","psi","sigma"), pos=NULL, group="all",
  trace=c("imputation","burnin","all"), thin=1, smooth=3, n.Rhat=3,
  export=c("none","png","pdf"), dev.args=list(), ...)

Arguments

Details

The plot method generates a series of plots for the parameters of the imputation model which can be used for diagnostic purposes. In addition, a short summary of the parameter chain is displayed.

Setting print to "beta", "psi" and "sigma" will plot the fixed effects, the variances and covariances of random effects, and the variances and covariances of residuals, respectively. Each plotting window contains a trace plot (upper left), an autocorrelation plot (lower left), a kernel density approximation of the posterior distribution (upper right), and a posterior summary (lower right). The summary includes the following quantities:

EAP:

Expected value a posteriori (i.e., the mean of the parameter chain)

MAP:

Mode a posteriori (i.e., the mode of the parameter chain)

SD:

Standard deviation of the parameter chain

2.5%:

The 2.5% quantile of parameter values

97.5%:

The 97.5% quantile of parameter values

Rhat:

Estimated potential scale reduction factor (hat{R})

ACF-k:

Smoothed autocorrelation at lag k, where k is the number of iterations between imputations (see summary.mitml)

The trace and smooth arguments can be used to forther influence how the trace plot is drawn, and what part of the chain should be used for it. The thin argument can be used for thinning the chain before plotting, in which case the number of data points reduces in the trace plot, and the autocorrelation is calculated up to lag k/thin, where k is the number of iterations between imputations. The n.Rhat argument controls the number of sequences that are used for calculating the potential scale reduction factor (hat{R}) in each plot (see summary.mitml). Further aguments to the graphics device are supplied using the dev.args argument.

The plot function calculates and displays diagnostic information primarily for the imputation phase (i.e., for iterations after burn-in). This is the default in the plot function and the recommended method for most users. However, note that, when overriding the default using trace="burnin", the posterior summary and the trace plots do not convey the necessary information to establish convergence. When trace="all", the full chain is displayed with emphasis on the imputation phase, and the posterior summary is calculated based only on iterations after burn-in as recommended.

Value

None (invisible NULL).

Note

The plots are presented one at a time. To proceed with the next plot, the user may left-click in the plotting window or press the "enter" key while in the R console. No plots are displayed when exporting to file.

Author(s)

Simon Grund

See Also

panImpute, jomoImpute

Examples

## Not run: 
data(studentratings)

# * Example 1: simple imputation

fml <- ReadDis + SES ~ ReadAchiev + (1|ID)
imp <- panImpute(studentratings, formula=fml, n.burn=1000, n.iter=100, m=5)

plot all parameters (default)
plot(imp)
plot(imp, print=c("beta","psi","sigma"))

plot fixed effects only
plot(imp, print="beta")

# export plots to file (using pdf device)
plot(imp, export="pdf", dev.args=list(width=9, height=4, pointsize=12))

# * Example 2: groupwise imputation

fml <- ReadDis + SES ~ ReadAchiev + (1|ID)
imp <- panImpute(studentratings, formula=fml, group=FedState, n.burn=1000,
       n.iter=100, m=5)

plot fixed effects for all groups (default for 'group')
plot(imp, print="beta", group="all")

plot fixed effects for first group only
plot(imp, print="beta", group=1)

## End(Not run)

Results