A visual diagnostic of EM convergence from multiple overdispersed
starting values for an output from `amelia`

.

`disperse(output, m = 5, dims = 1, p2s = 0, frontend = FALSE, …)`

output

output from the function `amelia`

.

m

the number of EM chains to run from overdispersed starting values.

dims

the number of principle components of the parameters to display and assess convergence on (up to 2).

p2s

an integer that controls printing to screen. 0 (default) indicates no printing, 1 indicates normal screen output and 2 indicates diagnostic output.

frontend

a logical value used internally for the Amelia GUI.

…

further graphical parameters for the plot.

This function tracks the convergence of `m`

EM chains which start
from various overdispersed starting values. This plot should give some
indication of the sensitivity of the EM algorithm to the choice of
starting values in the imputation model in `output`

. If all of
the lines converge to the same point, then we can be confident that
starting values are not affecting the EM algorithm.

As the parameter space of the imputation model is of a
high-dimension, this plot tracks how the first (and second if
`dims`

is 2) principle component(s) change over the iterations of
the EM algorithm. Thus, the plot is a lower dimensional summary of the
convergence and is subject to all the drawbacks inherent in said
summaries.

For `dims==1`

, the function plots a horizontal line at the
position where the first EM chain converges. Thus, we are checking
that the other chains converge close to that horizontal line. For
`dims==2`

, the function draws a convex hull around the point of
convergence for the first EM chain. The hull is scaled to be within
the tolerance of the EM algorithm. Thus, we should check that the
other chains end up in this hull.

Other imputation diagnostics are
`compare.density`

, `disperse`

, and
`tscsPlot`

.