MI for time-series: Amelia vs mice

[joh4nd]

Thanks for the possibility to raise questions here.

According to the AJPS article, "the best available imputation methods work poorly with the time-series cross-section data structures".

What is it that makes Amelia better than other software packages such as mice for MI on time-series data structures? Is it its explicit treatment of data structures, its algorithm, its multivariate normal distribution assumption and execution on all variables simultaneously, and/or anything else?

	Amelia	mice	implications for time-series
Explicit about data structures	yes 2010 AJPS article, Journal of Statistical Software, CRAN, documentation	no CRAN, readme, github, homepage, except time raster imputation and multi-level	self-evident
Algorithm/method	Expectation-Maximization Bootstrapping	Chained Equations and e.g. Predictive Mean Matching	?
How are variables used in one computation	assumes Multivariate Normal Distribution	computes Fully Conditional Specification	?

The available performance comparisons I could find do, except from a slide show, not show much of a difference in the results as far as I can tell:

• the review in the journal of statistical software
• election results
• "MICE PMM (Predictive Mean Matching) is worryingly overconfident".

Is this because the performances are not compared on data structures that maximizes performance differences like time-series data structures? E.g. 'the most common specification which excludes time'.