I am thinking about ways to make MI faster. One bottleneck is the EM algorithm commonly used to initialize the MCMC chain. I have a few questions:
1. Is there guidance regarding how many iterations are typically needed for convergence? I have heard that the number of iterations increases with the fraction of missing information, but I haven't seen the relationship quantified. 2. Are the EM algorithms implemented in popular MI software the fastest versions? 3. Is there any justification for the convergence criteria used by the EM algorithms implemented in popular software? Are they unnecessarily strict? Best wishes, Paul von Hippel
