4

Does anybody have any good research papers for me to read on whether using bootstrapping and maximum likelihood (ML) estimation together is a good idea, particularly when ML is being used with a relatively small proportion of missing data.

I wanted to use evidence to suggest it is a good idea or a bad idea (hopefully a good idea given that this was my plan of action for a Confirmatory Factor Analysis and a separate path analysis).

I’ve tried searching on Google Scholar, etc but keep coming up with papers that have combined the two but that don’t comment or test the appropriateness of the two. I was hoping to compare this with bootstrapping and multiple imputation which doesn’t seem as common or as well understood.

Dave
  • 62,186
SimonsSchus
  • 779
  • 3
    ML operates under some distributional assumption, which will generally perform well when that assumption holds (or very close to it). Bootstrapping often tends to have reasonable properties (at least in large samples) when that's not the case, so that you get some robustness to assumptions when they're a bit further from being true -- when the estimator may be convenient, but is no longer ML, or even nearly ML. – Glen_b May 17 '15 at 06:28
  • 1
    Bootstrap is non-parametric maximum likelihood, see https://stats.stackexchange.com/questions/91953/what-inferential-method-produces-the-empirical-cdf – kjetil b halvorsen Jun 20 '23 at 19:44
  • 1
    you may be interested in the search term "Parametric Bootstrap". – John Madden Jun 20 '23 at 20:16

0 Answers0