As a practitioner I am also a bit skeptical about the usefulness of MI for data analysis . Nevertheless I came across this paper on Firms' energy efficiency investments where MI seems to provide a useful statistical tool:

Investment in Energy Efficiency: Do the Characteristics of Firms Matter?

Stephen J. DeCanio and William E. Watkins. http://www.jstor.org/stable/2646732

I don't have the definitive answer by far, but can point you to a conference poster/paper I did with a couple colleagues a few years back (was while in grad school). We found that it often doesn't make much of a difference if by difference you mean "would I have made a different substantive interpretation about a relationship, model, or coefficients". This wasn't a simulation, so we didn't test many possible data scenarios, and it wasn't even that complex of an analysis example. Here is the link. https://www.researchgate.net/publication/260591261_Imputation_for_Missing_Physiological_and_Health_Measurement_Data_Tests_and_Applications?ev=prf_pub

I've been really curious about this, too...where does the statistically-arcane meet the useful for applied researchers. I'm looking forward to others' responses. Thanks for asking.

Conference Paper Imputation for Missing Physiological and Health Measurement ...

As a practitioner I am also a bit skeptical about the usefulness of MI for data analysis . Nevertheless I came across this paper on Firms' energy efficiency investments where MI seems to provide a useful statistical tool:

Investment in Energy Efficiency: Do the Characteristics of Firms Matter?

Stephen J. DeCanio and William E. Watkins. http://www.jstor.org/stable/2646732

I have an example from the educational arena. It is with data from the Trends in Mathematics and Science Study that is organized by the IEA (www.iea.nl). The specific example is about plausible values, which can be seen as an application of multiple imputation. The concrete example is the slide 43, but detailed information about the concepts involved is in the rest of the presentation.

I hope it helps!

Cheers,

MI can indeed be very useful, not only theoriticaly but also in practice. I used it in the malaria in pregnancy context, where the objective was to analyse the relation between malaria infection in early pregnancy and newborn's birth weight. No association was observed when we analysed only the complete cases (without missing data) and we were able to show a (limit) significant association when using the MICE method (Multiple Imputation by Chained Equation). Further details can be found at http://www.ajtmh.org/content/76/5/849.long

Hope it can help

Thanks for these - are there any more out there?

You also be interested in

http://missingdata.lshtm.ac.uk/

which is an extensive and very useful resource.

Kelvyn

kelvyn  - I routinely use mi these days - this was an instance where - from memory - results differed. Probably not in paper but could search  out non mi results if crucial 

http://llcsjournal.org/index.php/llcs/article/view/138

useful thread!

Thanks - good to hear from you -. I have been encouraging PHd students to use MI but finding it not making much difference - I was looking for a straightforward convimcing example where it did.

Kelvyn 

Ranjit Lall re-analysed the last 5 years of studies from 2 journals in Political Science. 

"in almost half of the studies, key results “disappear” (by conventional statistical standards) when reanalyzed"

http://scholar.harvard.edu/files/ranjitlall/files/pa_website_version_1.pdf

Lee

Thanks for the Lall paper - I enjoyed reading it - and it is convincing as  the missingness relates to countries at various stages of development.