Most of you know that the parameters of the imputation model are usually
drawn from a Bayesian posterior distribution. That's necessary to ensure
that Rubin's within-between (WB) formula will consistently estimate the
standard error of a multiple imputation point estimate. Unfortunately,
getting posterior draws can substantially increase runtime and
computational complexity.

What if you don't take posterior draws? What if you just estimate the
parameters of the imputation model using maximum likelihood (or something
equally efficient). Everything will run faster, with less coding, but
Rubin's WB formula will underestimate the standard errors.

Is there another way to estimate the standard errors? In a new article
for *Statistical
Science*
<https://urldefense.com/v3/__https://www.e-publications.org/ims/submission/STS/user/submissionFile/30892?confirm=5d63f917__;!!Dq0X2DkFhyF93HkjWTBQKhk!Dd1B2VYgozF-ixE63Ez1UAHi8fHP9vW0p9FvFbd-UhSp_Cb6hYSURd7OpiWZi2eamIBTRhSwIrGvcg$
 >,
Jonathan Bartlett and I derive three standard error estimators that work
when the imputation parameters are estimated by maximum likelihood. One
standard error estimator is a WB formula that combines the within and
between variances in a different way than Rubin. One is a decomposition of
the score function inspired by Robins and Wang. And one -- probably the
most useful -- is a new application of the bootstrap to calculate variance
components due to sampling and imputation. Jonathan implemented the methods
in R, and I just finished implementing them in Stata.

Theoretically, it's the deepest statistical article I've ever written, and
I hope you learn as much from reading it as we did from writing it. It does
not appear to be paywalled. Have a look!

von Hippel, P. T., & Bartlett, J. (2020). Maximum likelihood multiple
imputation: Faster imputations and consistent standard errors without
posterior draws
<https://urldefense.com/v3/__https://www.e-publications.org/ims/submission/STS/user/submissionFile/30892?confirm=5d63f917__;!!Dq0X2DkFhyF93HkjWTBQKhk!Dd1B2VYgozF-ixE63Ez1UAHi8fHP9vW0p9FvFbd-UhSp_Cb6hYSURd7OpiWZi2eamIBTRhSwIrGvcg$
 >.
*Statistical Science*.


-- 


Best wishes,
Paul von Hippel
Associate Professor, LBJ School of Public Affairs
University of Texas, Austin
PaulvonHippel.com 
<https://urldefense.com/v3/__http://paulvonhippel.com__;!!Dq0X2DkFhyF93HkjWTBQKhk!Dd1B2VYgozF-ixE63Ez1UAHi8fHP9vW0p9FvFbd-UhSp_Cb6hYSURd7OpiWZi2eamIBTRhQcARMIrA$
 >
@PaulvonHippel 
<https://urldefense.com/v3/__https://twitter.com/paulvonhippel__;!!Dq0X2DkFhyF93HkjWTBQKhk!Dd1B2VYgozF-ixE63Ez1UAHi8fHP9vW0p9FvFbd-UhSp_Cb6hYSURd7OpiWZi2eamIBTRhSMIUHkzw$
 >

Reply via email to