A few years ago on this list, Adriaan Hoogendoorn shared the results of a simulation showing that multiple imputation could be biased in small samples <https://urldefense.proofpoint.com/v2/url?u=http-3A__www.mail-2Darchive.com_impute-40listserv.it.northwestern.edu_msg00467.html&d=CwIBaQ&c=yHlS04HhBraes5BQ9ueu5zKhE7rtNXt_d012z2PA6ws&r=N9mDDDuK1isnKK-Q36bwNuZl066Rn4cNQtKxtVKMnWBnZ5yXlXHty3gF6wWXsdE6&m=8hTrgclP6vEYux_BkpmJcK_CGuNLlZLAbr_8HEkT5Y4&s=h44UFzswQAHPcs2YnfdWkroZwUMAqAnonZ-HhfNGwvM&e= >. An animated discussion followed. I wanted to make sure that Adriaan and others saw my recent article explaining the reason for the bias. It turns out that the situation simulated by Adriaan -- both his choice of prior and the pattern of missing data -- was just about perfect for illustrating the bias. The bias is maximal in this situation. The article discusses how to reduce the bias by changing the prior, and shows that the bias is smaller when other estimators -- such as full information maximum likelihood -- are chosen.
The published and paywalled article is here <https://urldefense.proofpoint.com/v2/url?u=http-3A__www.tandfonline.com_doi_abs_10.1080_10705511.2015.1047931&d=CwIBaQ&c=yHlS04HhBraes5BQ9ueu5zKhE7rtNXt_d012z2PA6ws&r=N9mDDDuK1isnKK-Q36bwNuZl066Rn4cNQtKxtVKMnWBnZ5yXlXHty3gF6wWXsdE6&m=8hTrgclP6vEYux_BkpmJcK_CGuNLlZLAbr_8HEkT5Y4&s=EqhuNTBRFwEj4EMjtHOhEif9bsiHX6bQ3IdRivHpnGw&e= >; and there's a free version here <https://urldefense.proofpoint.com/v2/url?u=https-3A__arxiv.org_ftp_arxiv_papers_1307_1307.5875.pdf&d=CwIBaQ&c=yHlS04HhBraes5BQ9ueu5zKhE7rtNXt_d012z2PA6ws&r=N9mDDDuK1isnKK-Q36bwNuZl066Rn4cNQtKxtVKMnWBnZ5yXlXHty3gF6wWXsdE6&m=8hTrgclP6vEYux_BkpmJcK_CGuNLlZLAbr_8HEkT5Y4&s=65xY-3N4vHNphktasM3Jvc7_f5ryqCaMYa_sNi3kHfQ&e= >. Best wishes, Paul von Hippel
