Thanks! There are really two separate questions: 1. How many imputations *should *analysts use? I actually have a new working paper <https://urldefense.proofpoint.com/v2/url?u=http-3A__arxiv.org_abs_1608.05406&d=CwIFaQ&c=yHlS04HhBraes5BQ9ueu5zKhE7rtNXt_d012z2PA6ws&r=N9mDDDuK1isnKK-Q36bwNuZl066Rn4cNQtKxtVKMnWBnZ5yXlXHty3gF6wWXsdE6&m=1Yr68G57ywVRTg6HN98m6mXb2dPdE_cjn5V1vy4Lowo&s=VMxk3JmVaz0yH7ck3nMCvpVHi_cR4xb1sGyzkwDIKfY&e= > on this subject. 2. How many imputations *do* analysts use in practice? This is the question that I meant to ask.
It sounds like we know less about (2) than about (1). If nobody's looked at (2), I wonder what the most efficient way would be to get an answer. Best wishes, Paul von Hippel LBJ School of Public Affairs University of Texas, Austin On Mon, Sep 12, 2016 at 7:21 AM, Frank Harrell <[email protected]> wrote: > I think that Royston and Wood is the best advice we have at present. > > What is really needed is a rule of thumb recognizing that having lots of > complete cases can help to insulate against missings somewhat, as can the > total sample size. > > ------------------------------ > Frank E Harrell Jr Professor and Chairman School of Medicine > > Department of *Biostatistics* *Vanderbilt University* > > On Mon, Sep 12, 2016 at 2:50 AM, Hoogendoorn, Adriaan < > [email protected]> wrote: > >> Originally, small numbers of imputations (3 or 5) were suggested, but >> currently there seems to be preference for larger numbers. >> >> Many recent studies that do report numbers of imputation used between 20 >> and 100 data sets (subjective view), but I’ve also seen as many as a >> thousand. >> >> I guess that it depends on the amount of Monte Carlo error (the loss of >> power to for testing an association) you are willing to accept. >> >> White, Royston and Wood (2010) suggest in their Statistics in Medicine >> journal article as a rule of thumb to use “at least the percentage of >> incomplete cases” (100 times the fraction of missing information), but they >> also state that you might need more in specific settings. >> >> >> >> >> Adriaan W. Hoogendoorn, PhD >> >> Senior Researcher – Statistician >> >> M: Room MB.02, A.J. Ernststraat 1187, 1081 HL Amsterdam, The Netherlands >> T: + 31(0)20 788 4649 >> >> E: [email protected] >> >> Office hours: Mo-Tu-We-Th >> >> >> >> >> >> GGZ inGeest, Onderzoek en innovatie >> Locatie A.J. Ernststraat, Amsterdam >> www.ggzingeest.nl >> <https://urldefense.proofpoint.com/v2/url?u=http-3A__www.ggzingeest.nl_&d=CwMGaQ&c=yHlS04HhBraes5BQ9ueu5zKhE7rtNXt_d012z2PA6ws&r=N9mDDDuK1isnKK-Q36bwNuZl066Rn4cNQtKxtVKMnWBnZ5yXlXHty3gF6wWXsdE6&m=uar2zvBt-CyCEgwFinsosPoZgBH_o6rJZAvxvVrWpeY&s=rOarCICWmLfxjAFAXCAKhMwIX7Cj-WlWK8neT8IgvxM&e=> >> >> >> * GGZ inGeest, samen op eigen wijze* >> >> >> >> >> >> >> >> >> >> *Van:* Impute -- Imputations in Data Analysis [mailto: >> [email protected]] *Namens *DAVID R JOHNSON >> *Verzonden:* maandag 12 september 2016 0:13 >> *Aan:* [email protected] >> *Onderwerp:* Re: Typical number of imputations >> >> >> >> My experience is that authors who use MI seldom report the number of >> imputations they used. >> >> >> >> ------------------------------------------------------------ >> ------------------ >> >> David R. Johnson >> Professor of Sociology, Human Development and Family Studies, and >> Demography >> Department of Sociology >> 413 Oswald Tower >> The Pennsylvania State University >> University Park, PA 16802 >> 814-865-9564 >> [email protected] >> ------------------------------------------------------------ >> ------------------- >> >> >> ------------------------------ >> >> *From: *"Allison, Paul D" <[email protected]> >> *To: *"Impute" <[email protected]> >> *Sent: *Saturday, September 10, 2016 7:58:55 AM >> *Subject: *Re: Typical number of imputations >> >> >> >> Good question, but I'm not aware of any studies. >> >> >> >> >> >> Paul D. Allison, Professor >> >> Department of Sociology >> >> University of Pennsylvania >> >> 581 McNeil Building >> >> 3718 Locust Walk >> >> Philadelphia, PA 19104-6299 >> >> 610-715-5702 >> >> 419-818-1220 (fax) >> >> www.pauldallison.com >> <https://urldefense.proofpoint.com/v2/url?u=http-3A__www.pauldallison.com_&d=CwMFAw&c=yHlS04HhBraes5BQ9ueu5zKhE7rtNXt_d012z2PA6ws&r=N9mDDDuK1isnKK-Q36bwNuZl066Rn4cNQtKxtVKMnWBnZ5yXlXHty3gF6wWXsdE6&m=WZhkIrvV_nPE4MPWJCbcERwh_uwCepoMW9A1CIlGono&s=jhCMas8sQAOyoeHxUbvqK4sCPETcmNiEkxGRSgzeBj8&e=> >> >> >> ------------------------------ >> >> *From:* Impute -- Imputations in Data Analysis < >> [email protected]> on behalf of Paul von Hippel < >> [email protected]> >> *Sent:* Friday, September 9, 2016 11:05 PM >> *To:* [email protected] >> *Subject:* Typical number of imputations >> >> >> >> Are there studies documenting how many imputations analysts typically use >> in MI? I know the recommendations, but I'm interested in what users are >> actually doing -- and whether users are using more imputations now than >> previously. >> >> >> Best wishes, >> Paul von Hippel >> LBJ School of Public Affairs >> Sid Richardson Hall 3.251 >> University of Texas, Austin >> 2315 Red River, Box Y >> Austin, TX 78712 >> >> >> >> ------------------------------ >> Dit e-mailbericht is uitsluitend bestemd voor de geadresseerde. Als dit >> bericht niet voor u bestemd is, wordt u verzocht dit aan de afzender te >> melden en het bericht te vernietigen. Het is niet toegestaan de inhoud van >> dit bericht verder te verspreiden of te gebruiken. Voor meer informatie >> over GGZ inGeest: www.ggzingeest.nl >> <https://urldefense.proofpoint.com/v2/url?u=http-3A__www.ggzingeest.nl&d=CwMFaQ&c=yHlS04HhBraes5BQ9ueu5zKhE7rtNXt_d012z2PA6ws&r=N9mDDDuK1isnKK-Q36bwNuZl066Rn4cNQtKxtVKMnWBnZ5yXlXHty3gF6wWXsdE6&m=R7jmu7lqPjB302OBAh_CkE93a76UNyZ9N28tKtxjyRg&s=49GTxMj99cJfVudpYucVtQ-F31qU4WHDUDQBdZdkAT8&e=>. >> Denk aan het milieu voordat u deze e-mail print. >> > >
