Hi Seppo,
I read, but rarely add to this list, but thought I'd add a short note to this one. On Wed, 1 Jun 2005, Laaksonen Seppo wrote: > The number of imputations is not most important for me. I am always trying to > minimise the bias in point estimates. A fine objective as long as you are aware of the associated interval estimate too. > If I believe that this is successful > reasonably, I will continue to estimate the best possible variances (as > unbiased as possible, a small overestimate is not not so bad than an > underestimate, since it is obvious that our estimate is not completely > unbiased). Here, I assume that you mean you are concerned with the correct coverage of interval estimates, in the Neyman (1935) sense, and therefore are far more worried about under-coverage than over-coverage; e.g., a nominal 95% interval estimate with 98% coverage is of less concern to you than a 95% interval with 93% coverage. Because of your stated real life concern with skew distns, etc., you cannot be solely focused on unbiased estimation of sampling variances, which is only an intermediate criterion used as a guide for accurate coverage of intervals -- don't forget Neyman!!! > MI does not automatically give any guarantee to the unbiasedness, > its 'single basis' should be unbiased. I'm puzzled here. Everything I've seen supports the claim that in practice, MI (even based on only a few imputations) does remarkably well, and generally better than all alternatives, at the two criteria you care most about: unbiased point estimation, and accurate coverage of interval estimates. Even in highly artificial examples designed to "break" MI, as in the ISI Review article by Nielsen, MI comes out not too bad with respect to these two criteria. I appears that Paul's experience is similar to mine. > The number of MI-based imputations > depends so much on how complex is the variability of the statistic being > estimated. In real life the distributions are often complex (skewed, ouliers, > etc.). It follows that the number of imputations needs to be higher, 5 or 7 > give rarely a satisfactory result. This comment I find puzzling too. My experience is the opposite. Can you provide real, or realistic, examples where MI based on 5 or 7 doesn't work (and a real competitor does)? It could be that you do have different types of examples in your experience than I have in mine, or it could be that you are not doing MI correctly. When having a "competition" between methods, it is usually wise to make sure the methods are being applied as well as possible, and then to see what happens when they are applied by non-experts. If one method wins when done by an expert, but doesn't when done by a non-expert, it implies that the method has superior potential but may not work without extra training -- which is a legitimate criticism of the method, at least in current practice if the extra training is demanding. I look forward to seeing your examples. Best wishes, Don > Seppo > > > > Paul von Hippel (28.5.2005 1:57): > >I'm looking for work that relates the fraction of missing information > >(gamma) to other properties of the data -- e.g., the correlation matrix and > >the fraction of values that are missing. > > > >Any references most appreciated. > > > >Thanks! > >Paul > > > >Paul von Hippel > >Department of Sociology / Initiative in Population Research > >Ohio State University > > > > > >_______________________________________________ > >Impute mailing list > >[email protected] > >http://lists.utsouthwestern.edu/mailman/listinfo/impute > > > > _______________________________________________ > Impute mailing list > [email protected] > http://lists.utsouthwestern.edu/mailman/listinfo/impute > -- Donald B. Rubin John L. Loeb Professor of Statistics Department of Statistics 1 Oxford Street Harvard University Cambridge MA 02138 Tel: 617-495-5498 Fax: 617-496-8057
