Pat, Why not just do the bootstrap from the missing data set directly and obtain direct ML estimates (under MAR assumptions) from each of your bootstrap samples? Stock software won't do this, but it can be done pretty easily and seems to work pretty well as long as your N is decent (and assuming you accept the premises underlying the bootstrap). The only advantage that multiple imputation gives you is that MAR may be more plausible if you impute using a superset of the variables in your substantive model. However, John Graham has an in press paper on including extraneous variables in the missing data model for direct ML estimation, so I don't think that MI keeps that advantage if you use Graham's method to estimate the quantity of interest in each of your bootstrap samples. But perhaps others will see something that I'm missing. Hope this helps, Craig Enders
-----Original Message----- From: Patrick S. Malone [mailto:[email protected]] Sent: Wednesday, November 20, 2002 9:44 AM To: [email protected] Subject: IMPUTE: Bootstrapping with imputed data Has anyone looked at it? I'm imagining a situation where you need to bootstrap to get at a quantity of interest, but for whatever reason imputation is the missing data solution of choice. One could just create the m imputed data sets and draw the bootstrap samples of size n from the overall pool of m*n observations. Does this work? Meaning, have desirable properties? Thanks, Pat -- Patrick S. Malone, Ph.D., Research Scholar Duke University Center for Child and Family Policy Durham, North Carolina, USA e-mail: [email protected] http://www.duke.edu/~malone/
