home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!gatech!paladin.american.edu!auvm!BIOSTAT.MC.DUKE.EDU!LHM
- Message-ID: <9207211302.AA23969@biostat.mc.duke.edu>
- Newsgroups: bit.listserv.sas-l
- Date: Tue, 21 Jul 1992 09:02:37 EDT
- Reply-To: "Lawrence H. Muhlbaier" <lhm@BIOSTAT.MC.DUKE.EDU>
- Sender: "SAS(r) Discussion" <SAS-L@UGA.BITNET>
- From: "Lawrence H. Muhlbaier" <lhm@BIOSTAT.MC.DUKE.EDU>
- Subject: Re: missing data analysis
- Comments: To: sas-l%vtvm2.BITNET@ncsuvm.ncsu.edu
- Lines: 50
-
- From an unidentified requestor on the SAS-L list:
-
- > From DEMPO@mc.duke.edu Mon Jul 20 15:47:24 1992
- >
- > has anybody out there had any experience doing analysis on missing data?
- >
- > our dataset will have the following characteristics:
- >
- > 1) multiple observations per subject
- > 2) subjects nested within a large number of groups
- > 3) some variables measured on the subjects some measured on the groups
- > (a group would be something like a census tract)
- >
- > data are missing when:
- >
- > 1) a subject lacks one or more variable measurements
- > 2) a group is lacking the required number of subjects
- > 3) a group is lacking one or more variable measurements
- >
- > has anyone used or developed routines that will estimate the missing
- > values (perhaps EM algorithm). BMDP has some routines for missing data
- > but don't seem to handle all our types of missingness.
- >
-
- I am unaware of actual routines that can directly handle the questions
- that you are raising. There are, however, some techniques that may have
- some promise. I am wrestling with some significant missing data problems
- myself, so this advice reflects my current study, rather than definitive
- answers.
-
- I would first look at Don Rubin's book on Multiple Imputation. It has
- some ideas that should work for variables that are missing (in either
- cases or groups).
-
- Imputing enitre observations is more difficult. I don't have any ideas
- there except to look at this idea of 'hierarchical modeling' that is
- an empirical Bayes technique.
-
- One final note: I have had some experience with the EM algorithm
- that was not entirely positive -- it would converge to local maxima,
- and the only way that I could tell it was 'local' was by using
- different starting points and converging to other answers. It
- would not provide any other indication that there were convergence
- problems.
-
- Lawrence H. ('Doc') Muhlbaier muhlb001@mc.duke.edu
- Assistant Research Professor muhlb001@dukemc
- Duke University Medical Center 919-286-3322 (office)
- DUMC 3865 919-286-9534 (home)
- Durham, NC 27710-7510 919-286-2947 (FAX)
-