Survey imputation

[scott.raynaud]

I'm looking at the article, "The Importance of Modeling the Sampling Design in Multiple Imputation for Missing Data," (Survey Methodology, December 2006 Vol. 32, No. 2, pp. 143149 Statistics Canada, Catalogue No. 12001) here: http://www.publications.gc.ca/Collection-R/Statcan/12-001-XIE/12-001-XIE2006002.pdf.  In the example on p. 147 the stratum and cluster variables are dummied up.  I'm wondering why this is necessary.  Doesn't SAS automatically create the dummies when everything is handed off from IVEware?  Is this not handled in a similar manner internally in SRCware?  Why even go to the trouble of creating dummies when the software does this for you?

[Trivellore Raghunathan]

This was necessary because we used R to tun the simulation. In IVEWARE, however, declaring the stratum and cluster variables as categorical will automatically create dummy variables. Also, using interaction feature in IVEWARE, you can create dummy variables for stratum-cluster combination.

Raghu

On Fri, Jan 29, 2016 at 10:33 AM, scott.raynaud <scott....@cookchildrens.org> wrote:

I'm looking at the article, "The Importance of Modeling the Sampling Design in Multiple Imputation for Missing Data," (Survey Methodology, December 2006 Vol. 32, No. 2, pp. 143149 Statistics Canada, Catalogue No. 12001) here: http://www.publications.gc.ca/Collection-R/Statcan/12-001-XIE/12-001-XIE2006002.pdf.  In the example on p. 147 the stratum and cluster variables are dummied up.  I'm wondering why this is necessary.  Doesn't SAS automatically create the dummies when everything is handed off from IVEware?  Is this not handled in a similar manner internally in SRCware?  Why even go to the trouble of creating dummies when the software does this for you?

--
You received this message because you are subscribed to the Google Groups "IVEware" group.
To unsubscribe from this group and stop receiving emails from it, send an email to iveware+u...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

Trivellore Raghunathan (Raghu)

Director, Survey Research Center

Research Professor, Institute for Social Research

Professor, Department of Biostatistics

University of Michigan

Phone: (734)-764-8365

Fax: (734)-763-9831

www.isr.src.umich.edu

www-personal.umich.edu/~teraghu

"A good life is filled with selfless actions full of compassion knowing well that we are all one"

[scott.raynaud]

Thanks for your response.  That makes much more sense.  Here's a follow-up.  If my model of interest is a random effects model, and I want to generate missing data for NHIS survey data where the data are stratified and then clusters sampled in each strata, how can that be done?  There's a hint in the first paragraph on p. 673 here: http://www.denveremresearch.org/phocadownload/outcomes/articles/advanced_statistics_missing_data_in_clinical_researchpart_2_multiple_imputation.pdf.  "...separate MI models are generated for each strata within the sample, with clustering accounted for within each stratified MI analysis."  Does that mean PSU must be treated as a random effect in each stratum during the imputation or is everything a fixed effect?

[Trivellore Raghunathan]

If the sample sizes within clusters are large then doing separate imputation makes sense. We have a paper just published in the Journal of Survey Statistics and Methodology

 http://jssam.oxfordjournals.org/cgi/reprint/smv031?ijkey=Mu3rWG3DVnbJs8X&keytype=ref

that uses the finite population Bayesian Bootstrap method to "uncomplex" the complex design and then imputes the missing values. We are working on getting a SAS MACRO wrapper to IVEware to implement this method. Building such wrappers are going to become easier with the new version of IVEWARE (0.3) soon to be released that will work with SAS, STATA, SPSS and R.

Raghu

Raghu  

www.src.isr.umich.edu