Announcement

Collapse
No announcement yet.
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • mi impute chained with survey data

    Dear Statalist

    i am working with survey data and i would like to impute some independent variables using mi impute chained. I have read that you cannot use the svy commands with mi impute chained. is this correct? and if so is there a way to account for strata and psu other that using them as categorical variables in the imputation models?

    thank you for considering my request

  • #2
    Where did you read that svy commands are incompatible with mi chained imputation? I don't think that's correct. Use mi svyset just as you would do for any other survey analysis of MI data. For a worked example, see: http://www.ats.ucla.edu/stat/stata/faq/mi_svy_logit.htm Adding the survey design to the imputation model seems like a good idea in any case (Reiter et al (2006).

    Reference:

    Reiter, Jerome P, Trivellore E Raghunathan, and Satkartar K Kinney. 2006. The importance of modeling the sampling design in multiple imputation for missing data. Survey Methodology 32, no. 2: 143. http://publications.gc.ca/collections/Collection-R/Statcan/12-001-XIE/12-001-XIE2006002.pdf#page=29.



    Last edited by Steve Samuels; 28 Jul 2014, 20:57.
    Steve Samuels
    Statistical Consulting
    [email protected]

    Stata 14.2

    Comment


    • #3
      I read that limitation as well (http://www.ssc.wisc.edu/sscc/pubs/stata_mi_models.htm), but have not been able to confirm in the Stata manual.

      Comment


      • #4
        Thanks for the link, LML. It does indeed state that:

        However, svy: cannot be used with mi impute chained.
        This statement is manifestly false, disproved by the UCLA example of svy estimation following mi impute chained.

        If you want to be a regular participant in Statalist, I suggest that you change your user-name to your full real name, as requested in the registration page and FAQ (you can do it with the "Contact Us" button at the bottom of the page). This is long-standing Statalist etiquette and has made Statalist a very satisfying experience over the years.
        Last edited by Steve Samuels; 30 Jul 2014, 18:45.
        Steve Samuels
        Statistical Consulting
        [email protected]

        Stata 14.2

        Comment


        • #5
          I wonder if the Wisconsin page mixed up ice and mi -- or if the statement was true at one point but mi got changed somewhere along the way. Most of the Wisconsin Stata pages are great. I've sent their help desk a message about this.
          -------------------------------------------
          Richard Williams, Notre Dame Dept of Sociology
          Stata Version: 17.0 MP (2 processor)

          EMAIL: [email protected]
          WWW: https://www3.nd.edu/~rwilliam

          Comment


          • #6
            I agree that the page is very out-of-date, despite a revision in January, 2013.

            The page's statement following the one I quoted is:
            You can apply weights (e.g. [pweight=weight]) but not correct for other elements of survey structure like strata or PSU.
            This hasn't been true at least since Stata 11, which came out in 2009.
            Steve Samuels
            Statistical Consulting
            [email protected]

            Stata 14.2

            Comment


            • #7
              Richard,
              Thanks for letting us (the sscc at UW) know about the issue with the webpage. We will address it. Thanks again.
              Mark

              Comment


              • #8
                I think there's some confusion here between the actual command "mi impute chained" and working with data that has been imputed using mi impute chained. You can use svy with many mi commands, later articles in our MI series discuss doing so. However, you cannot use svy with mi impute chained itself. Here's a do file to demonstrate.

                //
                // Example of how mi and svy do and do not get along
                //

                // Load a simple data set I created which already has the preliminaries taken care of:

                use http://ssc.wisc.edu/~rdimond/misvy.dta

                // Note that it is has been both mi set and svyset, but nothing has been imputed:

                mi set

                mi svyset

                // I will impute using mi impute chained,but attempts to prefix it with svy: fail:

                capture noisily svy: mi impute chained (regress) x1 x2 y, add(10)

                // On the other hand, I can specify pweights:

                mi impute chained (regress) x1 x2 y [pweight=wt], add(10)

                // But now that I've imputed, I can use svy along with mi estimate:

                mi estimate: svy: reg y x1 x2



                I'll add something to the article to make it clearer. Suggestions are welcome.

                Russell Dimond
                Statistical Computing Specialist
                Social Science Computing Cooperative
                University of Wisconsin-Madison

                Comment


                • #9
                  To answer Marilena's question, I think the main alternative to including strata and PSU as categorical variables in the imputation model (and perhaps interacting them) is to impute each combination of them separately using by(). That would depend on having sufficient observations in each group, of course.

                  Russell Dimond
                  Statistical Computing Specialist
                  Social Science Computing Cooperative
                  University of Wisconsin Madison

                  Comment


                  • #10
                    Thanks Russell & Mark. Looking at the UCLA example, I see that it indeed does not combine svy and mi impute chained. If anything, the UCLA example may be flawed because it does not include pweights on the mi impute chained command. I should have known better than to doubt my alma mater.

                    I assume that the issue isn't limited to mi impute chained -- it would apply to any mi impute command. Perhaps the wording in your FAQ could be broadened a bit. Maybe include a quick example at that point.
                    -------------------------------------------
                    Richard Williams, Notre Dame Dept of Sociology
                    Stata Version: 17.0 MP (2 processor)

                    EMAIL: [email protected]
                    WWW: https://www3.nd.edu/~rwilliam

                    Comment


                    • #11
                      Russell, the svy: mi impute chained statement fails because the svy: prefix is valid only if precedes a survey-aware estimation command.
                      Last edited by Steve Samuels; 31 Jul 2014, 09:53.
                      Steve Samuels
                      Statistical Consulting
                      [email protected]

                      Stata 14.2

                      Comment


                      • #12
                        Originally posted by Steve Samuels View Post
                        Russell, the svy: mi impute chained statement fails because the svy: prefix is valid only if precedes a survey-aware estimation command.
                        Correct. And that is why the Wisconsin FAQ is correct. You can't combine svy and mi impute but you can combine mi impute with pweights. And then use the imputed vars with an svy/mi combo.
                        -------------------------------------------
                        Richard Williams, Notre Dame Dept of Sociology
                        Stata Version: 17.0 MP (2 processor)

                        EMAIL: [email protected]
                        WWW: https://www3.nd.edu/~rwilliam

                        Comment


                        • #13
                          I think it's a bad idea to weight imputation models. In my view, the goal is to find a model that predicts for members of the sample, not for the population. In the same vein, Little and Vartivarian (2003) argue against weighting non-response models. From their abstract:

                          We show by simulations that weighting the response rates by the sampling weights to adjust for design variables is either incorrect or unnecessary.

                          Reference:
                          Little, RJ, and S Vartivarian. 2003. On weighting the rates in non-response weights. Stat Med 22, no. 9: 1589-1599.
                          Last edited by Steve Samuels; 31 Jul 2014, 10:31. Reason: added quotation
                          Steve Samuels
                          Statistical Consulting
                          [email protected]

                          Stata 14.2

                          Comment


                          • #14
                            Steve,

                            mi impute chained is a hybrid of sorts: it runs estimation commands (most of which are survey-aware), but is not one itself. It would be both sensible and useful if it could take the svy prefix, though I'm not sure anyone has worked out the statistics involved. At any rate I think we agree that svy cannot be used with the mi impute chained command, which is all our article says.

                            Russell Dimond
                            Statistical Computing Specialist
                            Social Science Computing Cooperative
                            University of Wisconsin-Madison

                            Comment


                            • #15
                              Steve,

                              I see our responses crossed. I don't have a strong opinion on whether weighting is actually a good idea--make that "arguably useful" in the above post.

                              Russell

                              Comment

                              Working...
                              X