SPSS Statistics

 View Only
Expand all | Collapse all

Appropriate way to use post-stratification weights when running statistical tests SPSS

  • 1.  Appropriate way to use post-stratification weights when running statistical tests SPSS

    Posted Mon May 15, 2023 06:10 PM
    Edited by Sonia Singh Tue May 16, 2023 09:28 AM

    I have used Complex Samples in SPSS (and SUDAAN in SAS, Survey in R) when working with survey data that were collected using a sampling design that was not random. For example, when an oversample was included in the data collection. Complex Samples incorporates the sample design into statistical tests, providing more accurate estimates and standard errors.

    I am now working with data that were collected from a survey panel and includes post-stratification weights to balance the sample with the population on gender, age, race/ethnicity, region, and education. Since this is not a design weight, I am wondering how to apply weights when running a statistical test. Is it sufficient to simply "weight cases" and perform standard statistical tests, or should I treat the post-stratification weights in Complex Samples the same way I would treat stratification weights if I had oversampled in the design?

    Very much appreciate any assistance.



  • 2.  RE: Appropriate way to use post-stratification weights when running statistical tests SPSS

    IBM Champion
    Posted Mon May 15, 2023 09:52 PM
    You have hit on a very difficult question, and I don't think that there is a solid consensus on what to do other than for estimating simple population averages.  The best action depends on what sort of analysis you plan to do.  You can rely on just making a sample more representative, but if you are modeling something, you would probably choose a different strategy.

    I have attached two short articles that discuss the issues, and, for modeling purposes, I tend to favor Gelman's proposal about incorporating the weight determinants into the regression equation.

    One other approach not mentioned in these pieces is using effective base weighting.  This is available in CTABLES and might be the best choice if tabulations are what you want.

    Another caution - maybe it is to some degree a comfort - is that in the modeling approach, if the model is correctly specified, the results should not vary greatly depending on the weighting scheme.  If they do vary a lot, that's a clue that the model may be misspecified.   And the first-moment results are okay in all these approaches, but the second-moment validity will vary.

    Good luck.
    --



    Attachment(s)

    pdf
    GelmanAndCarlin.pdf   232 KB 1 version
    docx
    TypesOfWeights.docx   15 KB 1 version


  • 3.  RE: Appropriate way to use post-stratification weights when running statistical tests SPSS

    IBM Champion
    Posted Mon May 15, 2023 09:56 PM
    I got a strange message when I tried to post a response, and I think maybe it did not post because of the attachments.  If you want to send me an email (jkpeck@gmall.com), I'll reply directly with my attempted post.

    --