SPSS Statistics

 View Only
  • 1.  Propensity Score Matching

    Posted Tue July 05, 2022 01:06 PM
    Hi, does anybody knows how to solve a problem with propensity score matching? 
    It estimates the PScore, then it starts FUZZY and gives the following error message.
    It seems that Phyton3 and the Fuzzy extension are installed. I do not understand what kind of error stops the procedure.
    Marco

    xxxx Messages <built-in function GetCaseValue> returned a result with an error set
    xxxx Delete Variables Text: M0.9560342718892494 Command: DELETE VARIABLES
    xxxx An undefined variable name, or a scratch or system variable was specified in a variable list which accepts only standard variables. Check spelling and verify the existence of this variable.
    xxxx Execution of this command stops.
    xxxx Dataset Activate Unknown dataset D0.9478274870593494.
    xxxx ADD FILES The filename is not valid.
    xxxx ADD FILES The working file has been restored, and subsequent commands may access the working file.
    xxxx Dataset Close Unrecognized dataset name D0.9478274870593494.

    ------------------------------
    Marco Spampinato
    ------------------------------


  • 2.  RE: Propensity Score Matching

    Posted Wed July 06, 2022 02:03 PM
    When you run this syntax, do you get the logistic regression output?  What does it show?  If that fails due to data conditions, the psm process will fail.

    ------------------------------
    Jon Peck
    ------------------------------



  • 3.  RE: Propensity Score Matching

    Posted Thu July 07, 2022 04:42 AM
      |   view attached

    Dear John,

    I also have an issue with a propensity score matching (PSA) for clinical observational studies in medicine. There is supposedly an extension for SPSS v28 with R but the last version of R (4.2.0) that I used does not work with SPSS v28.

    In SPSS, there is another possibility for PSA in Data menu, but it gives very different results than the procedure done in R when using R matchIt. A validated module in clinical practice

    In my case, the Group indicator for PSA is GROUP, and predictors are ASA, AGE, Diabetes, CKD, LVEF40, WIfI, and GLASS. All predictors are categorical variables (binary).

    Could you run PSA with SPSS on this file? I will be interested to see the results of your analysis and compare it with my results.

    One problem with PSA on SPSS is the "Match Tolerance" that has little sense when analysing categorical data (binary)

    I send you the SPSS file.

    Thank you for your help.

    JBR



    ------------------------------
    Jean-Baptiste Ricco, MD, PhD
    ------------------------------

    Attachment(s)

    sav
    CLTIJuly4.sav   59 KB 1 version


  • 4.  RE: Propensity Score Matching

    Posted Thu July 07, 2022 10:10 PM
    I will dig into this further, but I was able to run this job.  I did get a warning message that I need to investigate.

    This is what I got.
    image.png

    You didn't specify the match tolerance you are using, so I used .1.

    I see that all your predictors are categorical but binary, so the logistic regression is okay.  The match tolerance setting does make sense here, since it applies to the predicted probabilities.  They will, of course, be grouped according to the variable patterns, but that is not unreasonable.

    The R matchit procedure, which is not from IBM, is no longer available.  The underlying R packages that it requires were removed from CRAN.  I don't know why that was done.  It is not something that IBM can control.

    The way that the procedure used in matching works is doubtless different from what matchit did, but the test is whether you get matches that are similar enough for your analysis.  Look at the distributions of the variables by group.  They should be very similar, although since the matches are based on the p value, there is no specific bound on any of the predictors.

    An alternative would be to use the Case Control Matching procedure, which is on that same menu.  Instead of basing the matches on similarity of the probabilities, you specify bounds on the individual variables (FUZZ).  Since yours are all binary, that would mean exact matches on those characteristics, but there might be fewer matches.  You might have to drop some of the match controls to get enough matched cases.


    --





  • 5.  RE: Propensity Score Matching

    Posted Wed July 06, 2022 06:27 PM
    Can you send me the data?

    ------------------------------
    Jon Peck
    ------------------------------



  • 6.  RE: Propensity Score Matching

    Posted Wed July 06, 2022 06:46 PM
    If you can't send the data, can you send me an spv file with the full output from the command?  Also, can you send a histogram of the propensities from the logistic regression? (jkpeck@gmail.com)

    (Something is messing up this forum, so I can't reply in the usual way.)

    ------------------------------
    Jon Peck
    ------------------------------