SPSS Statistics

 View Only
  • 1.  Generalized Estimating Equations (GEE) Problem With Nested Factors

    Posted Wed December 08, 2021 04:41 PM
      |   view attached
    Hello! I am trying to run an analysis with Generalized Estimating Equations (using SPSS version 28.0) to predict idea ratings of individuals, which are nested within teams. When I ran the analysis with another individual who is more familiar with GEE, everything worked until we added the nested term: TeamNumber(Participant). We are not sure why this is happening, other than the team number and participant numbers being repeated multiple times because we have multiple idea ratings per individual (our study looks at predicting idea RATINGS binned in 3 different categories, where participants could generate as many ideas as they'd like in a 20-minute period...we already rated the ideas and validated the ratings, but the DV is ordinal). We are also using one IV (continuous) and one moderator variable (which is also continuous, which we added as a covariate in the model, so it is more or less an IV). There are also three more control variables (design task, instructor, semester, where semester could also be part of the nested term, but I want to figure out what is the problem with adding a nested factor). This is the error we got:
    "An unexpected error has occurred during estimation that suggests the specified model may be inappropriate for the data. You may be able to diagnose the problem by: checking for (quasi) complete separation in the data, checking for multicollinearity in the predictors, and running the model without a REPEATED subcommand (if one was specified) to determine whether the simpler model runs without warnings. No output will be displayed.
    Execution of this command stops."

    I have shared the dataset with someone who has time to look at it, as I am drawing a blank as to what is going on. I am running this analysis for my dissertation, so any help would be greatly appreciated!


    ------------------------------
    Courtney Cole
    ------------------------------

    Attachment(s)



  • 2.  RE: Generalized Estimating Equations (GEE) Problem With Nested Factors

    Posted Wed December 08, 2021 04:46 PM
    Hi. Could you please additionally share your syntax? Thanks. I can pass it on to a couple of statisticians who may have time to look at it and see if they can spot some irregularity.

    Also, please download and use version 28.0.1. There are fixes in various algorithms that may be related. Thanks.

    ------------------------------
    Rick Marcantonio
    Quality Assurance
    IBM
    ------------------------------



  • 3.  RE: Generalized Estimating Equations (GEE) Problem With Nested Factors

    Posted Wed December 08, 2021 06:09 PM
    Hello Rick! Thank you very much for your response. My syntax is shown below:

    DATASET ACTIVATE DataSet1.
    * Generalized Estimating Equations.
    GENLIN Rating_P (ORDER=ASCENDING) BY DesignTask Semester Instructor TeanNumber ParticipantNumber
    (ORDER=ASCENDING) WITH IndividualKAI IndividualPsychSafety
    /MODEL DesignTask Semester Instructor IndividualKAI IndividualPsychSafety
    Semester(TeanNumber(ParticipantNumber)) DesignTask*IndividualKAI
    DISTRIBUTION=MULTINOMIAL LINK=CUMLOGIT
    /CRITERIA METHOD=FISHER(1) SCALE=1 MAXITERATIONS=100 MAXSTEPHALVING=5 PCONVERGE=1E-006(ABSOLUTE)
    SINGULAR=1E-012 ANALYSISTYPE=3(WALD) CILEVEL=95 LIKELIHOOD=KERNEL
    /REPEATED SUBJECT=UniqueIdeaNumber WITHINSUBJECT=count SORT=YES CORRTYPE=UNSTRUCTURED
    ADJUSTCORR=YES COVB=ROBUST MAXITERATIONS=100 PCONVERGE=1e-006(ABSOLUTE) UPDATECORR=1
    /MISSING CLASSMISSING=EXCLUDE
    /PRINT CPS DESCRIPTIVES MODELINFO FIT SUMMARY SOLUTION WORKINGCORR.

    As you can see, I added in some interaction terms as well, but these run with no problem. I really just need to account for participants nesting within teams, which also are nested within semesters (as these teams only appear once per semester). I can also leave it to just participants nesting within teams, to simplify the analysis (the nesting within semesters is technically correct, but for the sake of interpretation/the analysis actually working, I am OK with omitting it).

    As for the SPSS version 28.0.1, do you know how I can go about downloading this without having to have some kind of special login? I am using a license from my academic institution, and after doing some searching, I could not find a page that would allow me to update to the latest version. I am also using Mac OS.

    Thank you!

    Courtney

    ------------------------------
    Courtney Cole
    ------------------------------



  • 4.  RE: Generalized Estimating Equations (GEE) Problem With Nested Factors

    Posted Thu December 09, 2021 02:23 PM
    Hi, Courtney. I just received this feedback and thought I would pass it on:

    The first problem that jumps out at me is that the SUBJECT specification on REPEATED is UniqueIdeaNumber, which is actually unique across all cases in the file. That means that there's only one observation per subject specified, so if this is right, there's really no repeated measures element involved here. If I take out the REPEATED subcommand, I get output, but that indicates a quasi-complete separation issue, which is not surprising, given the number of factors specified and their numbers of levels.

    I think the nested structure specified is actually backwards, as it should be ParticipantNumber(TeanNumber(Semester)), based on what I see in the data, but even so, I think this is what should be on the SUBJECT specification on REPEATED, which requires it to be specified like an interaction, or ParticipantNumber*TeanNumber*Semester. (Note that either way, it delineates all the unique combinations of the three factors.) When I take that effect off the MODEL subcommand and specify it as the SUBJECT effect using asterisks, things go much better. However, GENLIN still doesn't converge on estimating the working correlation matrix with the UNSTRUCTURED specification, even with 1000 iterations. That's not too surprising, as it's a 20x20 matrix, so there are 210 parameters to estimate. If I substitute a simpler structure, such as AR(1), EXCHANGEABLE, or MDEPENDENT, it does converge.

    If I add back that nested term (specified correctly or in the other direction), I get the original error. I don't think the data support estimating that effect with the specified structures. It's specifying a separate parameter for each subject. If I keep it on the MODEL and take out the REPEATED subcommand, I get rid of the original error, but again there's a quasi-complete separation, so unique estimates don't exist. I think the following (with possible adjustment of the working correlation structure) is probably the best option here:

    GENLIN Rating_P (ORDER=ASCENDING) BY DesignTask Semester Instructor TeanNumber ParticipantNumber 
    (ORDER=ASCENDING) WITH IndividualKAI IndividualPsychSafety
    /MODEL DesignTask Semester Instructor IndividualKAI IndividualPsychSafety DesignTask*IndividualKAI 
    DISTRIBUTION=MULTINOMIAL LINK=CUMLOGIT
    /CRITERIA METHOD=FISHER(1) SCALE=1 MAXITERATIONS=100 MAXSTEPHALVING=50 PCONVERGE=1E-006(ABSOLUTE) 
    SINGULAR=1E-012 ANALYSISTYPE=3(WALD) CILEVEL=95 LIKELIHOOD=KERNEL
    /REPEATED SUBJECT=ParticipantNumber*TeanNumber*Semester WITHINSUBJECT=count SORT=YES CORRTYPE=EXCHANGEABLE
    ADJUSTCORR=YES COVB=ROBUST MAXITERATIONS=100 PCONVERGE=1e-006(ABSOLUTE) UPDATECORR=1
    /MISSING CLASSMISSING=EXCLUDE
    /PRINT CPS DESCRIPTIVES MODELINFO FIT SUMMARY SOLUTION WORKINGCORR.
    ​


    ------------------------------
    Rick Marcantonio
    Quality Assurance
    IBM
    ------------------------------



  • 5.  RE: Generalized Estimating Equations (GEE) Problem With Nested Factors

    Posted Thu December 09, 2021 09:30 PM
    Edited by Courtney Cole Thu December 09, 2021 09:31 PM
      |   view attached
    Hello Rick,

    Thank you so much for this. I see that this is something I could only specify in the syntax for the repeated subject, and this makes sense to me to have all three of these here.

    I would just like to make sure that this accounts for the nesting of participants in teams, and teams in semesters. I have working with GEE for a month, and the way it has been explained to me is that it assumes nesting within the data and is a good way to run a multi-level analysis. As long as that is the case here, I will work with this method for the remainder of my analysis.

    I was also told that the Threshold in Parameter Estimates on the output does not mean much, even if there is a significant finding here (I attached an image here). Is this safe to assume? I also ran the syntax with the interaction effect removed to see what just the main effects were showing, so it is not the exact output as the syntax you sent, but slightly modified.

    Thank you!

    Courtney

    ------------------------------
    Courtney Cole
    ------------------------------