SPSS Statistics

 View Only

  • 1.  Significance tests vs total in custom tables

    Posted Fri April 05, 2024 10:22 AM

    Hi team,
    we are using significance tests in custom tables.
    Next to countries as columns we also include a total column and want to test all other columns not only against each other but also against total.
    Can you kindly advise how to do that in syntax?

    CTABLES
    /VLABELS VARIABLES=Q34 dcountry
    DISPLAY=LABEL
    /TABLE Q34 [MEAN] BY dcountry
    /CATEGORIES VARIABLES=dcountry ORDER=A KEY=VALUE EMPTY=INCLUDE total=yes
    /CRITERIA CILEVEL=95
    /COMPARETEST TYPE=MEAN ALPHA=0.05 ADJUST=BONFERRONI ORIGIN=COLUMN INCLUDEMRSETS=YES
    CATEGORIES=ALLVISIBLE MEANSVARIANCE=ALLCATS MERGE=NO SHOWSIG=NO
    /TITLES title="Subscription Metrics".

    Thanks much for your support.



    ------------------------------
    Christine Loof
    ------------------------------


  • 2.  RE: Significance tests vs total in custom tables

    Posted Fri April 05, 2024 12:15 PM
    Edited by Kirill Orlov Fri April 05, 2024 01:55 PM

    There is no such option. Because - tell us - which way to go to (say) "compare a percentage of a yes-response in a group with that in the total sample".

    Treating the total's percent as a fixed number and do the one-sample test? But total sample isn't the population, its percent has its error. (And, besides, mind that the same every respondent who effected the group's percent participates one more time to form the total sample percent. This violates the assumption that the large population must be independent from the sample from it in the sense that it is infinitely larger.)

    Alternatively, comparing a group with the total would be comparing that group with all the rest groups joint. This way, two groups are compared. It is legal. However, each group is thus gets compared with a nonfixed group, the group percent - compared with "total" percent which changes each time, each comparison.

    The third option is to break - in advance - each of your groups (subsamples) randomly into two parts: one (a bigger part) will still represent that group; the other (smaller and proportional to its share in the population) - will be sent to join with the other such parts from the other groups - to form the "group representing the population". So, having k groups initially, you end with k+1 groups and compare the k ones with that quasi-population representative one. You are comparing each group with one fixed group, nice; but, as your groups are of smaller n now, you've lost some power.

    Unfortunately, CTABLES does only all pairwise comparisons, and not each-with-one. But you could still request all pairwise without adjustment for multiple comparisons, and then apply the actual adjustment you need yourself.

    P.S. I was speaking of comparing proportions - as an example - but the same logic applies to comparing means.

    ------------------------------
    Kirill Orlov
    ------------------------------

    Original Message


  • 3.  RE: Significance tests vs total in custom tables

    IBM Champion
    Posted Fri April 05, 2024 12:50 PM

    As Kirill points out, the problem with testing against the total is multiple counting.  Imagine that there are two groups.  The total comparison would effectively be a two-group t test.

    You could aggregate the columns with k-1 in one column and the remaining column and do the column means test if one column is of particular interest.

    But you might want to think about this differently.  Considering averages rather than totals, this is a search for outliers.  There are some sig tests for outliers, but these are iffy.  Rather, one might look at graphical displays for patterns without sig testing.  You might also look for large residuals in a two-way crosstab.



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

    Original Message


  • 4.  RE: Significance tests vs total in custom tables

    Posted 8 days ago

    I know what you mean, but too many non-professionals need significance tests - the differences between "total" and "every category". 

    I find a way, syntax is :

    COMPUTE ttl = 0 .
    EXECUTE .
    MRSETS /MCGROUP NAME=$test VARIABLES=ttl test LABEL='test' .
    CTABLES
      /VLABELS VARIABLES=$anal $test DISPLAY=DEFAULT
      /TABLE $anal [C][COLPCT.COUNT PCT40.1, TOTALS[COUNT F40.0]] BY $test [C]
      /SLABELS POSITION=ROW
      /CATEGORIES VARIABLES=$anal EMPTY=INCLUDE TOTAL=YES POSITION=BEFORE LABEL="BASE"
      /CATEGORIES VARIABLES=$test EMPTY=INCLUDE
      /COMPARETEST TYPE=PROP ALPHA=0.05 ADJUST=NONE ORIGIN=COLUMN INCLUDEMRSETS=YES CATEGORIES=ALLVISIBLE MERGE=NO .

    But, the significance tests result has errors. One of the error examples: "Total" has 500 cases, and "success" is 249. "Target Category" has 193 cases, "success" is 107, and "CTABLE" tell me: they have significance difference .
    By the way, I have another idea, "Merge File With Cases", and forge a category of "Total", a magical thing happened, and there was no significant difference in the results this time.
    So, I am very confused about the testing method for "CTABLES". If possible, I hope you can list out the formulas. If necessary, I can also send you my data.


    ------------------------------
    Zhiying Na
    ------------------------------

    Original Message


  • 5.  RE: Significance tests vs total in custom tables

    Posted Fri April 05, 2024 01:31 PM
    Edited by Kirill Orlov Fri April 05, 2024 01:50 PM

    Jon's mentioning departures from the central tendency, and especially chi-square residuals (if we speak of percentage tests) - is quite relevant. Imagine you have a frequency table with 2 rows: the yes-response of interest, and the "other responses" response, and several groups of respondents as the columns. Then the standardized or adjusted resuduals in the cells tell you the departure of each group from the total sample (which percent is the weighted average percent). Here is an answer on the two kind of residuals you can get in the Crosstabs procedure in SPSS: https://stats.stackexchange.com/a/178068/3277

    Note that you may do approximate significance inference out of the residual magnitude.

    I would append the "resisuals option" as the fourth one to the three I mentioned earlier.

    P.S. I am putting an accent on comparing proportions here, though - as I see now - the OP has given a table syntax analyzing specifically means.

    ------------------------------
    Kirill Orlov
    ------------------------------

    Original Message


  • 6.  RE: Significance tests vs total in custom tables

    IBM Champion
    Posted Fri April 05, 2024 04:33 PM
    For means, you could work off of the residuals from a two-way ANOVA.--



    Original Message