SPSS Statistics

 View Only

  • 1.  Partial rank correlations with more than one control variable

    Posted 24 days ago

    Partial rank correlations in SPSS

    Ibm remove preview
    Partial rank correlations in SPSS
    I wish to use SPSS to find the partial correlation of a pair of variables, controlling for a set of additional variables. However, I want the partial correlation to be a rank correlation, such as the Spearman rho, rather than a Pearson correlation. The Partial Corr procedure in SPSS does not provide a rank correlation option.
    View this on Ibm >

    Following this troubleshooting entry https://www.ibm.com/support/pages/partial-rank-correlations-spss I am using partial rank correlations for my data analysis (although i use Kendall Tau b instead of Spearman rho).

    However, I want to include more than one control variable in my data set. 
    So the relevant syntax line from the example would look like this for me

    PARTIAL CORR salbeg time BY edlevel age

    In my tests, SPSS delivered plausible results and tables. However, I couldn't find anywhere a definitive statement that this actually works in the program, and I am not just producing "artefacts".

    So. Can I include more than one control variable in this way in the given SPSS code?



    ------------------------------
    Moritz Schmid
    ------------------------------


  • 2.  RE: Partial rank correlations with more than one control variable

    Posted 22 days ago

    Hi Moritz

    I tried with a sample data 'banklloan.sav' in SPSS. It is all getting fine and correct. I started with finding the partial correlation between 3 variables "preddef1 preddef2  and preddef3 " by taking the control variable as "preddef3" and the output is attached in Screenshot1. The Syntax used is

    PARTIAL CORR
      /VARIABLES=preddef1 preddef2 BY preddef3
      /SIGNIFICANCE=TWOTAIL
      /MISSING=LISTWISE.

    Then I tried to find the partial correlation between the same variables but by Spearmann rank correlation and the output is attached in Screenshot2.  The Syntax is

    NONPAR CORR preddef1 preddef2 preddef3 
     /MISSING = LISTWISE
     /MATRIX OUT(*).
    RECODE rowtype_ ('RHO'='CORR') .
    PARTIAL CORR
      /VARIABLES=preddef1 preddef2 BY preddef3 
      /SIGNIFICANCE=TWOTAIL
      /MISSING=LISTWISE
      /MATRIX IN(*).

    Again, for the Spearmann rank correlation, I added one for control variable "income". and the output is attached in Screenshot3. The Syntax used is 

    NONPAR CORR preddef1 preddef2 preddef3 income
     /MISSING = LISTWISE
     /MATRIX OUT(*).
    RECODE rowtype_ ('RHO'='CORR') .
    PARTIAL CORR
      /VARIABLES=preddef1 preddef2 BY preddef3 income
      /SIGNIFICANCE=TWOTAIL
      /MISSING=LISTWISE
      /MATRIX IN(*).
    I have attached  all the Screenshots of the output for the above 3 cases. Please note that the correlation values are not same in these output. That is , we are getting the correct output using the above Syntax.
    Regards


    ------------------------------
    Bindu Krishnan
    Senior Statistician
    IBM SPSS Statistics
    ------------------------------

    Original Message


  • 3.  RE: Partial rank correlations with more than one control variable

    Posted 21 days ago

    Hello Bindu,

    thank you very much for confirming that it works with more than one control variable. The output looks the same as does mine, although I am using three control variables in my analysis.

    Cheers
    Moritz



    ------------------------------
    Moritz Schmid
    ------------------------------

    Original Message


  • 4.  RE: Partial rank correlations with more than one control variable

    Posted 21 days ago

    Are you sure, Moritz, such thing as "partial rank correlation" exist, in principle? The well known idea and formulas of partial correlation is genuine linear regression idea. It is tied with Pearson r. You could rank your data first and then compute partial r, as usual, but that will still be partial r, only on ranked data, it won't be "partial nonparametric\rank rho\tau". Check https://stats.stackexchange.com/q/40995/3277.



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

    Original Message


  • 5.  RE: Partial rank correlations with more than one control variable

    Posted 21 days ago
    Edited by Moritz Schmid 21 days ago

    Hello Kirill,

    I have seen your answering similarly to different questions similar to mine when looking for answers. My data is ordinal and non-linear, that's why i choose Kendall tau. Considering how many different sources talk about partial rank correlation with Kendall tau, I am pretty sure it's a viable method. Whether or not SPSS calculates it correctly, I trust the people from IBM to know what they are doing... 

    Cheers

    Moritz



    ------------------------------
    Moritz Schmid
    ------------------------------

    Original Message


  • 6.  RE: Partial rank correlations with more than one control variable

    Posted 21 days ago
    Edited by Kirill Orlov 21 days ago

    Moritz, did you see the PARTIAL CORR command SPSS Algorithms chapter? It never offered any sort of "Kendall tau partial correlation". Its formulas assume the zero order correlations (the input) are Pearson r.



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

    Original Message


  • 7.  RE: Partial rank correlations with more than one control variable

    Posted 21 days ago
    Edited by Moritz Schmid 21 days ago

    Thank you for your comment, I will give this a thought.

    The question then arises, whether or not SPSS can calculate the partial correlations for Kendall tau at all. 

    ------------------------------
    Moritz Schmid
    ------------------------------

    Original Message