IBM Support

PI78407: DIFFERENT RESULTS ON COEFFICIENTS TABLE FOR LINEAR REGRESSION ON32 BIT VERSUS 64 BIT FOR 24 FP1 ON THE SAME 64 BIT COMPUTER

Subscribe

You can track all active APARs for this component.

 

APAR status

  • Closed as user error.

Error description

  • You have downloaded IBM SPSS Statistics 24 release , 32 bit and
    installed it on a Windows 7 64 bit operating system.
    
    IBM SPSS Statistics Client 32-bit 24.0 Microsoft Windows
    Multilingual (CNF0FML )
    Date posted: 22 Sep 2016
    
    You also installed 24 FixPack 1 for the 32 bit application.
    You run a Linear Regression  in this SPSS installation.
    
    
    REGRESSION
    /MISSING LISTWISE
    /STATISTICS COEFF OUTS R ANOVA
    /CRITERIA=PIN(.05) POUT(.10)
    /NOORIGIN
    /DEPENDENT y
    /METHOD=ENTER
    x1
    x2_1
    x2_2
    x2_3.
    
    You saved the results in an output file. Later on you
    uninstalled the software and instead installed the 64 bit
    version plus FixPack 1 on this same 64 bit computer
    
    IBM SPSS Statistics Client 64-bit 24.0 Microsoft Windows
    Multilingual (CNF26ML ) -
    Date posted: 22 Sep 2016
    
    You run the same linear regression syntax as before for the same
    file. You save the output and compare the results for the
    Coefficients table displayed for the 64 bit computer with the
    ones you got in SPSs 24 32 bit.
    You see some differences for the values on both tables 32 bit
    compared to 64 bit, and wonder where they come from?
    
    This issue is functioning as designed, see below explanation.
    

Local fix

  • When you specify ENTER for a set of variables, REGRESSION's
    sweep algorithm enters them according to the tolerance values
    (highest first). For the first one, it always enters the one
    listed last, because the algorithm searches the diagonal of the
    matrix it's sweeping and replaces the current high value with
    the new one if it's greater than or equal to it. Before
    anything's been entered, the values are all 1, so the last one
    winds up being chosen.
    
    After that it proceeds by entering predictors one at a time
    until no more can be entered based on tolerance values. In this
    case the first three variables entered in all cases are x4_4,
    x3_2 and x2_12. The divergence comes in after this.
    
    The variables x4_1 and x4_2 both have what appears to be
    essentially the same tolerance value at that point, but looking
    at the RBHEX16 values, they're slightly different on 24.0.0.1
    64-bit, with that for x4_1 being very slightly larger. So the
    64-bit version of Statistics chooses x4_1 at that point.
    
    On 32-bit, the tolerances come up slightly unequal again, but in
    the other direction, with that for x4_2 slightly larger and it
    is thus chosen as the next variable to enter.
    
    So from a software-independent perspective, the solutions are
    equally valid, since a tie in the tolerances means neither
    selection is preferable. Given the rules that SPSS Statistics
    uses, the 32-bit solution is probably the "correct" one here,
    since it corresponds to what you would get by taking the last
    one in the set of tied values. However, trying to change the
    code so that the results are the same on both versions would
    almost certainly result in creating more problems like this with
    other data. This issue is therefore functioning as designed.
    

Problem summary

Problem conclusion

Temporary fix

Comments

  • When you specify ENTER for a set of variables, REGRESSION's
    sweep algorithm enters them according to the tolerance values
    (highest first). For the first one, it always enters the one
    listed last, because the algorithm searches the diagonal of the
    matrix it's sweeping and replaces the current high value with
    the new one if it's greater than or equal to it. Before
    anything's been entered, the values are all 1, so the last one
    winds up being chosen.
    
    After that it proceeds by entering predictors one at a time
    until no more can be entered based on tolerance values. In this
    case the first three variables entered in all cases are x4_4,
    x3_2 and x2_12. The divergence comes in after this.
    
    The variables x4_1 and x4_2 both have what appears to be
    essentially the same tolerance value at that point, but looking
    at the RBHEX16 values, they're slightly different on 24.0.0.1
    64-bit, with that for x4_1 being very slightly larger. So the
    64-bit version of Statistics chooses x4_1 at that point.
    
    On 32-bit, the tolerances come up slightly unequal again, but in
    the other direction, with that for x4_2 slightly larger and it
    is thus chosen as the next variable to enter.
    
    So from a software-independent perspective, the solutions are
    equally valid, since a tie in the tolerances means neither
    selection is preferable. Given the rules that SPSS Statistics
    uses, the 32-bit solution is probably the "correct" one here,
    since it corresponds to what you would get by taking the last
    one in the set of tied values. However, trying to change the
    code so that the results are the same on both versions would
    almost certainly result in creating more problems like this with
    other data. This issue is therefore functioning as designed.
    

APAR Information

  • APAR number

    PI78407

  • Reported component name

    SPSS STATISTICS

  • Reported component ID

    5725A54ST

  • Reported release

    O00

  • Status

    CLOSED USE

  • PE

    NoPE

  • HIPER

    NoHIPER

  • Special Attention

    NoSpecatt / Xsystem

  • Submitted date

    2017-03-17

  • Closed date

    2017-03-21

  • Last modified date

    2017-03-21

  • APAR is sysrouted FROM one or more of the following:

  • APAR is sysrouted TO one or more of the following:

Fix information

Applicable component levels

[{"Business Unit":{"code":"BU048","label":"IBM Software"},"Product":{"code":"SSCQ88K","label":"Statistics Desktop"},"Component":"","ARM Category":[],"Platform":[{"code":"PF025","label":"Platform Independent"}],"Version":"O00","Edition":"","Line of Business":{"code":"","label":""}}]

Document Information

Modified date:
21 March 2017