Linear Statistical Models: Regression

Partial and Semipartial Correlation


Experimental Control

  • Controling variances by using equal groups.
  • Randomly assigning subjects to groups.
  • Conditions for all groups identical except for independent variable.

    Statistical Control

  • Control variance by removing unwanted variance from other variables.
  • control by partialing.

    Partial Correlation

  • r12.3 is the correlation between variables 1 and 2 with variable 3 removed from both variables. To illustrate this, run separate regressions using X3 as the independent variable and X1 and X2 as dependent variables. Next, compute residuals for regression...

    Venn Diagram of Partial Correlation

    More Partial Correlation

    Example

  • Let r12 = 0.7;   r13 = 0.6;   r23 = 0.9.

    Using Multiple Correlations

  • Thus, squared partial correlations represent the ratio of incremental variance to the residual variance.

    Example

  • Let R21.23 = .4947 and R21.3 = .62 = .36 (from the above example)

    Higher Order Partial Correlation

    and so on...

    Suppressor Variable

  • A special case when the partial correlation is larger than the zero-order correlation.
  • Zero or close to 0 correlation with the dependent variable.
  • Correlated with one or more independent variables.
  • Serves to suppress or control irrelevant variance.

    Example

  • Let r12 = 0.3;   r13 = 0.0;   r23 = 0.5.

  • Let r212 = 0.09;   r212.3 = 0.12;

  • Note: r212.3 is greater than r212 even though r13 = 0.
  • Standardized regressions Coefficients: β2 = .4 and β3 = -.2
  • Note: suppressor variable receives a negative coefficient.

    Causal Relationships

  • Partial correlation as a control method must be predicated on a sound theoretical framework.
  • Routine presentation of all higher-order partial correlations is a sign that theory explaining the relationship among the variables is missing (Gordon, 1968).
  • With only three variables there are many possible causal models.

    Measurement Error

  • Measurement error leads to biased estimates of zero-order and partial correlations.
  • Measurement error attenuates zero-order correlations.

    Correction for Attenuation

    Example

  • When r12 = .7; r11 = r22 = .8

    Correction for Attenuation in Partial Correlation

  • Partial correlation when correcting for unreliability in X3 only.

    Example

  • When r12 = .7; r13 = .5; r23 = .6; r33 = .8;

    and r12.3 = .58

    Another Example

  • When r12 = .7; r13 = .8; r23 = .7; r33 = .8;

    and r12.3 = .33

    Correction for Attenuation in Partial Correlation

  • Partial correlation when correcting for unreliability in all measures.

    Example

  • When r12 = .7; r13 = .5; r23 = .6; r11 = r22 = r33 = .8;

    and r12.3 = .58

    Semipartial Correlation

  • AKA -- part correlation.
  • r1(2.3) is the correlation between variables 1 and 2 with variation from variable 3 removed only from variable 2. Compute residuals for X2 using X3 as the independent variable...

    Venn Diagram of Semipartial Correlation

    Semipartial Correlation Formulas

  • Thus, the squared semipartial correlation represents the proportion of variance of the dependent variable accounted for by a given independent variable after another variable has already been taken into account.

    Example

  • r12 = .7; r13 = .6; r23 = .9

    Remember

  • r21(2.3) does not necessarily equal r21(3.2)

    since r21(2.3) = R21.23 - R21.3

    and r21(3.2) = R21.23 - R21.2

    Higher Order Semipartials

    Regression & Semipartials

  • When independent variables are uncorrelated:

  • When independent variables are correlated:

    Multiple Partial Correlation

  • R21.23(4) is the squared multiple correlation of variable 1 with variables 2 and 3 after the variation due to variable 4 has been partial out from all variables.

    Multiple Semipartial Correlation

  • R21(23.4) is the squared multiple correlation of variable 1 with variables 2 and 3 after the variation due to variable 4 has been partial out only from variables 2 and 3.

    Partial & Semipartial Correlation in Stata

    pcorr write read math science female
    
    (obs=200)
    
    Partial and semipartial correlations of write with
    
                   Partial   Semipartial      Partial   Semipartial   Significance
       Variable |    Corr.         Corr.      Corr.^2       Corr.^2          Value
    ------------+-----------------------------------------------------------------
           read |   0.2567        0.1756       0.0659        0.0308         0.0003
           math |   0.2925        0.2022       0.0855        0.0409         0.0000
        science |   0.2792        0.1922       0.0779        0.0369         0.0001
         female |   0.4239        0.3094       0.1797        0.0957         0.0000
    
    partcorr write read, part(math science female)  /* findit partcorr */
    
           Response Variable: write
       Predictor Variable(s): read
         Partial Variable(s): math science female
               Number of obs: 200
    
                               Coef       Coef Squared
        Mutiple Correlation = .59677648  .35614217
        Partial Correlation = .25672403  .06590723
    Semipartial Correlation = .17559606  .03083398
    
    partcorr write read math, part(science female)
    
           Response Variable: write
       Predictor Variable(s): read math
         Partial Variable(s): science female
               Number of obs: 200
    
                               Coef       Coef Squared
        Mutiple Correlation = .66641064  .44410315
        Partial Correlation = .47497762  .22560374
    Semipartial Correlation = .35680815  .12731206
    
    


    Linear Statistical Models Course

    Phil Ender, 24sep10, 28Jan98