Factor Rotation

Back to the adolescent data -- let's look at different rotations of the three factors with > 1.00. Varimax rotation tends to produce multiple group factors maintaining orthogonality often results in increased multivocality (loadings of variables on "primary factors" is decreased a bit and loadings on "secondary factors" is raised a bit) It is easy and reasonable to criticize orthogonal rotations for being simplistic or artificial. After all, most things we study are at least somewhat related. For example, if I told you that I had several anxiety and several depression measures and that they factored into 2 factors , youd not be surprised. But if I told you that the two factors were depression and anxiety, youd not be surprised. If I told you that those 2 factors were uncorrelated, youd probably be surprised. Oblique solution help to reduce this artificiality.

Total Variance Explained Initial Eigenvalues % of Variance Cumulative % 38.097 38.097 21.363 59.459 16.746 76.205 7.953 84.158 6.036 90.194 4.244 94.438 3.000 97.438 2.562 100.000 Extraction Sums of Squared Loadings Total % of Variance Cumulative % 3.048 38.097 38.097 1.709 21.363 59.459 1.340 16.746 76.205 Rotation Sums of Squared Loadings Total % of Variance Cumulative % 2.304 28.798 28.798 2.093 26.163 54.961 1.699 21.244 76.205

Component 1 2 3 4 5 6 7 8

Total 3.048 1.709 1.340 .636 .483 .340 .240 .205

Extraction Method: Principal Component Analysis.

Notce the variance "spreads out" across the 3 factors with this rotation -- common with Varimax. "sad" is a classic example of increased simple structure goes from multivocal to univocal so do several others note however, "extreme verbal abuse" goes from univocal to multivocal seems it would be very important to properly interpret why this variable is multi-vocal!

a Component Matrix

a Rotated Component Matrix

1 physical aggression property damage theft extreme verbal abuse sad anxious self-confidence compliance .758 .693 .362 .826 .540 .654 -.349 -.580

Component 2 3 .413 1.164E-03 .489 -.199 .656 -.204 6.589E-02 .235 -.510 .441 -.335 .507 .539 .669 .450 .551

1 physical aggression property damage theft extreme verbal abuse sad anxious self-confidence compliance .807 .853 .751 .562 -8.06E-02 .110 3.339E-02 -.165

Component 2 3 .301 -6.51E-02 9.875E-02 -.148 -.186 6.583E-02 .645 -.105 .846 -.155 .884 -5.44E-02 -2.60E-02 .927 -.209 .879

Extraction Method: Principal Component Analysis. a. 3 components extracted.

Extraction Method: Principal Component Analysis. Rotation Method: Varimax with Kaiser Normalization. a. Rotation converged in 4 iterations.

Quartimax rotation tends to produce a general factor and additional smaller multiple group factors

Quartimax rotation wont make strong group factors go away but you know you have a weak factor structure if varimax & quqrtimax give importantly different solutions one source of a weak factor structure that gives different solutions from different rotations is over factoring

Total Variance Explained Initial Eigenvalues % of Variance Cumulative % 38.097 38.097 21.363 59.459 16.746 76.205 7.953 84.158 6.036 90.194 4.244 94.438 3.000 97.438 2.562 100.000 Extraction Sums of Squared Loadings Total % of Variance Cumulative % 3.048 38.097 38.097 1.709 21.363 59.459 1.340 16.746 76.205 Rotation Sums of Squared Loadings Total % of Variance Cumulative % 2.337 29.214 29.214 2.083 26.040 55.254 1.676 20.951 76.205

Component 1 2 3 4 5 6 7 8

Total 3.048 1.709 1.340 .636 .483 .340 .240 .205

Extraction Method: Principal Component Analysis.

Still got some "spreading around" of the variance across the factors -- looks much the same as the varimax Pretty strong sign for "group factors" got them from both varimax and quartimax Remember, "strong" solutions won't be "hidden" by the rotation you select

a Component Matrix

a Rotated Component Matrix

1 physical aggression property damage theft extreme verbal abuse sad anxious self-confidence compliance .758 .693 .362 .826 .540 .654 -.349 -.580

Component 2 3 .413 1.164E-03 .489 -.199 .656 -.204 6.589E-02 .235 -.510 .441 -.335 .507 .539 .669 .450 .551

1 physical aggression property damage theft extreme verbal abuse sad anxious self-confidence compliance .814 .856 .746 .576 -6.10E-02 .129 2.063E-02 -.181

Component 2 3 .285 -4.99E-02 8.321E-02 -.135 -.203 7.244E-02 .634 -8.73E-02 .850 -.142 .882 -3.86E-02 -4.15E-02 .927 -.220 .873

Extraction Method: Principal Component Analysis. a. 3 components extracted.

Extraction Method: Principal Component Analysis. Rotation Method: Quartimax with Kaiser Normalization. a. Rotation converged in 4 iterations.

Direct Oblimin rotation Tends to produce varimax-looking factors, but which are oblique

Delta is a parameter that "controls" the extent of obliqueness amongst the factors. Negative values "decrease" factor correlations "0" is the default Positive values (don't go over .8) "permit" additional factor correlation

Delta = -.8
Component Correlation Matrix Component 1 2 3 1 1.000 .147 -.174 2 .147 1.000 -7.53E-02 3 -.174 -7.53E-02 1.000

Delta = -.6
Component Correlation Matrix Component 1 2 3 1 1.000 .149 -.181 2 .149 1.000 -7.79E-02 3 -.181 -7.79E-02 1.000

Extraction Method: Principal Component Analysis. Rotation Method: Oblimin with Kaiser Normalization.

Extraction Method: Principal Component Analysis. Rotation Method: Oblimin with Kaiser Normalization.

Delta = -.4
Component Correlation Matrix Component 1 2 3 1 1.000 .154 -.191 2 .154 1.000 -8.09E-02 3 -.191 -8.09E-02 1.000

Delta = -.2
Component Correlation Matrix Component 1 2 3 1 1.000 .160 -.207 2 .160 1.000 -8.58E-02 3 -.207 -8.58E-02 1.000

Extraction Method: Principal Component Analysis. Rotation Method: Oblimin with Kaiser Normalization.

Extraction Method: Principal Component Analysis. Rotation Method: Oblimin with Kaiser Normalization.

Delta = 0
Component Correlation Matrix Component 1 2 3 1 1.000 .171 -.237 2 .171 1.000 -9.52E-02 3 -.237 -9.52E-02 1.000

Delta = .2
Component Correlation Matrix Component 1 2 3 1 1.000 .194 -.307 2 .194 1.000 -.121 3 -.307 -.121 1.000

Extraction Method: Principal Component Analysis. Rotation Method: Oblimin with Kaiser Normalization.

Extraction Method: Principal Component Analysis. Rotation Method: Oblimin with Kaiser Normalization.

Delta = .4
Component Correlation Matrix Component 1 2 3 1 1.000 .329 -.509 2 .329 1.000 -.300 3 -.509 -.300 1.000

Delta = .6

Delta = .8 didn't converge

Component Correlation Matrix Component 1 2 3 1 1.000 .777 -.841 2 .777 1.000 -.793 3 -.841 -.793 1.000

Extraction Method: Principal Component Analysis. Rotation Method: Oblimin with Kaiser Normalization.

Extraction Method: Principal Component Analysis. Rotation Method: Oblimin with Kaiser Normalization.

When you do an oblique rotation you get two different matrices that can be used for interpretation: the structure matrix holds the correlations between each variable and each factor (same as with orthogonal rotations) the pattern matrix holds the beta weights to reproduce variable scores from factor scores There is considerable disagreement about which of these is the better basis for factor interpretation: Those who like using the structure matrix point out the long history of naming or interpreting factors (and other composite variables -- ldf, canonical correlation) in terms of the variables with which they correlate. Those who like using the pattern matrix point out that there is often "simpler structure" in the pattern matrix Those who like using the structure matrix point out that the apparent "simpler structure" (i.e., fewer multivocal items) in the pattern matrix is an illusion, made possible because of the correction for collinearity by the beta weights. Hopefully, the interpretation based on the two matrices will be similar Notice that the choice of Delta influences the factor structure/pattern as well as the factor intercorrelations. Higher correlations come with greater multivocality and both make it more difficult to interpret the factors
Pattern Matrixa Component 2 .787 .847 .774 .511 -.154 3.827E-02 6.719E-02 -.120

Delta = -.8
3 -2.49E-02 -.123 7.099E-02 -4.61E-02 -9.73E-02 1.194E-02 .939 .872

Structure Matrix Component 2 .826 .861 .735 .605 -.021 .168 .006 -.203

physical aggression property damage theft extreme verbal abuse sad anxious self-confidence compliance

1 .249 3.221E-02 -.231 .611 .852 .888 6.259E-02 -.116

physical aggression property damage theft extreme verbal abuse sad anxious self-confidence compliance

1 .370 .179 -.129 .694 .846 .892 -.091 -.286

3 -.128 -.192 .053 -.191 -.234 -.146 .923 .901

Extraction Method: Principal Component Analysis. Rotation Method: Oblimin with Kaiser Normalization. a. Rotation converged in 16 iterations.

Extraction Method: Principal Component Analysis. Rotation Method: Oblimin with Kaiser Normalization.

Delta = 0
Structure Matrix Component 2 .829 .862 .731 .613 -.010 .180 .000 -.211

Pattern Matrixa Component 2 3 .787 -1.36E-02 .848 -.119 .779 6.824E-02 .507 -2.52E-02 -.163 -7.26E-02 3.050E-02 3.954E-02 7.397E-02 .949 -.113 .875

1 physical aggression property damage theft extreme verbal abuse sad anxious self-confidence compliance .241 1.783E-02 -.240 .608 .858 .896 9.405E-02 -8.61E-02

1 physical aggression property damage theft extreme verbal abuse sad anxious self-confidence compliance .379 .191 -.123 .701 .847 .891 -.118 -.313

3 -.146 -.203 .051 -.218 -.261 -.176 .919 .906

Extraction Method: Principal Component Analysis. Rotation Method: Oblimin with Kaiser Normalization. a. Rotation converged in 9 iterations.

Extraction Method: Principal Component Analysis. Rotation Method: Oblimin with Kaiser Normalization.

Pattern Matrixa Component 2 1.155 1.271 1.325 .599 -.548 -.206 .515 .243

Delta = .6
3 .525 .252 .380 .573 .291 .602 1.683 1.358

Structure Matrix Component 2 .814 .806 .526 .759 .323 .456 -.350 -.567

1 physical aggression 9.692E-02 property damage -.341 theft -.641 extreme verbal abuse .791 sad 1.418 anxious 1.466 self-confidence .605 compliance .343

1 physical aggression property damage theft extreme verbal abuse sad anxious self-confidence compliance .553 .435 .070 .774 .747 .800 -.411 -.611

3 -.472 -.469 -.132 -.567 -.467 -.469 .766 .877

Extraction Method: Principal Component Analysis. Rotation Method: Oblimin with Kaiser Normalization. a. Rotation converged in 20 iterations.

Extraction Method: Principal Component Analysis. Rotation Method: Oblimin with Kaiser Normalization.

Promax Rotation An oblique rotation that tends to produce group factors that look like Direct Oblimin & Varimax Promax computations are more much quicker, so it is commonly used with very large factorings