4
$\begingroup$

This is my initial output of Principal Component Analysis (PCA) using Stata and correlation matrix (because different scales and measurement units of inputs):

Principal components/correlation                 Number of obs    =        350
                                                 Number of comp.  =          5
                                                 Trace            =          5
    Rotation: (unrotated = principal)            Rho              =     1.0000

    --------------------------------------------------------------------------
       Component |   Eigenvalue   Difference         Proportion   Cumulative
    -------------+------------------------------------------------------------
           Comp1 |      2.80769      1.79023             0.5615       0.5615
           Comp2 |      1.01746      .281177             0.2035       0.7650
           Comp3 |      .736282      .413602             0.1473       0.9123
           Comp4 |      .322679      .206788             0.0645       0.9768
           Comp5 |      .115892            .             0.0232       1.0000
    --------------------------------------------------------------------------

Principal components (eigenvectors) 

    ------------------------------------------------------------------------------
        Variable |    Comp1     Comp2     Comp3     Comp4     Comp5 | Unexplained 
    -------------+--------------------------------------------------+-------------
               A |   0.5627    0.0500   -0.1329   -0.2992   -0.7574 |           0 
               B |  -0.0466    0.9662   -0.2391    0.0725    0.0425 |           0 
               C |   0.5490   -0.0071   -0.1761   -0.5088    0.6393 |           0 
               D |   0.5036   -0.1168   -0.3114    0.7899    0.1091 |           0 
               E |   0.3552    0.2241    0.8928    0.1500    0.0628 |           0 
    ------------------------------------------------------------------------------

After orthogonal rotation (Varimax) I have these outputs:

Principal components/correlation                 Number of obs    =        350
                                                 Number of comp.  =          5
                                                 Trace            =          5
    Rotation: orthogonal varimax (Kaiser off)    Rho              =     1.0000

    --------------------------------------------------------------------------
       Component |     Variance   Difference         Proportion   Cumulative
    -------------+------------------------------------------------------------
           Comp1 |      1.00001  2.59031e-06             0.2000       0.2000
           Comp2 |            1  2.53877e-06             0.2000       0.4000
           Comp3 |            1  2.40356e-06             0.2000       0.6000
           Comp4 |      .999997  2.28153e-06             0.2000       0.8000
           Comp5 |      .999995            .             0.2000       1.0000
    --------------------------------------------------------------------------

Rotated components 

    ------------------------------------------------------------------------------
        Variable |    Comp1     Comp2     Comp3     Comp4     Comp5 | Unexplained 
    -------------+--------------------------------------------------+-------------
               A |   1.0000    0.0000   -0.0000    0.0000   -0.0000 |           0 
               B |   0.0000    0.0000    1.0000   -0.0000    0.0000 |           0 
               C |   0.0000    0.0000   -0.0000   -0.0000    1.0000 |           0 
               D |  -0.0000    0.0000    0.0000    1.0000    0.0000 |           0 
               E |  -0.0000    1.0000   -0.0000   -0.0000   -0.0000 |           0 
    ------------------------------------------------------------------------------

Component rotation matrix

    ----------------------------------------------------------------
                 |    Comp1     Comp2     Comp3     Comp4     Comp5 
    -------------+--------------------------------------------------
           Comp1 |   0.5627    0.3552   -0.0466    0.5036    0.5490 
           Comp2 |   0.0500    0.2241    0.9662   -0.1168   -0.0071 
           Comp3 |  -0.1329    0.8928   -0.2391   -0.3114   -0.1761 
           Comp4 |  -0.2992    0.1500    0.0725    0.7899   -0.5088 
           Comp5 |  -0.7574    0.0628    0.0425    0.1091    0.6393 
    ----------------------------------------------------------------

Here are some rows of datasets:

enter image description here

All options are Stata default options as we can see here:

enter image description here

Why we have strange outputs (specially in proportion and cumulative variances and rotated components) after rotation? How can I select between Orthogonal and Oblique rotation and rotation method (Varimax,Quantimax etc.)? Is any test to help selecting method? What is the problem of results?

PS 1.

After set maximum number of components to 3 I have these results:

Principal components/correlation                 Number of obs    =        350
                                                 Number of comp.  =          3
                                                 Trace            =          5
    Rotation: orthogonal varimax (Kaiser off)    Rho              =     0.9123

    --------------------------------------------------------------------------
       Component |     Variance   Difference         Proportion   Cumulative
    -------------+------------------------------------------------------------
           Comp1 |      2.53555      1.51519             0.5071       0.5071
           Comp2 |      1.02036     .0148549             0.2041       0.7112
           Comp3 |      1.00551            .             0.2011       0.9123
    --------------------------------------------------------------------------

Rotated components 

    ----------------------------------------------------------
        Variable |    Comp1     Comp2     Comp3 | Unexplained 
    -------------+------------------------------+-------------
               A |   0.5700    0.0944    0.0550 |      .09537 
               B |  -0.0005   -0.0067    0.9964 |     .001904 
               C |   0.5753    0.0370    0.0102 |       .1309 
               D |   0.5866   -0.1272   -0.0627 |       .2027 
               E |  -0.0005    0.9867   -0.0070 |     .007721 
    ----------------------------------------------------------

Component rotation matrix

    --------------------------------------------
                 |    Comp1     Comp2     Comp3 
    -------------+------------------------------
           Comp1 |   0.9319    0.3602   -0.0440 
           Comp2 |  -0.0446    0.2340    0.9712 
           Comp3 |  -0.3601    0.9031   -0.2341 

Ps2:

I compared MATLAB outputs with above results with this code in MATLAB:

[coeff ,score, latent, tsquared, explained, mu] = pca(data,'Centered','on','VariableWeights','variance');
[L,T] = rotatefactors(coeff);

Results:

out1 =

   -0.0000   -0.0000   -0.0000    0.0000   -0.0473
    0.0000    0.5293   -0.0000   -0.0000   -0.0000
    0.0634   -0.0000   -0.0000   -0.0000   -0.0000
    0.0000    0.0000   -0.0000    0.1088    0.0000
   -0.0000   -0.0000   -0.1285   -0.0000    0.0000

>> out2

out2 =

    0.5490   -0.0466   -0.3552    0.5036   -0.5627
   -0.0071    0.9662   -0.2241   -0.1168   -0.0500
    0.1761    0.2391    0.8928    0.3114   -0.1329
   -0.5088    0.0725   -0.1500    0.7899    0.2992
    0.6393    0.0425   -0.0628    0.1091    0.7574

Compared with Stata we have different rotated outputs!

Data: LINK (after normalization using a sample values as denominator of other samples because some theoretical concepts- I used mapstd and mapminmax in MATLAB but the behavior is the same + I removed outliers based on bigger than 2 standard deviations (abs(X-mean(x))>=2*SD) in this data-set.

$\endgroup$
  • 1
    $\begingroup$ It is always adviced is cases similar to yours to show the data and syntax as well, not only the results. $\endgroup$ – ttnphns May 28 '15 at 9:53
  • $\begingroup$ @ttnphns . Thank you for comment. I edited question. $\endgroup$ – user2991243 May 28 '15 at 10:15
  • 1
    $\begingroup$ You did not quite understand me. When I said "show data" I meant give the whole data (which the results correspond to) - either insert the values as text or - if the dataset is big - leave a link to the external storage. The idea is that a reader would be able to redo your analysis. $\endgroup$ – ttnphns May 28 '15 at 12:01
  • $\begingroup$ I added my database $\endgroup$ – user2991243 May 28 '15 at 12:05
5
$\begingroup$

I rerun your analysis in SPSS (I don't have Stata, and I didn't rerun it in Matlab this time).

The sweet pulp of your mistaken analysis is that you somehow managed to rotate eigenvectors, whereas rotations are normaly done of loadings. Please read my recent answers about eigenvectors/loadings and about rotations.

Your first analysis extracted all 5 components. I can confirm (in SPSS) the eigenvalues and the eivenvectors you displayed. Then one would expect that you request loadings (which are the eigenvectors scaled up to the respective eigenvalues) which are:

      Component
       1       2       3       4       5
V1   .943    .050   -.114   -.170   -.258
V2  -.078    .975   -.205    .041    .014
V3   .920   -.007   -.151   -.289    .218
V4   .844   -.118   -.267    .449    .037
V5   .595    .226    .766    .085    .021

Then this matrix after varimax rotation will be:

      Component             
       1       2       3       4       5
V1   .831    .247    .371    .012    .334
V2  -.014    .014   -.044    .999    .002
V3   .924    .188    .300   -.032   -.142
V4   .442    .124    .886   -.063    .027
V5   .215    .970    .107    .015    .021
 Rotation Method: Varimax without Kaiser Normalization. 

with the rotation transformation matrix:

       1       2       3       4       5
1    .760    .387    .513   -.050    .078
2    .018    .225   -.105    .968    .021
3   -.251    .884   -.317   -.235   -.011
4   -.595    .132    .790    .066   -.005
5    .066    .025    .038    .019   -.997

You rotated the matrix of eigenvectors, not loadings. We know that the eigenvector matrix in PCA is itself a special case of orthogonal rotation matrix. Its column sums-of-squares are 1, row sums-of-squares are 1 and cross-products of the columns are 0. Such a matrix, when it is rotated orthogonally to a "simple structure" - such as by varimax method - will inevitably turn into a very simple view like the one you got in rotated components table, with 0 and 1 values only. Each column contains only one 1 and each row contains only one 1, but you may shuffle the exact position of the 1s, that simple structure equivalently persists. For example SPSS varimax rotation gave me this in your place:

      Component             
       1       2       3       4       5
V1   .000    .000    .000   1.000    .000
V2   .000   1.000    .000    .000    .000
V3   .000    .000   1.000    .000    .000
V4  1.000    .000    .000    .000    .000
V5   .000    .000    .000    .000   1.000
 Rotation Method: Varimax without Kaiser Normalization.

In your second analysis you retained and rotated 3 of the total 5 components. Since you discarded two last columns in eigenvector matrix, the row SS were no longer 1 and so varimax gave you simple structure which consists of values fractional, not 0 and 1. But the sweet pulp remains: you again rotated the wrong matrix. You ought to have rotated loading matrix, not eigenvector matrix.

Also, in most cases it is better not to switch off Kaiser normalization when doing loadings rotation.


P.S. Stata documentation clearly states it that pca function computes and rotates only eigenvectors. It does, though, compute and rotate loadings in a special post-function:

Remark: Literature and software that treat principal components in combination with factor analysis tend to display principal components normed to the associated eigenvalues rather than to 1. This normalization is available in the postestimation command estat loadings; see [MV] pca postestimation.

$\endgroup$
  • $\begingroup$ Thank you for answer. 1: which of the results (stata or MATLAB) has the wrong rotation problem? because in stata we only select rotation and set rotation method. 2: you said but you may shuffle the exact position of the 1s, that simple structure equivalently persists. You shuffled data? I didn't get that. Why the position of ones are different in report of me and your report? $\endgroup$ – user2991243 May 28 '15 at 14:09
  • 1
    $\begingroup$ 1. I can't say, I'm not Stata user. There must be an option to rotate / display rotated loadings. 2. Even when you rotate loadings, the results by different software may differ in regards to the sequence of the columns. Remember that rotated components do not correspond to the unrotated ones. E.g. 1st component after rotation is not "rotated 1st component"! Thus, your table (with 0 and 1) is equivalent with mine: they differ only by the order of their columns. $\endgroup$ – ttnphns May 28 '15 at 14:22
  • 1
    $\begingroup$ What's with that? I see no problem. You may present the unrotated matrix. Then present the rotated matrix. We need a rotation for simple-structure style interpretation of factors (or components, if you wish to). Either you interpret the unrotated results and use them - for further speculations in your study - or the rotated ones and use them for that. Not both at the same time. $\endgroup$ – ttnphns May 28 '15 at 14:40
  • 2
    $\begingroup$ Strange: it seems that if varimax is applied to a 2x2 matrix of eigenvectors then it does not do anything at all (see here stats.stackexchange.com/questions/232281). Only for 3x3 or larger does it yield a matrix of zeros and ones. Can you confirm? Why is that? $\endgroup$ – amoeba Aug 31 '16 at 9:45
  • 1
    $\begingroup$ @amoeba, No, SPSS does it correct as I've just commented in here stats.stackexchange.com/q/232281/3277. Yesterday I simply was tired and looked in a wrong table when were checking it as you asked me. So come back to erase our last comments above. $\endgroup$ – ttnphns Sep 1 '16 at 15:05
1
$\begingroup$

PCA and CFA are highly subjective techniques with many heuristics, options and rules of thumb. For instance, there is no test for choosing between oblique and orthogonal rotations. It's simply a matter of analyst preference that can have significant downstream implications as a function of the choice made. For another, one commonly applied rule of thumb is that the component eigenvalues should have a minimum value of 1.0, the logic being that any preserved component should contribute at least as much to the overall variance as a single variable. For your data, this would give only two components, not three.

The next thing is that OLS PCA is not scale invariant. You state that you "normalized" your data, including removing some outliers. There may be a problem with terminology here but, for me, "normalizing" refers to mean centering only where "standardizing" refers to transforming the data into an orthonormal basis function that is mean centered with a standard deviation of one. If you only mean centered your data, then your results would be erroneous in OLS PCA.

The other thing is that removing outliers is always a bad idea. There's simply too much information in these outliers to warrant deleting them. Not to mention that deleting a first pass of outliers typically creates a new round of outliers, and so on, producing a fruitless, pointless infinite series of outlier deletions. 20th c statistical techniques such as Winsorizing, trimming and the whole boatload of techniques involving grooming data to a more "normal" PDF is really dumb. Use robust techniques instead.

$\endgroup$
  • $\begingroup$ Thank you for answer. What are robust techniques which can handle outliers in PCA (or FA)? Can you name some of them? + i used mapstd function in MATLAB which Process matrices by mapping each row's means to 0 and deviations to 1. $\endgroup$ – user2991243 May 28 '15 at 14:50
  • 2
    $\begingroup$ This answer raises some interesting problems/warnings in PCA, but it does not seem to address the specific OP question. The question was about the strangeness/discrepancy of the rotation results produces by different software. $\endgroup$ – ttnphns May 28 '15 at 14:56
  • $\begingroup$ Some of the robust techniques for PCA are e.g. projection pursuit PCA. You can check the R package {pcaPP} from cran.r-project.org/web/packages/pcaPP/pcaPP.pdf and references therein $\endgroup$ – Delyan Savchev May 28 '15 at 15:01
  • $\begingroup$ ttnphns-not being a Matlab user, I'm not qualified to answer the specific OP question regarding discrepancies. To me there may be so many fundamental flaws in the solution as to make the discrepancies in results sort of irrelevant. To user2991243, here are a couple of papers that have extensive references to more approaches to robust PCA: 1) Robust Principal Component Analysis? by Emmanuel J. Candes, Xiaodong Li, Yi Ma, and John Wright 2) CAUCHY PRINCIPAL COMPONENT ANALYSIS by Pengtao Xie & Eric Xing $\endgroup$ – Mike Hunter May 28 '15 at 15:09
  • $\begingroup$ presumably your "winzoring" was a typo for "Winsorizing". I've edited. $\endgroup$ – Nick Cox May 28 '15 at 18:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.