Discussion: View Thread

Factor analysis

  • 1.  Factor analysis

    Posted 05-15-2014 17:57
    I may be preparing a factor analysis for a dataset with continuous and categorical variables.

    My reading of the literature so far, is to use a polychoric method and it seems I can prepare that in both SAS and R.

    I'd appreciate pointers/suggestions to current methodology in factor analysis, or alternative methods, for a mix of variables.


    -------------------------------------------
    Chris Barker, Ph.D.
    Consultant and
    Adjunct Associate Professor of Biostatistics
    www,barkerstats.com

    ---
    "In composition you have all the time you want to decide what to say in 15 seconds, in improvisation you have 15 seconds."
    -Steve Lacy
    -------------------------------------------


  • 2.  RE:Factor analysis

    Posted 05-15-2014 18:17
    Chris,

    Am I correct that the categorical measures are at least ordinal.  If so, I personally have been leaning toward use of ML methods while using a scree test to determine the number of factors.

    -------------------------------------------
    David Mangen
    Owner
    Mangen Research Associates, Inc.
    -------------------------------------------








  • 3.  RE:Factor analysis

    Posted 05-15-2014 18:27
    Its a mix of all kinds, nominal, ordinal, categorical.
    Its a dataset collected by a surgeon for his transgender patients in the process of undergoing gender change.


    -------------------------------------------
    Chris Barker, Ph.D.
    Consultant and
    Adjunct Associate Professor of Biostatistics
    www,barkerstats.com

    ---
    "In composition you have all the time you want to decide what to say in 15 seconds, in improvisation you have 15 seconds."
    -Steve Lacy
    -------------------------------------------








  • 4.  RE:Factor analysis

    Posted 05-15-2014 18:38
    Try using Latent Class Models which are analogous to factor analysis using categorical data.  Penn State makes the SAS code available for free.  I haven't checked out the software in awhile but I think that they have expanded the use to ordinal variables. 

    The Methodology Center
    204 East Calder Way, Suite 400
    State College, PA 16801

    Phone: 814-865-3253
    Fax: 814-863-0000

    methodology.psu.edu



    -------------------------------------------
    Ihor Kowalysko
    -------------------------------------------








  • 5.  RE:Factor analysis

    Posted 05-15-2014 18:54
    Yes, I was leaning toward some of the latent class models as well, but note that latent class is probably a bit more like cluster analysis (for objects) than factor analysis (for variables).

    I believe that the Penn State center also makes their R-code available as well.

    The other though that I had was to ask you about the theoretical implications of the nominal level variables.  I have often found it theoretically compelling to suggest that the nominal measures have a statistical interaction effect on the measurement model.  As such I might move to the LISREL multi-group model and estimate a ML model within each group, attempting to constrain to equality parameters across-groups, relaxing those constraints only as necessary for the measurement model.
    -------------------------------------------
    David Mangen
    Owner
    Mangen Research Associates, Inc.
    -------------------------------------------








  • 6.  RE:Factor analysis

    Posted 05-15-2014 19:14

    I have done numerous factor analyses and principal components analyses using categorical variables, both nominal and ordinal with satisfactory (i.e., interpretable factors) results. However, I have not done them with multiple variable types in the same analysis. Since both FA and PCA are based on correlations there shouldn't be any reason why it can't be done with a mixture of variable types. That doesn't mean, however, that the resulting factors will be interpretable. My first thought, which experience has taught me is often not even close, is that the nominal variables may group together, and the same for the ordinal variables, and the interval/ratio variables.

    Sent from my iPad
    -------------------------------------------
    Daniel Mundfrom
    Professor of Applied Statistics and Chair
    Eastern Kentucky University
    -------------------------------------------





  • 7.  RE:Factor analysis

    Posted 05-16-2014 08:04
    I would use CATPCA in SPSS.  It is designed to use nominal, ordinal, and continuous variables in the same analysis.  It has built in comparison of results assuming different levels of measurement. It is often helpful to see whether ordinal variables are severely discrepant from interval level or not.

    -------------------------------------------
    Arthur Kendall
    Social Research Consultants
    -------------------------------------------








  • 8.  RE:Factor analysis

    Posted 05-16-2014 10:13
    Thank you. The online SPSS materials I found on a google search don't give a citation to the method.

    I have  SAS or R for the software.

    -------------------------------------------
    Chris Barker, Ph.D.
    Consultant and
    Adjunct Associate Professor of Biostatistics
    www,barkerstats.com

    ---
    "In composition you have all the time you want to decide what to say in 15 seconds, in improvisation you have 15 seconds."
    -Steve Lacy
    -------------------------------------------








  • 9.  RE:Factor analysis

    Posted 05-16-2014 12:23
    I just Googled CATPCA and found the manual
    http://pic.dhe.ibm.com/infocenter/spssstat/v20r0m0/index.jsp?topic=%2Fcom.ibm.spss.statistics.help%2Fidh_cpca.htm
    The manual had a link to the  algorithms
    http://pic.dhe.ibm.com/infocenter/spssstat/v20r0m0/index.jsp?topic=%2Fcom.ibm.spss.statistics.help%2Fidh_cpca.htm



    You can also try
    SPSS's TWOSTEP clustering which clusters cases based on variables with mixed levels on measurement.

    Am I correct in assuming that the reason for the factor analysis is to reduce the number of dimensions because there are too few degrees of freedom to use the variables as is?

    Were the variables chosen to measure a smaller set of constructs?
    Or is an ad hoc way to reduce the number of variables?

    If you have the data in SAS and all cleaned up etc, you could download the trial version of SPSS and just read the SAS data in.





    -------------------------------------------
    Arthur Kendall
    Social Research Consultants
    -------------------------------------------








  • 10.  RE:Factor analysis

    Posted 05-16-2014 11:07
    For SAS users there is a procedure called PRINQUAL whose description sounds similar to how Dr. Kendall describes CATPCA in SPSS.

    With best wishes,

    Tor Neilands

    -------------------------------------------
    Torsten Neilands
    Professor of Medicine
    UCSF Center for AIDS Prevention Studies
    -------------------------------------------








  • 11.  RE:Factor analysis

    Posted 05-16-2014 02:13
    Dear Chris,

    I've used Mplus for these sorts of situation.  There apparently are R packages that as well but my memory is now fuzzy on the particulars.

    Best regards,
    Kevin

    -------------------------------------------
    Kevin Gray
    Cannon Gray LLC
    -------------------------------------------