Obtaining factor scores with inputed matrix

[polisonica]

I am running EFA with SPSS using a polychoric correlation matrix. The factor analysis runs perfectly until I ask the program to save the factor scores. Then an error appears: "Factor scores cannot be computed with matrix input."
Has anyone enountered this problem before? Is there any walk-around? I know I can calculate the factor scores myself, but do not feel confident about it since most of my variables are categorical and have very different sdv.

[Art Kendall]

You do not have case by case information as input to FACTOR when you have already summarized the data into a correlation matrix.
Therefore, FACTOR cannot create scores   for each case.


Please give more detail so list members can help.
What is the context? Where did the set of items come from? a previously developed scale? new sets of items designed to measure specific constructs?
An available jumble of items that you are exploring? etc.

What are your cases? How many are there? Do you have the raw case by case data or do you only have the correlation matrix?

What do you want to do with the new variables? After you have chosen a number of factor to retain (NOT using Kaiser criterion of eigenvalues greater than 1.00) if you are creating summative scales, if you want the measures to have discriminative validity you would typically look at the varimax rotated loadings, and use unit weights for cleanly loading variables dropping splitters, etc.. There are a few situations where you would want "factor scores". 


What is (are) the response scale(s) on your items? If your items are not severely different from interval level, did you try ordinary correlations?  Do the results look very different? If you feel the intervals are severely discrepant did you try CATPCA?

What stopping rule are you using?







On 10/31/2012 9:28 AM, polisonica wrote:

I am running EFA with SPSS using a polychoric correlation matrix. The factor
analysis runs perfectly until I ask the program to save the factor scores.
Then an error appears: "Factor scores cannot be computed with matrix input."
Has anyone enountered this problem before? Is there any walk-around? I know
I can calculate the factor scores myself, but do not feel confident about it
since most of my variables are categorical and have very different sdv.




--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Obtaining-factor-scores-with-inputed-matrix-tp5715966.html
Sent from the SPSSX Discussion mailing list archive at Nabble.com.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD

[Ruben Geert van den Berg]

Dear all,

I'm not sure but couldn't you do something like

1. Compute a correlation matrix of your original variables

2. Run a factor analysis on some matrix

3. Insert the factor loadings (=correlation matrix between factor and raw variables), transposed factor loadings and the correlations among factors (identity matrix if rotation = orthogonal rather than oblique) into the original correlation matrix

4. Regress the factors one by on on the raw variables, with the expanded correlation matrix as input

5. Use the regression coefficients to calculate predicted values (=default factor scores) for the factors in your original data?

Some years ago we did something similar: assigning factor scores to new respondents (in real time). These were based on already existing factor loadings from a previous study. However, this situation was less complicated and involved only Pearson correlations.

Best,

Ruben

---

Date: Wed, 31 Oct 2012 10:17:48 -0400
From: [hidden email]
Subject: Re: Obtaining factor scores with inputed matrix
To: [hidden email]

You do not have case by case information as input to FACTOR when you have already summarized the data into a correlation matrix.
Therefore, FACTOR cannot create scores   for each case.


Please give more detail so list members can help.
What is the context? Where did the set of items come from? a previously developed scale? new sets of items designed to measure specific constructs?
An available jumble of items that you are exploring? etc.

What are your cases? How many are there? Do you have the raw case by case data or do you only have the correlation matrix?

What do you want to do with the new variables? After you have chosen a number of factor to retain (NOT using Kaiser criterion of eigenvalues greater than 1.00) if you are creating summative scales, if you want the measures to have discriminative validity you would typically look at the varimax rotated loadings, and use unit weights for cleanly loading variables dropping splitters, etc.. There are a few situations where you would want "factor scores". 


What is (are) the response scale(s) on your items? If your items are not severely different from interval level, did you try ordinary correlations?  Do the results look very different? If you feel the intervals are severely discrepant did you try CATPCA?

What stopping rule are you using?







On 10/31/2012 9:28 AM, polisonica wrote:

I am running EFA with SPSS using a polychoric correlation matrix. The factor
analysis runs perfectly until I ask the program to save the factor scores.
Then an error appears: "Factor scores cannot be computed with matrix input."
Has anyone enountered this problem before? Is there any walk-around? I know
I can calculate the factor scores myself, but do not feel confident about it
since most of my variables are categorical and have very different sdv.




--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Obtaining-factor-scores-with-inputed-matrix-tp5715966.html
Sent from the SPSSX Discussion mailing list archive at Nabble.com.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD

[polisonica]

Thank you for the responses so far. I see the issues is that SPSS works uniquely with the (polychoric) correlation matrix I am providing and ignores the case by case information available when running the EFA.
I have been doing some reading about CATPCA and feels like it may be an appropriate alternative. I am adding some of the information that has been requested. With that information, would CATPCA still be advisable? (first time I look into it, so not feeling 100% confident)

Context: I have a set of indicators that are usually associated with different dimensions of poverty (income, social support, participation, etc.). I am running an exploratory analysis to assess if the dimensions identified theoretically match the data I have.
I am using 16 variables, and working on a dataset with the case by case information for around 1800 cases representing individual respondents.
My variables are both binary and ordinal, ranging from 2 to 7 response categories.

I initially run the factor analysis with ordinary correlations, and the resulting factors look quite different to me. Previous studies on my area of research advise against doing so, and strongly suggest the use of polychoric correlations, which I have calculated and then saved the matrix in a different file.

Finally, to specify that I am not creating a summative scale. I would like to use the factor scores to a) check their correlation against accepted measures of poverty b) identify population groups whose mean scores are significantly different.
I am considering a 5 factor solution.    

So, two questions:
a) Is there any way I can tell SPSS to use both the polychoric correlation matrix and the case by case info to create factor scores?
b) Would CATPCA be a better alternative?