I have 24 items including 0 and 1. I used TetMat* program to obtain tetrachoric correlations. Well, I prepared the SPSS data set which was exampled in the help files. But I get some errors.
* Uebersax JS. TetMat User Guide (v. 1.0). Computer program documentation.
June 2006.
Hi,
I used a programme called CEFA (comprehensive exploratory factor analysis) to assess the factor structure of items on a scale that were dichotomously scored. I think its freely available and quite easy to use.
Rizwana
I'm not an SPSS user, but I am curious what your error message was? Was it something to the effect of indicating that you have a "not positive definite" matrix? This is often a problem when polychoric correlations are used. You might try using another factor extraction method like minimum residual estimation if SPSS supports this (I assume SPSS uses maximum likelihood by default).
Well...thank you Rizwana. now i am trying CEFA software. i am on it. trying build up .inp file. and paul you are right main problem is not positive definite matrix. i am using AMOS software for conducting SEM. but mine two factor model give an error as you mentioned before "not positive definite matrix". now i am going to try CEFAtool 3.04. Lets see what happen:
I am not sure if you have found a solution yet. If not, you can try this:
First, bear in mind that factor analysis works on correlation or covariance matrices (I will assume correlation matrix here). The not positive definite nature of a matrix refers to the fact that it has negative eigenvalues. The eigenvalues of a correlation matrix are proportional to the proportion of variance explained by the corresponding eigenvectors. If an eigenvalue is negative, that is a problem, because variance and proportion of variance should be positive (even in the case of negative correlations).
The standard way to fix this is basically to alter your correlation matrix to be positive definite. For a sophisticated approach to this problem, see the following article: http://imajna.oxfordjournals.org/content/22/3/329.full.pdf+html
For an easy approach, you can do the following:
Perform an eigenvalue decomposition on the correlation matrix. Find the negative eigenvalues and replace them with small positive values (under 0.1 and less than any pre-existing positive value). Now, if you multiply the eigenvector matrix by the eigenvalue vector and then by the transposed eigenvector matrix, you get the matrix you started with. Here, you will multiply by the vector of your modified eigenvalues. This will produce a positive definite correlation matrix, because you have designed it to be so. Also, stick with an estimation method like principal axis factoring in this case, which SPSS supports.
You should report that you have done this in the manuscript. It is playing with the data, but the bias should not be too big.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS