I did E.Factor analysis for 55 items in math, there is problem in data which is with non-normal distribution, so I did it with FACTOR , and I choose :
Robust Diagonally weighted least squares, with O.Imp. Parallel Analysis, I had 2 factors, with - - RMSEA 0.080 ......H0 : RMSEA < 0.05 ; P = 1.00
- Non-Normed Fit Index (NNFI; Tucker & Lewis) = 0.877; BC Bootstrap 95% confidence interval = ( 0.861 0.909)
- Comparative Fit Index (CFI) = 0.892; BC Bootstrap 95% confidence interval = ( 0.878 0.920) (lower than 0.900 : poor)
- GFI = 0.935
- AGFI = 0.926
- Root Mean Square of Residuals (RMSR) = 0.0832
- Expected mean value of RMSR for an acceptable model = 0.0423 (Kelley's criterion)
I HAVE ALSO, Weighted Root Mean Square Residual (WRMR) = 0.0895 (values under 1.0 have been recommended to represent good fit; Yu & Muthen, 2002)
I did the same thing with HULL METHOD FOR SELECTING THE NUMBER OF COMMON FACTORS (HULL) Lorenzo-Seva, Timmerman, Kiers (2011), I had 1 factor,
- Root Mean Square Error of Approximation (RMSEA) = 0.059
- Non-Normed Fit Index (NNFI; Tucker & Lewis) = 0.933
- Comparative Fit Index (CFI) = 0.954
- Goodness of Fit Index (GFI) = 0.963
- Adjusted Goodness of Fit Index (AGFI) = 0.947
- Root Mean Square of Residuals (RMSR) = 0.0567
So, which one is better??? or may be there is another method to extract factors?????
Thanks a lot..
Hi Baghdad BRAHIM Mohammed
It seems your dataset have better fit with unidimensonal structure. Your parameter estimation method choice is ok i think. But you can judge firstly number of dimensions on your data. Except Parallel Analysis and Hull method, you can try DETECT (R software) and DIMTEST (DIMTEST software). Another step is examine the modification indices to increase fit indices. Last step is confirmatory analysis like CFA or for dimensionality confirmatory DETECT analysis.
I hope it helps to you
It is a good method to use the factor program especially for one-dimensional tests with a large number of items in your sample, its indexes are the confirmatory model IFC-TLI NNFI comparative indices are greater than .90 SRMR-rmsa less than 0.06 so it confirms the unidimensionality of the model, there is also a method of plotting to group the one-dimensional tests formed by a large number of items in plots that provide better results in this type of instruments.
Translated with www.DeepL.com/Translator (free version)
Hello Brahim,
There's always a danger in trying alternate analytic methods with the same data, then opting for the one with the most "pleasing" results--that of overfitting to your (specific) data set.
The first order of business, I think, would be to decide conceptually on whether one or two factors makes better theoretical sense for your variable set. If you're using confirmatory statistics to characterize model-data fit, then it seems to me you should have a target structure in mind.
That said, the one-factor solution results certainly seem to be promising (as Hakan Koğar suggests).
Good luck with your work.
- Badges
- Science topic
- Similar topics
- Indexes