Hi, I am testing the factorial construct validity of an instrument which includes 15 Likert-type items. These are non-symmetrical. My sample size is around 600.
Using Confirmatory Factor Analysis (CFA), what is the best estimation method given my data? The robust ML or the ADF?
Other suggestions?
Thank you.
I did a factor analysis years ago, 10 years ago or more on SF36, eight domains and this was done in SPSS. Please remind me of what ML and ADF stand for ?
Hi,
for using ADF, your sample size is too small. Hence, robust ML (Satorra-Bentler correction or Yuan-Bentler correction) is the way to go.
HTH
Holger
Hi, ML is maximum likelihood, ADF is asymptotically distribution free. Thank you.
Dear all,
There are 3 main methods used to obtain estimates in CFA: ML, GLS and ULS.
Both ML and GLS are scale invariant or free, ie, that parameters estimates obtained will not change if different units of measurement are used. The additional advantage is that we can use overall shi-saquare statistics for testing the adequacy of a given mode. The most popular is ML.
When the vriables distribution depart markedly from normality, ML and GLS produce unreliable goodness of fit statistics and standar errors.
Special CFA techiniques are better to analyse ordinal and nonnormal variables (Muller, 1993).
In terms of sample size, some goodness of fit indexse, like NFI are distorted for small samples whereas NNFI fit accuracy at all sample sizes.
Omiting important factor loading lowers dramatically the value of goodness of fit indexes...and model misspecification gives an improper solution with unique error negative, parameter estimates very large or factor intercorrelations escedd 1 (Bagozzi & Yi, 1988).
Helena
Dear Michael,
as far as I know (please correct me if wrong), WLSMV (as its sister DWLS) are estimation methods to treat categorial/ordinal data, but not non-normality. Both estimators assume that the latent response variable that underlies the observed ordinal response is normaly distributed AFAIK. For nonnormality, the correction methods proposed by Albert Satorra and colleagues could be more fruitful.
Finney, S. J., & DiStefano, C. (2006). Non-normal and categorical data in structural equation modeling. In G. R. Hancock & R. O. Mueller (Eds.), Structural equation modeling: A second course (pp. 269–314). Greenwich, CT: Information Age.
Hu, L.-T., Bentler, P. M., & Kano, Y. (1992). Can test statistics in covariance structure analysis be trusted? Psychological Bulletin, 112(2), 351–362.
Best,
Holger
My recommendation is go for robust ML if you have numerical indicators, if indicators are categorical use WLSMV you can find both estimatiors in MPLUS or in lavaan package for R
I prefer robust maximum likelihood estimator since this estimator is the most powerful in terms of unbiased, consistency, efficient and suffieciency
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS