I am planning to use Mplus for EFA. However, the response to my individual items on the questionnaire are skewed. Can I still use Mplus to perform EFA for such data? Also, what is the thumb rule to calculate efficient sample size for a scale of 20 items.
If you have continuous items that are non-normal, you can use the MLM estimator for EFA. There is no rule of thumb for sample size. It depends on many factors. You could do a Monte Carlo study to see what would be needed for data like yours.
finnigan posted on Tuesday, May 04, 2010 - 6:07 am
I have a sample size of 113 individuals who have completed a 56 item survey using using a 5 point likert scale. The survey items have reported problems in the literature in that 18-25 items do not load on any factor. The purpose of the factor analysis is to identify items that do not load with a view to running a multiple indicator growth model with fewer construct indicators.
I tried to estimate an EFA , but MPLUS issued a warning that it could not calculate chi squared under MLR ; to be expected, I suppose, and the covaraiance matrix was not positive definite was reported under ML.
Is there any other estimator besides ML and MLr that might work given the sample size and non normality. I tried to use the Analysis = general and tech 12/13 to obtain mardias coefficent but I'm not sure if the warning encompassed output for mardia
I have posted on Semnet and bollens 2stage least squares estimator was suggested.
I'm writing the message to kindly ask your opinion on sample size even though I understand that there is no rule of thumb for sample size.
I have a relatively small sample of survey data, of which sample size is 401 from 75 groups (firms). The largest sample size of a group is 61, and the smallest of a group is 1. Survey instruments are composed of 35 items. I expect 4-5 factors.
Is multilevel EFA appropriate for this small number of cases per group? Do I need to remove some groups which has such a small number of cases, i.e., 1? If not, what would be appropriate analysis for this dataset?
Clusters of size 1 do not need to be eliminated from the analysis.
I don't think your sample size is large enough for the following reasons:
1. You need at least one cluster with more observations than within-level parameters. Your largest cluster size is 61. The number of parameters in an EFA extracting 4 or 5 factors using 35 items is well over that amount. See the formula on Slide 104 of the Topic 1 course handout.
2. You should not have more between-level parameters than the number of clusters. You have 75 clusters and if you want to do the same EFA model on the between level, you would have more parameters than clusters.