I am trying to perform a CFA on data (10 indicators: $n=300$) that is severely non-normal but continuous (counts of a clinical behavior over a period of weeks): many cases are at zero, a fair few between 1 and 30, and a handful at values sometimes much higher (even $> 1000$). I am using lavaan in R with the MLM estimator to accommodate for this distribution. I am primarily comparing a single common factor model with a model containing three correlated factors.
If I use all cases, I get warnings indicating a non-positive definite matrix, and negative error variance (i.e. Heywood case). Because of the severity of the outliers, in line with Bollen's (1987) recommendations, I explored whether the extreme values were the cause of the negative error variance. I removed all cases 3 standard deviations above the mean ($n=16$) and this led to admissible solutions when I re-ran my models, and sensible parameter estimates. Am I interpreting these errors appropriately, and does this sound like a reasonable procedure to have taken in these circumstances?
On a related note, is there a limit to how high the Satorra-Bentler scaling correction factor can be before one should be concerned? I see values of around 4 in some cases (unsurprisingly given the non-normality of the data), although I cannot locate any literature that comments on this issue.
