Hello. I was wondering if in the event that you are performing an EFA and receive a warning that there is a Heywood case, is computing the EFA in a CFA framework the only way you can deal with this? Thank you.
As a follow-up question I also wanted to know if in the event that Mplus tells you which specific variable is causing a Heywood case, then should you delete that particular variable? If this is incorrect, what are one's other options?
Computing an EFA in a CFA framework or using our new EFA feature described in the Version 5.1 Examples Addemdum would still result in a negative residual variance. If it is small and non-signficant, you could fix it to zero in this setting. Removing the variable will probably not help. You should reduce the number of factors or increase your sample size.
Mat Weldon posted on Friday, August 17, 2012 - 8:52 am
I have conducted an EFA with count indicators. I do not appear to have a Heywood case, but I have a near Heywood case in that all of the residual error variances are vanishingly small and insignificant. The count indicators themselves have very small means (ranging from 1/5 to 1/20), but the sample size is large (5570). The same situation occurs for one, two and three factors, and when I conduct a CFA the residual variances are not included in the output. Otherwise, the estimates seem reasonable.
Are these results to be trusted? Is there any way to improve the estimates?