We don't use a particular cutoff. This issue usually comes up in EFA because there are no standard errors. We recommend doing EFA in a CFA framework to obtain standard errors and thereby be able to asses statistical significance.
yang posted on Wednesday, April 12, 2006 - 12:18 pm
I am doing EFA on 12 continuous indicator variables, and the results suggest a 2-factor structure. However, there are 2 indicators (PN; DS) that have loadings (Promax Rotated) smaller than 0.40 (0.276, 0.390; 0.156, 0.321) on both of the 2 factors, while another indicator (CO) has loadings greater than 0.40 (0.413, 0.541) on both of the 2 factors. As to the estimated residual variances of these 3 indicators, the former two have pretty large value (0.803, 0.623), and the latter one has 0.276. PN has Skewness equals to -2.3 and Kurtosis equals to 7.2, CO has standard deviation obviously larger than others (13.6 vs. 1.0-5.0), and DS is approximately normally distributed.
I have 3 questions: 1. How to deal with the two indicators that have small loadings on both of the 2 factors?
2. Should I scratch out PN as some references suggested for its ¡°Cross loadings¡±? I am still worrying about its non-normality, although I am using MLM in the estimation option.
3. Is there any ¡°Golden Standard¡± of the cutoff for the factor loadings?
you suggested above to do "efa in a cfa framework" -- can expand on what specifically you need to do when doing a "efa in a cfa framework". what values in the efa would tell you that a loading is significant? if a loading is < .30 on a factor, but it is significant, you are saying you would keep the item????
It is no longer necessary to do an EFA in a CFA framework to obtain standard errors for factor loadings. These are provided in Version 5. The ratio of the factor loading to the standard error tells you whether the factor loading is significantly different from zero. The size of the loading speaks to the practical significance of the loading.
1. To test pattern coefficients (loadings) for significance, is it necessary to do a Bonferroni correction for the critical point fort he z-statistic, as discussed in Cudeck and O'Dell (1994), or just use the absolute value of z greater than 2 for alpha = .05?
2. Why are structure coefficients rarely discussed?