I would not make the decision about item removal solely based on the loadings. You should also think about which item(s) are good indicators of your factor from a substantive (content) perspective. Can the item with the highest loading be seen as a marker variable for the intended factor/construct in terms of its content?

The inhomogeneity of the loadings (.88 vs. .5 or .6) in your case may indicate that the items are rather heterogeneous (they may contain large amounts of item-specific variance that is not shared with the other items; in addition, some of the items may share specific variance with others, e.g., due to similar wording).

Assuming this was a CFA, what was the overall model fit? Was it acceptable? If not, have you checked standardized residuals and/or modification indices for signs of residual associations among certain items (correlated error variables between items with similar content)? This can give you a sense for which items are most strongly related. Also, looking at the observed correlation matrix of the items can sometimes be helpful to understand what's going on (which items are highly correlated? Which items are not so highly correlated?)

Alireza--I am wondering if you have carefully examined the observed correlation matrix for these items, given that the model-implied correlations (using the tracing rule) seem to suggest this high-loading item is more-related to the remaining items than the remaining items are to one another? And, is there still significant error variance (unexplained variance) in the item with the highest loading?

~Sarah

Dear Christian Geiser and Sarah D. Newton

Thank you both for your very informative response. based on your suggestions , and while checking the results of cross-loadings I found that the mentioned item was loaded onto two factors. Thus, I guess this was the main reason for high factor loading in compare with other items.

Best of luck

Alireza

Item loading of 0.8 is within acceptable limits of factor loading because higher value of 0.9 is an indication of discriminate validity issues. However if it's removal improves the AVE then I feel that item may not be a good indicator for that factor. I think you that is your justification.

Taiwo O. Oladipo-Abodunwa

Thank you for reply .

I have had exactly the same experience. I wonder if there's a statistical explanation for this; and also some way to justify the decision to remove the highest loading item?

Francois van der Bank -- I am not sure if this relates to your particular situation, too. But in Alireza shabani shojaei's case, having high factor loadings on two different factors suggests that the item is multidimensional. (Part of the variance in the item can be explained by at least two different underlying factors.) This multidimensionality could definitely be an argument for item removal, given the desirability/necessity of unidimensional items (i.e., items that only measure a single factor). Remember that Cronbach's alpha assumes unidimensionality, right? So, I would think that this multidimensionality could potentially impact the resulting reliability estimate.

Dear Sarah D. Newton - Thank you for the quick response!

Yes, I did initially have an issue with multidimensionality but that was due to two other items with very similar wording. Deletion of the one item sorted out that problem. In my case, the highest-loading item (.862) loaded on a single factor. It was further clear from the observed correlation matrix that the item correlated more with the remaining items than the remaining items correlated with one another. Based on the content of the items, there is no reason to believe that the item is a “marker variable” for the construct. Deletion of this item ensured a 1-factor model with a decrease in the number of nonredundant residual correlations with absolute values greater than .05.

Thank you for sharing, Francois van der Bank !