I am developing a scale based on a theoretical model, which now has 73 items, representing 10 clusters (so about 7 times per cluster). I want to perform a CFA,with n=2100. But as the number of items is high, I would like to divide the items in three groups (n=700) so that each respondent has only around 24 items to score (likert-scale).

Is it statistically possible/wise to do that or should each respondent score each item?
Any comments?

Thank you,