Hello Arjumand,

I'm not sure I understand the thinking underlying the statement. As a result, I've got more questions than answers, so here are some thoughts which come to mind:

Why would someone want to discard useful identifying, observable indicators of a factor (latent variable)?

if the purpose is to generate "short forms" of a measure, please note that the literature tends to show that such efforts lead to: (a) scores with lower reliability; (b) differential relationships with other factors/variables.

If the purpose is to discuss the minimum number of indicator variables needed for identification of a factor model (as in confirmatory factor analysis), then three or more is the preferred number.

Of course, why stop at two? One could extend the pruning to just a single indicator variable, but that's hardly evidence for a common factor!

Good luck with your work.

Hi,

Ideally you are not supposed to delete more than 20% of the items per construct. If you are deleting more items that means you are challenging existing theory.

Regards,

Uday Bhale

David Morse thanks for giving an elaborative response. Actually, I need a reference to cite in my work as I have dropped many indicators on their weak loadings and want to retain at least 3 par construct as Uday Arun Bhale mentioned in his response one should not drop much indicators.

It depends on context, and as David Morse says, one is often fine. If I ask people their age in years, that variable is influenced by the latent variable of the age that the respondent wants to convey to the interviewer. It is not perfect, but asking multiple related questions is unlikely to be worth the time. The breadth of other constructs that you want to measure would require multiple questions to span this breadth. So, without more info it is difficult to give an answer ... though with lots of the people recently answering questions on RG just using ChatGPT someone may post a long detailed rambling answer broken into steps!