Discriminant validity

Question

My question is about the need for this validation. I have highly correlated constructs in a cfa model fitted using the lavaan library in R, so the correlation is greater than the root of the AVE.

From what I've researched, this validation is to assess whether constructs that shouldn't be correlated really aren't.

But let's say that the nature of the questions that will be used to generate the constructs makes some of them highly correlated. Something expected by the researcher.

Some doubts arise: 1 - can I consider that there is discriminant validity since there is only high correlation in what I expected there to be? 2 - or not, does the high correlation in the constructs violate some assumptions of the cfa model and should I adopt some intervention to solve this such as grouping two or more constructs into one or eliminating the construct that presents high correlation?

Answer 1

Rönkkö & Cho (2020) have extensively discussed this issue. The main take-home message of their paper is to assess the degree to which two or more variables are related to each other, rather than relying on cut-off points (i.e., “yes” or “no”). You can find the proposed classification system in their paper. In summary, one way is to look at the upper confidence interval of the correlations between your (latent)variable, and see whether it go beyond 0.8. (see "CI CFA(sys)" technique in their paper): https://doi.org/10.1177/1094428120968614