This message often points to improper solutions/out-of-bounds parameter estimates/Heywood cases. You can check the completely standardized solution ("std.all" in lavaan, see https://rdrr.io/cran/lavaan/man/standardizedSolution.html) for latent correlation estimates > |1|. This warning message may also point to zero or negative latent variance estimates (less likely for a CFA model).

Also common when factors are highly correlated: this can be due to dependencies among more than 2 latent variables (one latent variable may be a perfect linear function of a set of other factors in the model, i.e., it is perfectly predictable from other factors). In that situation, there may be no single improper estimate in the output.

Do you have highly correlated factors in your model? It could be that some factors are so highly correlated (lack discriminant validity) that they could be collapsed into one factor.

Hello Mahdi,

To elaborate on Christian Geiser's comment, your five factors are indeed highly correlated in the "troublesome" adolescent sample:

F1,F2: .72

F1,F3: .82; F2,F3: .90

F1,F4: .68; F2,F4: .94; F3,F4: .79

F1,F5: .79; F2,F5: .82; F3,F5: .92; F4,F5: .89

So, the collinearity among the factors is clearly an issue. With relationships that strong, you may want to reconsider the factor structure. Either a second-order model or a single factor model might be viable options, given your results.

Good luck with your work.