Globally-optimized latent variable extraction in formative-reflective models