Hierarchical latent variable models for dimensionality reduction: an application on composite indicators