<p>Network analysis is becoming routine in many fields. For some practical settings, for example, the analysis of gene expression data, the interconnections among variables can be classified as direct and indirect, where the indirect interconnections can be attributed to other variables, and the direct interconnections can reflect more essential properties and be of unique interest. For the estimation of direct network interconnections, two families of approaches have been developed. The first family assumes that the contributing variables (to indirect interconnections) are completely unobserved and treated as latent variables. The second family focuses on the scenario where the contributing variables are fully observed. In this study, we consider the more realistic scenario where the contributing variables are partly observed. To facilitate interpretation and improve estimation, we jointly consider two graphical models: one incorporating a set of unobservable latent variables and the other incorporating the observed contributing variables. We propose a hierarchical structure in sparsity, which naturally leads to a joint estimation. For the proposed approach, the consistency properties are rigorously established. In simulation, it demonstrates competitive performance. In the analysis of a lung cancer gene expression study, it leads to sensible findings different from the alternatives.</p>