The Steiner tree problem is a classical NP-hard optimization problem with a wide range of practical applications. In an instance of this problem, we are given an undirected graph G=(V,E), a set of terminals R, and non-negative costs c_e for all edges e in E. Any tree that contains all terminals is called a Steiner tree; the goal is to find a minimum-cost Steiner tree. The nodes V R are called Steiner nodes. The best approximation algorithm known for the Steiner tree problem is due to Robins and Zelikovsky (SIAM J. Discrete Math, 2005); their greedy algorithm achieves a performance guarantee of 1+(ln 3)/2 ~ 1.55. The best known linear (LP)-based algorithm, on the other hand, is due to Goemans and Bertsimas (Math. Programming, 1993) and achieves an approximation ratio of 2-2/|R|. In this paper we establish a link between greedy and LP-based approaches by showing that Robins and Zelikovsky's algorithm has a natural primal-dual interpretation with respect to a novel partition-based linear programming relaxation. We also exhibit surprising connections between the new formulation and existing LPs and we show that the new LP is stronger than the bidirected cut formulation. An instance is b-quasi-bipartite if each connected component of G R has at most b vertices. We show that Robins' and Zelikovsky's algorithm has an approximation ratio better than 1+(ln 3)/2 for such instances, and we prove that the integrality gap of our LP is between 8/7 and (2b+1)/(b+1).