This paper investigates the impact of the embeddedness of politicians in a local elite network on government performance in decentralized and centralized political systems. Formal political decision-making among a set of legislators is modeled via a mean voter decision rule derived from a modified non-cooperative legislative bargaining game of a Baron-Ferejohn type. Legislators’ policy preferences are derived endogenously from political support maximization based on legislators’ beliefs how a rural development policy translates into the welfare of the agrarian and non-agrarian population. Legislators are generally uncertain regarding the political technology, i.e. the welfare changes induced by a policy. Accordingly, legislators communicate with the local elite to learn more about the true political technology and hence to undertake better informed political decisions. However, local elites might be biased in favor of a specific population group, i.e. communication might also bias political beliefs. A trade-off between more efficient policy learning and an increased policy bias induced by an increased embeddedness in local elite networks is identified. Policy bias is attenuated in centralized when compared to decentralized systems, while vice versa the speed of policy learning through local elite networks is c.p. higher in decentralized when compared to centralized systems. Moreover, within a constitutional system elite network structures such as local size, clustering or centralization have an impact on overall efficiency of political decsion-making.