My dissertation focuses on how presidents allocate discretionary funds to localities. I study national leaders who often face a dilemma about how to allocate limited distributional resources to their greatest political advantage. I argue that the president's resource allocation strategy depends on the level of decentralization in a country. Presidents wish to allocate resources to their supporters in the electorate, and in centralized countries, they can do so directly, without any interference from local politicians. In decentralized countries, however, presidents must rely on local politicians as brokers, intermediaries between themselves and voters. Accordingly, presidents target localities with co-partisan mayors, even if most voters in the municipality did not support the president in the last election in decentralized settings. By analyzing 12 years of spending data in 100,000 municipalities, I show that presidents indeed allocate more to municipalities with presidential core voters in centralized Argentina and Venezuela, but mostly deliver resources to municipalities with local allies in decentralized Brazil and Mexico. I also show that local politicians have more influence among voters in decentralized countries, and much less so in centralized cases. Using survey experiments and electoral observation data from Argentina, Brazil, Mexico and Venezuela, I show that a co-partisan mayor helps the president's approval and vote share in decentralized, but not in centralized systems. By increasing the power of local politicians, decentralization creates an incentive for the president to support co-partisan mayors. In centralized countries, mayors are neither a threat nor an aid to presidential electoral prospects. By shaping presidential incentives, the degree of centralization affects short-term resource allocation and long-term efforts at party-building. These results help us to understand why in Venezuela and in Argentina presidents were able to deliver targetable goods to their voters in a higher pace than Brazil and Mexico presidents.

Longitudinal government expenditure data are often used in political science to assess politicians' agenda, presidents' distributive strategies, and coalition bargaining. Typically, this type of data are semicontinuous; contains some temporal dependence; and present a positive correlation between the probability of expenditure disbursed and the expected level of expenditure by each unit. In this thesis I use a Bayesian two-part model to separately estimate the likelihood of municipalities receiving national investments over time and mean expenditure trajectories among those chosen to receive national investments. I use elicited priors from fieldwork interviews with national and local politicians to estimate more precise posterior distributions. This approach extends to government expenditure data desegregated by states or municipalities. Applications to data from Brazil and Mexico uncovers two temporal patterns in government expenditures: funds are targeted to municipalities governed by co-partisan mayors, regardless of the president's vote share. Regions lacking either supportive voters or local allies receive little or no distributive funds whatsoever.