This thesis focuses on the effects of anti-corruption and, more generally, oversight policies in Brazil. The results shed light on a complex interaction between corruption control and policy outcomes, often missing from the relevant literature. I find robust evidence of unintended short-term impacts of anti-corruption initiatives, which are a direct consequence of increased oversight of the bureaucracy. These outcomes are not likely to be a result of reducing corruption itself, rather a response to higher requirements of compliance, aggravated by a context of low state capacity and limited resources. After an introductory chapter contextualising the following work, three substantive chapters provide empirical results and theoretical models documenting the methodology and the consequent findings. Chapter 2 examines the effects of random audits on municipal budget. It is shown that upon being audited, local administrations reduce spending for the following years; the gap between budgeted amounts at the beginning of the year and actual budget execution grows substantially reaching to 10-15% of the total discretionary expenses. I propose a political delegation model which explains why, in a low policy capacity context, bureau- crats are likely to distance themselves from the politicians’ goals when facing higher risk of sanctions. Chapter 3 focuses on public procurement, a crucial and often overlooked mechanism of policy delivery. It finds that after sanctioning companies found guilty of not complying with procurement regulation, government agencies reduce the number of future tenders. I propose a model that combines the administrative response to these sanctions with the bidding behaviour of suppliers to explain these effects. In Chapter 4, I propose a novel machine learning model that predicts the probability of a company bidding on a tender using only network data. This graph model combines a regularised variational autoeconder and a multinomial distribution to generate alternative procurement networks. The model is able to produce interpretable results that assist on the analysis of these large networks.