Learning about corruption: a statistical framework for working with audit reports

Mendes, Eduardo Fonseca
Título da Revista
ISSN da Revista
Título de Volume

Quantitative studies aiming to disentangle public corruption effects often emphasize the lack of objective information in this research area. The CGU Random Audits Anti-Corruption Program, based on extensive and unadvertised audits of transfers from the federal government to municipalities, emerged as a potential source to try to fill this gap. Reports generated by these audits describe corrupt and mismanagement practices in detail, but reading and coding them manually is laborious and requires specialized people to do it. We propose a statistical framework to guide the use of text data to construct objective indicators of corruption and use it in inferential models. It consists of two main steps. In the first one, we use machine learning methods for text classification to create an indicator of corruption based on irregularities from audit reports. In the second step, we use this indicator in a regression model, accounting for the measurement error carried from the first step. To validate this framework, we replicate an empirical strategy presented by Ferraz et al. (2012) to estimate effects of corruption in educational funds on primary school students’ outcomes, between 2006 and 2015. We achieved an expected accuracy of 92% on the binary classification of irregularities, and our results endorse Ferraz et al.. findings: students in municipal schools perform significantly worse on standardized tests in municipalities where was found corruption in education.

Área do Conhecimento