Detecting Corruption: Evidence from a World Bank project in Kenya

Abstract

Corruption is a major problem in development aid, in part because areas with the greatest need for development assistance often have weak governance. In these environments, traditional anti-fraud measures such as audits or criminal enforcement are limited in their effectiveness. Moreover, aid organizations face incentives to downplay bad outcomes for fear of alienating donors, which has led to the suppression of negative findings related to development aid fraud.

In this paper, we develop new statistical tests to uncover strategic data manipulation consistent with fraud, which can help identify falsified data and facilitate monitoring in difficult-to-audit circumstances. We apply this method to a World Bank community driven development project in Kenya. Our statistical tests rely on the fact that human-produced digits and naturally occurring digits have different digit patterns: unmanipulated digits follow the Benford’s Law distribution. We improve upon existing digit analysis techniques by being sensitive to the value of digits reported, which helps distinguish between intent to defraud and error, and by improving statistical power to allow for finer partitioning of the data. We also produce simulations that demonstrate the superiority of our new tests to the standards in the field, and we provide a new R package for conducting our statistical tests.

Our study finds substantial evidence of fraud, validated by qualitative data, a forensic audit conducted by the World Bank, and replication with a separate dataset for external validity. We uncover higher levels of fraud in a Kenyan election year when graft also had political value and in harder to monitor sectors. This methodology also has broad applications to many forms of data beyond those encountered in development aid.