Bunching and Learning in Targeting Poverty: Evidence from Vietnam

Abstract

Manipulation of the eligibility criteria is one reason that could increase the number of non-poor participants in anti-poverty programs in developing countries. Despite ample evidence that households manipulate these criteria, little is known about how such behaviors evolve over time in a long-term program. Using data from Vietnam, I find that, early on in each phase of its National Anti-Poverty Program, about 1-2% of the population (or 8-18% relative to the program size) bunch at the official income cutoff in order to appear eligible. However, this fraction falls by 60-100% towards the end of the phase, only to increase yet again when a new phase starts with a new income cutoff. To explain this temporal pattern of bunching, I develop a model in which over time the program staff learn to rely on housing conditions, a less-manipulable criteria, to select households. This refined information, in turns, discourages households from manipulating their income. I find that an increase of 0.5 standard deviation in the housing quality index further reduces the chance of being accepted to the program by 25.11% after two years. Meanwhile, other criteria, including reported income and asset holdings, do not contribute any additional predictive power to the program status over the same period. Without this learning process, the program would have misallocated about 1.7%, or equivalently 32.3-36.4 million USD (PPP), of its budget to non-poor households during the first phase of the program.

Jade Ngoc Nguyen

Senior Consultant

I’m a PhD economist who loves applying causal inference and data analytics to solve practical problems.