Cash transfer programs (CTPs) have spread in the last decade to help fight extreme poverty in different parts of the world. However, these programs face challenges related to both exclusion errors, where households below the poverty line do not receive the intended support, and inclusion errors, where cash reaches households that do not require assistance. A key issue here is to ensure that the cash is distributed to the targeted beneficiaries in an efficient and egalitarian manner.

Big data and machine learning have been used recently by several CTPs to target the right beneficiaries. We demonstrate how these targeting methods can be integrated into the cash allocation problem to synthesize the impact of targeting errors on the design of the allocation rules. We design allocation rules to minimize the squared gap of the shortfall between the consumption level and the poverty line. Considering the distribution of targeting errors, we develop three classes of allocation rules: predictive, stochastic, and robust allocation.

We show that a simple predictive allocation rule is already optimal when the targeting errors are “well calibrated”, whereas the robust allocation rule is preferred when targeting accuracy is compromised. We validate this finding using real data from a CTP in Malawi. Furthermore, we show that the robust allocation model can be suitably modified to ensure equitable treatment for beneficiaries near the poverty line, reducing the undesired bunching effect. As a result, the refined robust policy leads to a “smoother” wealth distribution and pushes more beneficiaries above the designated poverty line.