Governments running social assistance often have to identify poor households without reliable information on their incomes. Earnings fluctuate, many people work informally, and administrative records only partially capture household economic circumstances. To estimate welfare despite these gaps, poverty targeting has moved from simple household classifications towards statistical models that rank households by estimated welfare.
Machine learning now enables analysis of larger, more varied datasets.
These advances could reduce targeting errors and administrative costs, especially where income records are thin. However, predictive accuracy is not a good measure of whether targeting has improved, for two reasons. It does not fully account for households missing from the data in the first place, and it does not capture how households respond to the criteria used to determine eligibility.
First, consider the data gaps problem. For decades, developing countries without reliable income or consumption records have used the proxy means test, which estimates welfare from observable household characteristics such as housing, education, occupation and assets. It is cheap and practical, but like any model built on imperfect proxies, it produces both inclusion and exclusion errors. Machine learning can find more complex patterns in the same survey data and draw on sources that surveys cannot reach, including satellite imagery and mobile phone records.
During the pandemic, Togo used machine learning to target emergency cash transfers under its Novissi program. Satellite imagery helped identify poorer cantons, while mobile phone records were used to predict welfare among subscribers. In the rural program, the approach identified 47% of poor individuals, compared with 33% under geographic targeting, reducing exclusion errors by 21% in relative terms. The improvement was greater in rural areas, where geographic targeting was less effective at distinguishing households within already-poor communities. However, the program required access to a mobile phone and a valid subscriber identity module card, putting the 15% of individuals living in households without a phone at risk of exclusion, regardless of the model’s accuracy.
The evidence from Bangladesh is more mixed. In a 2025 working paper, Emily Aiken and four co-authors found that mobile phone-based targeting identified 32% of the poorest households, compared with 52% using a proxy means test. Both approaches used machine learning, but household survey characteristics predicted consumption better than phone activity. Phone-based targeting was cheaper, however, and could be more cost-effective when screening large populations.
These findings illustrate the limitations of poverty targeting, including methods using machine learning. Someone in regular salaried work leaves a fairly consistent record through wages, taxes or social insurance contributions. Someone selling food from home, working seasonally on a farm at harvest or helping in a family business without a wage may leave almost no record. Machine learning can make better use of available information, but households with limited or missing data may still be excluded. Average predictive accuracy can rise while the same households keep being missed.
The second problem arises even when the data are good enough. The way governments assess eligibility can itself influence behaviour.
A review of seven randomised evaluations in six countries found no systematic evidence that cash transfers discourage adults from working. But overall employment figures can hide changes in labour-market decisions, particularly movements between formal and informal work.
In my Labour Economics paper, I examined Indonesia’s large-scale unconditional cash transfers, introduced in 2005, 2008 and 2013 to compensate households for fuel subsidy cuts. In the first year following the later programs, employment declined by about 2% relative to its average, while formal employment fell by around 8%. More revealing, recipients were less likely to move into formal employment, with no significant increase in those leaving it.
These effects were modest and short-lived. The findings suggest that labour-market adjustments occurred primarily through reduced entry into formal employment, rather than widespread withdrawal from work or exits from existing formal jobs.
The likely reason is eligibility. Indonesia’s transfers were unconditional and temporary, but later rounds ranked households using employment status, occupation and other characteristics. Although benefits did not decline with earnings, households near the eligibility threshold may have feared that formal employment, which makes income more visible to the government, would jeopardise future assistance. A similar pattern emerges in evidence from Uruguay, where formal employment fell because eligibility was explicitly tied to registered earnings.
That raises a question for poverty targeting using machine learning. The programs I studied relied on observable household characteristics, although beneficiaries may not have known precisely how these affected their eligibility.
Machine-learning models can incorporate more variables and complex relationships, potentially making eligibility decisions even harder to anticipate. Could this uncertainty reduce behavioural distortions by making it harder for households to adjust their decisions around eligibility criteria? Or could it encourage more cautious behaviour (and greater distortions) if households avoid economic opportunities for fear of losing future assistance? Little is known about how algorithmic eligibility decisions influence household behaviour.
Machine learning could make eligibility reassessments cheaper and more frequent by incorporating updated employment and income records. If households expect changes in their economic circumstances to affect future assistance, they may adjust their decisions to preserve eligibility, creating further behavioural distortions.
The Togo and Bangladesh studies show that machine learning can improve poverty prediction, but its performance depends on the data available, while households hardest to observe may also be more likely to be poor. The Indonesian and Uruguayan evidence shows that eligibility rules can influence household behaviour, although the effects of less transparent algorithmic decisions remain uncertain. Hence, a higher accuracy rate does not, by itself, constitute better social protection. If targeting becomes more accurate on average while the same households continue to be missed, it has not necessarily become fairer.
