Few ideas have spread through development policy as quickly as the nudge. A text-message reminder costs a fraction of a cent. A redesigned default costs little beyond the meeting where it is agreed. When a randomized trial shows that either one shifts savings, vaccination, or school enrollment, the implied cost-effectiveness looks extraordinary. Funders, governments, and dedicated behavioral units have responded by building portfolios around these tools. Some of that enthusiasm is earned.

The appeal goes beyond cost. Behavioral economics replaced a thin model of poor households as either uninformed or simply short of cash. Present bias, limited attention, and the cognitive load of scarcity offered a richer account of why people with clear interests in immunizing children or buying fertilizer so often fail to act. That account generated testable designs, and the experimental movement in development economics tested many of them.

Fifteen years of accumulated trials now allow a more disciplined question than whether nudges work. We need to ask which mechanisms matter, under what conditions, at what scale, and compared with what other use of the same budget and political attention. Answering that requires separating three things the policy conversation routinely blurs. Some effects are real and economically meaningful. Some are statistically detectable but trivial for welfare. Others shrink or vanish once publication bias and scaling are taken seriously.

The clearest case for commitment devices comes from the SEED experiment in Mindanao, the Philippines. Nava Ashraf, Dean Karlan, and Wesley Yin offered bank clients a savings account that restricted withdrawals until a chosen date or target amount. Only about 28 percent of those offered the product took it up. Yet after twelve months, average balances in the treatment group had risen roughly 81 percent relative to control. Because that is an intent-to-treat estimate, the effect on actual adopters was considerably larger.

Agriculture produced an equally instructive result. In western Kenya, Esther Duflo, Michael Kremer, and Jonathan Robinson observed that farmers who said they planned to use fertilizer often had no money left by planting season. Offering a small, time-limited benefit, free delivery, immediately after harvest, when cash was on hand, raised fertilizer use substantially. The effects were comparable to much larger subsidies offered later in the season. The design followed directly from a model of present-biased farmers procrastinating on a profitable investment.

Health behaviors show similar patterns. In rural Udaipur, Banerjee, Duflo, and colleagues found that reliable monthly immunization camps raised full immunization from about 6 percent to 18 percent. Adding a kilogram of lentils per visit and a set of plates on completion raised it to 39 percent. It also lowered cost per fully immunized child, because fixed nurse time was spread across more children. In Kenya, weekly SMS reminders improved antiretroviral adherence. Daily reminders, perhaps easier to ignore, did not.

Education and savings interventions also produced effects at strikingly low cost. Robert Jensen told Dominican boys the measured returns to secondary schooling, which they had substantially underestimated. This increased completed schooling at almost no marginal expense. Karlan, McConnell, Mullainathan, and Zinman found that monthly text reminders raised savings balances by about 6 percent across several countries. When the message referenced the client's specific savings goal, the increase was roughly 16 percent. Salience, not new information about the account, appeared to drive the difference.

The common thread is diagnostic. In each successful case, the target behavior was one people already endorsed. The intervention addressed a specific gap in timing, attention, or self-control, and it aligned the decision moment with available resources. None of these programs persuaded anyone to want something new. They closed the distance between stated intention and action, and the experimental evidence measured that distance precisely enough to design around it.

Takeaway

The behavioral interventions that work best do not change what people want. They close the gap between an intention someone already holds and the moment they are able to act on it.

The first limitation concerns what we read. Stefano DellaVigna and Elizabeth Linos compared 126 trials run by two US nudge units with the published academic literature. Academic papers reported average effects of about 8.7 percentage points. The nudge-unit trials, which were larger and reported regardless of outcome, averaged around 1.4. Most of the gap reflected selective publication and small samples. These were not development trials, but the lesson transfers: the published behavioral literature is a filtered sample, and the filter favors large effects.

The second concerns take-up and selection. Commitment devices help only those who adopt them, and adoption is typically low. SEED's 28 percent was considered a success. More troubling, later work in the Philippines by Anett John found that many clients who chose commitment savings contracts subsequently defaulted and paid penalties, leaving them worse off. Partially sophisticated individuals know they have self-control problems but underestimate how severe those problems are. The device designed to protect them then becomes a source of loss.

The third concerns magnitude. A 6 percent increase in balances that start at a few dollars is precisely estimated, yet it does essentially nothing for consumption smoothing, asset accumulation, or vulnerability to shocks. Reminder effects also frequently decay once messages stop, and few studies follow participants long enough to show whether habits form. Statistical significance against a zero null is the wrong benchmark. The relevant question is whether the change comes close to what a household needs to escape the constraint it actually faces.

The fourth concerns behavioral predictions that simply failed. A prominent hypothesis held that charging for insecticide-treated bed nets would increase usage through sunk-cost or screening effects. Jessica Cohen and Pascaline Dupas tested it in Kenya and found no such effect. Those who paid used nets no more than those who received them free. Meanwhile, demand fell by roughly 60 percent when the price rose from zero to about 60 US cents. Liquidity and price, not psychology, were the binding constraints.

Finally, scaling compresses effects. Pilot results reflect motivated implementers, carefully chosen sites, and close supervision. Government delivery introduces message fatigue, weaker targeting, and competing administrative priorities. John List's work on the voltage effect formalizes a pattern that evaluators of behavioral programs have seen repeatedly. None of this means nudges fail. It means the expected effect at scale is usually smaller than the headline estimate, and cost-effectiveness calculations built on pilot coefficients overstate the case.

Takeaway

A statistically significant effect is not the same as a meaningful one. The right benchmark is not zero but the change a poor household would need to escape the constraint it actually faces.

Reread the successes and a pattern emerges: each depended on a structural component already in place. In Udaipur, the rise from 6 to 18 percent full immunization came from solving supply. Nurses reliably showed up, in a setting where health workers were often absent. The lentils worked on top of that fix. Without a functioning camp, no incentive could have moved immunization. Likewise, the Kenyan fertilizer nudge worked because fertilizer was available, profitable on those plots, and affordable with harvest income.

There is also reason to think some behavioral biases are partly caused by poverty itself. Mani, Mullainathan, Shafir, and Zhao found that the same Indian sugarcane farmers performed worse on cognitive tests before harvest, when money was tight, than after it. If scarcity taxes attention and self-control, relieving the underlying constraint may reduce the very biases that nudges are designed to work around. Treating the bias as a fixed trait risks addressing the symptom while leaving the cause untouched.

The strongest evidence for integration comes from bundled programs. The multi-country graduation evaluation by Banerjee and colleagues combined asset transfers, consumption support, training, savings, and regular coaching, and it produced durable gains in consumption and assets across six countries. Coaching and savings are behavioral components embedded in a structural transfer. In Morocco, Benhassine and coauthors showed that a small unconditional cash transfer labeled for education raised school participation substantially. A behavioral frame attached to real resources achieved much of what formal conditions would have.

For program design, the implication is sequential. First diagnose the binding constraint: price, supply, credit, information, or self-control. Then deploy behavioral tools at the last mile, such as take-up of existing entitlements, timing of payments, and adherence to treatment already provided. For evaluation, this argues for factorial designs that cross structural and behavioral arms, so that complementarities are estimated rather than assumed. Outcomes should also be measured long enough to show whether welfare actually changes.

There is a political risk worth naming. Because nudges are cheap and visible, they can be offered as substitutes for funding the clinics, roads, and transfers that address structural constraints. Cost-effectiveness ratios calculated on tiny cost figures make this substitution look efficient. But an intervention that multiplies existing capacity cannot replace that capacity. A portfolio that shifts budgets toward reminders while infrastructure deteriorates may report impressive ratios while delivering little poverty reduction.

Takeaway

Ask what the nudge is attached to. A behavioral intervention multiplies whatever structural capacity sits beneath it, and a multiplier applied to zero still yields zero.

The behavioral evidence in development is neither a revolution nor a fad. A meaningful set of rigorous trials shows commitment devices, well-timed offers, and small incentives producing real changes in savings, input use, and immunization, often at remarkably low cost. That record deserves to be taken seriously.

Yet those successes share conditions that are easy to overlook. The behavior was already desired, the friction was specific, and a functioning structural base existed beneath the nudge. Where those conditions were absent, effects were small, faded, failed to replicate, or were overwhelmed by price and liquidity constraints. Publication bias and scaling losses further shrink the expected return.

The productive stance for designers and evaluators is to treat behavioral tools as complements whose value depends on what they are attached to. Diagnose the binding constraint, invest in the structural fix, and use behavioral design to ensure that capacity is actually used. Then test the combination, at scale, against outcomes that matter for poverty.