Does foreign
aid work?

Foreign aid has been arguing with itself for thirty years. One camp (Jeffrey Sachs, the UN Millennium Development Goals apparatus) says a large enough transfer can break the poverty trap and pay for itself in growth. Another (William Easterly, Dambisa Moyo's Dead Aid) says aid props up bad governments, crowds out local institutions, and may even slow the growth it's meant to speed up. This run builds a World Bank panel — Net ODA received as a share of GNI, matched to GDP growth, 4,747 country-years across 162 aid-recipient countries, 1990–2023 — and tests the claim five ways, each spec built to close a gap the previous one left open.

The naive, same-year comparison looks like it settles the fight against aid. Pooled across every country-year, with standard errors clustered by country: every additional point of ODA/GNI tracks with -0.044 points less GDP growth that same year, 95% CI [-0.082, -0.006], p=0.024 — excludes zero. But the two countries doing the most work in that number are wars, not aid failures: Iraq in 1991 received ODA equal to 137.6% of its GNI the same year its economy shrank 64.0% (the Gulf War); Rwanda in 1994 received 94.9% of GNI in aid the same year its economy collapsed 50.2% (the genocide). Aid arrives because a country is in crisis — a same-year correlation cannot tell a dollar that responded to a catastrophe from a dollar that caused one.

Drop the literal war/collapse years and the same-year gap survives, but it's cut in half and it's barely there. Excluding the 97 country-years with GDP growth beyond ±20% (the crisis outliers, not a cherry-picked cut): -0.023 points, 95% CI [-0.046, -0.001], p=0.042 — still technically excludes zero, but the upper bound of that interval sits at -0.0008, a hair's width from it. Shift to a one-year lag — this year's aid against next year's growth, with year fixed effects, removing the same-year crisis-response channel entirely — and the sign flips positive: +0.039, 95% CI [-0.005, +0.082], p=0.081 — contains zero. Whatever the same-year number was picking up, it does not survive removing the possibility that aid is reacting to, not producing, the growth number next to it.

The design the actual academic literature on aid effectiveness uses finds nothing either. Burnside & Dollar (2000) and, more skeptically, Rajan & Subramanian (2008) test aid's effect on growth by measuring aid in an early window and growth in a later, non-overlapping one — exactly what a same-year panel can't do. Rebuilt here: mean ODA/GNI 1990–2005 against mean GDP growth 2006–2023, one row per country, 140 countries with sufficient coverage in both windows. Naive: +0.020 points of later growth per point of earlier aid, 95% CI [-0.033, +0.073], p=0.466 — contains zero. Add each country's own 1990 income as a control (the standard convergence term in growth regressions — poorer countries mechanically tend to grow faster, which could confound a raw aid comparison since aid also flows disproportionately to poorer countries): the aid slope flips sign again, -0.021, CI [-0.065, +0.023], p=0.343 — still contains zero. The control itself is doing real, expected work — initial income alone predicts subsequent growth at -0.785 points per log-dollar, CI [-1.058, -0.512], p=1.7e-08, a textbook convergence effect — so this is a regression finding a real signal where one exists and still finding nothing for aid.

Named, concretely: neither the heaviest aid recipients nor the lightest stand out. São Tomé and Príncipe took in the most aid in the early window, 45.9% of GNI on average, and grew at 3.37%/year afterward — almost exactly the 140-country sample's own middle. Palau, the second-heaviest recipient (37.6%), shrank −0.83%/year. Meanwhile Brazil and Libya, which received essentially no aid at all in the early window (0.03% and 0.02% of GNI), grew at 2.07% and 2.02% — unremarkable, not obviously worse or better for having skipped the aid the theory says should have helped them. A scatter of all 140 countries (right) shows why: there is no visible slope in either direction.

Read plainly: this desk can confirm neither side of the aid debate. The one spec that clears zero — the same-year comparison — is also the one most contaminated by aid flowing toward crises, and it stops clearing zero comfortably the moment the crisis years are removed or the timing is lagged by a single year. The design built specifically to avoid that contamination, cross-country and cross-time, finds an interval containing zero whether or not initial income is controlled. That is not proof aid does nothing for any single country in any single program — it means thirty-four years of the World Bank's own aid and growth data, at this level of aggregation, cannot distinguish a real effect from none.