
You know that optical illusion where a ball appears to roll uphill, but the whole table is tilted? That's what conventional two-way fixed effects (TWFE) does to aid evaluation. Aid projects land in poorer areas. Those areas were already on a different trajectory. TWFE compares them to richer clusters that never got projects and calls the convergence a 'treatment effect.' Tilt the table flat — use an estimator that accounts for staggered treatment timing — and much of the uphill roll disappears. The committed claim: when you replace TWFE with the de Chaisemartin and D'Haultfoeuille (dCdH) switcher-stayer estimator across 2,166 DHS clusters in 35 African countries from 2002 to 2013, most donor-sector combinations show no statistically reliable wealth improvement. Neither the World Bank nor China uniformly improves local material living standards. Effects are sector-specific and fragile. The paper's architecture is clean causal inference, not prediction. The authors build a balanced panel linking geocoded AidData project locations to satellite-imputed International Wealth Index (IWI) estimates — a household-centered measure that doesn't privilege infrastructure the way nighttime lights do. They run the standard TWFE event-study, then rerun with dCdH, which avoids 'contaminated comparisons' where already-treated units serve as controls for newly-treated ones. The delta between the two estimators is the paper's real finding. The ladder is where this gets interesting. Under TWFE — the method most prior subnational aid studies use — Chinese Energy Generation and Supply looks strongly positive. Under dCdH, it falls to approximately zero. That's not a small discrepancy; it's the difference between 'aid works' and 'aid placement correlates with pre-existing trends.' For the World Bank, Health survives the methodological upgrade with positive evidence. Education shows positive but noisier gains. For China, Water Supply and Sanitation and Other Social Infrastructure and Services retain positive associations, but with residual selection concerns the authors flag honestly. The integrity picture is mixed but creditable. The authors use publicly available DHS data and AidData geocoding — both community resources, not proprietary datasets. They run pre-treatment diagnostics showing that project placement is selective, which is the right move: you should show your own method's vulnerability before claiming results. The key weakness is that this remains observational. No randomization, no instrumental variable. The dCdH estimator handles staggered timing better, but it cannot eliminate all selection bias. The authors acknowledge this. The absence of code availability and pre-registration is standard for the field but still a gap. The successor question is obvious: run this with individual-level DHS microdata instead of cluster-level IWI, and extend the panel past 2013. The AidData geocoding now covers projects through the late 2010s. My read: the authors stopped at 2013 because DHS survey waves thin out and the balanced panel requirement becomes binding. They likely tried extending and couldn't maintain balance. A second obvious experiment — applying the same dCdH framework to DAC bilateral donors (USAID, DFID, GIZ) — would either generalize the finding or isolate something specific about World Bank and Chinese project selection. That's probably the next paper. The practical takeaway for anyone who reads aid-effectiveness literature: when a study uses TWFE on geocoded projects, you should mentally discount the headline effect by 30-70%. The methodological bias runs in one direction — it inflates treatment effects — and this paper gives you the calibration. That's a lens you can carry to every paper in this subfield.
The analogy
You know that optical illusion where a ball appears to roll uphill, but the whole table is tilted? Aid projects target poorer areas. Those areas were converging anyway. TWFE calls the convergence a treatment effect. Tilt the table flat with a better estimator, and the ball stops moving.
Tilted table illusion
A ball appears to roll uphill because the entire surface is secretly sloped — the motion is the table, not the ball
TWFE aid bias
Aid projects target poorer clusters already on convergence trajectories — TWFE attributes the pre-existing trend to the intervention
Honest assessment
This is a methodological correction paper, not a discovery paper. Its value is showing that the standard tool in geocoded aid evaluation (TWFE) systematically inflates treatment effects, and demonstrating the deflation on two major donors across 35 countries. The finding that Chinese Energy Generation goes from strongly positive to ~zero under dCdH is the single most striking result. But the paper doesn't close the question — it opens a methodological lane. The balanced-panel constraint limits the sample period to 2002-2013, and the 'positive' survivor results (WB Health, China WASH) haven't been tested against multiple-comparison corrections. Anyone reading geocoded aid studies should use this paper as a calibration tool: mentally discount TWFE estimates by a large margin.
What Is CLAIMS?
63
/ 100
BedrockNews applies a structured analytical framework called CLAIMS to academic papers and technical research releases.
How CLAIMS worksClaim
3/5
Ladder
4/5
Architecture
4/5
Integrity
3/5
Milestone
2/5
Successor
3/5
What is this paper actually committing to?
Solid but iterative. 'Better estimator deflates treatment effects' is a known template from labor econ (Goodman-Bacon, Sun & Abraham). Applying it to geocoded aid is useful execution, not a new category of result.
The field fight
Does geocoded foreign aid measurably improve local living standards in recipient countries?
Side A
Yes — subnational geocoded studies consistently show positive effects near project sites
Held by: Dreher et al., BenYishay et al., AidData TWFE studies
Side B
Those results are inflated by TWFE bias and nighttime-light proxies that favor infrastructure; better estimators shrink effects toward zero
Held by: This paper, Goldsmith-Pinkham et al. (methods), de Chaisemartin & D'Haultfoeuille
Milestone tracker
Unlocks: Definitive answer on whether dCdH deflation holds at scale or whether longer time horizons reveal delayed positive effects
Current
2,166
clusters · 12yr
this paper
Next
3,000+
clusters · 18yr
panel through 2020
~2-3 years depending on DHS wave releases and AidData 3.0 coverage
Verified Facts (7)
balanced panel of 2,166 DHS clusters across 35 countries from 2002 to 2013
Chinese Energy Generation and Supply appears strongly positive under TWFE but falls close to zero under dCdH
For the World Bank, positive evidence is strongest in Health
clusters that later receive projects often begin from weaker relative positions before treatment onset
switcher-stayer estimator of de Chaisemartin and D'Haultfoeuille (dCdH)
satellite-imputed International Wealth Index estimates, a household-centered measure of material living standards
For China, Water Supply and Sanitation and Other Social Infrastructure and Services show positive associations with local wealth
Practical Implications (4)
If you evaluate aid projects using TWFE event-studies on geocoded data, expect your positive estimates to shrink 30-100% under staggered-treatment estimators — budget for this in study design
Nighttime lights as an outcome measure favors infrastructure-heavy projects; switching to IWI or similar household-centered measures changes which sectors look effective
For donors: sector-specific evaluation is non-negotiable. Donor-wide 'does our aid work?' questions are not answerable with current methods
For policymakers citing Chinese aid effectiveness in energy: that finding appears to be a TWFE artifact, not a real effect
More latest news
Read all