Chinese vs. World Bank Development Projects: Insights from Earth Observation and Computer Vision on Wealth Gains in Africa, 2002-2013

Leverages computer vision on satellite imagery to impute wealth outcomes and control for confounding variables, enabling a continent-scale evaluation of development programs across 9,899 neighborhoods in 36 African countries and covering about 88 percent of the population with 6.7 km International Wealth Index grids.

What does the paper study?

This paper asks whether Chinese and World Bank development projects improve neighborhood-level living conditions in Africa between 2002 and 2013. It provides a continent-wide, sector-specific comparison across 9,899 neighborhoods in 36 countries, representing about 88 percent of the population.

What data power the analysis?

The outcome is the International Wealth Index (IWI), imputed from satellite imagery and DHS surveys, yielding a 1990-2020 panel at 6.7 km resolution. Treatment data come from AidData and cover Chinese and World Bank projects during 2002-2013.

Multi-sector projects are counted in each relevant sector.

How are impacts estimated?

The study combines pre-treatment daytime satellite imagery with tabular covariates to estimate treatment propensities and then applies inverse-probability weighting to recover sector-specific average treatment effects. Outcomes are measured one 3-year period after project commitment to capture near-term changes in IWI.

What are the headline results?

Across sectors, both donors are associated with positive average wealth gains, while China shows larger and more consistent effects. Image-augmented models typically produce smaller effects than tabular-only models, indicating stronger confounding control when imagery is included.

World Bank — Sector extremes (median ATEs)

Largest: Trade and Tourism (330) at +12.29 IWI points. Smallest: Government and Civil Society (150) at -0.16 IWI points.

China — Sector extremes (median ATEs)

Largest: Emergency Response (700) at +15.15 IWI points. Smallest: Agriculture, Forestry and Fishing (310) at +1.21 IWI points.

What does the paper find about project placement?

Assignment-mechanism analyses show that project placement is systematically predictable from satellite imagery, and World Bank placement is often more predictable than Chinese placement. The cross-sector canonical correlation of placement-salience profiles is 0.182, indicating limited alignment in how the two donors target sectors once geography is accounted for.

How are the findings stress-tested?

Robustness checks include unit fixed effects at a spatial resolution about 67 times finer than prior fixed-effects analyses, adding lagged nightlights, changing the outcome grid scale, and tightening treatment definitions to precise locations. Baseline and strict-precision sector ATEs remain positively correlated at 0.61.

Which figures should you look at first?

Start with the sector ATE figure to see the full distribution of effects and where Trade and Tourism (330) and Emergency Response (700) sit within it. Then check the placement-predictability figure to compare how well imagery explains donor targeting.

• Sector-specific ATEs for World Bank and China, highlighting extremes by sector.
• Out-of-sample placement predictability across sectors and donors.
• Visualization of the 6.7 km neighborhood grid and IWI measurement.

Frequently asked questions

What is the unit of analysis?

Each unit is a 6.7 km by 6.7 km neighborhood grid cell anchored on DHS cluster locations. The outcome is IWI measured one 3-year period after project commitment.

How large is the sample?

The analysis spans 9,899 neighborhoods in 36 African countries, representing about 88 percent of the population. The study period is 2002-2013.

How many projects and sectors are covered?

China includes 18 sectors with 8,557 projects across 6,798 locations, while the World Bank includes 13 sectors with 51,303 projects across 38,135 locations. Multi-sector projects are counted in each relevant sector.

What is the outcome data source?

The International Wealth Index is imputed from satellite imagery and DHS surveys, producing a 1990-2020 panel. The index is measured at 6.7 km resolution.

Figures

References

@article{daoud2026chinese,
  title={Chinese vs. World Bank Development Projects: Insights from Earth Observation and Computer Vision on Wealth Gains in Africa, 2002-2013},
  author={Daoud, Adel and Cindy Conlin and Connor T. Jerzak},
  journal={World Development},
  year={2026},
  volume={202},
  pages={107328},
  publisher={}
}

Related Work

@article{pettersson2026debiasing,
  title={Debiasing Machine Learning Predictions for Causal Inference Without Additional Ground Truth Data: 'One Map, Many Trials' in Satellite-Driven Poverty Analysis},
  author={Pettersson, Markus B. and Connor T. Jerzak and Adel Daoud},
  journal={Proceedings of the AAAI Conference on Artificial Intelligence (AAAI-2026), Special Track on AI for Social Impact},
  year={2026},
  volume={40},
  pages={39106-39115},
  publisher={}
}

–

@article{daoud2026+planetary,
  title={Planetary Causal Inference: Understanding the Environment, Society, and Economy through Earth Observation and AI Systems},
  author={Daoud, Adel and Connor T. Jerzak},
  journal={A Book Project},
  year={2026+},
  volume={},
  pages={},
  publisher={Under contract with Cambridge University Press}
}

–

@article{fucheng2025optimizing,
  title={Optimizing Multi-Scale Representations to Detect Effect Heterogeneity Using Earth Observation and Computer Vision: Applications to Two Anti-Poverty RCTs},
  author={Fucheng, Warren Zhu and Connor T. Jerzak and Adel Daoud},
  journal={Proceedings of the Fourth Conference on Causal Learning and Reasoning (CLeaR), Proceedings of Machine Learning Research (PMLR)},
  year={2025},
  volume={},
  pages={},
  publisher={}
}

–

@article{jerzak2023image,
  title={Image-based Treatment Effect Heterogeneity},
  author={Jerzak, Connor T. and Fredrik Johansson and Adel Daoud},
  journal={Proceedings of the Second Conference on Causal Learning and Reasoning (CLeaR), Proceedings of Machine Learning Research (PMLR)},
  year={2023},
  volume={213},
  pages={531-552},
  publisher={}
}

@article{nicolas audinet de pieuchon 2025benchmarking,
  title={Benchmarking Debiasing Methods for LLM-based Parameter Estimates},
  author={Nicolas Audinet de Pieuchon and Adel Daoud and Connor T. Jerzak and Moa Johansson and Richard Johansson},
  journal={Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year={2025},
  volume={},
  pages={},
  publisher={},
  doi={}
}

–

@article{sakamoto2026scoping,
  title={A Scoping Review of Earth Observation and Machine Learning for Causal Inference: Implications for the Geography of Poverty},
  author={Sakamoto, Kazuki and Connor T. Jerzak and Adel Daoud},
  journal={Hall, Ola and Ibrahim Wahab (eds.), Geography of Poverty},
  year={2026},
  volume={},
  pages={},
  publisher={Edward Elgar Publishing (Cheltenham, UK)}
}