Target at the right level: aid, spillovers, and growth in sub-saharan Africa

ABSTRACT

Previous aid effectiveness literature is subject to aggregation bias and does not discuss aid spillover effects. Using spatial analysis and data from geocoded World Bank aid projects, this article investigates international aid effectiveness and aid spillovers at the sub-national level in 3,764 second-order administrative divisions (ADM2) in 48 countries in Sub-Saharan Africa over the period of 1995–2014. The empirical analysis is based on an instrumental variable approach and relies on nightlights data as a proxy for economic activity. The empirical results reveal three previously undocumented findings on aid effectiveness. First, we find that aid at the local level (ADM2) promotes economic growth at an economically and statistically significant level. Second, we uncover significantly positive aid spillovers across adjacent localities (ADM2). Third, aid flows at more aggregate levels (ADM1 and country level) have the opposite effect and reduce economic growth. The net effect of all aid variables is near zero and is within the range of coefficient estimates reported at the country level by previous papers. These results suggest that targeted aid projects can be effective in promoting economic growth.

KEYWORDS:

• Aid effectiveness
• economic growth
• growth spillovers
• nightlights
• sub-saharan Africa

JEL CLASSIFICATION:

• O1
• O4
• O55
• F35

Acknowledgments

We thank Niels-Hugo Blunch, Gregory Burge, Todd Fagin, Pallab Ghosh, John Harris, Daniel Hicks, Shaomeng Jia and Thomas Neeson; seminar participants at the University of Oklahoma in 2016, and Marshall University in 2018; and session participants at the 2016 SEA and the 2017 EEA conferences for their comments and suggestions on earlier drafts of this paper. Firat Demir is thankful for the Carnegie Corporation of New York Grant (G-20-57642, G-22-59079), the Fulbright Commission, and the Vilnius University Faculty of Economics and Business Administration for his Fulbright visit in 2022. The views and any remaining errors are ours.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

¹ Askarov and Doucouliagos (2015b) is the only paper we are aware of that examines aid spillovers. In their country-level macro analysis, they report a positive growth effect in aid-recipient countries but a negative spillover effect in others. Instead, we focus on micro-level spillovers here.

² Note that unlike Dreher and Lohmann (2015), we use a dynamic growth model here, which controls for path dependency and convergence dynamics.

³ The period averages are over 1995–1998, 1999–2002, 2003–2006, 2007–2010, and 2011–2013 where 1995 is the first year for aid projects in the dataset. We also use nightlight data for the period of 1991–1994 to gain an additional period as we use the lagged values of aid in Equation (1)(1) $\mathit{Growt}{h}_{\mathit{ijt}}={\alpha }_0+{\alpha }_1\ast \mathit{Ligh}{t}_{\mathit{ijt}-1}+{\beta }_1\ast \mathit{Ai}{d}_{ijt-1}^{ADM2}+{\beta }_2\ast \mathit{Ai}{d}_{-ijt-1}^{ADM2}+{\beta }_3\ast \mathit{Ai}{d}_{ijt-1}^{ADM1}+{\beta }_4\ast \mathit{Ai}{d}_{ijt-1}^{Co\mathrm{untry}}+\mathit{\gamma }{ }^{\mathrm{\prime }}{X}_{\mathit{ijt}-1}+{\delta }_i+{\delta }_t+{\epsilon }_{\mathit{i},\mathit{t}}$(1) . Using four-year averages is common in the aid effectiveness and growth literature and allows examining medium- and long-run effects (Burnside and Dollar 2000; Collier and Dollar 2002; Clemens et al., 2012).

⁴ The aid variables at the ADM2 level are with precision levels 1–3. Aid at the ADM1 level is with precision level 4, and aid at the country level is with precision levels 5–8. More details are in the Appendix. For robustness, however, we also experimented with varying aggregation levels, and found similar results (Tables 3 and 5, column 8).

⁵ Unlike Brückner (2013), we did not use the international commodity prices as an IV as they do not vary across ADM2s (or even countries) and therefore are absorbed by the ADM2 and time fixed effects.

⁶ The intuition is that aid has an endogenous and an exogenous part. We would like to capture the endogenous part in Equation (2)(2) $\mathit{Ai}{d}_{\mathit{ijt}-1}={\tau }_1\ast \mathit{Growt}{h}_{\mathit{ijt}}+{\delta }_i+{\delta }_t+{\text{isin}}_{\mathit{ijt}}$(2) , and remove it in Equation (3)(3) $\mathit{Ai}{d}_{ijt-1}^{\ast }=\mathit{Ai}{d}_{\mathit{ijt}-1}-{\tau }_1\ast \mathit{Growt}{h}_{\mathit{ijt}}$(3) . The residual is then assumed to be exogenous and can be used as an IV for aid. For more details, see Brückner (2013).

⁷ Growth rates based on income and nightlights can differ as the income elasticity of lights may be different than one, and the light-output ratio may change overtime. Also, nightlights are measured by different satellites in different years, effecting sensor quality and mechanics. In addition, sensor sensitivity is likely to diminish by age. Cloud cover, humidity and other weather conditions can also affect light diffusion. However, by using the growth rate of nightlight density, we difference out the location-specific fixed effects. Any remaining time, satellite or location specific factors, including increases in electric connectivity, are controlled by the use of ADM2 fixed effects and time fixed effects. The residual part is then treated as a measurement error. To address some of these issues, we also conducted additional robustness tests by limiting the sample to a subgroup based on the initial level of lights. For an extensive discussion of using nightlight data as a measure of income and growth, see Henderson, Storeygard, and Weil (2012) and Donaldson and Storeygard (2016).

⁸ For details of the dataset, see Tierney et al. (2011). The data does not separate different types of aid such as humanitarian or infrastructure. For a similar approach, see Briggs (2018).

⁹ While we know that increasing temperatures tend to increase crop yields, the effects of precipitation are not clear. First, because many sub-Saharan African countries have rainforest climate, where the nutritious soil is accumulated on the top; increasing precipitation can wash the nutrients away and reduce the productivity of the soil (Sachs 2001). Second, increasing moisture, cloud cover or spreading of pests can lower crop yields and therefore growth rates, especially in countries that lack the necessary infrastructure to deal with these problems.

¹⁰ We exclude observations with extreme (the top 5% and bottom 5%, also the top 10% and the bottom 10% as an alternative) standard deviations (also standard deviations relative to the mean as an alternative) of yearly air temperature and precipitation as robustness tests. Excluding air temperature and precipitation fluctuation outliers does not change any of our results.

¹¹ We also experimented with alternative clusters for the standard errors, including no clustering, ADM1, country, ADM2-time, ADM1-time, and country-time. The results are reported in Appendix Table A23 and are consistent with our benchmark findings.

¹² We should note that the growth effects of aid are likely to be heterogenous across different sectors. For example, aid flows to infrastructure projects are likely to have different short-, medium-, and long-run effects than those targeting education, for example. However, because the same aid project frequently involves multiple sectors and that we lack data on its sub-sectoral distribution, we are unable to test the local growth effects at the sectoral aid level.

¹³ However, due to data limitations, we are unable to empirically test the exact mechanisms, through which aid affects growth at the local level.

¹⁴ The mean of (log) aid variables, $\mathit{Ai}{d}_{ijt-1}^{ADM2}$,$\mathit{Ai}{d}_{-ijt-1}^{ADM2}$, $\mathit{Ai}{d}_{ijt-1}^{ADM1}$, $\mathit{Ai}{d}_{ijt-1}^{co\mathrm{untry}},$ are 1.448, 3.323, 2.687 and 1.612 during 1995–1998 but are 2.938, 5.436, 5.131 and 2.676 during 1999–2002, respectively.

¹⁵ In the Appendix, we repeat all regressions including the ICRG variable and find similar results.

¹⁶ To preserve sample size, we do not include ICRG or PolityIV in the baseline regression. Also, both variables in the sample are very stable and are mostly absorbed by country fix effects.

¹⁷ The 48 countries are: Angola, Benin, Botswana, Burkina Faso, Burundi, Cameroon, Central African Republic, Chad, Cape Verde, Comoros, Congo, Rep, Congo (Dem. Rep.), Côte d’Ivoire, Eritrea, Equatorial Guinea, Ethiopia, Gabon, Gambia, Ghana, Guinea, Guinea-Bissau, Kenya, Lesotho, Liberia, Madagascar, Malawi, Mali, Mauritania, Mauritius, Mozambique, Namibia, Niger, Nigeria, Rwanda, São Tomé and Principe, Senegal, Seychelles, Sierra Leone, Somalia, South Africa, South Sudan, Sudan, Swaziland, Tanzania, Togo, Uganda, Zambia, Zimbabwe..

¹⁸ See, for example, ‘Night Lights and ArcGIS: A Brief Guide’, https://darrylmcleod.com/wp-content/uploads/2016/06/Night-Lights-and-ArcGIS-A-Brief-Guide.pdf.

¹⁹ Otherwise no value would be returned if the polygon does not touch the centre of any raster cells.