Chinese development assistance and household welfare in Sub-Saharan
Africa
1 Introduction
An “earth-shaking rise” is how the Diplomat, an international online news magazine, described China’s evolution from a poor country to a global superpower within no less than 30 years1. The
international community, however, is torn between admiration for the country’s achievements and criticism for its authoritarian leadership style, human rights violations, and business practices. To some extent, this skepticism also applies to China’s new role as a bilateral donor. Formerly an aid recipient itself, China is now probably the largest emerging donor and a major official creditor worldwide (Horn et al., 2019). China is especially active in Sub-Saharan Africa where it has been acting as a “donor” since the 1950s (Bräutigam, 2009; Morgan and Zheng, 2019), but only brought its financial engagement to recognizable scale after the turn of the millennium. This is when the international community started paying close attention to the country’s activities across the African continent (Kobayashi, 2008; Bräutigam 2009).
Assessing the allocation and effectiveness of Chinese aid is challenging: the Chinese government barely publishes information on its foreign assistance. Yet, data availability has improved since the AidData research lab first issued the Chinese Official Finance Dataset in 2017, giving rise to a new body of quantitative research on Chinese foreign assistance. Various studies focus on the drivers of Chinese aid allocation (see e.g., Strange et al., 2013; Strange et al., 2014; Dreher et al., 2015; Dreher and Fuchs, 2015; Kilama, 2016; Dreher et al., 2019). Another strand of literature explores the effects of Chinese aid on governance outcomes. A study by Isaksson and Kotsadam (2018) finds a positive relationship between Chinese aid and local corruption. Gehring et al. (2018) investigates the relationship between Chinese aid and conflict in Sub-Saharan Africa. They find that, if anything, Chinese aid tends to reduce conflict. Research on the effectiveness of China’s aid on development outcomes is conducted too (see e.g., Dreher et al., 2016; Dreher et al., 2017; Bluhm et al., 2018) but mostly focused on economic outcomes. Dreher et al. (2016) and Dreher et al. (2017) respectively finds a positive effect of Chinese aid on growth in recipient countries, including at the sub-national level. Bluhm et al. (2018) explore the effect of Chinese aid projects on the distribution of subnational economic activity in low- and middle-income countries and finds that infrastructure projects reduce economic inequality between subnational localities.
In view of China’s growing influence as an international donor, especially in Sub-Saharan Africa2, it is important to explore its development impact beyond economic outcomes. In this study,
we investigate the impact of Chinese aid projects on three social development dimensions: education, child health and nutrition. First, and foremost we focus on these specific outcomes due to their
1
https://thediplomat.com/2011/02/understanding-chinas-global-impact/2
Sub-Saharan Africa, which is still the poorest region in the world, is the largest recipient of Chinese foreign aid and a major business partner for China when it comes to infrastructure investments and the exploitation of natural resources (Brautigam, 2009; Sanfilippo, 2010).relevance for economic development and human capital formation in the developing world. An extensive literature documents the importance of education and early childhood development for individuals’ life outcomes (World Bank 2018). Child health and nutrition play a key role for human development (Walker et al. 2011, Page and Pande 2018). Despite measurable international progress in these areas, children in Sub-Saharan Africa are still at a comparative disadvantage. Enrolment in primary schooling is 22% in Sub-Saharan Africa compared to 79% in OECD countries (World Bank 2018). Moreover, the quality of services ensuring children’s healthy development, such as health centers, is often insufficient and needs to be addressed to achieve the Sustainable Development Goals (SDGs) (World Bank 2018). Second, and to the best of our knowledge, quantitative empirical evidence on the effects of Chinese assistance on health and education outcomes is largely missing from the literature3. Third, although China is mainly known for its focus on business and
infrastructure projects in Sub-Saharan Africa, it is very active in education and health as well (Shajalal et al., 2017; Strange et al., 2014; King, 2014; King, 2010; Bräutigam, 2009; Morgan and Zheng, 2019)4. Between 2009 and 2013, China increased its health-related assistance to Sub-Saharan
Africa by 146% (Shajalal et al., 2017). It is assumed to rank among the top 10 bilateral global health donors to Africa (Grépin et al., 2014). China’s engagement in both education and health includes the provision of infrastructure and the transfer of skills and knowledge including, for example, training cooperation and providing foreigners with scholarships to study in China (Shajalal et al., 2017; King, 2010; King, 2014). Moreover, the long-term deployment of Chinese Medical Teams has been a core instrument of Chinese health assistance to Africa (Kadetz and Hood 2017). Even though, and as alluded to earlier, literature on Chinese development assistance has generated a considerable amount of new knowledge about the activities of this emerging donor, the focus has largely been put on economic- or governance related-outcomes. So, to the extent that Chinese foreign assistance leads to improvements in education as well as child health and nutrition, it can make a valuable contribution to closing development gaps between Sub-Saharan Africa and the rest of the world.
A further important aspect of our study, and a reason why we believe that this analysis is important, is to understand if some of the characteristics that make Chinese aid different from traditional donors can lead to different implications for social outcomes. The modus operandi of Chinese development cooperation, tying aid to more commercial purposes and co-locating projects
3
Grépin et al. (2014) is the only study we are aware of that investigates Chinese engagement in the health sector using quantitative data. Based on the same dataset (AidData) that we use for this study, the authors provide a descriptive analysis of 255 health and health-related projects that China allocated to 46 African countries between 2000 and 2012. The study compares Chinese assistance to health-related projects from “traditional” OECD donors. It finds that over half of HPWS projects target infrastructure and 40% target human resource development; providing equipment and drugs is also common. A qualitative study by Kadetz and Hood (2017) investigates Chinese health aid to Madagascar and draws the conclusion that China “failed to improve capacity and local health sector sustainability”. We are not aware of any quantitative empirical studies on the effects of Chinese aid on education and nutrition outcomes.4
An historic reconstruction of the distribution of Chinese projects in Africa made by Morgan and Zheng (2019) shows that the social sector has always been prominent. Taken together, health and education account for 35% of the total number of Chinese projects into Africa since 1955.geographically (Chin and Gallagher, 2019), might be driven by non-altruistic considerations and be subject to a higher degree of political capture by African leaders (see, for instance, Dreher et al., 2019).
Chinese aid projects can influence social outcomes for African citizens both directly and indirectly. First, improvements in education, health and nutrition may be directly affected by the involvement of Chinese foreign assistance in these sectors. In addition, education and health outcomes may be affected by other projects, including those providing large-scale infrastructure (King 2010). Improvements in health can result indirectly from projects with relevant spillover effects such as access to electricity, piped water, road infrastructure and improved household wealth (Kotsadam et al. 2018). Similarly, transportation infrastructure such as roads, which makes up at least 21% of Chinese aid, can facilitate access to education for many. Finally, and to the extent that Chinese development assistance stimulate local economic growth (as found by Dreher et al., 2016), it might be argued that increased wealth can improve households’ ability to cover school or health fees.
Our analysis is based on a sample of 13 countries for which we analyze whether Chinese aid projects increase household welfare in the following areas: average years of education; child health and nutrition. Our main analysis pools all Chinese projects available in the AidData data set to estimate their effect on our outcomes of interest. In addition, we perform a sector level analysis that simultaneously considers the impact of social and economic projects on each outcome variable. This serves to test treatment effect heterogeneity across projects in different sectors and whether social sector projects successfully target education and health outcomes. Last, we explore whether the effects of Chinese aid vary with its scale, i.e. the treatment intensity and according to the age of the projects.
Any study using the AidData set faces two important limitations. First, it must rely on observational data that is prone to selection bias: aid projects are unlikely to be randomly allocated across countries or regions within countries. Second, project data entries in the data set stem from different sources, many of which have not been officially confirmed by the Chinese government (see Section 2.1 for details on the data set). Like previous studies (e.g., Dreher et al. 2016, Dreher et al. 2017, Bluhm et al. 2018), we can only address the first problem. To obtain reliable estimates of the effect of Chinese aid on health and education outcomes, we apply the following identification strategy (see Section 2 for details). We match georeferenced data on Chinese aid projects available from
AidData to georeferenced household data in the Demographic and Health Survey (DHS) at two points
in time: before and after the inflow of Chinese aid - our treatment variable. Then, we use difference-in-differences (DiD) estimations to compare the wellbeing of households located within 25 kilometers of at least one aid project at any point in time between the two DHS survey waves to the wellbeing of control households located outside of this 25 km radius. Instead of comparing individual households, we aggregate all variables at the area level (see Section 2.3), meaning that we compare the same
household areas5 before and after treatment6. We use inverse probability weighting (IPW) to further
reduce selection bias between treatment and control areas. Since data on the number of Chinese projects on the ground as well as on project finance volumes is patchy our main analysis relies on a binary treatment dummy to indicate the presence of a project. However, we also want to consider treatment intensity and therefore extend our main analysis, first, by replacing the treatment dummy with the number of projects and, second, by project finance volume.
Our study contributes to the literature on aid effectiveness as follows. First, it adds to the growing body of cross-country research that uses project-based data to study aid effectiveness at the sub-national level (see Kilby, 2000; Dollar and Svensson, 2000; Denizer et al., 2013; Metzger and Günther, 2015; Briggs, 2018, Kotsadam et al., 2018). Within this field, we add to recent literature that combines georeferenced data with quasi-experimental methods to causally identify the local effects of aid (Kotsadam and Tolonen, 2016; Isaksson and Kotsadam, 2018; Kotsadam et al., 2018).7 Studies of
this type can simultaneously consider the influence of micro- and macro-level factors on development outcomes, thus, bridging the gap between macro- and microeconomic research. Second, our study contributes to a growing body of economic literature on the effectiveness of Chinese aid (see e.g., Dreher et al., 2016; Dreher et al., 2017; Bluhm et al., 2018). Third, we provide evidence on key social sectors, education, health and nutrition, which have rarely been examined by the literature on aid effectiveness (Arndt et al., 2015), but whose improvements are essential to the development process. To the best of our knowledge, this study is the first to quantitatively assess the micro level impact of Chinese project aid on education, child health and nutrition in a larger sample of Sub-Saharan African countries. To our knowledge it is also the first study to investigate the long-term impact of Chinese aid projects on education, health and nutrition. The question of whether Chinese development assistance affects social or economic outcomes in the long term is of major research interest. Yet, the existing literature mainly focuses on short-term impacts using annual data. Corresponding results might be biased by annual fluctuations in outcome variables as well as annual fluctuations in aid. Our analysis mitigates this problem by using “long differences” between DHS waves (on average 12.6 years). This strategy may be particularly relevant to capture changes in social outcomes: indeed, investment in education takes years to bear fruits while changes in health and nutritional outcomes are known to be slow and path dependent.
Our results point to a positive impact of Chinese aid on social development outcomes. First, we find that households in areas receiving Chinese aid projects tend to stay in school longer and experience a reduction in child mortality. However, we do not detect a significant effect on child
5
We construct these areas based on their geographical proximity to a Chinese aid project. These areas are not identical to predefined DHS household clusters.6
Since the DHS survey is not a panel, we cannot compare the same households at two different points in time, namely before and after treatment. However, we can compare households located in same area at two different points in time, allowing us to achieve a high comparability between households across survey waves.7
Benyishay et al. (2017) provide a review on how increasingly available disaggregated geospatial data allows for applying a variety of impact evaluation methods to assessing the effectiveness of aid, among others.nutrition. Second, a sector level analysis confirms these findings and adds that social sector projects specifically improve education and reduce child mortality - suggesting that Chinese aid successfully targets social development goals. Moreover, improvements in social outcomes also seem to be indirectly driven by increased economic wellbeing. Third, when taking the number of Chinese aid projects and their financial volume into account, i.e. their distribution and scale, we find that our results holds for lower treatment intensity levels only, which suggests a non-linear effect of Chinese aid. Fourth, taking into consideration the timing of the project, it seems that results in education are largely driven by projects taking place in the last five-year period before the endline while health outcomes are mainly driven by older projects. Overall, our results show that Chinese aid exhibits patterns that are akin to “traditional aid”: previous research on Western donors has shown that education- and health-related aid has positive effects on development outcomes in these areas (see e.g. Riddel and Niño-Zarazúa 2016 and Kotsadam et al. 2018), and that development aid more generally can have non-linear, u-shaped effects (see Daalgard et al., 2004).
2 Data and Methods
Our analysis is based on the combination of two datasets: (1) Aid Data’s Chinese Official Finance to
Africa dataset, and (2) the Demographic and Health Survey (DHS). Our sample consists of 13
Sub-Saharan African countries for which at least two DHS survey waves are available and for which the first survey wave corresponds to a period that is characterized by little or no Chinese aid activities (before and/ or close to the year 2000). These countries are: Benin, Cote d’Ivoire, Ethiopia, Ghana, Guinea, Kenya, Malawi, Namibia, Nigeria, Senegal, Togo, Uganda and Zimbabwe.
2.1 Aid Data
The Chinese Official Finance to Africa dataset reports geocoded information on Chinese aid projects in the region. Projects can be implemented in multiple locations. Overall, the dataset comprises 1,955 projects in 50 African countries, spanning 3,545 locations and covering the years 2000 to 2012. The countries included in our sample account for 852 projects in 1,745 locations, which are widely distributed within countries and over time (see Figure 1). Since our unit of observation is the project-location pair we will use the term project throughout the text to identify each unique project-project-location included in our database.
For each project, the database provides detailed information on its precise location, the sector it belongs to classified following the OECD Creditor Report System (CSR) purpose codes, financial
volume, the type of flow8 (e.g., ODA or other official flows, OOF), and project status (e.g., planned or
implemented).
Figure 1. Mapping of Chinese Aid Projects by Destination Countries
Source: Authors’ elaboration on AidData
We make two main adjustments to the data. First, a large majority (67.4%) of projects are classified either as “completed” or “implemented”, while the rest is categorized as “in the pipeline” (i.e. planned). Projects in the pipeline have not yet formally started; 11.3% of them are pledged only. Since their implementation is uncertain, we exclude these projects from the analysis. Second, we only consider projects with a relatively precise location, since it is key to our identification strategy to define control and treatment areas as precisely as possible. The AidData data set provides a precision code for each georeferenced project, indicating the level of granularity of the GPS data. Following the approach of previous studies (see e.g. Dreher et al., 2015; Briggs, 2018), we consider projects with a precision code of up to 4, meaning that the project location is – in the least granular case – analogous
8
In China’s case, it is important to distinguish between official development assistance (ODA)-like flows and other official flows (OOF). According to the DAC definition, ODA is (a) provided by official agencies to developing countries, (b) aimed at promoting economic development and welfare, and (c) contains a grant element of at least 25 percent. OOF are also funded by government agencies, but not primarily aimed at development goals and/or not sufficiently concessional to categorise as ODA. Results from an empirical work by Dreher et al. (2015) show that Chinese ODA to Africa is not significantly correlated with the characteristics of national political institutions, while OOF-like flows are more likely to go to countries with higher corruption levels.to a first order administrative division such as a province, state or governorate.9 67% of projects in our
sample fall into this category.
The final sample counts 878 projects (project-location pairs). Transport, communication and energy projects account for 44% of all projects in our sample. Education and health projects account for about 22% of all projects (Figure 2).
Figure 2. Distribution of Chinese aid projects (#) by sector
23.23% 20.62% 10.93% 10.48% 7.40% 7.29% 5.35% 5.01% 5.01% 4.67% Communications Transport and Storage Education
Health
Government and Civil Society Energy Generation and Supply Water Supply and Sanitation Other Social infrastructure and servi..
Others Not allocated
Source: Authors’ elaboration on AidData
Most projects in our sample are categorized as ODA-like (53.2%); 41.8% are classified as Vague Official Finance, since information for a more detailed categorization is not available. Only 4.3% of all projects are classified as other official flows (OOF): commercial activities mostly related to the communication sector and loans to build telecom infrastructure. Moreover, above 46% of project financing is provided in the form of loans.
As mentioned earlier, the AidData dataset has important limitations (Strange et al., 2013). First, China does not officially report data on its foreign assistance. Project-related data is compiled from various sources, including government reports, the media, the private sector, and civil society organizations. Consequently, the data might be not fully representative of Chinese assistance to Sub-Saharan African countries. Second, a bias in China publicly reporting certain projects but not others
cannot be ruled out (Strange et al., 2013). Third, since China’s foreign assistance does not adhere to the OECD-DAC Creditor Reporting System (CRS), it cannot be categorized as precisely as aid from OECD-DAC donors, reducing the comparability between aid flows from China and OECD-DAC members (Horn et al., 2019). Finally, financial disbursement data is not available at the project site level. It is only available at the overall project level. Therefore, we do not use the financial disbursement data in our main analysis, but indicate the presence of a project in any given geographical location with a dummy variable instead. This is our treatment variable. To still take treatment intensity into account, in section 3.3, we repeat our main analysis by replacing the treatment dummy with (i) the number of projects going to certain areas, and (ii) the project finance volume. Since financial disbursement information is not available at the local level, we follow the approach taken by previous studies and divide the total financial disbursement to a project by the number of its locations (see e.g., Dreher et al., 2016 and Gehring et al., 2018).
2.2 Demographic and Health Survey (DHS)
The DHS is a large-scale survey program. Data are nationally representative and include information on a wide range of indicators of households’ wellbeing, including assets, education, health, and nutrition. All DHS surveys are based on largely standardized questionnaires, which allows us to analyze and compare the effect of Chinese aid on household welfare across countries. For this study, we use DHS survey data on 13 Sub-Saharan African countries. We consider two data points for each country: before and after treatment with Chinese aid. We define the pre-treatment wave to be before or very close to the year 2000. Until 1999, China was a recipient of development assistance (Page and Pande, 2018) itself and not very active as a donor in Sub-Saharan Africa. This changed with the establishment of the Forum on China-Africa Cooperation (FOCAC) in 2000 and the financial pledges China made to African governments within this framework (Brautigam, 2009). Financial flows picked up from 2006, when China significantly increased its development assistance to Sub-Saharan Africa and installed the China-Africa Development Fund for 5 billion dollars. Figure A1 in the Appendix illustrates this development: the provision of Chinese loans, a major component of China’s overall aid flows, were nearly zero around 2000 for the continent and then increased significantly afterwards. Consequently, we define the treatment wave to start well after the year 2000. Pre- and post-treatment waves vary for every country in the sample. The countries and their corresponding survey years are listed in Table A1 in the Appendix. Taken together, the surveys include information on 297,459 households.
To assess the impact of Chinese aid projects on household welfare in Sub-Saharan Africa, we focus on three social relevant dimensions: education, child health and nutrition. We proxy each dimension with the following indicators taken from the DHS.
(i) Average years of education. This is one of the common measures for education levels and a proxy
to track the stock of human capital in society. Average years of education impact positively on equity and economic development and labor productivity (Barro and Lee, 2013). In our analysis, the average years of education in each household considers only members over 25 years of age.
(ii) Child mortality. Child mortality is considered a valid indicator for the development of a country as
it is associated with multiple factors including quality and access to the healthcare system and maternal health (MacDorman and Mathews, 2009). This indicator is also used to understand health inequalities within the society and the capacity of governments to invest in the next generation (Martorano et al., 2014). In our analysis, it is calculated as the number of children between 12 and 60 months died in the last 5 years as a share of all the children born over the same period.
(iii) Weight for age Z-score. This anthropometric measure relates weight to age for children under the
age of five. This indicator is used to monitor malnutrition – underweight especially - and to assess the growth process. Malnutrition is associated with increased health risks (Adamson, 2013), as well as poor educational outcomes and lower labor productivity (Behrman et al., 2004). Z-scores are calculated as standard deviations from the median values computed by WHO (WHO 2010).
2.3
Empirical Strategy
To link household data from the DHS to project data in AidData we use georeferenced information. Household clusters surveyed in the DHS and Chinese aid projects in AidData both come with GPS point coordinates; hence, we match household and project data based on those GPS coordinates. The
AidData data set moreover contains information on the point in time in which an aid project was
established. Combining data from both data sets enables us to compare the welfare of control and treatment household areas before and after the implementation of a Chinese aid project.10 Our
empirical strategy leans on recent studies that evaluate the impact of aid projects at the sub-national level (Kotsadam and Tolonen, 2016; Isaksson and Kotsadam, 2018; Wegenast et al., 2019; Kotsadam et al., 2018). However, it differs from previous research in a number of aspects.
Previous studies apply the DiD strategy under the assumption that both DHS waves (before and after treatment with a Chinese aid project) are based on the same underlying population. However, since the DHS is not a panel, different survey waves are based on different sets of clusters and households within clusters. The chance that the same households are interviewed in both survey waves is extremely low. To deal with this limitation, which can potentially bias our results, we move the analysis from the level of household clusters to the area level. Areas are defined as unique
10
Each DHS cluster is randomly reassigned a GPS location that falls within a specified distance of the actual location (0-2 km for urban DHS clusters and 0-5 km for rural clusters, with one percent of clusters randomly selected to be displaced by up to 10km). This is done to preserve the anonymity of survey respondents. While we cannot control directly for this, our strategy to work with larger buffer zones (as detailed in the rest of the section) is in line with standard practice in the literature and should partially attenuate this problem (see Gollin et al. 2017).boundaries sharing the first decimal place for both latitude and longitude coordinates.11 For each DHS
survey wave, we identify the household clusters located within such an area. Relying on the first decimal grants enough flexibility to balance the number of clusters with the degree of homogeneity of households located in each area.
Figure 3 provides a graphical representation of this procedure, using Malawi as an example. Map 1 reports the geographic distribution of the newly constructed household areas in both survey waves. Map 2 in Figure 3 zooms in one newly constructed area located in the surroundings of the Blantyre district. Red and blue dots are the original clusters from each DHS survey wave (2004 pre-treatment and 2015 post-pre-treatment respectively for Malawi). Each newly constructed area is comprised of several DHS household clusters. The number of clusters varies by area. For each of these areas, we compute the average value of all household characteristics and outcome variables that we use in our analysis.12 We do this for every survey wave and then match the same areas across the
two waves.13 For Malawi, there are 232 household areas for which we can match information from
pre-treatment surveys to post-treatment surveys; these are the orange circles in Figure 3. For all countries included in our final sample, we have 1,229 household areas that we observe across two time periods (see also Figures A2-A3 in the Appendix).14 Within each area, we have on average 42
households and 219 individuals at the baseline and 66 households and 309 individuals at the endline (see Table A2 in the Appendix).
11
Each decimal roughly corresponds to 11.1 km.12
When doing this, we use weights provided by each DHS survey.13
Wegenast et al. (2019) propose a similar method since they re-aggregate information from DHS clusters to the district level, computing mean values to get to a dataset that varies by district and year. They do not, however, provide the additional steps we make in our analysis to ensure a more precise comparability between treated and control areas.14
Benin: 79 units; Cote d’Ivoire: 38; Ethiopia: 84; Ghana: 105; Guinea: 61; Kenya: 191; Malawi: 232; Nigeria: 96; Namibia; 87; Senegal: 61; Togo: 90; Uganda: 61; Zimbabwe: 44.Figure 3. Procedures to construct areas
Malawi (Map 1) Blantyre, Malawi (Map 2)
Source: Authors’ elaboration on AidData and DHS
Second, we use inverse probability weighting (IPW), a matching method, to adjust for selection bias that may result from the fact that we are using non-randomized, observational data (Horvitz and Thompson, 1952; Wooldridge, 2007). IPW aims at constructing a credible counterfactual from the data in two stages by identifying and then matching quasi identical units in the treatment and control groups15. Recent work by Busso et al. (2014) shows that reweighting performs well compared
to other matching estimators especially in cases, like ours (see Figure 4), in which the overlap of covariate distribution between treatment and control group is good. We use a Probit model to estimate the probability of being a recipient of Chinese aid for each area. Variables (measured at the baseline) in the model include the demographic composition of the households living in the area: the share of female-headed households, average age of household members, average size of households, the average number of males, the average number of females, the average number of children aged between 0 and 4, the average number of children aged between 5 and 14, the average number of children aged between 15 and 17, and the average dependency ratio. Finally, we add the log number of households included in each area to control for population density and region fixed effects. Results are reported in Table A3 in the Appendix.
15
Our main results are robust to different matching methods. We have tried for instance a matching based on a Kernel estimator, and this does not affect the final results reported in Section 3. Results of this alternative analysis are not reported here for reasons of space but are available upon request to the authors.Table A4 in the appendix reports the differences in mean demographic characteristics between treatment and control households at baseline. After matching, control variables are balanced across the two groups excluding for the average number of children aged between 0 and 4, the average number of children aged between 5 and 14, the dependency ratio and the population density. The overall balance is satisfied. This is shown by the Hotelling test: after matching, we do not observe a significant difference in means between control and treatment group (Prob > F (85, 943) = 0.8655). The graph depicting the weighted propensity scores for control and treatment groups also indicates the goodness of the overall balance since the two distributions get very close one each other after the matching (Figure 4, right panel).
Figure 4. Distribution of propensity scores, before (left panel) and after (right panel) matching
Third, we apply a DiD model where we reweight the control sample by the IPW to make it as comparable as possible to the treatment sample. The model specification is:
Yxit=β1Tit+β2Dxi+β3
(
Tit∗Dxi)
+β4Zxit+δx+θit+ε (2)Where Yxit represents the outcome of interest (see Section 3.2) computed at the area level x in country i at time t. The validity of our DiD approach crucially depends on the parallel trend assumption which
posits that the average change in the comparison group represents the counterfactual change in outcomes on the treatment group if there were no treatment. We visually check this assumption for various education, health and nutrition indicators and confirm that it holds for three outcomes that we employ in our analysis. Details are provided in Figure A4 in the Appendix.
D is the treatment dummy which equals 1 if an area was treated with a Chinese aid project. Tit
is a time dummy that equals 1 for the post-treatment DHS survey wave. Zxit is a vector of
characteristics measured at the area level. These characteristics include (a) the demographic composition of the households, i.e., the average number of persons living in the household, household
members’ average age, the dependency ratio, a dummy variable (=1) if the household head is female, and the (log) number of households in each area. Furthermore, we add time varying country ( θit )
and area (
δ
x ) fixed effects to control for time trends and area-specific characteristics that may bespuriously correlated with our dependent and independent variables. This allows us to control for confounding factors and omitted variable bias at a disaggregate geographic level. Additionally, time varying country fixed effects account for factors potentially affecting the treatment and outcome variables at the national level and over time. Standard errors are originally clustered at the country level, and then corrected using wild bootstrap methods to overcome the potential bias due to the small number of countries in our sample (Roodman et al., 2019). Table 1 reports descriptive statistics on the main variables used in the regression analysis.
Table 1: Descriptive Statistics
Variables Obs Mean Std. Dev. Min Max
Dependent Variables
Average Years of Education (N) 292,702 4.95 3.17 0 16
Child Mortality 297,349 0.01 0.02 0 1
Weight for age Z-score 292,188 -0.88 0.59 -3.75 4.72
Treatment Variables
Treated with Project 297,459 0.31 0.46 0 1
Number of Projects 297,459 2.51 7.58 0 52
Total Amount in $ Millions (Financial Volume) 297,459 44.30 251.00 0 5,490 Household Characteristics
Female-Headed Household 297,459 0.27 0.15 0 1
Age of Household Members 297,459 44.65 5.30 27.04 68
Household Size 297,459 4.86 1.47 1 19
Dependency Ratio 297,459 0.51 0.11 0.17 1
(Ln of) No. of Households per area 2,484 3.62 0.75 1.61 7.39
3 Results
We begin with a discussion of our main analysis on the effect of Chinese aid projects on education and child health and nutrition (see Section 3.1). This part of the analysis aggregates projects across all sectors. In Section 3.2, we discuss the results of a sector level analysis which simultaneously assesses the effect of social and economic projects on each outcome variable. The sector analysis investigates whether there are heterogeneous treatment effects across different (broadly defined) sectors. In Section 3.3, we discuss the results of an analysis that explores to what extent variance in the distribution and scale of Chinese aid affects our outcome variables. To this end, we replace our binary treatment dummy with (i) the number of projects within an area, and (ii) the financial volume of the
project. As discussed before, our main analysis is not based on these metrics since their reliability is limited. Finally, we account for the temporal distance between the timing of treatment and the post treatment survey wave to assess to what extent treatment effects may take more or less time to materialize (Section 3.4).
3.1 Main analysis
The results of our weighted DiD estimations are reported in Table 216, with (columns 1-3) and
without controls (columns 3-6). We have two main findings. First, the DiD coefficients in columns 1 and 4 indicate that Chinese aid has a positive and significant effect on household areas’ average years of education: the presence of Chinese aid projects results in approximately half a year of additional schooling. This is equivalent to an almost 10% increase of average education years for the country-areas in our sample (see Table 1). This result fits with studies that demonstrate positive effects of “traditional” aid on education outcomes (see e.g., Riddel and Niño-Zarazúa 2016). Second, we find that areas receiving Chinese aid experience a rather small, but significant reduction in child mortality rates (Table 2, column 2 and 5): the presence of Chinese aid reduces child mortality by 0.9 percentage points.17 Kotsadam et al. (2018) provides a relevant reference point for this result. He investigates the
effect of USAID projects (pooled across various sectors) on infant and child mortality rates in Nigeria using a DiD strategy and a 25 km radius around project locations to assign households to the treatment group. His results indicate that USAID projects significantly reduce child mortality rate (1-5) in project areas by about 2.5 percentage points. Third, we find that Chinese aid projects do not have a significant impact on weight-for-age outcomes in treatment areas (Table 2, columns 3 and 6). In sum, our findings seem to suggest that Chinese aid is similar to other “traditional” donors regarding its effects on education and health outcomes.
16
In Table A4 in the Appendix we have run our main analysis based on the unweighted sample that includes a repeated cross section of households (i.e. data are not aggregated at the area level). The treatment coefficients are always statistically significant and with the expected sign. As discussed in Section 2.3, this approach does not allow us to control for confounding factors and omitted variables at a disaggregated geographic level and may lead to wrong conclusions about the impact of Chinese aid on social outcomes.17
We also tested the effect of Chinese aid projects on neonatal mortality and infant mortality. Infant mortality is widely used in the literature while neonatal mortality is one of the new targets indicators introduced to monitor SDG number 3. We do not find a statistically significant effect of Chinese aid on any of these variables (results are available upon request). This might be explained by the fact that determinants of child mortality differ by age (Martorano et al., 2014). As suggested by Woods (2009: 3), survival chances in the early days are affected by different factors, including prematurity and the complications of childbirth. Moreover, demographic factors including the sex of the new-born, the length of time between births, the birth order, the age and the health status of the mother are also key to explaining the survival rate particularly in his/her earliest stages of life (Pozzi and Fariñas, 2015). By contrast, mortality and especially after the first year of life of the child -seems to be more sensitive to socio-economic and environmental factors (Matthiessen and McCann, 1978). Therefore, different causes might require different interventions. While it might be more immediate to improve socio-economic factors, health as well as bio-demographic factors may require structural interventions. The main implication is that infant mortality and child mortality are not expected to move in tandem, being related to separate histories. Indeed, the existing evidence shows that child mortality started to decrease far earlier and faster than infant mortality (Pozzi and Fariñas, 2015).Several explanations may be advanced for these results. Improvements in social outcomes can be a direct result of Chinese projects in the areas of education and health. After all, educational and health projects each account for about 11% of the total project portfolio (Figure 2). As far as the health sector is concerned, China is currently involved with building an African public health system and concentrates activities on setting up infrastructure including hospitals, equipment, and medicine. It also trains health teams and makes efforts to combat malaria and maternal, neonatal, and child health (Shajalal et al., 2017). Our results show that China has a positive impact on child mortality, indicating that its projects are effective and reinforcing positive trends in the region. Although Sub-Saharan Africa is among the regions with the highest risk of child death in the first month of life it has seen a fast decline in its under-five mortality rates, with the annual rate of reduction doubling between 1990-2000 and 2000-2011.18 A small but positive causal differential effect of Chinese Aid on this
outcome thus might occur because additional resources coming in from China are supporting positive trends and complementing ongoing efforts. However, our results also show that Chinese aid does not improve nutrition outcomes. This might be owed to the fact that achieving this goal is widely known to require specific behavioral changes by children’s parents and, consequently, interventions which are designed along those lines (see e.g., Sanghvi, 2016 and Kennedy et al., 2018).19 In other words,
nutrition-based outcomes might require specific interventions targeting behavioral and nutritional factors simultaneously to fighting causes of malnutrition. To the extent that Chinese aid concentrates on health infrastructure and staff training it may fail to address these issues effectively. As Shajalal et al. (2017) point out, China has paid limited attention towards disease prevention, health promotion and awareness initiatives, and health education. The latter three might be particularly relevant for achieving behavioral change among parents in order to improve malnutrition. In addition, a debate on designing new and better interventions to address nutrition outcomes have rather recently regained focus in the international community given that child nutrition seems to be a more stubborn development problem to solve (Akombi et al., 2017). Sub-Saharan Africa has one of the highest levels of child malnutrition globally (stunting, wasting, and underweight) and efforts made in this direction may take time to materialize (Akombi et al., 2017). It is unclear to what extent China has been or will be engaging its foreign assistance in this area.
Regarding education outcomes, we observe a statistically and practically significant effect of Chinese aid on average years of education. As King (2014) points out, China has been successful in providing nine-year compulsory education even before the 2001 creation of the Millennium Development Goals. It is reasonable to assume that the country has managed to successfully transfer its expertise and experience in this area to the African countries it is targeting with its assistance. In addition, King (2014) discusses that China’s stand-alone education and training projects in Africa are
18
https://www.usaid.gov/sites/default/files/documents/1860/Africa%20Key%20Facts%20and%20Figures.pdf, last accessed December 3, 2019.19
See also: https://www.usaid.gov/sites/default/files/documents/1864/at-scale-nutritionSBCC-technical-guidance-brief-edit-508.pdf, last accessed July 10, 2019.very much driven by partner country priorities. Due to the progress that has been made in improving education outcomes over the past 30 years and the maturity of this development sector, it might be that African partner countries set the right project priorities with China, which in turn results in successful education projects (in terms of years of schooling).
In sum, our findings regarding the effects of Chinese aid seem reasonable in the overall context of health and education trends in Sub-Saharan Africa. Yet, we want to emphasize that we interpret our findings with caution, since detailed information on delivery modes of Chinese health and education projects is scarce.
Improvements in education and health may also result from projects with relevant spillover effects, such as access to electricity, roads, or piped water. For instance, transportation infrastructures, which makes up at least 21% of Chinese aid, can facilitate access to education of health services for a large number of pupils. Improvements in social outcomes may be also indirectly driven by increased economic welfare resulting from Chinese aid projects, for example by increasing local economic opportunities. Indeed, as the analysis presented in Table A6 shows, we find a significant and positive effect of Chinese aid on proxies for households’ economic wellbeing and development. Areas that have been treated by at least one Chinese aid project have experienced an increase in their economic outcomes, measured either as the size of night time light (NTL) intensity20 , or as an asset index
calculated from DHS data21 (as in Filmer and Pritchett, 1999; Filmer and Pritchett, 2001). These
results are indicative of a potential indirect effect of Chinese aid that goes through broader economic improvements for local households. They also support evidence provided by Dreher et al. (2016), who finds a positive effect of Chinese aid on economic growth (proxied by per capita NTL) at the sub-national level using the same data we use but for a larger sample covering the whole African continent.22
20
NTL intensity is computed at the area level with information from the National Centers for Environmental Information of the U.S. National Oceanic and Atmospheric Administration ‘NOAA’. In our dataset we consider both "raw" and "stable" night-time light measurements. The former contains the average of the visible band digital number values with no further filtering. The latter contains the lights from cities, towns, and other sites with persistent lighting.21
The asset index is computed using principal component analysis based on household information on assets such as the type of flooring, electricity, radio and car. The index is comparable across countries since it is based on assets that are reported in all the DHS waves and are measured according to the same scale. The continuous scale of the index is adjusted by adding a constant such that it contains only positive values; then it is transformed into its natural logarithm.22
Though based on the same data, there are several differences between our analysis and the work by Dreher et al. (2016). First, their coverage is higher both in terms of countries and years. Second, their measure of aid is based on the estimated value of aid projects. Third, they apply a standard panel estimator and control for endogeneity of aid using an IV. Last, they aggregate information at higher geographic levels (ADM1 and 2) compared to our areas.Table 2: Results, full sample
1 2 3 4 5 6 Average Years of Education ChildMortality age Z-scoreWeight for
Average Years of Education
Child
Mortality age Z-scoreWeight for
Treated*Time (D*T) 0.459*** -0.009*** 0.007 0.301*** -0.009*** -0.009 [0.107] [0.002] [0.031] [0.082] [0.002] [0.038] Time (T) 0.232*** -0.006*** 0.773*** 0.132 -0.011*** 0.765*** [0.059] [0.001] [0.017] [0.108] [0.002] [0.036] Female-Headed HH -0.290 0.003 0.036 [0.523] [0.009] [0.166] Age Of HH members -4.030*** 0.048** -0.014 [0.655] [0.017] [0.173] HH Size 0.243 0.001 0.000 [0.408] [0.009] [0.072] Dependency Ratio 7.596*** 0.061*** 1.224*** [1.174] [0.015] [0.384] N. of HH -0.145* -0.003* -0.032 [0.070] [0.002] [0.032] Constant 5.528*** 0.019*** -1.480*** 17.401*** -0.182*** -1.930** [0.002] [0.000] [0.001] [3.145] [0.056] [0.718] Adjusted p-value of D*T
using wild bootstrap test 0.0350** 0.0150** 0.8218 0.0140** 0.0120** 0.8068
Observations 2,031 2,058 2,027 2,031 2,058 2,027
R-squared 0.278 0.073 0.123 0.521 0.088 0.159
*** p<0.01, ** p<0.05, * p<0.1;
Note: All the variables are calculated at the area-level. See Section 2.2 for a description of how geographic areas have been constructed. All regressions include area fixed effects and country*year fixed effects. Standard errors are originally clustered at the country level and adjusted using the wild bootstrap method.
3.2 Sectoral Analysis
The sectoral analysis links Chinese projects more directly to their intended outcomes. We group projects based on their Creditor Report System (CRS) sectoral codes. Our main distinction is among social and economic sector projects.23 We re-estimate the entire set of regressions from Table 2, but
include separate treatment dummies for social and economic projects, as well as a dummy for projects that do not fall in any of these two categories24. Including all the treatment dummies in the same
23
Economic projects include: Transport and Storage (210); Communications (220); Energy Generation and Supply (230); Banking and Financial Services (240); Business and Other Services (250); Agriculture, Forestry and Fishing (310); Industry, Mining, Construction (320); Trade and Tourism (330). Social projects include: Education (110); Health (120); Population Policies (130); Water Supply and Sanitation (140); Government and Civil Society (150); Other Social Infrastructure and Services (160); Women in Development (420); Developmental Food Aid (520); Non-Food Commodity Assistance (530). CRS codes are provided in parentheses.24
This category includes about 10% of total projects in our sample (see also Figure 2). Among them, are projects that do not fit into any of the previous categories due to their generic sectoral allocation (e.g., CRS codes 430 “Other Multisector”, 600 “Action related to Debt” and 700 “Emergency Response”) or projects thatmodel enables us to account for heterogeneous response to different types of aid, complementing the previous analysis on treatment effects. Moreover, it excludes a potential omitted variable bias which might result from the fact that the decision to provide both types of projects is mutually dependent. As stated before, Chinese aid is known for the interlinkages it seeks to create between different sectors (Chin and Gallagher, 2019).
The results presented in Table 3 confirm the findings shown in Table 2: Chinese aid projects improve education and reduce child mortality, but have no significant effect on nutrition. Importantly, they show that social sector projects specifically improve social outcomes25. Economic projects, on
the other hand, have low explanatory power. However, as we discussed in the previous section, they still may affect social outcomes indirectly through an increase in economic welfare.
These results are consistent with macro-level studies that find a positive relationship between social sector projects, especially education aid, provided by Western donors and educational outcomes in developing countries (see e.g. Michaelowa and Weber, 2007; Birchler and Michaelowa, 2016; Riddel and Niño-Zarazúa, 2016 for a review of findings on the effectiveness of education aid). Furthermore, it is consistent with micro- and project-level studies which find a positive relationship between social interventions and corresponding outcomes (see Riddel and Niño-Zarazúa, 2016; Heynemann and Lee, 2017).
have not been allocated in any CRS code. The latter are mostly loans, some of which linked to large scale projects (for instance a 3 billion loan to support infrastructure construction in Ghana in 2009).
25
Due to the small number of observations, it is not possible to make the analysis even more granular to investigate the effect of education projects on educational outcomes, as well as on health projects on child mortality.Table 3. Sectoral Analysis 1 2 3 Average Years of Education Child Mortality Weight for age Z-score
Treated (economic projects)*Time (T) 0.140 -0.006 0.003
[0.104] [0.004] [0.048]
Treated (social projects)*Time (T) 0.275** -0.015*** 0.035
[0.123] [0.004] [0.069]
Treated (other projects)*Time (T) 1.324* -0.015 -0.718
[0.609] [0.016] [0.549] Time (T) 0.103 -0.019*** 0.773*** [0.109] [0.003] [0.048] Constant 17.718*** -0.174*** -1.951** [2.915] [0.056] [0.721] Observations 2,031 2,058 2,027 R-squared 0.525 0.092 0.168 *** p<0.01, ** p<0.05, * p<0.1.
All the variables are calculated at the area-level. See Section 2.2 for a description of how geographic areas have been constructed. All regressions include controls (Female HH; Age of HH members; HH size; dependency ratio and the number of HHs in the area); area fixed effects and country*year fixed effects. Standard errors are clustered at the country level. Regressions include as well interaction terms of all the treatments with each other. This allows to account for cases in which projects in different sectors co-locate.
3.3 Distribution and scale of Chinese project aid
In our main analysis, we focus on whether a project is implemented or not and refrain from using less reliable information on the number of projects or project finance volume. In the analysis presented in this section, we take treatment intensity into account by replacing the binary treatment dummies with (a) the number of projects going to certain areas, and (b) the actual financial volume of the project. Moreover, we test for non-linear effects of aid that have been identified by earlier macro- and project-level studies. To do this, we apply the propensity score-based marginal mean weighting through
stratification (MMWS) method to our data (Hong, 2010; Hong, 2012; Huang et al 2005).
MMWS combines propensity score stratification and inverse probability of treatment
weighting to reduce selection bias, and is applicable to so-called multivalued treatments which can
take on more than one value (Linden, 2014). Based on the distribution of the number of projects (see Table 4) we consider two, hence multivalued, ordinal treatment levels: (1) equal or lower than the median value; and (2) higher than the median value (i.e. 2 projects). Based on the distribution of
financial project size (Table 4), we consider the following two treatment levels: (1) equal or lower than the median value and; (2) higher than the median value (i.e. 149.92 $ million)26.
Table 4. Number of projects (#) and total project amount (constant $ million)
Mean Median 25th percentile 75th percentile
Number of projects 6.45 2 1 6
Total amount 1428.87 149.92 68.57 540.04
As before, areas not located within a 25km radius of a Chinese aid project serve as the control group. MMWS reduces selection bias by achieving balance on baseline (and observable) characteristics between all treatment levels (Linden, 2014). The empirical strategy relies on four steps. First, we compute the generalized propensity score for each area. To this end, we use an ordered logistic regression and regress our ordinal treatment variable on the same set of covariates used in the previous selection model (see Section 2.2). Results are reported Table A3 (Models 2 and 3) in the Appendix. Second, the generalized propensity scores that we computed for each area is stratified into three strata (zero projects, below and above median values). Third, we compute the MMWS weights for each area corresponding to its stratum and treatment level. Finally, we re-estimate our main specifications using the MMWS as probability weights.
The results presented in Table 5 (panel A) indicate that the treatment intensity in terms of the number of projects matters: the effect on our educational and health outcome variables is non-linear. More specifically, Chinese aid projects have a significant impact on social outcomes in areas receiving one or two projects. By contrast, the effect is no longer statistically significant in areas with a higher number of projects. The results are not significant on nutrition. Table 5 (panel B) reports the regressions that include treatment intensity based on financial project volume. The results mirror those reported in Table 5 (panel A) in the case of education but not for health: the impact of Chinese assistance is only statistically significant below median values of financial disbursement (i.e. around 150 $ million).
In summary, the results in Table 5 is an important addition to our work. On the one hand, they support our core results by pointing to a positive relation between Chinese aid and social related outcomes that are robust to a treatment indicator that accounts for the scale of the projects. On the other hand, they show that the relationship between treatment level and outcome variables is non-linear and that the effectiveness of Chinese aid on social outcomes seem confined to lower scale interventions.
26
In additional tests, we have used an alternative set of thresholds. According to our results, it seems that the tipping point on which the net effect becomes insignificant is around two projects - or between $ 150 and 200 millions in terms of financial volume. Results are available upon request.This is consistent with existing evidence, given that u-shaped non-linear effects of aid have been already documented in the literature (see Dalgaard et al., 2004). For example, Michaelowa and Weber (2007) provide evidence for diminishing returns to aid especially for primary and secondary education, which can be motivated by issues related to the recipient countries’ absorptive capacities. Moreover, looking at the literature on the effectiveness of aid on economic growth, these results are consistent with the view that high levels of aid might generate political capture and deterioration of governance (Bräutigam and Knack, 2004). Overall, however, these are speculative statements that we can advance given our limited information about the specific delivery modes of Chinese aid projects.
Table 5: Results – Multilevel Treatment
1 2 3 Average Years of Educatio n Child
Mortality age Z-scoreWeight for
Panel A. Number of Projects
Treat_Level 1*Time (D*T) 0.258** -0.007** 0.106 [0.098] [0.003] [0.097] Treat_Level 2*Time(D*T) 0.119 -0.006 0.139 [0.119] [0.005] [0.096] Time (T) 2.365*** -0.011*** 0.245*** [0.096] [0.001] [0.062] Constant 1.068 0.020 -0.668 [1.681] [0.043] [0.629]
Controls Yes Yes Yes
Observations 2,423 2,483 2,416
R-squared 0.513 0.107 0.174
Panel B. Financial Project Volume
Treat_Level 1*Time(D*T) 0.746*** -0.003 0.196 [0.231] [0.006] [0.122] Treat_Level 2*Time (D*T) -0.089 -0.010 0.030 [0.139] [0.006] [0.060] Time (T) 2.251*** -0.012*** 0.251*** [0.116] [0.003] [0.057] Constant 1.332 0.029 -1.303** [2.232] [0.029] [0.470]
Controls Yes Yes Yes
Observations 2,423 2,483 2,416
R-squared 0.494 0.140 0.171
Note: All the variables are calculated at the area-level. See Section 2.2 for a description of how geographic areas have been constructed. All regressions include controls (Female HH; Age of HH members; HH size; dependency ratio; the number of HHs in the area); area fixed effects and country-year fixed effects. Standard errors are clustered at the country level.
3.4
Timing of the intervention
Next, we address an important feature of our data: the temporal distance between the year in which the treatment occurs and the endline survey wave which provides us with the post-treatment measurement of our outcome variables. One potential issue of our main specification is, in fact, that we cannot determine whether the effects that we find are due to projects implemented in more recent years, or are consequence of interventions implemented earlier on. Hence, by considering the variance in the temporal distance between the timing of treatment and the post-treatment survey wave, we can assess to what extent treatment effects may take more or less time to materialize, or die down over time.
To account for this timing effect, we employ a multilevel treatment approach. More precisely, we apply the propensity score-based MMWS method described in the previous section. Based on the distribution on the temporal distance between the timing of treatment and the post-treatment survey across countries, we consider two treatment levels: (1) equal or lower than the median value (i.e. 5 years) and; (2) higher than the median value. The results, summarized in Table 6, show that while education seems to be largely driven by projects taking place in the last five year period before the endline survey, health outcomes are mainly driven by projects older than 5 years27.
27
As a further check to test if the results are sensitive to different time lengths, we have tried to run our analysis on different sub-samples on the basis of the overall length across the available DHS waves. The two sample are composed, respectively, by countries with pre- and post-treatment survey distance between 10 and 13 years (i.e. Ethiopia, Guinea, Kenya, Malawi, Namibia and Nigeria) and countries with pre and post-treatment survey distance between 14 – 16 years (i.e. Benin, C. d'Ivoire, Ghana, Senegal, Togo and Zimbabwe). Note that we have removed Uganda from this exercise because the distance between the two surveys in this country is only 6 years. Results of separate regressions on the two groups (not reported for reasons of space, but available upon request to the authors) remain very similar to our baseline specification.Table 6: Results – Timing of the project applying a multilevel treatment
1 2 3
Average Years of
Education Child Mortality Weight for ageZ-score
Projects <=5 yrs*Time (D*T) 0.391** -0.004 0.149 [0.168] [0.006] [0.093] Projects>5 yrs*Time(D*T) 0.146 -0.012*** -0.024 [0.191] [0.002] [0.036] Time (T) 2.398*** -0.017*** 0.235*** [0.102] [0.002] [0.034] Constant 0.633 0.016 -1.591** [1.743] [0.031] [0.576]
Controls Yes Yes Yes
Observations 2,111 2,171 2,105
R-squared 0.493 0.113 0.149
*** p<0.01, ** p<0.05, * p<0.1; 1HH
Note: All the variables are calculated at the area-level. See Section 2.2 for a description of how geographic areas have been constructed. The two levels of the treatment refer, respectively, to projects implemented during the last 5 years and to projects older than 5 years. All regressions include controls (Female HH; Age of HH members; HH size; dependency ratio; the number of HHs in the area); area fixed effects and country-year fixed effects. Standard errors are clustered at the country level.
4 Robustness
We check for the robustness of our main results in various ways. First, we check the validity of the results discussed above using alternative definitions of the control group. Second, we test the validity of the main estimation results using alternative definitions of the dependent variable. Third, we exclude potential outliers i.e. areas that received many projects. Fourth, we test whether the results are sensitive to enlarging the buffer zone that is used to assign areas to treatment and control groups from 25 km to 50 km. Fifth, we check the validity of the results discussed above considering only areas with projects completed. Sixth, we run a placebo test based on project implementation status: we re-estimate our main regressions from Table 2 with projects that are categorized as being in the “pipeline”. Those projects have either not yet been started or completed. If our identification strategy works, we should not find any treatment effects when using the sample of projects that have not yet been started or implemented. Last, we repeat our analysis excluding areas benefitting from World Bank aid.
4.1 Alternative definition of the control group
In this section, we test the robustness of the analysis using additional variables to compute the inverse probability weights. More specifically, we augment the selection model described in Section 2.3 to account for economic factors calculated at the baseline in order to exclude that Chinese projects end up targeting more developed areas within a country. We measure the latter either as the size of night time light (NTL) intensity, or as an asset index calculated from DHS data (see footnotes 20 and 21 for more information on these two variables). The regressions reported in Table A7 in the Appendix show that the results are very close (both in their magnitudes and levels of significance) to those reported in the baseline estimation (see Table 2). Note however that, in both cases, we observe a worsening of the overall balance compared to our baseline specification, especially in the case of NTL.28
4.2 Alternative definition of the dependent variables
We test the sensitivity of our results to alternative measures for education, health and nutrition. In the case of education, we replace the average years of education with the average educational attainment. Both variables proxy the level of the stock of human capital in the society. Average educational attainment consists of six categories: (0) “no education”, (1) “incomplete primary”, (2) “complete primary”, (3) “incomplete secondary”, (4) “complete secondary”, and (5) “higher education”. In the case of health, child mortality ratio is replaced by considering the absolute number of children between 12 and 60 months died in the last 5 years, rather than their share on the total . Finally, on the nutrition dimension, we replace the weight for age z-score with the other traditional anthropometric measures such as the weight for height z-score, the height for age z-score and BMI for age z-score.29
As shown in Table A8 in the appendix, results are consistent with our baseline estimations (see Table 2).
4.3 Excluding outliers
The distribution of projects across areas shows that the 25th percentile benefitted from only 1 project while the 95th percentile benefited from more than 36 projects on average. Therefore, we test to what degree our results may be driven by areas recording a large number of projects. For this purpose, we exclude areas recording more than 36 projects on average. The regressions reported in Table A9 in the Appendix show that the results are robust to the exclusion of outlier areas.
28
The Hotelling test is Prob>F(86,942) = 0.3897 in the case of NTL and Prob>F(86,942) = 0.8578 in the case of the asset index.29
Height for age score is measured considering height (in cm) and age of the children. Weight for height z-score is computed multiplying weight (in kg) for standing height (in cm). Finally, the Body Mass Index (BMI) is calculated dividing weight (in kg) over standing height (in cm).4.4 Altering the buffer zone
We test to what extent choosing a buffer zone of 50 km instead of 25 kilometers to assign areas to control and treatment groups affects our main findings. The results presented in Table A10 in the Appendix indicate that Chinese aid projects still have a positive impact on the social welfare of households (areas) that live up to 50km away from the project site. However, the effect such projects on outcomes apparently wears off with increasing distance. This seems reasonable: education-related interventions as well as health-related interventions are likely local in nature (such as the construction of schools, hospitals or provision of other services), reaching a limited number of people. This, in turn, may limit spillover effects beyond a certain geographical distance.
4.5 Including only areas with completed projects
As discussed in Section 2.1, in our analysis we consider aid projects whose status is either completed or implemented. It might be, however, the case that some of the projects originally defined as “implemented” were not completed during the period considered, and that local governments had to commit their own resources to match aid pledges, indirectly affecting our outcomes of interest. Therefore, we check whether our results hold to a stricter definition of the treatment that includes only completed projects. Results, reported in Table A11 in the Appendix, show that our findings remain substantially unaffected to changes in the definition of the treatment.
4.6 Placebo test
As discussed in Section 2.1, a relatively large share of Chinese aid projects (around 32%) were not yet completed or implemented during the period covered by our analysis. Such projects, classified as being “in pipeline” can be used to run a placebo test. Since potential effects of these projects should not yet have materialized, they should not significantly affect our outcomes variables. To run this test, we constructed an alternative database following the same procedure as described in Section 2, but only including pipeline projects. We then repeat our analysis based on this project sample. We find no effect of pipeline projects on our outcome variables. This indicates that our identification strategy seems valid and that observed effects are not based on a spurious relationship between treatment and outcome variables. The regression results are presented in Table A12.
4.7 Excluding areas with WB/IBRD projects
Last, a further investigation on the distribution of projects across areas shows that around 38 per cent of the treated areas benefitted from both Chinese aid projects and World Bank aid projects (see Figure A5 in the Appendix). Hence, one possible concern is that our main results are not picking up the specific contribution of Chinese aid. To investigate this, we re-estimate our baseline model excluding areas that benefitted from both types of projects. Geolocalized data on World Bank Aid projects are
