Version classiqueVersion mobile

The Effect of Foreign Aid on Sub-national Development

 | 
Dumebi Ochem

3. Data

Texte intégral

  • 1 All variables used in the estimations were obtained from the PRIO-GRID dataset, with the exception (...)

1The data framework employed in this paper is derived from the PRIO-GRID 2.0 dataset, which is based on a standardized quadratic grid cell structure covering all terrestrial areas of the world (Tollefsen et al., 2012). The spatial unit of analysis is a single grid cell at a 0.5 x 0.5 decimal degree resolution, which corresponds to a cell size of approximately 55 x 55 kilometres at the equator. Each grid cell then contains cell-specific information on socio-economic conditions, climate characteristics, local conflict events and foreign aid, where the latter is spatially linked to the grid cell structure.1

2The main advantage of this approach consists in its ability to mitigate some of the endogeneity issues faced in the aid effectiveness literature. Contrary to administrative regions such as provinces and municipalities, grid cells are insensitive to political influences, and are completely exogenous to the variables in my analysis. As such, it is unlikely that aid donors would allocate funds according to economic conditions within grid cells, addressing concerns related to reverse causality or sample selection. In addition, the high degree of spatial resolution allows for a more accurate analysis of the impact of aid projects, as one should be able to separate the effect within a particular grid from an unrelated effect in a more distance cell.

  • 2 Due to differences in panel size between variables and the lag specification in the baseline model, (...)

3Given that this investigation focuses solely on African countries, the panel covers a cross-section of 10,674 grid cells representing all 54 fully recognised sovereign states, over the period 1992-2014.2 Table 1 provides descriptive statistics for the relevant variables used in the estimation.

3.1 Outcome variable

  • 3 Image and data processing by NOAA’s National Geophysical Data Center. DMSP data collected by US Air (...)

4Night-time luminosity data at the grid cell-level is derived from the DMSP-OLS Nighttime Lights Time Series Version 4 (Average Visible, Stable Lights, & Cloud Free Coverages), for the period of 1992 to 2012.3 The data are calibrated to account for inter-satellite differences and inter-annual sensor decay, as per the methodology in Elvidge et al. (2014), rendering it suitable for time-series analysis. Values are standardized between 0 and 1, where 1 indicates the highest observed nightlight intensity in the time-series, and 0 represents the lowest. This approach allows the indicator to be comparable across grids, countries and time, as opposed to other sources such as income data obtained from household surveys.

5The use of satellite data as a proxy for economic activity and development has become widespread in recent years. In fact, night-time lights have been shown to be a suitable alternative for sub-national measures of real GDP, particularly in low-income countries where such information is seldom available or accurate. In addition, many studies have estimated an “exchange rate” between night-time lights and real GDP, allowing one to make interpretations on the effects of foreign aid in economic terms.

6For instance, Henderson et al. (2012) and Hodler & Raschky (2014) argue for a linear relationship between nightlights and GDP at the country and regional levels respectively, and calculated an elasticity of approximately 0.3. Hu & Yao (2019) instead found the relationship to be non-linear and dependent upon geographic location.

7It is important to note that the nightlight distribution across grid cells and time exhibits a high degree of positive skewness, as a result of over 70% of the sample displaying very low night-time light activity. I account for this issue by applying a logarithmic transformation to the data, to which I also add a small number of 0.01 to retain all zero values.

8In Section 5.2, I instead employ an inverse hyperbolic sine (IHS) transformation, which has been argued to be a more desirable alternative to the natural log (Friedline et al., 2015).

3.2 Treatment variable

  • 4 The IDA offers grants and highly concessional loans (i.e. Official Development Assistance) to the p (...)

9Data on the foreign aid projects is obtained from the World Bank Geocoded Research Release (Version 1.4.2) published by AidData. This geo-referenced dataset includes all World Bank projects from both the International Bank of Reconstruction and Development (IBRD) and International Development Association (IDA) lending lines (AidData, 2017).4 It covers a total of 5,881 projects, of which 5,684 are geocoded, approved between 1995 and 2014 and distributed across 61,243 project locations. The AidData dataset provides information on the coordinates of each project location, recipient countries, main sector, duration, as well as commitment and disbursement flows.

10In addition, each geocoded location is assigned a location class code and a geographic exactness code according to the IATI standard to describe the level of granularity (AidData Research and Evaluation Unit, 2017). The location class specification comprises of 4 discrete categories to differentiate geographic features; these include: administrative regions (e.g. state, province, independent political entity), populated place (e.g. village, city), structure (e.g. building, road, bridge), and other topographical features (e.g. national park, river). The binary geographic exactness specification describes whether the coordinates correspond to an exact location, or an approximation.

11For the main empirical analysis, I restrict the sample to only include project locations classified as either a populated place or structure (location code 2 and 3), which reduces the sample to 47% of its original size. Using AidData’s original eight-point precision code geocoding methodology, this corresponds to locations that are exactly geocoded (precision code 1) or within 25km of an exact location (precision code 2).

12One problem with using this dataset is the incidence of missing observations, particularly for the disbursements and project end date variables.

  • 5 One project with 194 locations in Ethiopia (Road Sector Development Program Support Project) was re (...)

13To address the first issue, I refer to a separate AidData document outlining project- and year-specific transactions. I argue that for those project locations for which there is a recorded total commitment amount but missing disbursements, it actually signifies that there were no disbursements made, and should therefore correspond to zero values.5

  • 6 The results remain fairly unchanged when these observations are instead excluded from the estimatio (...)

14For the second issue, I assume that projects that have a start date but no recorded end date represent projects that were dropped by the respective governments, or were never implemented. This hypothesis is supported by the fact that all such projects received no disbursements. I therefore re-classify their project status as “Dropped”, and set the end date equal to the start date. Those projects would therefore be considered as active only for the year coinciding with the time a commitment was made – which may itself have an economic effect.6

15Using the projects’ start and end years, I then spatially overlay each project location on the PRIO-GRID grid cells by means of Geographic Information Systems (GIS) techniques, in order to obtain a panel of active projects per grid cell, per year. Based on this panel, I produce four separate indicators of foreign aid.

16First, I construct a binary variable indicating whether at least one project was located in a particular grid cell in a given year. Second, I construct a variable indicating the total number of active projects in a given grid cell and year.

17Finally, I construct two continuous variables indicating the total annual financial flows allocated to each grid cell, one using the commitment amounts and the other using disbursements values. Both variables were constructed using information on the actual transaction dates, rather than the project duration dates, and are expressed in millions of US dollars, at constant 2011 prices. In addition, because financial information is only provided at the project level, I follow previous methodologies and assume that both commitment and disbursement amounts are distributed uniformly across project-specific locations.

18Similarly to the outcome variable, I employ both the logarithmic and IHS transformations to account for the right-skewness in the distribution of the aid flows indicators.

3.3 Explanatory variables

19Population count. Estimates for the total population size within grid cells (i.e. at a resolution of 30 arc-minutes) are obtained from the Gridded Population of the World Version 3 (GPWv3) database, published by CIESIN-CIAT (2005). These are available for the years 1990, 1995, 2000, 2005, 2010, and 2015. I thus employ a linear interpolation between each interval to construct an annual population count measure across grids for the whole sample period.

20The more recent GPWv4 database (CIESIN, 2018) would serve as a more suitable population measure, as it supersedes its predecessor. Unfortunately, estimates are only available from the year 2000. To maximise sample size, I chose to use the GPWv3 indicator for the main estimations.

  • 7 The GPCP combined precipitation data were developed and computed by the NASA/Goddard Space Flight C (...)

21Climate characteristics. Yearly observations of total precipitation (in millimetres) in each grid cell are derived from the GPCP v.2.2 Combined Precipitation Data Set (Huffman et al., 2012).7 The original data only reported daily averages for each month; as a result, this was multiplied by the number of days in each month to obtain approximate monthly totals, from which yearly totals were estimated.

  • 8 Both the GPCP and GHCN Gridded V2 data were provided by the NOAA/OAR/ESRL PSD, Boulder, Colorado, U (...)

22Annual mean temperature estimates (in degrees Celsius), are obtained based on monthly meteorological statistics from GHCN/CAMS, developed at the Climate Prediction Center, NOAA/National Weather Service (Fan & van den Dool, 2008).8

23The severity of drought for the entirety of a cell’s rainy season (defined as the three consecutive months in which it on average rained the most during a year) is given by the Standardized Precipitation and Evapotranspiration Index (SPEI-3) value for the last month of the rainy season, from the SPEI Global Drought Monitor (Beguería et al., 2014). For each month, the SPEI-3 index measures deviations from long-term normal rainfall during the three preceding months. Values are standardized with mean 0 and standard deviation 1, where negative values indicate below local average rainfall, and vice versa.

24Conflict. The UCDP/PRIO Armed Conflict Dataset Version 18.1, published by Sundberg & Melander (2013), provides a comprehensive account of all individual events involving organized lethal violence, across grid cells and time. Events are categorised according to three types of conflict – state-based, non-state and one-sided violence – and each location is assigned a geo-reference precision code.

25In line with the aid indicators, I only include those observations with an associated precision code of 1 and 2. From this sample, I construct two grid- and year-specific conflict indicators: one indicates the number of unique conflict events which took place in a given grid and year, while the other indicates the total number of deaths across conflicts.

26Natural resources. I identify the presence of natural resources with a binary variable indicating whether petroleum, diamond or gem deposits where discovered within a given grid cell (Gilmore et al., 2005; Lujala et al., 2005; 2007; Lujala, 2009).

27Other indicators of poverty. To depict the standards of health across grid cells, I use two measures indicating the rate of infant mortality and the prevalence of child malnutrition, based from raster data from the SEDAC Global Poverty Mapping project (CIESIN, 2005a; 2005b). The former gives the number of children per 10,000 live births that die before reaching their first birthday (converted to a percentage), while the latter represents the percent of children under the age of 5 that are malnourished. Both indicators are a snapshot for the year 2000 only.

28In addition, I capture the degree of regional urbanization across grid cells using two proxies, which could be important factors in controlling for urban bias in aid allocation.

29The first measure denotes the total travel time (in minutes) by land transportation from the grid’s centroid to the nearest major city with at least 50,000 inhabitants (Uchida & Nelson, 2009). It is derived from a global high-resolution raster map of accessibility developed for the EU, using a combination of several sources collected between 1990 and 2005.

30The second indicator measures the spherical distance in kilometres from the grid cell centroid to the national capital city in the corresponding country (Weidmann et al., 2010). It is based on coordinate pairs of capital cities from the cShapes dataset v.0.4-2, and captures changes over time wherever relevant.

Table 1: Descriptive statistics of main variables across 10,674 grid cells (1995-2012)

Table 1: Descriptive statistics of main variables across 10,674 grid cells (1995-2012)

Note: Each observation is a grid/year. The travel time, child malnutrition and infant mortality variables are time-invariant.

Notes

1 All variables used in the estimations were obtained from the PRIO-GRID dataset, with the exception of the aid and population data. At the request of the developers, the original source of each variable will be cited in the main text.

2 Due to differences in panel size between variables and the lag specification in the baseline model, the effective sample size is reduced to the period 1996-2012, equating to a total of 181,458 observations.

3 Image and data processing by NOAA’s National Geophysical Data Center. DMSP data collected by US Air Force Weather Agency.

4 The IDA offers grants and highly concessional loans (i.e. Official Development Assistance) to the poorest countries, while the IBRD offers non-concessional loans (i.e. Other Official Flows) to middle-income and creditworthy countries. Unfortunately, the dataset does not distinguish between the two arms, and other ancillary data do not have consistent information to this regard, making it impossible to study potential differences in the effects on poverty outcomes between the two types of development assistance.

5 One project with 194 locations in Ethiopia (Road Sector Development Program Support Project) was removed from the final sample as there were missing data on both commitments and disbursements.

6 The results remain fairly unchanged when these observations are instead excluded from the estimations.

7 The GPCP combined precipitation data were developed and computed by the NASA/Goddard Space Flight Center’s Laboratory for Atmospheres as a contribution to the GEWEX Global Precipitation Climatology Project.

8 Both the GPCP and GHCN Gridded V2 data were provided by the NOAA/OAR/ESRL PSD, Boulder, Colorado, USA, from their website at http://www.esrl.noaa.gov/psd.

Acheter

Rechercher dans OpenEdition Search

Vous allez être redirigé vers OpenEdition Search