Mapping the interaction between development aid and stunting in - PDF document

Bosco et al. , 2017 28th IUSSP INTERNATIONAL POPULATION CONFERENCE Mapping the interaction between development aid and stunting in Nigeria Claudio Bosco 1,2* , Natalia Tejedor-Garavito 1,2 , Daniele de Rigo 3 , Carla Pezzulo 1,2 , Linus Bengtsson 1,2,4 , Andrew J Tatem 1,2 and Tomas J Bird 1,2 * Presenter Abstract For meeting sustainable development goals (SDGs) an improved 1 WorldPop, Department of Geography and Environment, understanding of geographic differences in health status, wealth and access to University of Southampton, resources is crucial. The equitable and efficient allocation of international aid Southampton, UK relies on knowing where funds are needed most. For instance, aid for poverty Full list of author information alleviation or financial access improvement requires knowledge of where the poor is available at the end of the are. Unfortunately, detailed, reliable and timely information on the spatial article distribution and characteristics of intended aid recipients in many low income countries are rarely available. This lack of information also hinders assessments of the impacts of aid; when presented at national scales, development and health indicators conceal important inequities, with the rural poor often least well represented. High-resolution data on key social and health indicators are therefore fundamental for targeting limited resources, especially where development funding has recently come under increased pressure. In this study, we show how modern statistical approaches can be used to maps for the distribution of indicators with a level of detail that can support geographically stratified decision-making. Using predictive modelling techniques, the rates of stunting in children under the age of five from Demographic and Health Surveys (DHS) geolocated cluster data were exploited to predict high-resolution maps (2008 – 2013) in Nigeria. An array of different modelling techniques was applied to produce prediction maps. These included Bayesian geostatistical models and machine learning techniques. An ensemble model was also exploited for aggregating the different modelling results. By combining these maps with information on the disbursement of aid for stunting alleviation in Nigeria (AidData database - http://aiddata.org/), we quantified both the distribution of aid relative to population characteristics related to stunting, and how aid disbursement interacts with changes in this index. In spite of the lack of exhaustive information related to aid disbursement, the results here demonstrate the potential of this approach. Keywords: modelling; development aid; malnutrition; machine learning; artificial neural networks; Bayesian geostatistical models Data The nationally representative indicator we investigated in this research comes from DHS survey data (http://dhsprogram.com, NPC, 2009, 2014). The DHS is a pro- gram of national household surveys implemented across various low- and middle-

Bosco et al. , 2017 Page 2 of 11 income countries (LMICs), collecting and analysing data on population coming from more than 300 surveys in over 90 countries. The measure of stunting we used in our research comes from the height-for-age Z-scores. Children whose height-for-age Z-score is below minus two standard devi- ations (-2 SD) from the median of the WHO (2006) reference population are con- sidered short for their age (stunted) and chronically malnourished. The geolocated cluster-level proportions of stunted children were used in our analyses (Figure 1a). We looked at stunting in children as this indicator can be linked to environmental factors leading to low caloric intake. Specifically, factors such as poverty and agricul- ture yields have been linked to stunting in past work (Gething et al. , 2015). Exploiting the relationship with covariate layers and accounting for spatial auto- correlation, we predicted stunting at locations where survey data were not available (Alegana et al. , 2015, Bosco et al. , 2017a, Golding et al. , 2017, Sedda et al. , 2015). Several physical (e.g. topography, aridity, potential evapotranspiration, land cover) and some social (population, ethnicity) covariate grids derived from public avail- able datasets were assembled and converted to a common 1km 2 spatial grid suitable to integration into the modelling architecture. We mainly focused on factors that have been shown to correlate with stunting (Bosco et al. , 2017a,b, Kinyoki et al. , 2016). Information on the distribution of aid moneys for alleviation of stunting in Nigeria came from the AidData database. AidData has a portal (http://aiddata.org/) of open data (Murray-Rust, 2008, Stallman, 2005) where the locations of investments at the sub-national level are available. For this research we used the 1.3.1 version of the Level 1 product, of all geocoded projects from the Development Assistance Database (DAD) Aid Information Management System (AIMS) managed for Nige- ria (AidData, 2016). This data set consists of geographical locations of the different investments identified in Nigeria, with (in most cases) their financial commitment and disbursement. The precision of the investment locations vary depending on the information available, ranging from data on the exact location of disbursement through to records that only indicate that the funds were allocated somewhere within the country as a whole, where in the latter case it is likely that the funding went to a government ministry or financial institution. Material and Methods Machine learning (Artificial neural networks, ANNs, Breiman, 2001, de Rigo et al. , 2005, Hornik et al. , 1989, Kreinovich, 1991) and Bayesian geostatistical (BGS) tech- niques (Gelman and Hill, 2006, Zaslavsky, 2002) exploiting nationally representative geolocated surveys and gridded spatial layers of covariates were applied to predict stunting in Nigeria at high spatial resolution. In order for the uncertainty to be mitigated, a robust ensemble model based on the stacked generalization (Wolpert, 1992) was proposed to aggregate different maps of stunting related to the year 2008. The stacking involves training a learning algorithm to combine the predictions of several other learning algorithms.

Bosco et al. , 2017 Page 3 of 11 The computational modelling architecture was implemented with free software (Stallman, 2009) mostly in a GNU/Linux computing environment. GNU Bash tools (Free Software Foundation et al. , 2010) were used to connect various intermediate data-transformation modules (D-TMs, de Rigo, 2013, 2015). ANNs relied on GNU Octave (Eaton et al. , 2008, Eaton, 2012, nnet package [1] ) and GNU R (AMORE package Castej´ on Limas, 2010, Venables et al. , 2018, also using a Windows ver- sion [2] ). BGS analysis exploited the package R-INLA (Rue et al. , 2009). The pro- cessing of the various arrays of data and models was based on a semantic modelling approach to split the analysis in a chain of simpler tasks and corresponding D-TMs. In particular, this array-based approach follows the Semantic Array Programming (SemAP) paradigm (de Rigo, 2012a,b, 2015). SemAP semantic checks were sys- tematically introduced in the code to mitigate inconsistencies between input data, parameters and outputs. Geospatial analysis was performed with ESRI ArcGIS [3] . To seamlessly integrate geospatial and array-based semantics, the SemAP applica- tion to geospatial problems (Geospatial Semantic Array Programming, GeoSemAP, de Rigo et al. , 2013, de Rigo, 2015) was exploited for its flexibility to easily cope with different geospatial scales and uneven arrays of data (Bosco et al. , 2015, Bosco and Sander, 2015, Caudullo, 2014, Mubareka et al. , 2014). The ensemble approach is a reproducible D-TM applied to the results of the array of models. Each map was obtained by applying heterogeneous models (ANNs and BGS), so as to increase design diversity. Having a robust gridded raster estimation of the distribution of stunting in different time intervals enabled quantification of the distribution of aid relative to need in space and time. Using our modelling approaches, we investigated if any spatio- temporal variation in stunting in children under age of five was related to the magnitude and distribution of aid (Figure 3). Due to a lack of spatial information within the AidData datasets (only less than 40 % of the projects registered locational [1] https://octave.sourceforge.io/nnet/ . [2] https://cran.r-project.org/bin/windows . [3] http://www.esri.com/arcgis . Table 1 Summary of database for all data available and those projects related to stunting in Nigeria. Data All projects Related to stunting Year of start (min) 1988 2002 year of end (max) 2020 2017 Donors 28 12 Total Projects 621 271 (263 with disbursement or commitment) Geocoded Projects 595 103 Locations 1843 483 Total Disbursements $6,255,493,636 $5,575,541,874 Total Commitments $2,144,374,320 $1,538,242,430 Geocoded Disbursements $6,093,125,384 $2,833,812,856 Geocoded Commitments $2,116,331,293 $717,178,638