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 1WorldPop, Department of

Geography and Environment, University of Southampton, Southampton, UK Full list of author information is available at the end of the article

Abstract For meeting sustainable development goals (SDGs) an improved understanding of geographic differences in health status, wealth and access to resources is crucial. The equitable and efficient allocation of international aid relies on knowing where funds are needed most. For instance, aid for poverty alleviation or financial access improvement requires knowledge of where the poor

• are. Unfortunately, detailed, reliable and timely information on the spatial

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 1km2 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

• pen 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

• f 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´

• n 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

• btained 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

Bosco et al., 2017 Page 4 of 11

information on money disbursement at admin1 level (Table 1)) different scenarios for re-allocating aid distribution were explored. Money without spatial information were re-allocated following, for example, proportions found in projects where spatial information was available, or on the basis of the number of people living in each of the districts.

Results and conclusions

The results highlight that accurate high resolution maps of stunting in children under age of five (Figure 1) can be produced. The accuracy of each of the ap- plied models was measured through the proportion of variance explained and the value of RMSE and MAE. The value of the explained variance through validation is higher than 60 % in mapping stunting in 2013 and close to 55 % in 2008. Because

• f their high predictive power these maps were summarized to policy-relevant ad-

ministrative units to support decision makers for planning and resource allocation (Figure 2). One of the main challenges in the AidData database lies in the lack of geocoding in the data. Many projects are not geocoded indicating either a lack of information, or allocation to a government department. Some scenarios to re-allocate money were tested, however in many cases it is impossible to test whether this reallocation of funds represents a plausible scenario. In each case we therefore had to make a best approximation. By analysing the maps of predicted stunting in Nigeria, rural areas present an higher proportion of stunted children than urban areas (Figure 1) with population centres showing the lowest values. A north-south trend is evident by comparing 2008 and 2013 maps. The north-west of the country shows an increasing in the proportion of children stunted while the south goes towards a consistent stunting alleviation. The uncertainty associated with both these maps is generally low. Greatest uncertainty is present in those regions with sparse data. Unfortunately when comparing aid disbursement with stunting in Nigeria, there was a significant lack of detailed geolocated information on aid disbursement (Table 1). As such, we were able to perform our analysis only at admin 1 level – the same level at which DHS data for the country are representative. The limited availability

• f data made a comparison of the aid spending between periods challenging. Above

average disbursement (relative to regions with similar stunting levels) was found in Sokoto and Borno states. The higher degree of expenditure in Borno is logical given the high degree of general poverty in this area, which is frequently subject to food shortages and intensive international aid work, as well as being subject to frequent violence from the Boko Haram militant group. The Sokoto region is also relatively impoverished, but no more so than other northern states such as Yobe, which we would also expect to receive substantial funding for development. Another stand-out state is the Federal Capital Territory of Abuja. Again, being a capital region it is likely that Abuja receives a substantial portion of support aimed at building logistic infrastructure, or as a hub from which aid is disbursed for minor projects.

Bosco et al., 2017 Page 5 of 11

In terms of places with lower funding than expected, the states of Yobe, Jigawa and Gombe are all located in areas of relatively high poverty. In the period 2007- 2012, many other regions have zero (or close to zero) reported aid disbursements. Many southern states such as Lagos, Kogi, Ondo, Osun and Ebonyi are regions with relatively lower stunting (especially in 2013) and so lower per capita aid is

• expected. However the very low values reported here seem excessively so and we

suspect either misreporting in many of these regions. In the period 2003-2007, the near-zero per-capita disbursement for most states is almost certainly false, probably linked with the lack of detailed spatial information for most of the projects. This work shows the value of exploiting advanced modelling architectures for com- bining geo-located household survey data with geospatial covariates layers. It al- lowed us to quantify the distribution of stunting in a low-income country and to estimate the associated uncertainty. With household surveys regularly undertaken, monitoring of stunting at country level and at a larger scale (continental or global) every few years is possible. This research presents analysis to help clarify where aid for stunting alleviation is needed most, how well this need has been covered by past funding and ultimately, whether the giving of aid is related to observable changes in stunting rates. From the perspective of funders, this level of end-to-end information will allow greater responsibility and responsiveness in the allocation of aid. At the most basic level, the use of maps in showing where aid is needed most will be a valuable tool for budgeting and allocation.

Author details

1WorldPop, Department of Geography and Environment, University of Southampton,

Southampton, UK.

2Flowminder Foundation, Stockholm, Sweden. 3Maieutike Research

Initiative, Milano, Italy.

4Karolinska Institute, Dept. of Public Health Sciences, Stockholm,

Sweden.

References

[1] AidData, 2016. Nigeria AIMS geocoded research release, version 1.3.1. In: AidData

• Datasets. AidData, Williamsburg, VA and Washington, DC.

http://aiddata.org/data/nigeria-aims-geocoded-research-release-level-1-v1-3-1, http://aiddata.org/research-datasets , INRMM-MiD:14546755 [2] Alegana, V. A., Atkinson, P. M., Pezzulo, C., Sorichetta, A., Weiss, D., Bird, T., Erbach-Schoenberg, E., Tatem, A. J., 2015. Fine resolution mapping of population age-structures for health and development applications. Journal of The Royal Society Interface 12 (105), 20150073+. https://doi.org/10.1098/rsif.2015.0073 , INRMM-MiD:14546782 [3] Bosco, C., de Rigo, D., Dewitte, O., Poesen, J., Panagos, P., 2015. Modelling soil erosion at European scale: towards harmonization and reproducibility. Natural Hazards and Earth System Science 15 (2), 225-245. https://doi.org/10.5194/nhess-15-225-2015 , INRMM-MiD:13508255 [4] Bosco, C., Sander, G. 2015. Estimating the effects of water-induced shallow landslides on soil erosion. IEEE Earthzine 7, 910137+. http://earthzine.org/?p=910137, https://doi.org/10.1101/011965 , INRMM-MiD:13455081

Bosco et al., 2017 Page 6 of 11

[5] Bosco, C., Alegana, V. A., Bird, T., Pezzulo, C., Bengtsson, L., Sorichetta, A., Steele, J., Hornby, G., Ruktanonchai, C. W., Ruktanonchai, N. W., Wetter, E., Tatem, A. J.,

• 2017. Exploring the high-resolution mapping of gender-disaggregated development
• indicators. Journal of The Royal Society Interface 14 (129), 20160825+.

https://doi.org/10.1098/rsif.2016.0825 , INRMM-MiD:14332252 [6] Bosco, C., Alegana, V., Bird, T., Pezzulo, C., Hornby, G., Sorichetta, A., Steele, J., Ruktanonchai, C., Ruktanonchai, N., Wetter, E., Bengtsson, L., Tatem, A. J., 2017. Mapping indicators of female welfare at high spatial resolution. Tech. rep., WorldPop project, Flowminder Foundation, Stockholm, Sweden. http://data2x.org/wp-content/uploads/2017/02/Mapping-Indicators-of-Female-Welfare- at-High-Spatial-Resolution.pdf , INRMM-MiD:14335578 [7] Breiman, L., 2001. Statistical modeling: the two cultures (with comments and a rejoinder by the author). Statistical Science 16 (3), 199-231. https://doi.org/10.1214/ss/1009213726 , INRMM-MiD:7796484 [8] Castejon Limas, M., Ordieres Mere, J. B., Gonzalez Marcos, A., de Pison Ascacibar, F. J., Pernia Espinoza, A. V., Alba Elias, F., Perez Ramos, J. M., 2014. AMORE: A MORE flexible neural network package. In: The Comprehensive R Archive Network. https://cran.r-project.org/web/packages/AMORE/index.html , INRMM-MiD:14565897 [9] Caudullo, G., 2014. Applying Geospatial Semantic Array Programming for a reproducible set of bioclimatic indices in Europe. IEEE Earthzine 7 (2), 877975+. http://www.earthzine.org/?p=877975 https://doi.org/10.1101/009589 , INRMM-MiD:13385094 [10] de Onis, M., Onyango, A., Borghi, E., Siyam, A., Pinol, A., Garza, C., Martines, J., Martorell, R., Victora, C. G., Bhan, M. K., Ara´ ujo, C. L., Lartey, A., Owusu, W. B., Bhandari, N., Norum, K. R., Bjoerneboe, G.-E. A., Mohamed, A. J., Dewey, K. G., Belbase, K., Black, M., Chumlea, W., Cole, T., Frongillo, E., Grummer-Strawn, L., Shrimpton, R., Van den Broeck, J., Pan, H., Rigby, R., Stasinopoulos, M., van Buuren, S., Albernaz, E., Tomasi, E., de C´ assia Fossati da Silveira, R., Nader, G., Sagoe-Moses, I., Gomez, V., Sagoe-Moses, C., Taneja, S., Rongsen, T., Chetia, J., Sharma, P., Bahl, R., Baerug, A., Tufte, E., Rudvin, K., Nysaether, H., Alasfoor, D., Prakash, N. S., Mabry, R. M., Al Rajab, H. J., Helmi, S. A., Nommsen-Rivers, L. A., Cohen, R. J., Heinig, M. J.,

• 2006. WHO child growth standards: length/height-for-age, weight-for-age,

weight-for-length, weight-for-height and body mass index-for-age - Methods and

• development. World Health Organization, Geneva, Switzerland, 312 pp.

http://www.who.int/childgrowth/standards/technical report/en/ , INRMM-MiD:14547506 [11] de Rigo, D., Castelletti, A., Rizzoli, A. E., Soncini-Sessa, R., Weber, E., 2005. A selective improvement technique for fastening neuro-dynamic programming in water resources network management. In: Proceedings of the 16th IFAC World Congress (IFAC Praha 2005). http://folk.ntnu.no/skoge/prost/proceedings/ifac2005/Papers/Paper4269.html (archived at http://www.webcitation.org/5zuFRunw5) , INRMM-MiD:10793225 [12] de Rigo, D., 2012. Semantic Array Programming for environmental modelling: application of the Mastrave library. In: Seppelt, R., Voinov, A. A., Lange, S., Bankamp,

• D. (Eds.), International Environmental Modelling and Software Society (iEMSs) 2012

International Congress on Environmental Modelling and Software - Managing Resources of a Limited Planet: Pathways and Visions under Uncertainty, Sixth Biennial Meeting. pp. 1167-1176. https://scholarsarchive.byu.edu/iemssconference/2012/Stream-B/69/ , INRMM-MiD:12227965 [13] de Rigo, D., 2012b. Semantic Array Programming with Mastrave - Introduction to semantic computational modelling. The Mastrave project. http://mastrave.org/doc/MTV-1.012-1/ , INRMM-MiD:11744308 [14] de Rigo, D., 2013. Software uncertainty in integrated environmental modelling: the role of semantics and open science. Geophysical Research Abstracts 15, 13292+. https://doi.org/10.6084/m9.figshare.155701 , INRMM-MiD:12794802

Bosco et al., 2017 Page 7 of 11

[15] de Rigo, D., Corti, P., Caudullo, G., McInerney, D., Di Leo, M., San-Miguel-Ayanz, J.,

• 2013. Toward open science at the European scale: Geospatial Semantic Array

Programming for integrated environmental modelling. Geophysical Research Abstracts 15, 13245+. https://doi.org/10.6084/m9.figshare.155703, RePEc:pra:mprapa:44194 , INRMM-MiD:11977126 [16] de Rigo, D., 2015. Study of a collaborative repository of semantic metadata and models for regional environmental datasets’ multivariate transformations. Ph.D. thesis, Politecnico di Milano, Milano, Italy. http://hdl.handle.net/10589/101044 , INRMM-MiD:13769492 [17] Eaton, J. W., Bateman, D., Hauberg, S., 2008. GNU Octave manual version 3: a high-level interactive language for numerical computations. Network Theory. http://www.network-theory.co.uk/docs/octave3/ , INRMM-MiD:9115371 [18] Eaton, J. W., 2012. GNU octave and reproducible research. Journal of Process Control 22 (8), 1433-1438. https://doi.org/10.1016/j.jprocont.2012.04.006 , INRMM-MiD:10799158 [19] Free Software Foundation, Abernathy, W., Anderson, G., Anhalt, C., Bai, L., Baptista, V., Barakat, M., Batini, E., Bavier, E., Becher, J., Becher, V., Biberg Kristensen, L., Boyle, T., Brown, P., Catkan, B., Charzat, S., Cherlin, E., Compall, S., Davies, P., de Rigo, D., Devarajan, D., Dorrington, M., Fern´ andez Pi˜ nas, D., Fortin, P., Gauland, M., Gillmor, D. K., Goh, T., Golin, R., Haichao, X., Hall, D., Hyde, A., Innoccenti, B., Jonsson, A., Kibbe, D., Kochenderfer, V., Krampis, N., Lee, M., Lockwood-Childs, S., Lucifredi, F., Magallon, M., Martin, A., McConnaughey, W., McNeely, M., Mengel, M., Merriam, W., Nair, S., Newell, M., Oliver, J., Oram, A., Oranen, J., Pawson, D., Prastowo, T., Ravikumar, S., Revilak, S., Schaumburg, S., Skwarecki, B., Sullivan, J., Tange, O., Tenney, K., Thahir, S., Walck, S., Weissmann, B., Wells, S., Williams, C., Woodacre, B., Woof, J., et al., 2010. Introduction to the Command Line. FLOSS

• Manuals. http://write.flossmanuals.net/command-line , INRMM-MiD:13644772

[20] Gelman A., Hill J., 2006. Data analysis using regression and multilevel/hierarchical

• models. Cambridge University Press. ISBN:9780521686891 , INRMM-MiD:9983048

[21] Gething, P., Tatem, A., Bird, T., Burgert-Brucker, C. R., 2015. Creating spatial interpolation surfaces with DHS data. No. 11 in DHS Spatial Analysis Reports. ICF International, Rockville, Maryland, USA. https://dhsprogram.com/publications/publication-SAR11-Spatial-Analysis-Reports.cfm , INRMM-MiD:14547328 [22] Golding, N., Burstein, R., Longbottom, J., Browne, A. J., Fullman, N., Osgood-Zimmerman, A., Earl, L., Bhatt, S., Cameron, E., Casey, D. C., Dwyer-Lindgren, L., Farag, T. H., Flaxman, A. D., Fraser, M. S., Gething, P. W., Gibson, H. S., Graetz, N., Krause, L. K., Kulikoff, X. R., Lim, S. S., Mappin, B., Morozoff, C., Reiner, R. C., Sligar, A., Smith, D. L., Wang, H., Weiss, D. J., Murray, C. J. L., Moyes, C. L., Hay, S. I., 2017. Mapping under-5 and neonatal mortality in Africa, 2000–15: a baseline analysis for the Sustainable Development Goals. The Lancet 390 (10108), 2171-2182. https://doi.org/10.1016/S0140-6736(17)31758-0 , INRMM-MiD:14546798 [23] Hornik, K., Stinchcombe, M., White, H., 1989. Multilayer feedforward networks are universal approximators. Neural Networks 2 (5), 359-366. https://doi.org/10.1016/0893-6080(89)90020-8 , INRMM-MiD:1700245 [24] Kinyoki, D. K., Kandala, N.-B., Manda, S. O., Krainski, E. T., Fuglstad, G.-A., Moloney, G. M., Berkley, J. A., Noor, A. M., 2016. Assessing comorbidity and correlates

• f wasting and stunting among children in Somalia using cross-sectional household

surveys: 2007 to 2010. BMJ Open 6 (3), e009854+. https://doi.org/10.1136/bmjopen-2015-009854 , INRMM-MiD:14546857 [25] Kreinovich, V. Y., 1991. Arbitrary nonlinearity is sufficient to represent all functions by neural networks: a theorem. Neural Networks 4 (3), 381-383. https://doi.org/10.1016/0893-6080(91)90074-f , INRMM-MiD:10833268

Bosco et al., 2017 Page 8 of 11

[26] Mubareka, S., Jonsson, R., Rinaldi, F., Fiorese, G., San-Miguel-Ayanz, J., Salln¨ as, O., Baranzelli, C., Pilli, R., Lavalle, C., Kitous, A., 2014. An integrated modelling framework for the forest-based bioeconomy. IEEE Earthzine 7 (2), 908802+. http://earthzine.org/?p=908802 https://doi.org/10.1101/011932 , INRMM-MiD:13494913 [27] Murray-Rust, P., 2008. Open data in science. Serials Review 34 (1), 52-64. https://doi.org/10.1016/j.serrev.2008.01.001 , INRMM-MiD:3753449 [28] National Population Commission (NPC) [Nigeria], ICF Macro, 2009. Nigeria demographic and health survey 2008. National Population Commission and ICF Macro, Abuja, Nigeria. https://dhsprogram.com/publications/publication-fr222-dhs-final-reports.cfm , INRMM-MiD:14547611 [29] National Population Commission (NPC) [Nigeria], ICF Macro, 2014. Nigeria demographic and health survey 2013. National Population Commission and ICF Macro, Abuja, Nigeria. https://dhsprogram.com/publications/publication-fr293-dhs-final-reports.cfm , INRMM-MiD:14547615 [30] Rue, H., Martino, S., and Chopin, N., 2009. Approximate Bayesian inference for latent Gaussian models using integrated nested Laplace approximations (with discussion). Journal of the Royal Statistical Society, Series B , 71(2):319-392. [31] Sedda, L., Tatem, A. J., Morley, D. W., Atkinson, P. M., Wardrop, N. A., Pezzulo, C., Sorichetta, A., Kuleszo, J., Rogers, D. J., 2015. Poverty, health and satellite-derived vegetation indices: their inter-spatial relationship in West Africa. International Health 7 (2), 99-106. https://doi.org/10.1093/inthealth/ihv005 , INRMM-MiD:14546815 [32] Stallman, R. M., 2005. Free community science and the free development of

• science. PLoS Med 2 (2), e47+. https://doi.org/10.1371/journal.pmed.0020047 ,

INRMM-MiD:11232893 [33] Stallman, R. M., 2009. Viewpoint: why ”open source” misses the point of free

• software. Communications of the ACM 52 (6), 31-33.

https://doi.org/10.1145/1516046.1516058 , INRMM-MiD:9506694 [34] Venables, W.N., Smith, D.M., R Core Team, 2018. An Introduction to R - Notes on R: a programming environment for data analysis and graphics, version 3.4.4 Edition. R Core Team, 105 pp. https://cran.r-project.org/doc/manuals/r-release/R-intro.pdf , INRMM-MiD:14557691 [35] Wolpert, D. H., 1992. Stacked generalization. Neural Networks 5 (2), 241-259. https://doi.org/10.1016/s0893-6080(05)80023-1 , INRMM-MiD:3157859 [36] Zaslavsky, A. M., 2002. Hierarchical Bayesian modeling. In: Press, S. J. (Ed.), Subjective and Objective Bayesian Statistics: Principles, Models, and Applications, Second

• Edition. John Wiley & Sons, Inc., Hoboken, NJ, USA, pp. 336-358.

https://doi.org/10.1002/9780470317105.ch14 , INRMM-MiD:14546817

Bosco et al., 2017 Page 9 of 11

Figures

Figure 1 Map of the cluster-level survey data (top row), for the indicator of stunting in children under age of five in Nigeria. Map of the stunting in Nigeria at 1 km2 resolution (middle row) and related uncertainty maps (interdecile) (bottom row). The map of stunting related to year 2008 was created by applying an ensemble approach on the output of an array of models.

Bosco et al., 2017 Page 10 of 11 Figure 2 Map of stunting in children under age of five at administrative level 2 (weighted by population) for the year 2013.

Bosco et al., 2017 Page 11 of 11 Figure 3 Top: Reported per capita disbursement for each state in Nigeria, sorted according to the percentage of stunted children under age of 5 (2008 values). Bottom: Proportion of the population weighted stunting in children under age of 5 at administrative 1 level for the year 2008 and 2013 in Nigeria.