WPS8688 Policy Research Working Paper 8688 Understanding the Geographical Distribution of Stunting in Tanzania A Geospatial Analysis of the 2015–16 Demographic and Health Survey George Joseph Peter W. Gething Samir Bhatt Sophie C.E. Ayling Water Global Practice January 2019 Policy Research Working Paper 8688 Abstract Tanzania is home to the third highest population of stunted geo-referenced Demographic and Health Survey in con- children in Sub-Saharan Africa, with about 2.7 million chil- junction with relevant spatially gridded covariate data, such dren under the age of five failing to reach their full potential as nighttime lights, water and sanitation access, vegetation of growth attainment compared with the reference pop- index, travel time, and so on. Geospatial techniques, such as ulation as per the World Health Organization standards. model-based statistics and Bayesian inference implemented Several studies have shown that stunted growth during using the INLA algorithm, along with appropriate model childhood entraps the future of children in a vicious circle validation exercises are employed to develop high-resolu- of recurrent diseases, reduced human development, and tion maps of stunting in Tanzania at 1×1-kilometer spatial lower earnings, thus increasing their likelihood of being resolution. The maps show that areas of consistently high poor when they grow up. To reduce stunting, the Govern- stunting rates tend to be more common in rural parts of the ment of Tanzania and development partners are introducing country, especially throughout the western and southwest- a convergence of multisectoral interventions adapted to ern border areas. There is high prevalence of low stunting local needs. However, the existing stunting data are repre- in the urban areas around Dar es Salaam, Arusha, and sentative only at higher administrative levels, thus making it Dodoma, as well as in the south of Lake Victoria. difficult to implement these efforts. The paper uses the 2016 This paper is a product of the Water Global Practice. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/research. The authors may be contacted at goseph@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Understanding the Geographical Distribution of Stunting in Tanzania: A Geospatial Analysis of the  2015–16 Demographic and Health Survey1  George Josepha, Peter W. Gethingi§,  Samir Bhattiii, Sophie C.E. Aylingii i. Water Global Practice Group, World Bank Group, 1818 H Street NW, Washington DC, 20433, USA. ii. Oxford Big Data Institute, Li Ka Shing Centre for Health Information and Discovery, Nuffield Department of Medicine, University of Oxford, Oxford OX3 7LF, UK. iii. Department of Infectious Disease Epidemiology, Imperial College London, London W2 1PG, UK. § Corresponding author JEL: O55; I14; C31  Key Words: Stunting; Geo‐statistics; Tanzania; Africa, Health Policy  1  This work was made possible by the Swedish International Development Cooperation Agency and benefited  from funding from the Government of Japan through the Japan Trust Fund for Scaling Up Nutrition.    Understanding the Geographical Distribution of Stunting in Tanzania: A Geospatial Analysis of the  2015–16 Demographic and Health Survey    Background  Stunting, defined as being more than two standard deviations below the World Health Organization  (WHO) Child Growth Standards median,1 arises from a range of factors including low breastfeeding  rates  and  malnutrition  among  mothers,  inadequate  health  care  or  poor  access  to  safe  water  and  sanitation leading to a range of acute or subclinical infections in early life. Globally, around 155 million  children are estimated to be stunted, and international targets have been established to reduce global  stunting by 40% by 20252,3  as one component of achieving the SDG goal 2.2 of eliminating all forms  of malnutrition by 2030.4 However, the current average annual rate of reduction in stunting of 2.3%  falls short of the 4% required if this goal is to be met.5   Recent  World  Bank  estimates  suggest  that  the  per  capita  income  penalty  incurred  among  contemporary workforces as a result of failure to eliminate stunting in their childhood amounts, on  average across all countries, to around 7% of gross domestic product (GDP) per capita. In Sub‐Saharan  Africa  and  South  Asia,  these  figures  rise  to  around  9%  to  10%.6    Several  studies  have  shown  that  stunted growth during childhood entraps the future of children in a vicious circle of recurrent diseases,  reduced human development7,8,9 and lower earnings.10,11 This, in turn, can increase the likelihood of  being poor and even lead to the transmission of poverty to the next generation.12,13  Like many low‐income countries, stunting is widespread in Tanzania. An estimated 2.7 million children  in Tanzania are stunted, making the country home to the third highest population of stunted children  in Sub‐Saharan Africa, after Ethiopia and the Democratic Republic of Congo. Although multisectoral  efforts to reduce stunting rates have yielded some progress (with prevalence among children under 5  declining from an estimated 42% in 201014 to 34% in 201615), there is a renewed call to accelerate this  progress and further reduce stunting nationwide through a host of nutrition‐specific and nutrition‐ sensitive  interventions,  particularly  focusing  on  the  first  1,000  days  of  life.16,17  These  multisectoral  interventions—spanning  the  health,  education,  social  protection,  water  supply,  and  sanitation  sectors—are intended to converge at the local level for maximum effectiveness.18 As part of these  efforts, there is an increasing recognition of the need to understand the geographic heterogeneity of  stunting across the country at a granular level. This granularity supports strategies that are optimally  2    targeted  to  populations  in  greatest  need  and  tailored  to  local  conditions,  thereby  promoting  the  efficiency and impact of program efforts.  The  emergence  of  standardized  cross‐sectional  household  surveys  such  as  the  Demographic  and  Health  Surveys  (DHS)  Program,19  which  now  routinely  include  georeferencing  (i.e.,  latitude  and  longitude coordinates), provides a new opportunity to understand subnational geographic variations  in a wide range of socioeconomic, health and demographic indicators, including stunting. A recent  DHS  survey  has  been  undertaken  in  Tanzania15  and,  like  nearly  all  others,  is  configured  to  provide  representative estimates of stunting rates at the first subnational unit level (regions in Tanzania), but  not  below.  However,  geospatial  methods  have  been  established  that  can  use  the  georeferenced  survey data alongside relevant spatially gridded covariate data to reliably estimate DHS indicators at  much finer levels of spatial disaggregation.20,21   The  use  of  georeferenced  survey  data  with  geospatial  models  to  generate  interpolated  surfaces  is  perhaps most established in the field of infectious diseases.22,23,24,25,26 In recent years, these methods  have also been developed and applied to a broader set of social, demographic, and health contexts.  This includes the use of both DHS and World Bank Living Standards Measurement Survey data, among  other  sources,  to  create  high‐resolution  poverty  maps  for  a  range  of  Sub‐Saharan  African  countries.27,28,29,30  Most  recently,  the  method  has  been  applied  to  mapping  water  and  sanitation  indicators in Nigeria based on the 2015 Nigeria Water and Sanitation Survey.31 This study describes  the use of such an approach to generate a new high‐resolution map of stunting prevalence in Tanzania.    Methods  Data  Data on stunting were obtained from the Tanzania DHS 2015‐16.15 This consisted of a two‐stage cluster  randomized  survey  design  and  a  final  sample  of  9,846  children  under  the  age  of  five  for  whom  anthropometric and age data were obtained and stunting rates for those children computed.  Geospatial  covariates  (spatial  gridded  data  for  which  values  are  known  at  all  1×1km  pixels  across  Tanzania) were also included in the model, as they may partially explain the observed spatial variation  in stunting rates and thus allow more accurate predictions at locations without DHS survey data. An  initial set was identified based on previous performance in predicting poverty, water and sanitation  indicators  and  other related  metrics  in  Sub‐Saharan African  countries.32,33 These  included  satellite‐ derived measures of vegetation density (Enhanced Vegetation Index, EVI), temperature (Land Surface  Temperature,  LST),  land  wetness  (the  Tasselled  Cap  Wetness  index,  TCW),  and  land  brightness  3    (Tasselled  Cap  Brightness  index,  TCB),  as  measured  by  NASA’s  MODIS  system  (http://modis.gsfc.nasa.gov/); aridity, as measured by the CGIAR project (http://csi.cgiar.org/Aridity/);  brightness of night‐time lights measured by the DMSP OLS project (https://ngdc.noaa.gov/eog/); and  a  delineation  of  urban  versus  rural  areas  constructed  by  the  Nelson  Institute  (https://nelson.wisc.edu/sage/data‐and‐models/schneider.php).  A  final  covariate  was  constructed  based on physical access to improved water sources. This covariate was created for the current study  by, first, using a ‘friction surface’ that estimates the time required to traverse each 1×1km pixel across  Tanzania.34 This varied according to the type of landcover, topography, and the layout of the road and  wider transport network across the country. The friction surface was then used in a least‐cost path  algorithm  to  estimate  the  likely  journey  time  from  the  center  of  each  1×1km  pixel  to  the  nearest  functioning water point (such as a well, bore hole, or pump). The latitude and longitude, as well as  level of functionality, of every such water point in Tanzania were recorded as part of the Water Point  Mapping project undertaken by The Ministry of Water and Irrigation, Government of Tanzania.35  Model‐based Geostatistics framework  The predictive approach used in this study to generate a fine‐scale map of stunting prevalence across  Tanzania was based on a body of statistical theory known as model‐based geostatistics (MBG)36,37.  MBG models are a class of generalized linear mixed model, with an approximation of a multivariate  Normal random field (i.e., a Gaussian process) used as a spatially autocorrelated random effect term.  The proportion of children who were recorded as being stunted , Y  xi  , at each location in Tanzania   was modeled as a transformation   of a spatially structured field superimposed with additional  random variation  . The count of individuals stunted   from the total sample of   in each survey  cluster was modeled as a conditionally independent binomial variate given the unobserved underlying  Y  xi   value. The spatial component was represented by a stationary Gaussian process   with  mean   and covariance  . The unstructured component   was represented as Gaussian with zero  mean  and  variance  .  Both  the  inference  and  prediction  stages  were  coded  using  the  INLA  framework,38 primarily in the R programming language.39  The mean component   was modeled as a linear function of the  n geospatial covariates described  above,    ,where    was  a  vector  consisting  of  a  constant  and  the  covariates  indexed  by  spatial  location  ,  and      was  a  corresponding  vector  of  regression coefficients. Each covariate was converted to z‐scores before analysis. Covariance between  spatial locations was modeled using a Matern covariance function:    4     1  d ( xi ; x j )   d ( xi ; x j )    C  d ( xi ; x j )    2  1  2  K  2      2         Where  d(xi ; xj ) is the geographical separation between two points;   , ,  are parameters of the  covariance function defining, respectively, its amplitude, degree of differentiability, and scale;  K is  the modified Bessel function of the second kind of order  , and   is the gamma function.  Bayesian  inference  was  implemented  using  the  INLA  algorithm  to  generate  approximations  of  the  marginal posterior distributions of the stunting prevalence  Y  xi  at each location on a regular 1×1 km  spatial grid across Tanzania and of the unobserved parameters of the mean, covariance function and  Gaussian random noise component. At each location, the posterior distribution was summarized using  the posterior mean as a point estimate and the width of the posterior 95% credible intervals (CIs) as  a metric of uncertainty, and maps were generated of each of these metrics in ArcGIS 10.4.  While 1×1km resolution maps provide the most fine‐grained picture of variation in stunting across the  country, it is also useful to summarize these patterns at higher levels of aggregation corresponding to  administrative unit levels at which program planning, implementation and decision making are carried  out. Aggregated stunting prevalence and the count of stunted children was therefore calculated at  regional  (1st  sub‐national  administrative  level)  and  district  (2nd)  levels,  using  an  underlying  gridded  population surface to provide pixel‐level population denominator values.40  The predictive performance of the model was assessed via out‐of‐sample validation. A four‐fold hold‐ out procedure was implemented whereby 25% of the data points were randomly withdrawn from the  data  set,  the  model  run  in  full  using  the  remaining  75%  of  data,  and  the  predicted  values  at  the  locations  of  the  hold‐out  data  compared  to  their  observed  values.  This  was  repeated  four  times  without replacement such that every data point was held out once across the four validation runs.  Standard validation statistics were computed as measures of model precision (mean absolute error)  and bias (mean error).  Results  Model validation  Figure 1 shows a scatterplot plotting values predicted in the out‐of‐sample validation exercise (x‐axis)  against the corresponding observed values that were held out (y‐axis). Points close to the 1:1 line were  5    therefore predicted more accurately than those further away. The plot reveals no systematic tendency  for over or under prediction at any level of observed prevalence. The validation statistics (Table 1)  further  indicate  a  relatively  well  performing  model,  with  reasonable  correspondence  between  predicted and observed values (correlation 63%), precision (mean absolute error 11%) and accuracy  (mean square error 2%). The latter two metrics are expressed in the same units as the predictions –  i.e. prevalence on a scale of 0‐100%.  Pixel‐level predictions of stunting prevalence  Figure 2 shows the predicted map of stunting prevalence in children under five years old, constructed  using the mean of the predicted posterior distribution within each 1×1km grid cell. Many areas of the  country are within the yellow part of the color scale, which is similar to the national mean stunting  prevalence reported in the DHS survey of 34% [DHS, 2016]. Areas of notably lower stunting prevalence  (shown in the greener shades) include the predominately urban areas around Dar es Salaam, Arusha  and Dodoma, as well as a broader region to the south of Lake Victoria centered around the city of  Tabora. Areas of consistently high stunting rates (red shades) tended to be more common in the more  rural parts of the country, especially throughout the western and southwestern border areas.  Figure 3 shows how the uncertainty of the stunting prevalence predictions varies geographically. The  uncertainty  metric  used  is  the width  of the  posterior  95% credible  interval  for  each grid  cell. Cells  where prevalence could be predicted with high certainty had narrow posterior distributions around  the  point estimate,  and thus the  range of  possible  prevalence  values between  the  2.5th and  97.5th  percentiles  of  the  distribution  was  small,  and  vice‐versa.  Areas  predicted  with  relatively  low  prevalence (green in Figure 2) tended to also have smaller uncertainty, represented by the blue shades  in Figure 3, while areas of higher stunting tended to be more uncertain, shown in the purple and pink  shades in Figure 3. Even in these regions, however, the 95% credible interval tended to be no wider  than 10%, meaning the model could predict stunting prevalence with 95% probability to within a range  of 10 percentage points.   Aggregated estimates of stunting   Figure 4 shows estimates of stunting prevalence and the number of stunted children aggregated to  the Tanzanian 1st (Regions) and 2nd (Districts) administrative levels. While the pixel‐level map shown  in Figure 2 provides the most granular, fine detailed, prediction of the landscape of stunting across  the country, the aggregated prevalence maps (top row) provide a summary of this spatial pattern at  different levels that may more directly inform programmatic planning. Both the District and Region  maps highlight how high‐prevalence areas are concentrated towards the southwest, with the lowest  6    levels in the central and northeastern parts of the country. The maps of estimated numbers of stunted  children (bottom row) show rather different patterns to the prevalence maps. This is because these  maps  combine  both  relative  frequency  of  stunting  with  the  absolute  number  of  children  living  in  different  Districts  and  Regions.  This  means  that  many  parts  of  Tanzania  with  high  prevalence,  particularly in the southwest, do not feature prominently because they are very sparsely populated,  while other regions of only intermediate prevalence nevertheless contain large numbers of stunted  children because of higher population density.   Discussion  Given the importance of reducing stunting through convergent multisectoral approaches, there is  a  growing need to generate high‐resolution predictions of stunting at the local level.  This study has  demonstrated  the  use  of  cross‐sectional,  georeferenced,  household  survey  data  in  a  geospatial  modeling framework to generate highly granular predictions of the landscape of stunting in Tanzanian  children.  The  map  reveals  the  considerable  spatial  heterogeneity  in  stunting  across  Tanzania,  in  particular  the  stark  urban‐rural  divide,  with  cities  having  notably  lower  average  levels  of  stunting  relative  to  even  nearby  rural  areas.  Similarly,  areas  proximal  to  major  transport  routes  across  the  country tended to have lower average levels of stunting than the most remote rural areas. A rigorous  model performance  assessment has demonstrated that the map is predicted with reasonably high  accuracy at the level of individual pixels, and an accompanying map of uncertainty indicates those  places where predictions could be made with more or less certainty.   Aggregation of this pixel‐level map to administrative units provides a simple summary of prevalence  rates within subsections of the country that may correspond to the units at which operational planning  to counter childhood stunting is carried out and programs implemented. This aggregation also allows  the  straightforward  estimation  of  the  number  of  children  stunted  within  each  unit  and,  taken  together, the two measures  (prevalence  and  number stunted) provide  distinct and complimentary  information for decision making.    7    References                                                              1. World Health Organization (2018). World Health Organization (WHO) Child Growth Standards  (http://www.who.int/childgrowth/en/, accessed 6 February 2018).  2. World Health Organization (2012). Resolution WHA65.6. Comprehensive implementation plan on maternal,  infant and young child nutrition. In: Sixty‐fifth World Health Assembly Geneva, 21–26 May 2012. Resolutions  and decisions, annexes. Geneva: World Health Organization.  3. World Health Organization (2018). Global targets 2025 To improve maternal, infant and young child  nutrition. (www.who.int/nutrition/topics/nutrition_globaltargets2025/en/, accessed 6 February 2018).  4. United Nations (2015). Transforming our world: the 2030 Agenda for Sustainable Development.  (https://sustainabledevelopment.un.org/post2015/transformingourworld, accessed 6 February 2018).  5. Development Initiatives (2017). Global Nutrition Report 2017: Nourishing the SDGs. Bristol, UK:  Development Initiatives.  6. Galasso E, and Wagstaff A. (2017). The Economic Costs of Stunting and How to Reduce Them. World Bank  Policy Research Note, PRN/17/05, March 2017.  7. Currie J. (2009). Healthy, wealthy, and wise: socioeconomic status, poor  health in childhood, and human  capital development. J. Econ. Lit., 47, 87–122.  8. Alderman H., Behrman JR (2006). Reducing the incidence of low birth weight in low‐income countries has  substantial economic benefits. World Bank Res. Obs. 21, 25–48.  9. Alderman H, Hoddinot J, Kinsey B (2006). Long term consequences of early childhood malnutrition. Oxford  Econ. Papers 58, 450–474.  10. Vogl T. (2014). Height, skills, and labor market outcomes in Mexico. J. Dev.Econ. 107, 84–96.  11. Lawson N, and Spears D (2016). What doesn't kill you makes you poorer: Adult wages and early‐life  mortality in India. Economics & Human Biology 22: 1‐16.  12. Harper C, Marcus R, Moore K (2003) Enduring poverty and the conditions of childhood: lifecourse and  intergenerational poverty transmissions. World Development, 31, 535‐554.  13. Grantham‐McGregor S, Cheung YB, Cueto S, Glewwe P, Richter L, Strupp B (2007). Developmental potential  in the first 5 years for children in developing countries. The lancet 369, 60‐70.  14. National Bureau of Statistics (NBS) [Tanzania] and ICF Macro (2011). Tanzania Demographic and Health  Survey 2010. Dar es Salaam, Tanzania: NBS and ICF Macro.  15. Ministry of Health, Community Development, Gender, Elderly and Children (MoHCDGEC) [Tanzania  Mainland], Ministry of Health (MoH) [Zanzibar], National Bureau of Statistics (NBS), Office of the Chief  Government Statistician (OCGS), and ICF (2016) Tanzania Demographic and Health Survey and Malaria  8                                                                                                                                                                                              Indicator Survey (TDHS‐MIS) 2015‐16. Dar es Salaam, Tanzania, and Rockville, Maryland, USA:MoHCDGEC,  MoH, NBS, OCGS, and ICF.  16  World Bank (2013). "Improving Nutrition Through Multisectoral Approaches." Washington DC: The  World Bank.   p://documents.worldbank.org/curated/en/625661468329649726/Improvingnutrition- through-multisectoral-approaches    17  Hoddinott, J.  The Economics of Reducing Undernutrition in Sub‐ Saharan Africa: Global Panel on   on Agriculture and Food Systems for Nutrition Working Paper     http://glopan.org/sites/default/files/Global_Panel_Working_Paper.pdf    18. UNICEF (1990). Strategy for improved nutrition of women and children in developing countries. A  UNICEF Policy Review. New York: UNICEF.  19. The Demographic and Health Surveys Program. (https://dhsprogram.com/, accessed 9 Feburary 2018).  20. Gething PW, Burgert CR (2017) The DHS Program Modeled Map Surfaces: Understanding the Utility of  Spatial Interpolation for Generating Indicators at Subnational Administrative Levels. DHS Spatial Analysis  Reports No. 15. Rockville, Maryland, USA: ICF.  21. Gething PW, Tatem AJ, Bird T, and Burgert‐Brucker CR (2015) Creating spatial interpolation surfaces with  DHS data. DHS Spatial Analysis Reports No. 11. Rockville, Maryland, USA: ICF International.  22. Bennett A, Kazembe L, Mathanga DP, Kinyoki D, Ali D, Snow RW, et al (2013) Mapping malaria transmission  intensity in Malawi, 2000–2010. Am J Trop Med Hyg. 89, 840–9.  23. Elyazar IRF, Gething PW, Patil AP, Rogayah H, Kusriastuti R, Wismarini DM, et al (2011). Plasmodium  falciparum malaria endemicity in Indonesia in 2010. PLoS One. 6, e21315.  24. Gemperli A, Vounatsou P, Kleinschmidt I, Bagayoko M, Lengeler C, and Smith T (2004) Spatial Patterns of  Infant Mortality in Mali: The Effect of Malaria Endemicity. American Journal of Epidemiology 159: 64‐72.  25. Gething PW, Patil A, Smith DL, Guerra CA, Elyazar IRF, Johnston G, et al (2011) A New World Malaria Map:  Plasmodium Falciparum Endemicity in 2010. Malaria Journal 10:378.  26. Gething PW, Elyazar IRF, Moyes CM, Smith DL, Battle KE, Guerra CA, et al (2012) A Long Neglected World  Malaria Map: Plasmodium Vivax Endemicity in 2010. PLoS Neglected Tropical Diseases 6:e1814.   27. Gething PW and Rosas N (2015). Developing a Poverty Map for Targeting of Social Safety Net Programs in  Sierra Leone. Report prepared for the World Bank, Washington DC.  28. Gething PW and Adoho F (2015). Developing a Poverty Map for Democratic Republic of Congo. Report  prepared for the World Bank, Washington DC.  29. Gething PW and Pop LB (2015). Developing a High Resolution Poverty Map for Afghanistan in 2011/12.  Report prepared for the World Bank, Washington DC.  9                                                                                                                                                                                              30. Gething PW and Molini V (2015). Developing an Updated Poverty Map for Nigeria. Report prepared for the  World Bank, Washington DC.  31. Gething PW, Dasgupta B, Andres LA (2017). Geospatial modeling of water and sanitation indicators in  Nigeria in 2015. Report prepared for the World Bank, Washington DC.  32. Gething PW and Rosas N (2015). Developing a High Resolution Poverty Map for Tanzania. Report prepared  for the World Bank, Washington DC.  33. Andrés LA, Bhatt S, Dasgupta B, Echenique JA, Gething PW, And Joseph G (2016). Geospatial modeling of  water and sanitation indicators in Nigeria in 2015. Report prepared for the World Bank, Washington DC.  34. Weiss DJ, Nelson A, Gibson HS, Temperley W, Peedell S, Lieber A, Hancher M, Poyart E, Belchior S, Fullman  N, Mappin B, Dalrymple U, Rozier J, Lucas TCD, Howes RE, Tusting LS, Kang SY, Cameron E, Bisanzio D, Battle  KE, Bhatt S, Gething PW (2018) A global map of travel time to cities to assess inequalities in accessibility in  2015. Nature, doi:10.1038/nature25181.  35. Water Point Mapping project undertaken by The Ministry of Health and Family Welfare, Govt. Of Tanzania   36. Diggle PJ and Ribeiro PJ (2007) Model‐Based Geostatistics: Springer Science & Business Media.  37. Diggle PJ, Tawn J and Moyeed R (1998) Model‐Based Geostatistics. Journal of the Royal Statistical Society:  Series C (Applied Statistics) 47, 299‐350.  38. 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,71,319‐392.  39. R Core Team (2018). R: A language and environment for statistical computing. R Foundation for Statistical  Computing, Vienna, Austria. (http://www.R‐project.org/, accessed 8 February 2018).  40. The Worldpop Project (2018).( www.worldpop.org, accessed 8 February 2018).  World Bank (2013). "Improving Nutrition Through Multisectoral Approaches." Washington DC: The  World Bank.   p://documents.worldbank.org/curated/en/625661468329649726/Improvingnutrition- through-multisectoral-approaches             10    Figures      Figure 1. Validation scatterplot. Each point compares a predicted value of stunting prevalence at a  given location with the observed value that was held out in the validation procedure. The 1:1 line is  shown for reference.  11      Figure 2. Predicted stunting prevalence in children under 5 years old in Tanzania in 2015‐16. The  value plotted in each 1×1km pixel is the point estimate (posterior mean) from a Bayesian posterior  predictive distribution. Regional boundaries are overlaid for reference. Note that the semi‐ autonomous island Region of Zanzibar was excluded from this analysis and is shown in grey.    12      Figure 3. Uncertainty associated with predicted stunting prevalence in children under 5 years old  in Tanzania in 2015‐16. The value plotted in each 1×1km pixel is the width of the posterior 95%  credible interval – larger values indicate higher uncertainty. Regional boundaries are overlaid for  reference. Note that the semi‐autonomous island Region of Zanzibar was excluded from this analysis  and is shown in grey.      13      Figure 4. Predicted stunting prevalence (top row) and number of children stunted (bottom row) by  District (left column) and Region (right column). Note that the semi‐autonomous island Region of  Zanzibar was excluded from this analysis and is shown in grey.  14    Tables    Validation performance    Correlation  0.63    Mean absolute error  11%    Mean square error  2%    Model coefficients   mean  (95% CI)  Intercept  2.25  (0.82, 3.68)  Aridity  ‐0.67  (‐1.48, 0.13)  Enhanced vegetation index  0.29  (‐1.09, 1.66)  Land surface temperature  ‐0.06  (‐0.09, ‐0.03)  Urban areas  ‐0.45  (‐0.82, ‐0.08)  Night‐time light brightness  ‐0.03  (‐0.09, 0.02)  TCB  ‐3.32  (‐6.08, ‐0.56)  TCW  ‐3.69  (‐7.41, 0.03)  Journey‐time to water point  0.00  (0.00, 0.00)    Table 1. Model validation performance and fitted fixed effect coefficients. Performance statistics  are from a four‐fold out‐of‐sample validation procedure, mean absolute error and mean square  error are in units of prevalence (%). Fitted fixed effect coefficients are summarised as the mean and  95% credible interval (CI) of the posterior distribution.      15    Region  Population  U5  N stunted U5  Prevalence stunting  Dodoma  221,970  69,255  31.20%  Arusha  191,785  55,487  28.93%  Kilimanjaro  170,832  56,471  33.06%  Tanga  203,677  76,906  37.76%  Morogoro  223,094  72,202  32.36%  Pwani  109,482  29,922  27.33%  Dar es salaam  342,205  43,148  12.61%  Lindi  93,501  31,566  33.76%  Mtwara  135,399  49,083  36.25%  Ruvuma  142,843  59,364  41.56%  Iringa  99,954  41,869  41.89%  Mbeya  279,582  119,621  42.79%  Singida  141,907  40,326  28.42%  Tabora  255,120  71,618  28.07%  Rukwa  103,645  51,097  49.30%  Kigoma  195,550  68,322  34.94%  Shinyanga  182,097  53,802  29.55%  Kagera  239,763  91,807  38.29%  Mwanza  264,753  90,507  34.19%  Mara  195,174  61,567  31.54%  Manyara  177,856  55,248  31.06%  Njombe  77,316  38,563  49.88%  Katavi  58,282  21,990  37.73%  Simiyu  189,483  58,838  31.05%  Geita  180,172  69,614  38.64%    Table 2. Predicted prevalence of stunting in children under 5 (U5) by Region. Shown are the total  U5 population, the estimated number stunted and the estimated prevalence of stunting. Predictions  were only made for Regions of mainland Tanzania.  16