RESEARCH ARTICLE Small area estimation of child undernutrition in Ethiopian woredas Thomas Pave Sohnesen1☯*, Alemayehu Azeze Ambel2☯, Peter Fisker3☯, Colin Andrews2‡, Qaiser Khan2‡ 1 Independent Consultant, Copenhagen, Denmark, 2 World Bank, Washington DC, United States of America, 3 Department of Food and Resource Economics, Changing Disasters, University of Copenhagen, Copenhagen, Denmark ☯ These authors contributed equally to this work. ‡ These authors also contributed equally to this work. * tpavesohnesen@worldbank.org a1111111111 a1111111111 a1111111111 a1111111111 Abstract a1111111111 Reducing child undernutrition is a key social policy objective of the Ethiopian government. Despite substantial reduction over the last decade and a half, child undernutrition is still high; with 48 percent of children either stunted, underweight or wasted, undernutrition remains an important child health challenge. The existing literature highlights that targeting OPEN ACCESS of efforts to reduce undernutrition in Ethiopia is inefficient, in part due to lack of data and Citation: Sohnesen TP, Ambel AA, Fisker P, updated information. This paper remedies some of this shortfall by estimating levels of stunt- Andrews C, Khan Q (2017) Small area estimation of child undernutrition in Ethiopian woredas. PLoS ing and underweight in each woreda for 2014. The estimates are small area estimations ONE 12(4): e0175445. https://doi.org/10.1371/ based on the 2014 Demographic and Health Survey and the latest population census. It is journal.pone.0175445 shown that small area estimations are powerful predictors of undernutrition, even compared Editor: Abdisalan Mohamed Noor, Kenya Medical to household characteristics, such as wealth and education, and hence a valuable targeting Research Institute - Wellcome Trust Research metric. The results show large variations in share of children undernourished within each Programme, KENYA region, more than between regions. The results also show that the locations with larger chal- Received: December 14, 2015 lenges depend on the chosen undernutrition statistic, as the share, number and concentra- Accepted: March 27, 2017 tion of undernourished children point to vastly different locations. There is also limited Published: April 14, 2017 correlation between share of children underweight and stunted across woredas, indicating that different locations face different challenges. Copyright: © 2017 Sohnesen et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Introduction Data Availability Statement: Both census and DHS survey data is third party data belonging to Child undernutrition is an important public health problem in developing countries, as the Central Statistical Agency in Ethiopia. Their data reflected by undernutrition being rated as the first priority among the world’s ten most impor- access policy and request for access to data can be tant challenges by the Copenhagen Consensus[1]. Its effect could be either immediate through found on this website: http://www.csa.gov.et/index. increased child morbidity and mortality or later in adult life by affecting health and labor mar- php/data-access-polices. They provide access to a ket outcomes. Despite considerable progress in reducing undernutrition—underweight (too 10 percent sample of the census and the full DHS survey, for research purposes. Requests can be low weight for age) fell from 41 to 25 percent of children, stunting (too short for age) fell from emailed to: Mr. Biratu Yigezu, CSA Director General 58 to 40 percent of children, and wasting (too low weight for height) fell from 12 to 9 percent (ybiratu@yahoo.com). of children from 2000 to 2014[2]—Ethiopia is still among the countries in the world with PLOS ONE | https://doi.org/10.1371/journal.pone.0175445 April 14, 2017 1 / 17 Small area estimation of child undernutrition in Ethiopian woredas Funding: This work was funded by the World Bank highest child undernutrition prevalence. Taken together, about 48 percent under the age of Promoting Basic Services Program Phase 3. five were undernourished (being either stunted, underweight, or wasted) in 2014, equivalent Additionally, World Bank provided support in the to approximately 6.3 million children[2]. form of salaries for authors TS, AA, CA, and QK. The funder did not have any additional role in the The government is aware of this challenge and has set out to reduce stunting to 30 percent study design, data collection and analysis, decision and wasting to 3 percent by 2015[3]. Programs to address these challenges include: increasing to publish, or preparation of the manuscript. The agricultural productivity, promoting girls’ education, immunization, integrated management specific roles of these authors are articulated in the of neonatal and childhood illnesses, improved access to water and sanitation, family planning, ’author contributions’ section. prevention of mother-to-child transmission of HIV, skilled birth delivery, and delaying of Competing interests: The authors received funding pregnancy[3]. from World Bank, a commercial company, for this The prevalence of different types of undernutrition differs by location in Ethiopia, indicat- study. Additionally, authors TS, AA, CA, and QK are ing that different locations face different challenges[4]. International evidence also shows that employed by World Bank, a commercial company. There are no patents, products in development or many different interventions improve undernutrition outcomes, but also that similar interven- marketed products to declare. This does not alter tions have different impacts in different settings and locations[5]. Various interventions and our adherence to all the PLOS ONE policies on correlations have also been identified in Ethiopia to matter for undernutrition outcomes, sharing data and materials. including: shocks and food aid[6, 7], maternal education and nutritional status[8–15], owner- ship of selected assets such as cows[16], food availability and diversity[17], access to trained and educated health professionals[18], access to safe water and feeding practices[19, 20]. Rajkumar et al.[4] evaluate targeting of efforts to reduce undernutrition in Ethiopia and conclude, among other things, that: 1. Most nutrition-related programs focus exclusively on a subset of the country’s woredas and these are difficult or impossible to identify accurately because of a poor nutrition informa- tion system. 2. Many programs use food insecurity as a proxy for nutrition insecurity, even though these are not highly correlated. Further, woreda targeting for food insecurity is often based on dated information. They conclude: “Ethiopia’s malnutrition rate could probably be much reduced by shifting some of the programs from the woredas with a high concentration of major programs into wore- das with high malnutrition”. This paper addresses this shortage of information by estimating levels of stunting and underweight for all woredas. As in many other developing countries, there is large urban-rural divide in prevalence in Ethiopia. Child stunting is 16 percentage points lower in urban (27.0%) than rural (42.6%) areas.[8] The urban-rural divide is large, but there are also notable variations, particularly in underweight, across regions[4]. The importance of the spatial dimension has been observed in Ethiopia[15] and in other countries, and the result is often strong even after controlling for other correlates[21]. Undernutrition regressions for Vietnam, South Africa, Pakistan, and Morocco suggest that community-level effects are of great importance[22]. Fuji[22] for instance, in his undernutrition regressions for Cambodia finds that individual-level and household-level variables explained 20 to 30 percent of the variation in the z-score, while he was able to increase the explanatory power of the model to about 40 to 60 percent by including geographic variables and interaction terms. Alderman and Christiaensen[9] is an exception, as they find in their analysis of Ethiopia that location does not matter significantly once they con- trol for other correlates. The local estimates of undernutrition in this paper is limited to underweight and stunting as an imputation model with reasonable accuracy was not found for wasting. In OLS regres- sions between z-scores for wasting and household characteristics, it was not possible to find explanatory variables that explained the z-score. A weaker correlation between household characteristics and z-scores for wasting is often observed, and previous country studies using PLOS ONE | https://doi.org/10.1371/journal.pone.0175445 April 14, 2017 2 / 17 Small area estimation of child undernutrition in Ethiopian woredas small area estimation (henceforth SAE) for undernutrition in Bangladesh[23], Cambodia[22, 24], Ecuador, Panama, Dominican Republic[25] and Tanzania[26], were also unable to find a suitable model for wasting. Other undernutrition SAE country studies include stunting in Bra- zil[27]; and stunting, underweight and wasting in Nepal[28]. The ability to model wasting depends both on the range of variables available in the census, and on the correlation between z-scores for wasting and household characteristics in each country. This paper contributes in two aspects. First, it addresses the shortage of updated evidence on local undernutrition prevalence by estimating undernutrition prevalence for all woredas in 2014, using SAE. These estimates are the first for Ethiopia. It hereby adds to the rich literature on who is undernourished by showing where to find the undernourished. Second, an evalua- tion of the SAE models shows that geographical targeting using woreda levels of undernutri- tion is a viable strategy, as the SAE models have as good or better predictive power than common known correlates of undernutrition. The SAE estimates show that different undernu- trition statistics point to different woredas having the largest problems. The different statistics, combined with the correlation between them, could be used to select the most appropriate type of intervention in different locations. The rest of the paper is organized as follows: Section 2 describes the data, Section 3 presents the SAE methodology, Section 4 shows results, while Section 5 concludes. Data The primary data source of undernutrition in this study is the Ethiopia Mini Demographic and Health Survey (EMDHS) 2014[2]. The EMDHS 2014 is a stratified nationally representa- tive survey, including 5,579 children under five years of age. The survey follows the Demo- graphic Health Surveys[29] standard survey design, although some sections, such as HIV and immunization, were not included. Detailed information about the sampling is available in the survey’s report[2]. Undernutrition is measured as share of children moderately or severely undernourished (below two standard deviations from the mean) based on z-scores for stunting (height for age), and underweight (weight for age). Z-scores were calculated using EMDHS and the 2006 WHO growth standards[30]. To obtain estimates of undernutrition prevalence at woreda levels, EMDHS is combined, through SAE, with the 2007 Census. The 2007 census has two formats—a long and a short for- mat. The long format is richer in terms of data and in addition to demographic information, it includes information on assets, housing characteristics, education, fertility, and mortality. A randomly selected 20 percent of the households received the long form, while the other 80 per- cent received the short form that only covered basic demographics. In addition to the census the agricultural zones and quality of soil[31] is added as explanatory variables. The EMDHS and a 10 percent sample of the 2007 census is available at request from the Ethiopian Central Statistical Agency[32]. Small area estimation methodology Woreda level estimates of undernutrition are estimated by SAE. The basic idea is to construct a prediction model between the z-score and observable household and location characteristics in EMDHS. This model is then used to predict a z-score for every child using the census. This paper relies on the SAE method developed in Elbers, Lanjouw, and Lanjouw[33] (henceforth ELL). This specific method is attractive, as it can be easily implemented in the freeware pro- gram PovMap2[34]. There are numerous technical variations of the SAE methodology. For PLOS ONE | https://doi.org/10.1371/journal.pone.0175445 April 14, 2017 3 / 17 Small area estimation of child undernutrition in Ethiopian woredas recent overview of variations in methodology and application of SAE see Rao et al. [35] and Pratesi [36]. The methodology relies on three steps. In step one, a set of explanatory variables which are similarly defined and distributed in the survey and the census are identified. Restricting the explanatory variables (X below) to those that are time invariant, ensures that the model esti- mates z-scores for 2014 (the year of the survey) and not 2007 (the year of the census). The sec- tion below, provides more details on step one. In step two, a z-score model is constructed in the survey data: survey zi ¼ Xih b þ Z 0 g þ uih ð1Þ where Xih 0 is the vector of explanatory variables for child i in household h, β is the vector of coefficients, Z0 is the vector of location specific variables, γ is the vector of coefficients, and uih is the error term due to the discrepancy between the imputed z-score and the actual value. Xih 0 is child and household level variables that have similar definitions and distributions in survey and census, while Z0 includes location-specific averages or other transformations of variables found in the census that cannot be found in the survey, and external variables, such as geo- spatial variables that can be added to both the survey and the census. In step three, undernutrition estimates and their standard errors are computed via simula- tions using the estimation model. There are two sources of errors involved in the estimation process: errors in the estimated regression coefficients (b ^, ^ g ) and the disturbance terms, both of which affect undernutrition estimates and the level of their accuracy. In ELL, a simulated z- score is calculated for each census child by a predicted z-score as in (2): survey ^ ^i ¼ Xih z b þ Z0^ gþu^ ih ð2Þ These simulations are repeated a large number of times (100 in the case of this application), with new values for β, γ, and uih drawn from their distributions, in each repetition. The distri- butions of β, γ are obtained through GLS, while the distribution of uih is obtained through a non-parametric approximation following Elbers and van der Weide[37]. All the estimation is automated within the PovMap2 program. For any given location (such as a woreda), each simulation is used to calculate the share of children underweight or stunted (a z-score below negative two standard deviations). The mean across the simulations of an undernutrition statistic provides a point estimate of the sta- tistic, and the standard deviation provides an estimate of the standard error. In most applications of ELL, the error term uih is decomposed into two independent com- ponents: uih = ηc + εih, where ηc is a cluster-specific effect and εih an individual effect. In this application, the error term has not been split into a cluster and individual effect, as no cluster effect was found. The lack of a cluster effect is likely closely related to the sampling design of EMDHS, as it is designed with few EAs within each region and relatively large samples within each EA. A cluster effect would often be based on EAs, and with limited number of EAs within each region, it was not possible to establish a distribution of cluster effects. Results Results are presented in three sections. Section one documents the application of the SAE methodology to Ethiopian data and evaluate the quality of the application, section two evalu- ates if SAE models, with the limited data are powerful predictors compared to other more common correlates of undernutrition, while section three illustrates the spatial patterns in undernutrition across Ethiopian woredas. PLOS ONE | https://doi.org/10.1371/journal.pone.0175445 April 14, 2017 4 / 17 Small area estimation of child undernutrition in Ethiopian woredas Application of SAE to Ethiopian data SAE estimates of stunting and underweight is done at woreda level. Woreda is the third administration layer after region and zone. Each woreda is also divided into kebeles, which are the smallest administrative units in Ethiopia. A woreda on average has around 12000 children, with a standard deviation of 8000, with around 2000 children at the 5 percentile and 28500 children at the 95 percentile. The SAE estimates are based on regional z-score models. Regional models have the advan- tage of fitting data relevant to local circumstances, although they limit the number of variables in each model as the number of observations decreases. The EMDHS 2014 survey sampled all woredas in the city regions of Harari, Dire Dawa, and Addis Ababa, and the survey estimates from EMDHS were used for those three regions as opposed to SAE estimates. The Ethiopian census has both a short and a long form, and to increase model fit, the mod- els use the long form, including enumeration area (EA) averages of all variables from the cen- sus whether found in the survey or not. Separate models for the long and short form are theoretically possible. Unfortunately, this is not possible using the PovMap software, as the program cannot handle the likely correlation between two such models, which would result in inaccurate estimates of undernutrition standard errors. To mitigate issues arising from the time interval between the 2007 Population Census and EMDHS 2014, only variables whose distributions are time invariant between the two are used in the z-score models. The PovMap program jointly shows the distributions from each data source for each variable, allowing a visual inspection of the similarity of the distributions. Pov- Map also provides a frequency chi-square test for categorical variables and a Kolmogorov- Smirnov test for continuous variables. These tools are used to judge similarity of the distribu- tions and if they are time invariant or not. A strict statistical cut off has not been applied as there is a tradeoff between model contribution and similarity of variables. Further, in some cases a categorical variable might seem perfectly similar, but one category when used as dummy might not pass the statistical threshold though the overall categorical variable does. In such a case, the one category dummy would be allowed to enter the model despite not passing the statistical test on its own. To avoid overfitting skewed dummy variables with a mean smaller than 0.05 in the survey are also excluded from the models. Variable selection into the estimation models was based on contribution to adjusted r-square, and correlation between variables (highly correlated variables were excluded). Further, all models include variables gen- erated at levels above the household level, as too high spatial correlation in the error term rela- tive to the overall error term can jeopardize accuracy of estimated standard errors of undernutrition indicators[38, 39]. The models also include variables as number of children in household and children close in age, which capture within household correlation. See also Jones and Haslett[28], as well as Fujii[22], for alternative ad hoc modeling solutions for within household correlation. S1 Table shows the survey and census means for all variables included in the final models. All models include variables that are based on EA or higher-level census means; these are not included in S1 Table, as they are comparable by definition. The models have adjusted r-squares similar to those of other undernutrition regressions, particular those used for undernutrition maps (0.06 to 0.27 for stunting and between 0.08 and 0.14 for underweight). Some studies have succeeded in developing undernutrition models for SAEs of undernutrition with R-squares as high as 0.6 and 0.7[22]. However, some express con- cern for overfitting with R-squares above 0.35[27]. The full regional regression models are enclosed in S2 Table. As a check on model accuracy, Table 1 shows the model-imputed undernutrition preva- lence and those measured in EMDHS at regional level. The table shows that measured PLOS ONE | https://doi.org/10.1371/journal.pone.0175445 April 14, 2017 5 / 17 Small area estimation of child undernutrition in Ethiopian woredas Table 1. Measured and imputed undernutrition prevalence, 2014. Region Imputed Survey 95% confidence interval for Observations in survey undernutrition rate undernutrition rate survey estimates Stunting Tigray 0.440 0.459 0.380 0.537 435 Afar 0.404 0.465 0.396 0.533 565 Amhara 0.413 0.432 0.376 0.488 518 Oromia 0.41 0.385 0.334 0.435 635 Somali 0.42 0.36 0.292 0.427 565 B. Gumuz 0.412 0.406 0.353 0.459 426 SNNP 0.421 0.443 0.391 0.496 673 Gambela 0.275 0.219 0.149 0.288 403 Underweight Tigray 0.351 0.312 0.262 0.364 445 Afar 0.496 0.455 0.408 0.514 608 Amhara 0.341 0.291 0.248 0.343 523 Oromia 0.271 0.229 0.196 0.286 656 Somali 0.403 0.386 0.331 0.46 591 B. Gumuz 0.327 0.282 0.212 0.364 438 SNNP 0.289 0.260 0.213 0.325 699 Gambela 0.224 0.199 0.126 0.273 412 Source: DHS 2014 and imputed values from models. https://doi.org/10.1371/journal.pone.0175445.t001 undernutrition prevalence in EMDHS and the estimated prevalence from census are similar, and none of the estimates are outside the 95 percent confidence interval of the measured levels. As also illustrated in Table 1, both measures from surveys and the estimates from models come with standard errors. The estimated standard errors of undernutrition levels across wor- edas are on average similar and a little larger than those observed at regional levels from the survey. Woredas with smaller populations of children generally come with larger standard errors. Fig 1 illustrates the standard errors in relation to the number of children in each wor- eda. It illustrates both how woredas with lower number of children have larger standard errors (the left tail), and how standard errors are larger for stunting than underweight. This is so Fig 1. Standard errors of underweight and stunting prevalence, and number of children in woredas. https://doi.org/10.1371/journal.pone.0175445.g001 PLOS ONE | https://doi.org/10.1371/journal.pone.0175445 April 14, 2017 6 / 17 Small area estimation of child undernutrition in Ethiopian woredas despite stunting models having higher r-square and is due to larger underlying variation in z- scores for stunting than for underweight. As example of the influence of standard errors: Sixty-five percent of woredas have underweight levels above the national average of 26.6, while only 41 percent of these are significantly above the national average (based on the wor- eda estimates lower bound of the 95 percent confidence interval being above the national mean). Similarly, 46 percent of woredas have stunting levels that are above the national aver- age, while only 14 percent are significantly above that level (again using the 95 percent confi- dence interval). Table 2. OLS regression of z-scores. Column 1 2 3 4 5 6 7 8 Dependent variable: z-score HAZ HAZ HAZ HAZ WAZ WAZ WAZ WAZ HH has access to protected water 0.03 -0.06 0.07 0.02 -0.03 0.05 (0.11) (0.12) (0.10) (0.07) (0.08) (0.07) HH has no access to toilet -0.10 -0.02 -0.06 -0.01 0.04 0.01 (0.10) (0.10) (0.08) (0.07) (0.06) (0.06) HH utilize health service 0.05 0.05 0.05 0.06 0.05 0.05 (0.10) (0.09) (0.09) (0.07) (0.07) (0.07) Mothers’ education (no education is the excluded category) Primary 0.21** 0.07 0.14 0.19** 0.13* 0.16** (0.09) (0.10) (0.09) (0.07) (0.08) (0.07) Secondary 0.73*** 0.59** 0.45 0.72*** 0.65*** 0.58*** (0.25) (0.27) (0.28) (0.15) (0.16) (0.15) Tertiary 1.42*** 0.71** 0.68** 1.11*** 0.75** 0.81** (0.29) (0.29) (0.29) (0.31) (0.33) (0.32) Wealth Index 0.00** 0.00 0.00* 0.00 (0.00) (0.00) (0.00) (0.00) Age of child (age 4 excluded category) Less than 1 1.70*** 1.45*** 0.94*** 0.46** (0.14) (0.27) (0.09) (0.19) 1 through 2 0.12 -0.21 0.17** 0.06 (0.13) (0.24) (0.08) (0.14) 2 through 3 -0.34*** -0.35** 0.03 0.02 (0.12) (0.17) (0.09) (0.13) 3 through 4 -0.25** -0.26 0.08 -0.03 (0.10) (0.16) (0.07) (0.13) Regional dummies x x x x x x x x SAE models x x x x Constant 0.45 -1.07*** -1.02*** -2.25*** -0.37 -1.21*** -1.31*** -1.31*** (0.68) (0.20) (0.23) (0.42) (0.34) (0.13) (0.14) (0.38) Number of observations 3,838 3,838 3,838 3,838 3,976 3,976 3,976 3,976 R2 0.241 0.023 0.176 0.247 0.152 0.040 0.107 0.164 The asterisks indicate the significance level: *** p<0.01, ** p<0.05, * p<0.1. Standard errors in parentheses. Regressions take sample design and household weights into account by using Stata’s svy command. The city regions of Harari, Dire Dawa, and Addis Ababa are excluded as there are no SAE model for those locations. Data: EMDHS 2014. https://doi.org/10.1371/journal.pone.0175445.t002 PLOS ONE | https://doi.org/10.1371/journal.pone.0175445 April 14, 2017 7 / 17 Small area estimation of child undernutrition in Ethiopian woredas Fig 2. Variation in share of children stunted and underweight across woredas in each regions. The bar in center of box is the median stunted and underweight prevalence across woredas in each region. The box is the 25th and 75 percentile; dots are outside values. https://doi.org/10.1371/journal.pone.0175445.g002 Are SAE estimates of undernutrition in woredas a useful targeting mechanism? As highlighted in the introduction, existing literature points to weaknesses in data for geo- graphical targeting, but can models utilizing only a reduced set of explanatory variables (as the Fig 3. Share of children underweight in each woreda. The figure shows share of children underweight. https://doi.org/10.1371/journal.pone.0175445.g003 PLOS ONE | https://doi.org/10.1371/journal.pone.0175445 April 14, 2017 8 / 17 Small area estimation of child undernutrition in Ethiopian woredas set of potential variables is reduced to those available in the census that are time invariant, see above) really predict undernutrition? Most variables highlighted in the literature and described in the introduction are not available in the census, and hence cannot be used in our predictive models. Given available data, a complete analysis of this question is not possible, but it is possi- ble to compare the predictive power of the SAE models with more traditional correlates of undernutrition found in the EMDHS survey. The literature on correlates of malnutrition include both variables that can be linked directly to causality and some with more indirect links to causality. For SAE models the focus in on predictive power and causality is not an requirement. To compare the SAE models predictive power to other more common correlates of undernutrition, an OLS regression of z-scores is estimated in the EMDHS with both com- mon undernutrition correlates and the SAE models. The regional SAE models are included in the national regression by multiplying variables in each model by regional dummies. In the interest of space, the SAE models are not shown in Table 2, but the complete models can be seen in S2 Table. The regressions further include the following common correlates of under- nutrition: a) access to health services approximated by households utilizing any pre- or post- natal services; b) access to clean water approximated through households utilizing protected water source; c) lack of access to toilet facilities; d) access to food and other necessities, and or knowledge approximated by wealth measured by the EMDHS asset index; e)mothers’ educa- tion; f)children’s age; and g)regional dummies. Among these common correlates of Fig 4. Share of children stunted in each woreda. The figure shows share of children stunted. https://doi.org/10.1371/journal.pone.0175445.g004 PLOS ONE | https://doi.org/10.1371/journal.pone.0175445 April 14, 2017 9 / 17 Small area estimation of child undernutrition in Ethiopian woredas undernutrition, only the children’s age (at individual level) and mothers’ education (at an aggregated level) are found in the census, and therefore among the potential variables in the SAE models. All commonly correlated variables for both stunting and underweight have the expected signs, though some of the insignificant variables change sign when controlling for wealth and children’s age (column 2–4 and 6–8, Table 2). Similar patterns have been observed in Ethiopia before, based on other data sets[40]. The common correlates/indicators of undernutrition (column 2–4 and 6–7) have r-squares of 0.18 and 0.11 for stunting and underweight respectively, with most of the explanatory power coming from the wealth index and children’s age. In comparison, the SAE models (col- umn 1 and 5 in Table 2) have r-squares of 0.24 and 0.15 for stunting and underweight. Hence, the SAE models have more about 50 percent more predictive power then “common undernu- trition correlates”. Evaluated jointly, the traditional correlates do add predictive power to the SAE models, but it is not a large contribution (column 4 and 8, Table 2). Spatial patterns in woreda estimates of undernutrition A first observation, based on the SAE-obtained woreda level estimates, is that only a few regions are homogenous with similar levels of undernutrition across woredas (Fig 2). This is Fig 5. Number of children underweight in each woreda. The figure shows the number of underweight children. Each dot represent 100 underweight children. https://doi.org/10.1371/journal.pone.0175445.g005 PLOS ONE | https://doi.org/10.1371/journal.pone.0175445 April 14, 2017 10 / 17 Small area estimation of child undernutrition in Ethiopian woredas Fig 6. Number of children stunted in each woreda. The figure shows the number of stunted children. Each dot represent 100 stunted children. https://doi.org/10.1371/journal.pone.0175445.g006 Fig 7. Share and number of underweight children in each woreda. The figure shows the number of underweight children by the height of the column and the share of children underweight by the color of each woreda. The color scale is the same as in Fig 3. https://doi.org/10.1371/journal.pone.0175445.g007 PLOS ONE | https://doi.org/10.1371/journal.pone.0175445 April 14, 2017 11 / 17 Small area estimation of child undernutrition in Ethiopian woredas Fig 8. Share and number of stunted children in each woreda. The figure shows the number of stunted children by the height of the column and the share of children stunted by the color of each woreda. The color scale is the same as in Fig 4. https://doi.org/10.1371/journal.pone.0175445.g008 Fig 9. Concentration of underweight children (number of underweight children per km2). The figure shows the number of underweight children per km2 in each woreda. https://doi.org/10.1371/journal.pone.0175445.g009 PLOS ONE | https://doi.org/10.1371/journal.pone.0175445 April 14, 2017 12 / 17 Small area estimation of child undernutrition in Ethiopian woredas illustrated in Fig 2 by the wide ranges of levels across woredas within each region. In almost all regions, woredas with higher levels have more than twice the level of those with lower levels. The variation in number of woredas in each region influences the graph, though it does not seem to drive results as both Gambella, with only 13 woredas, and Oromia, with 277 woredas, both have relatively small variation. The variation in undernutrition across woredas within regions indicates that there is scope for spatial targeting within most regions. Figs 3 and 4 show that woredas further away from the center of the country tend to have higher share of underweight and stunted children than woredas closer to the center (with some notable variation). This is in contrast to Figs 5 and 6 showing the number of undernour- ished children (each dot represents 100 undernourished). Figs 5 and 6 show that in many areas most stunted or underweight children live close to the center of the country. This combined pattern is illustrated jointly in Figs 7 and 8, with height of columns showing the number of underweight and stunted children, while the colors illustrate the share of underweight and stunted children. This shows how some areas, like the dry areas in Somali (to the southeast) have a relatively high share of undernourished children but also a small population, and, there- fore a low absolute number of undernourished children. Other areas, like the center of coun- try, have a large population with a low share of undernourished, captured through tall columns and light colors in Figs 7 and 8. A few areas, like the north west, seem to have both a high share of undernourished and a large absolute number of undernourished. Further, more urban areas, though having a relative low share of undernourished children [8], have a large population, and a large population that is spatially very concentrated. This concentration is illustrated by Figs 9 and 10 where the height of columns show that the num- ber of children per km2 is very high in some locations. Fig 10. Concentration of stunted children (number of stunted children per km2). The figure shows the number of stunted children per km2 in each woreda. https://doi.org/10.1371/journal.pone.0175445.g010 PLOS ONE | https://doi.org/10.1371/journal.pone.0175445 April 14, 2017 13 / 17 Small area estimation of child undernutrition in Ethiopian woredas Fig 11. Spatial correlation between share, number and concentration of underweight children. https://doi.org/10.1371/journal.pone.0175445.g011 Figs 3 to 10 illustrate that different statistics (share, number, and number of children per km2) all support an understanding of where the challenges are largest. Some woredas are among the worst by one statistic, but among the better off by another. That one statistic is insufficient for understanding the spatial distribution of underweight and stunting is further illustrated in Figs 11 and 12 that show the spatial correlation among the different woreda sta- tistics (share, number and number per km2) for underweight and stunting, respectively. Figs 11 and 12 show that for both underweight and stunting, there is very limited correlation between these statistics (share, number and number per km2). The spatial correlation between stunting and underweight is also of interest. Fig 13 shows that the correlation between share of children stunted and underweight in each woreda is noisy (image to the left). However, the correlation between number of children underweight and stunted (image in the center), and number of underweight and stunted per km2 (images in the center and to the right), is high. The pattern reflect that the variation in prevalence of stunting and underweight observed in the left panel is not sufficient to dominate the spatial concentration of the population of children. For instance, two small woredas of 1000 children each, could have a prevalence of 20 and 30, resulting in 200 and 300 undernourished children in each. Other two larger woredas of say 10000 children in each, with same prevalence would have 2000 and 3000 undernourished children. In this case, though some woreads have same prevalence, the two larger woredas with very different prevalence would have a higher correla- tion, as observed in the two left images in Fig 13. The pattern can also be observed in the maps, where Figs 9 and 10 look similar, while Figs 3 and 4 are less similar. Fig 12. Spatial correlation between share, number and concentration of stunted children. https://doi.org/10.1371/journal.pone.0175445.g012 PLOS ONE | https://doi.org/10.1371/journal.pone.0175445 April 14, 2017 14 / 17 Small area estimation of child undernutrition in Ethiopian woredas Fig 13. Location of underweight and stunted children. https://doi.org/10.1371/journal.pone.0175445.g013 Concluding remarks This paper contributes the first estimates of child undernutrition for all woredas in Ethiopia. Hence, the paper contributes to the rich literature on who is undernourished with the first esti- mates of where the undernourished resides. It further shows that SAE models of undernutri- tion have as much or more predictive power than common observed correlates of undernutrition. As such, the estimates can be used to address the concerns highlighted by Raj- kumar et al.[4] in the introduction “Ethiopia’s malnutrition rate could probably be much reduced by shifting some of the programs from the woredas with a high concentration of major programs into woredas with high malnutrition”. The results provide the data to proof or dis- proof Rajkumar et al.’s claim and assess whether current efforts are as effective as they can be, and where there might be need for new or additional efforts. The results also show that wore- das with “high malnutrition”, as referenced above, depends on the statistic utilized. Woredas with a high share of undernourished are often not the same as those with many undernour- ished, or a high concentration of undernourished (number of undernourished per km2). These aspects could inform the type of interventions used in different locations. The spatial correlation of underweight and stunting, could also be informative for planning purposes, as the number and concentration of undernourished children is high, while there is much less correlation between the share of undernourished children. Supporting information S1 Table. Mean in survey and census of variables in prediction models. (DOCX) S2 Table. Regional prediction models for stunting and underweight. (DOCX) Acknowledgments The authors acknowledge financial support from the World Bank. The authors would like to thank: Phyllis Ronek for editorial assistance; Tesfaye Hailu (Ethiopian Public Health Institute), Kalle Hirvonen (IFPRI), Silvia Kaufmann (World Bank), Lisa Saldanha (World Bank), and Henrik Hansen (Copenhagen University), as well two anonymous referees for valuable com- ments and suggestions. PLOS ONE | https://doi.org/10.1371/journal.pone.0175445 April 14, 2017 15 / 17 Small area estimation of child undernutrition in Ethiopian woredas Author Contributions Conceived and designed the experiments: TPS AA CA QK. Analyzed the data: TPS PF. Contributed reagents/materials/analysis tools: TPS AA PF CA QK. Wrote the paper: TPS AA. References 1. Consensus C. Copenhagen Consensus 2012: Expert panel findings10/10/2015. http://www. copenhagenconsensus.com/sites/default/files/outcome_document_updated_1105.pdf. 2. Agency CS. The Ethiopia Mini Demographic and Health Survey (EMDHS). Addis Ababa, Central Sta- tistical Agency.; 2014 3. Ethiopia GotFDRo. National Nutrition Programme June 2013-June 2015. Addis Ababa: Government of Federal Democratic Republic of Ethiopia.; 2013. 4. Rajkumar AS, Gaukler C, Tilahun J. Combating Malnutrition in Ethiopia: An evidence-based approach for sustained results: World Bank Publications; 2011. 5. Ainsworth M, Ambel X, Martin G, Sinha S, Huppi M. What Can We Learn from Nutrition Impact Evalua- tions. Washington DC: Independent Evaluation Group, the World Bank. 2010. 6. Quisumbing AR. Food aid and child nutrition in rural Ethiopia. World Dev. 2003; 31(7):1309–24. 7. Yamano T, Alderman H, Christiaensen L. Child Growth, Shocks, and Food Aid in Rural Ethiopia. Ameri- can Journal of Agricultural Economics. 2005; 87(2):273–88. 8. Ambel AA, Huang W-c. Maternal Education, Linkages and Child Nutrition in the Long and Short-run: Evidence from the Ethiopia Demographic and Health Surveys. International Journal of African Develop- ment. 2014; 1(2,). 9. Christiaensen L, Alderman H. Child malnutrition in Ethiopia: Can maternal knowledge augment the role of income? Econ Dev Cult Change. 2004; 52(2):287–312. 10. Girma W, Genebo T. Determinants of nutritional status of women and children in Ethiopia. Ethiopia Health and Nutrition Research Institute Calverton, Maryland USA 2002. 2002. 11. Negash C, Whiting SJ, Henry CJ, Belachew T, Hailemariam TG. Association between Maternal and Child Nutritional Status in Hula, Rural Southern Ethiopia: A Cross Sectional Study. Plos One. 2015; 10(11):e0142301. https://doi.org/10.1371/journal.pone.0142301 PMID: 26588687 12. N D. Determinants of Severe Acute Malnutrition among Under Five Children in Shashogo Woreda, Southern Ethiopia: A Community Based Matched Case Control Study. Journal of Nutrition & Food Sci- ences. 2014; 4(300). 13. Genebo T, Girma W, Haider J, and Demmissie T. The association of children’s nutritional status to maternal education in Ziggbaboto, Guragie Zone South Ethiopia. Ethiopian Journal of Health Develop- ment. 1999 13(1):55–61. 14. Ayele DG, Zewotir TT, Mwambi HG. Structured additive regression models with spatial correlation to estimate under-five mortality risk factors in Ethiopia. Bmc Public Health. 2015; 15(1):1–12. 15. Ayele DG, Zewotir TT. Childhood mortality spatial distribution in Ethiopia. Journal of Applied Statistics. 2016:1–16. 16. Hoddinott J, Headey D, Dereje M. Cows, Missing Milk Markets, and Nutrition in Rural Ethiopia. J Dev Stud. 2015; 51(8):958–75. 17. Motbainor A, Worku A, Kumie A. Stunting Is Associated with Food Diversity while Wasting with Food Insecurity among Underfive Children in East and West Gojjam Zones of Amhara Region, Ethiopia. Plos One. 2015; 10(8):e0133542. https://doi.org/10.1371/journal.pone.0133542 PMID: 26285047 18. Mekonnen A, Jones N, Tefera B. Tackling child malnutrition in Ethiopia: do the sustainable development poverty reduction programme’s underlying policy assumptions reflect local realities? Young Lives, An International Study of Child Poverty, Working Paper No, 19 2005. 19. Bantamen G, Belaynew W, Dube J. Assessment of factors associated with malnutrition among under five years age children at Machakel Woreda, Northwest Ethiopia: a case control study. Journal of Nutri- tion & Food Sciences. 2014;2014. PLOS ONE | https://doi.org/10.1371/journal.pone.0175445 April 14, 2017 16 / 17 Small area estimation of child undernutrition in Ethiopian woredas 20. Asfaw M, Wondaferash M, Taha M, Dube L. Prevalence of undernutrition and associated factors among children aged between six to fifty nine months in Bule Hora district, South Ethiopia. Bmc Public Health. 2015; 15:41. https://doi.org/10.1186/s12889-015-1370-9 PMID: 25636688 21. Kandala N-B, Madungu TP, Emina JB, Nzita KP, Cappuccio FP. Malnutrition among children under the age of five in the Democratic Republic of Congo (DRC): does geographic location matter? Bmc Public Health. 2011; 11(1):1–15. 22. Fujii T. Micro-Level Estimation of Child Undernutrition Indicators in Cambodia. The World Bank Eco- nomic Review. 2010; 24(3):520–53. 23. Haslett SJG, and Isidro M. Small-Area Estimation of Child Undernutrition in Bangladesh. Bangladesh: Bureau of Statistics, United Nations World Food Programme and International Fund for Agricultural Development; 2014. 24. Haslett S, Jones G, Sefton A. Small-area Estimation of Poverty and Malnutrition in Cambodia: National Institute of Statistics, Ministry of Planning, Royal Government of Cambodia and the United Nations World Food Programme, Cambodia, April 2013. 25. ´as K, Wilde P, Friedman D. Mapping Hunger: a report on mapping malnutrition Rogers BL, Wirth J, Macı prevalence in the Dominican Republic, Ecuador and Panama. Friedman School of Nutrition Science and Policy, 2007. 26. Simler KR. Nutrition mapping in Tanzania: an exploratory analysis. International Food Policy Research Institute (IFPRI), 2006. 27. ¸ ão In Brazil. American Gilligan DO, Veiga A. An Evaluation Of Geographic Targeting In Bolsa Alimentac Agricultural Economics Association (New Name 2008: Agricultural and Applied Economics Associa- tion), 2003. 28. Haslett S, Jones G. Small area estimation of poverty, caloric intake and malnutrition in Nepal. Nepal Central Bureau of Statistics/World Food Programme, United Nations/World Bank, Kathmandu. 2006;184:999337018–5. 29. www.dhsprogram.com 30. WHO M. Growth, Reference, Study, Group. 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 devel- opment Geneva: World Health Organization. 2006. 31. Development MoAaR. Agroecological Zones: Definition and Maps. Addis Ababa, Ethiopia 1998. 32. www.csa.gov.et. 33. Elbers C, Lanjouw JO, Lanjouw P. Micro-level estimation of poverty and inequality. Econometrica. 2003; 71(1):355–64. 34. http://iresearch.worldbank.org/PovMap/PovMap2/ 35. Rao JNK, Molina I. Small Area Estimation: Wiley; 2015. 36. Pratesi M. Analysis of Poverty Data by Small Area Estimation: Wiley; 2016. 37. Elbers C, van der Weide R. Estimation of normal mixtures in a nested error model with an application to small area estimation of poverty and inequality. The World Bank, 2014 Jul. Report No. 38. Tarozzi A, Deaton A. Using Census and Survey Data to Estimate Poverty and Inequality for Small Areas. Rev Econ Stat. 2009; 91(4):773–92. 39. Elbers C, Lanjouw P, Leite PG. Brazil within Brazil: Testing the Poverty Map Methodology in Minas Gerais. World Bank Policy Research Working Paper. 2008;No. 4513 40. Headey D. An analysis of trends and determinants of child undernutrition in Ethiopia, 2000–2011. Inter- national Food Policy Research Institute (IFPRI), 2014. PLOS ONE | https://doi.org/10.1371/journal.pone.0175445 April 14, 2017 17 / 17