Zinc deficiency is highly prevalent and spatially dependent over short distances in Ethiopia

Zinc (Zn) is an essential nutrient for human health. In Ethiopia, a high prevalence of Zn deficiency has been reported. To explore demographic variation and spatial dependencies in the Zn status of the Ethiopian population, we analyzed archived serum samples (n = 3373) from the 2015 Ethiopian National Micronutrient Survey (ENMS), a cross-sectional survey of young children, school-age children, women of reproductive age (WRA) and men conducted in all 9 regions and two city administration of Ethiopia. Serum Zn concentrations, measured using inductively coupled plasma-mass spectrometry (ICPMS), were compared to thresholds based on age, sex, fasting status, and time of blood collection, after adjusting for inflammation status. Median serum Zn concentration of the population was 57.5 μg dL−1. Overall, it is estimated that 72% of the population was Zn deficient, with high prevalence in all demographic groups. Spatial statistical analysis showed that there was spatial dependence in Zn status of WRA at distances of up to 45 km. Zinc deficiency is spatially dependent over short distances. Although WRA in most areas are likely to be Zn deficient, prevalence of deficiency varies at regional scale and between rural and urban inhabitants, suggesting there is scope to explore drivers of this variation, prioritize nutritional interventions, and to design more representative surveillance programs.

Serum Zn concentrations. The estimated population mean serum Zn concentration is 57.7 μg dL −1 (median of 57.5 μg dL −1 ) following adjustment for inflammation. Estimated population serum concentrations vary by Region. The Tigray Region had the lowest median serum Zn concentration (55.2 μg dL −1 ) while the highest median serum Zn concentration (68.3 μg dL −1 ) was recorded for participants from Addis Ababa. The disaggregated data of average serum Zn concentration by administrative zones for the larger regions is indicated in Supplementary Table S1. The median serum Zn concentrations varied by demographic groups, with 47.1 μg dL −1 observed in YC and 62.6 μg dL −1 in men (Table 2). Urban populations had a greater serum Zn concentration than the rural populations. Serum Zn concentration was greater in households whose head was educated.
Adjustments for inflammation decreased the estimated overall prevalence of Zn deficiency from 75.1 to 72.1% (Table 2). Those who were affected with inflammation had lower serum Zn status than those who were not affected. The reference values used to adjust for inflammation were > 5 mg L −1 for CRP and 1 g L −1 for AGP. These values are the cut-off points for determining inflammation 21 and as such, some observations which are classified as not inflamed were also adjusted for inflammation.
Among the population within this study, 31% had elevated CRP and/or AGP concentrations indicative of inflammation (Table 1). Therefore, exclusion of these groups from the analyses would reduce the sample size and may bias the results, as reported in other studies 7 .
We determined Zn deficiency prevalence across regional and demographic groups before and after correcting serum Zn concentration based on the BRINDA internal regression correction approach 22 . The prevalence of serum Zn deficiency was different across residence categories (55.3% in Urban vs 74.9% in Rural) ( Table 2). The highest prevalence of serum Zn deficiency was found in Oromia (76.7%), Tigray (75.5%), Amhara (71.5%) Regions, and the lowest in Addis Ababa (25.4%).

Correlations among Serum Zn, inflammation biomarker, age and incidence of diarrhea. Serum
Zn level was correlated with age of participant (r = 0.24). Serum Zn concentration was negatively correlated with inflammation (AGP) (r = − 0.14).
Spatial variation of Zn status. The spatial variation of Zn status was explored using the data for WRA, applying methods similar to those described in a previous study on Se 15 . Thus, estimates of the variogram were obtained using alternative approaches, with exponential models fitted by weighted least squares (Fig. 1).
The cross-validation errors for the model fitted to the estimates using Matheron (1962) 23 appear to be normally distributed (Fig. 2), with a median SSPE of 0.414. This model was therefore selected given that the SSPE There is therefore limited spatial dependence in the variation of serum Zn concentration. A large proportion of the overall variance is attributable to the spatially uncorrelated nugget effect, and the correlated variance is spatially correlated up to distances of 45 km, as seen in the variograms 24 , suggesting that Zn nutritional status of WRA in Ethiopia is determined by factors that vary over short distances.
The spatial variation predicted for serum Zn concentration for WRA across Ethiopia obtained by ordinary kriging is shown in Fig. 3a. The larger predicted serum Zn concentrations were found in some areas of the country including Addis Ababa while smaller predicted Zn concentrations were observed in Oromia Region, in line with the observations made on the population-weighted predictions aggregated by Region (Table 2). However, this approach illuminates intra-Region variation which is predicted, based upon the location of the samples. The kriging variances of the predictions of serum Zn concentration for WRA are shown in Fig. 3b; these are the expected squared errors of prediction which are minimized by the kriging prediction based on the variogram model.
As described previously for Se 15 , these interpolation errors are inevitable for all spatial predictions. This is because, given that serum Zn varies spatially, predictions are being made from sparse observations. However, by using the kriging variance (Fig. 3b), we can understand better where spatial predictions are reliable from a Table 1. Characteristics of study participants in the Ethiopian National Micronutrient Survey (ENMS), 2015. Stages of inflammation were categorized as: no inflammation (CRP ≤ 5 mg L −1 and AGP ≤ 1 g L −1 ); incubation (CRP > 5 mg L −1 and AGP < 1 g L −1 ); early convalescence (CRP > 5 mg L −1 and AGP > 1 g L −1 ); late convalescence (CRP ≤ 5 mg L −1 and AGP > 1 g L −1 ) based on Thurnham et al. (2010) 21 . www.nature.com/scientificreports/ particular distribution of observations. It is apparent that there is uncertainty in most parts of the country because of the short distance over which serum Zn is spatially correlated and the sparse distribution of observations 26 . In these conditions, the predicted value tends to the mean of the observations, and the kriging variance is large.
Where the kriging variance is large (shaded in light gray colors), further local sampling would be needed to determine Zn deficiency prevalence with greater confidence. Figure 4 shows a probability map that the mean serum Zn concentration of WRA is below the cut-off for adequacy (< 70 μg dL −1 ). The map legend includes the "calibrated phases'' of the Intergovernmental Panel for Climate Change (IPCC) 27 , to improve the communication of uncertainty to all stakeholders 28,29 . This shows that WRA in large parts of Ethiopia are predicted to be 'likely' through to 'virtually certain' to have deficient serum Zn status (Fig. 4). In central parts of Ethiopia and other small parts of Northern and Eastern Ethiopia, WRA are 'unlikely' through to 'exceptionally unlikely' to be Zn deficient.

Discussion
Reanalysis of archived serum samples from the ENMS using ICMPS, with data adjusted for inflammation status, revealed 72.1% of the population of the ENMS were Zn deficient. The median serum Zn concentrations was 57.5 µg dL −1 , which is low compared to data from other countries 20 , and the lower threshold considered indicative of optimal nutritional status. The highest prevalence of Zn deficiency was in Oromia (76.7%) and Tigray Regions (75.5%); the lowest prevalence was in Addis Ababa (25.4%). Serum Zn concentrations varied between demographic groups with median concentrations greatest among men (62.6 µg dL −1 ) and smallest among YC (47.1 µg dL −1 ). Age was significantly positively associated with serum Zn concentration (r = 0.24) ( Table 3). The increase in serum Zn concentration with age is consistent with studies in the USA and France 30,31 . Table 2. Serum Zn concentrations and prevalence of deficiency by study characteristics in the Ethiopian National Micronutrient Survey (ENMS). a The estimation was weighted using the ENMS sampling weight factor. b Zn deficiency was defined based on serum Zn considering age, sex, fasting status and time of blood collection 10 . c BRINDA internal regression correction approach, which accounts for both CRP and AGP, was applied to calculate the prevalence of Zn deficiency (Namaste et al. 2017) 22 . www.nature.com/scientificreports/ In this study, inflammation was detected among 11% and 28% of participants based on elevated serum CRP or AGP concentrations, respectively (9% of partcipants had elevated concentration of both). Mean serum Zn concentration was highest in participants with no inflammation (57.3 µg dL −1 ) and lowest in participants with inflammation in the early convalescence state (53.6 µg dL −1 ). The overall correlation between inflammation and serum Zn was relatively low (r = − 0.14 with AGP). The relationship between serum Zn concentration and inflammation biomarkers have not been widely studied 32 . However, it has been noted previously that infection can produce Acute Phase Reactants (APRs) and a decline of serum Zn concentration 17 . In a review of data from micronutrient surveys in thirteen countries, the lowest serum Zn concentrations were found in YC classified as being in the early convalescence state of infection, which is consistent with the finding of our current study 20 .
Along with inflammation, serum Zn concentration is also influenced by time of the day of blood collection in a fasting or nonfasting state. In this study the highest mean (61.7 µg dL −1 ) and median (61.7 µg dL −1 ) concentrations were observed in participants who had blood drawn in the fasting state, followed by samples taken in the morning from those in the non-fasting state. This is consistent with other studies 33 . Therefore, to determine Zn status in populations, inflammation biomarkers and time of the day of blood collection should be taken into consideration to estimate a reliable Zn deficiency prevalence.
The largest prevalence of deficiency was observed in Oromia and Tigray, followed by Amhara Regions. Tessema et al. 2018 34 reported that half of the households in Tigray (50%) and a quarter of households in Amhara (25%) live in areas where soil concentrations of Zn are low, potentially limiting entry of Zn into local diets. In their study, the correlation between serum Zn and soil Zn was significant in Ethiopia, but it was relatively low (r = 0.09), indicating that other determinants affect serum Zn.
The large proportion of the population we have estimated to have inadequate Zn nutritional status is a finding which is consistent with other approaches that have been published. In Ethiopia, a high national prevalence of Zn deficiency has been estimated previously based on Zn supplies in the food system, which reported that 81% of the population were at risk of deficiency due to inadequate dietary Zn supplies, estimated from food balance sheets www.nature.com/scientificreports/ and regional food composition data 11 . A high prevalence of Zn deficiency (> 20%) has also been reported based on serum Zn status and dietary Zn intake among young children and school-aged children 13,34-36 and women and pregnant women 12,37,38 . Several factors may contribute to a high prevalence of Zn deficiency in Ethiopia, where the consumption of Zn-rich foods such as meat is low 39 . Food systems are highly localized in Ethiopia, particularly in rural areas, with a large proportion of dietary intakes met through subsistence production or purchases of locally-produced food 40 . Thus, individuals' Zn status may reflect the soil types and landscapes where they reside. From the geostatistical analysis, the selected variogram (Fig. 1a, Matheron) and the fitted parameters show that about 60% of the variance is spatially dependent (i.e. the spatially dependent component as a proportion of the sum of the two variance components). However, spatial dependence is limited to about 45 km (i.e. 3 × the distance parameter). That is to say, two observations which are close together (less than 45 km) are more likely to be similar than two observations separated by a greater distance. The correlation between any two observations decays with distance, here approaching zero at about 45 km. This is a short distance, certainly relative to the size of Ethiopia, so while we can see this short-range spatial pattern in the kriged map, the spatial pattern is seen at fine spatial scales and more detailed sampling would be needed to resolve it and to develop targeted interventions should these be desired. Figure 3b also shows us the kriging variance is very high in most parts of the country. A larger proportion of the total variance is in the uncorrelated "nugget" effects and spatial dependence is seen over shorter distances. As noted above, this does have implications for the potential to design interventions that are tailored to local conditions, for example agricultural interventions that might influence Zn status of crops in smallholder subsistence settings, including use of inputs such as manure and fertilizer 41 .
These new detailed analyses of the ENMS could inform programs or policies to improve surveillance and potentially alleviate Zn deficiency. The joint WHO/UNICEF/IAEA/IZiNCG meeting on Zn status indicators agreed that in populations where the Zn deficiency prevalence rate is > 20% it should be considered a public health concern, which is the case in Ethiopia 42 . Potential strategies to alleviate Zn deficiency include the promotion of dietary diversity and biofortification, i.e. the application of Zn to crops via fertilizers 43,44 and the release of crop varieties containing higher concentrations of Zn 45 . The spatial analysis of these data show that, while there is spatial dependence in the biomarker, this is limited to short distances. This limits the value of spatial mapping for designing interventions to address Zn deficiency at national scale. However, it does suggest that, at local scale, there could be variations in the prevalence of Zn deficiency which mean that more efficient interventions could be better targeted on the basis of spatial information, based either on direct sampling of the biomarker or a better understanding of the environmental factors which drive the observed variability.
Strengths of our study include the large population coverage, comprising several demographic groups, using a cross-sectional sampling design. We improved on previous estimates of population Zn status by using more accurate and sensitive instrumental analyses to determine Zn concentration in serum samples, employing advanced geostatistical methods to predict the Zn status of the population across unsampled locations, and by adjusted serum Zn concentrations considering inflammation biomarkers using different cutoffs dependent on time of blood draw. However, in the ENMS, EAs are not evenly distributed between regions proportional to size of the population. In addition, in areas of small observations such as Afar, Somali and south-east Oromia regions the uncertainty of the prediction is large which warrants further sample analysis from the areas for predictive mapping work and estimating Zn nutrition status of the wider population. Rural electrification and access to roads is known to improve nutritional status through increasing employment, access to health care services, improving nutritional knowledge, reducing fertility, improved access to market, and creating wealth 46,47 . Exclusion of EAs www.nature.com/scientificreports/ in the present study due to lack of access to road and electricity may underestimate the magnitude of national and regional Zn deficiency prevalence, suggesting the importance of further studies in the areas not covered in the present study. Analyses of potential contributors to variation in serum Zn in different areas and among different

Conclusions
The present study shows that Zn deficiency is highly prevalent in Ethiopia, and provides the first estimation which accounts for the impact of inflammation on the assessment of Zn status in Ethiopia. Risk of Zn deficiency is spatially dependent within a short distance up to 45 km, which is likely to be due to environmental and food system factors including soil type, landscape features, and food production and distribution 11 . This study may inform national strategies to enhance Zn status of the Ethiopian population, and to improved micronutrient surveillance programs.  Table 3. Spearman's rank correlations between serum Zn, inflammation biomarkers, the incidence of diarrhea, and age of all demographic groups. AGP α-1-glycoprotein concentration, CRP C-reactive protein concentration.

Serum Zn (μg dL −1 ) AGP (g L −1 ) CRP (mg L −1 ) Age (years) Diarrhea
Serum Zn (μg dL −1 ) 1.00  Following the household census conducted in all selected EAs preceding the actual survey, 13 of them were excluded due to lack of infrastructure (e.g. EAs that can't be accessed by car or pack animals or are more than 5 h walking distance) and security issues. Within each selected EAs 11 households were randomly selected for enumeration. EAs contain on average 181 households (150-200) 49 . The present study uses two datasets: socio-demographic data collected from georeferenced households in the ENMS, and newly-generated data on Zn concentrations measured in archived serum samples from the ENMS. The data on Zn concentrations were obtained only for those individuals from the ENMS survey for whom sociodemographic data, inflammation biomarker and location information from GPS were available, and where there was at least 0.3 mL of archived serum sample. A total of 3373 out of 4700 (71.7%) of the archived serum samples contained sufficient material for analysis of Zn concentration.

Materials and methods
Data collection and analysis. Socio-demographic data. Socio-demographic information was collected using a structured questionnaire. The research assistants and supervisors were trained on data collection. The questionnaire was pilot tested in a cluster that was not selected for the actual survey. Questionnaires were refined based on the finding from the pilot testing before the actual data collection began 15 .
Collection, processing and analysis of serum Zn. Blood collection and processing methods are described elsewhere 14,15,34 . Briefly, venous blood samples were collected from participants by trained phlebotomists and processed as per the World Health Organization (WHO) blood collection guidelines 50 . Samples were aliquoted www.nature.com/scientificreports/ in the field and subsequently analyzed for a range of micronutrients 14 . One aliquot of each sample was stored at EPHI at − 80 °C as archived material in case analyses needed to be repeated or additional micronutrient analyses could be supported, as is the case with the present study. The concentration of Zn in serum samples was determined using ICP-MS (Thermo Fisher Scientific iCAPQ, Thermo Fisher Scientific, Bremen, Germany). Samples were introduced, via a single line, from an auto-sampler incorporating an ASXpress™ rapid uptake module (Cetac ASX-520, Teledyne Technologies Inc., Omaha, NE, USA) through a perfluoroalkoxy (PFA) Microflow PFA-ST nebulizer (Thermo Fisher Scientific, Bremen, Germany). All samples and external element calibration standards were diluted as 0.3 mL added to 6 mL of a solution containing: (1) 0.5% HNO 3 (Primar Plus grade), (2) 2.0% methanol (Fisher Scientific UK Ltd, Loughborough, UK) and (3) three internal standards including 72 Ge (10 µg L −1 ), 103 Rh (5 µg L −1 ), 193 Ir (5 µg L −1 ) (SPEX Certiprep Inc., Metuchen, NJ, USA). Zn calibration standards were prepared at 0, 20, 40, 100 µg L −1 (Claritas-PPT grade CLMS-2; SPEX Certiprep Inc., Metuchen, NJ, USA). The ICP-MS was operated in 'collision-reaction cell mode' .
The limit of detection (LOD) for Zn was measured as 3 × standard deviation of 10 operational blanks; the limit of quantification (LOQ) was calculated as 10 × standard deviation. The LOD and LOQ were 1.43 and 4.77 µg L −1 , respectively. Accuracy was verified by the use of two Seronorm™ reference materials: L-1 (Lot 1801802) and L-2 (Lot 1801803) (Nycomed Pharma AS, Billingstad, Norway). These were prepared in an identical way to the samples and typically run at the start and the end of each analytical run. Average Zn recovery (%; n = 24) when compared to accredited values determined across 10 analytical batches of blood serum was 88% and 86% for L-1 and L-2, respectively. Serum Zn concentrations were corrected for infection using the BRINDA internal regression approach 22 . This was done by running a linear regression model with lnZinc as the dependent variable; and lnCRP and lnAGP as the independent variables. This generated the slope (regression coefficient) of lnCRP (β1) and lnAGP (β2). These two parameters were then used to adjust for the effect of inflammation, for observations whose CRP and AGP values were above the reference values. A reference value for serum lnCRP and lnAGP is used to avoid over-adjusting the micronutrient biomarkers among individuals with low levels of inflammation. The reference values for this analysis were the maximum value of the lowest decile for lnCRP (p10) and that of lnAGP (p10).
The equation was thus, adjusted Zn concentration = exp [unadjusted lnZinc-β1(CRPobserved -maximum of lowest decile for CRP) -β2 (AGPobserved -maximum of lowest decile for AGP)]. The model was checked to ensure all assumptions were met by examining the plot of residuals, homogeneity of variance, and normality.
Descriptive statistical analyses were conducted in STATA (Version 14.0, StataCorp LLP, Texas, USA). Prevalence of Zn deficiency among all demographic groups in Ethiopia were determined as suggested by Hotz et al. 2003 33 . Zn deficiency defined as serum Zn concentration < 65 µg dL −1 for morning, non-fasting samples for YC and SAC under 10 years of age and < 57 µg dL −1 for afternoon, non-fasting samples. For those 10 years old and older, a cutoff of < 70 µg dL −1 was used for males and < 66 µg dL −1 was used for females for morning, non-fasting samples; the cutoff of < 61 µg dL −1 was used for males and < 59 µg dL −1 was used for females for afternoon, nonfasting samples. A cutoff of < 74 µg dL −1 was used for men and < 70 µg dL −1 was used for WRA for morning, fasting samples 10 .
Serum Zn concentrations were mapped at national level for WRA only. This is because WRA had the largest sample size in the ENMS and the sample covered more EAs than participants of other demographic groups in the survey. Inclusion of all age categories in the modelling could have clustering effect causing a non-reliable predictive mapping, and would have limited our scope to compare to appropriate biomarker thresholds which are specific to demographic groups due to the variation in concentration expected with age. Even though serum zinc concentration of the present study participants differs by age category, it isn't expected that the pattern of serum zinc concentration distribution to be different among these groups. This is because the study participants were either from the same households or villages dependent on local food sources. For example, our previous studies show that the spatial pattern of distribution of serum Se concentration among women of reproductive age and young children in the Amhara region was not different 15,51 . Ordinary kriging predictions were computed on the nodes of a 60-m square grid. In geostatistical terminology, a prediction is obtained from the prediction distribution of a random variable at an unsampled location. This is a conditional distribution, conditional on the geostatistical model (the variogram in the case of ordinary kriging) and the observed values at sample locations. In ordinary kriging, the prediction is the mean of the prediction distribution, and its variance is the ordinary kriging variance.
Data values were first aggregated to mean values for each EA, and mean coordinates were computed for each EA. One single value was a probable outlier according to the criteria of Tukey (1977) 25 . This was removed from the data for variogram estimation but was returned for spatial prediction by kriging. This is because estimates of the variogram are particularly susceptible to extreme values 26 but if, as here, there is no reason to think the observation erroneous then it should be considered for local prediction.
Because of the large extent of the sampled region, all spatial analyses were done using the latitude and longitude recorded for the observations, rather than on a rectilinear grid since no single projection would be suitable at all locations. Under the usual assumptions of stationarity in geostatistics, spatial dependence is modelled in www.nature.com/scientificreports/ terms of the vector that separates locations, or lag. In this setting, the lag distance between any two points was computed as the great circle distance on a spherical approximation. This was computed using the dist Vincenty Sphere function from the geosphere package for the R platform 52 . The function final Bearing from the same package was used to obtain bearings. Note that, when distances between locations are measured on the sphere, then not all variogram functions that are suitable for use on the plane can be applied. On the sphere the exponential model is authorized 53 , and so it was used for all variogram modelling in this study. Exploratory variogram analysis indicated that there was no marked directional dependence of the variogram (Supplementary Figure S1), so an isotropic variogram (Supplementary Figure S2) was estimated using the standard estimator due to Matheron (1962) 23 . In addition, variogram values were estimated using the robust estimators due to Dowd (1984) 54 and to Cressie and Hawkins (1980) 55 . These alternative estimators were considered because, although a marginal outlier (an unusual datum as it appears on the overall distribution) was removed, spatial outlier(observations unusual in its spatial context) can inflate estimates of the semivariance 26 . An exponential variogram model 24 was fitted by weighted least squares, and then tested by cross-validation. In cross-validation each observation was, in turn, removed and predicted from the remaining observations by ordinary kriging. The squared cross-validation error at each site was standardized by dividing by the ordinary kriging variance. The median standardized squared prediction error (SSPE) has an expected value of 0.455 in the case of a valid underlying variogram model with kriging errors that appear normally distributed 25 . The estimator due to Matheron (1962) 23 is more statistically efficient than the robust alternatives, so if the model fitted to these estimates appeared correct from the cross-validation results (median SSPE) then the alternatives were not considered. If the SSPE suggests that the model fitted to Matheron estimates are affected by outliers, then the models fitted to robust estimates are also cross-validated, and one is selected on the cross-validation results. This procedure follows the recommendations of Lark (2000) 26 .
Our spatial analysis is focused on WRA because this was the demographic group with the largest sample size with coverage of 315 out of 343 enumeration areas. The statistical summary for WRA is shown in Supplementary Figure S3.
Ethical approval. This study was conducted according to the national research ethics guideline by the Ethiopian Science and Technology Ministry. All procedures involving human subjects were approved by the National Health Research Ethics Review Committee at the Ethiopian Science and Technology Ministry (Reference 3.10/433/06) and support letters were obtained from the respective regional and local health bureaus. Adult participants gave informed consent for their participation. Informed consent was also obtained from mothers or guardians for their children to participate in the study. In addition, assent was obtained from participating school age children. A separate ethical approval was obtained for the present study from the Research Ethical Review Committee of the Ethiopian Public Health Institute (Protocol EPHI-IRB-140-2018). Archived serum samples were transferred from storage at EPHI to the University of Nottingham, UK for analysis under a Material Transfer Agreement.