Spatial distribution of iron rich foods consumption and its associated factors among children aged 6–23 months in Ethiopia: spatial and multilevel analysis of 2016 Ethiopian demographic and health survey

The data retrieved from the standard 2016 Ethiopian Demographic and Health Survey (EDHS) dataset. A nationally representative cross-sectional survey was conducted from January 18, 2016, to June 27, 2016, which is available from the standard measure Demographic and Health Survey database website. EDHS 2016 was the fourth nationally representative survey. Ethiopia is situated in the horn of Africa from 3⁰ to 14⁰N and 33⁰ to 48°E.

All living children aged 6–23 months were the source population, whereas, all selected/sampled living children aged 6–23 months living with their mother were the study population. In 2016 EDHS, a total of 645 clusters (Enumeration Areas) (202 urban and 443 rural) were selected with a probability proportional to each Enumeration Areas size and independent selection in each sampling stratum. Among the total of selected clusters with zero coordinates and clusters not had the proportion of consumption iron rich foods were excluded for the analysis. Finally, a total of 598 (185 urban and 413 rural) clusters were included for this study. In the selected Enumeration Areas (EAs) a total of 3055 weighted number of living children aged 6–23 months living with their mother were included for this study. The recorded data were accessed at www.​measuredhs.​com.

Ethiopian Demographic and Health Survey data were collected by two-stage stratified sampling. Each region of the country stratified into urban and rural areas, yielding 21 sampling strata. In the first stage, 645 Enumeration Areas selected with a probability proportional to the Enumeration Area size by independent selection in each sampling stratum. In the second stage of selection, a fixed number of 28 households per cluster selected with an equal probability systematic sampling from the newly created household listing. The detailed sampling procedure was available in the Ethiopian Demographic and Health Survey reports from measure DHS website (www.​dhsprogram.​com).

From the standard 2016 EDHS dataset, age of the mothers’, educational status of mother and father, occupational status of the mother and father, mother’s religion, family size, wealth status, media exposure, child sex, and age of the child considered as the individual level independent variables. Whereas, residence and administrative region of the country were taken as community (cluster) level independent variables.

The data were cleaned using STATA version 16.0/MP software and Microsoft Excel. Sample weighting was performed for further analysis.

Spatial autocorrelation (Global Moran’s I) statistic measure used to assess spatial heterogeneity for consumption iron rich foods among children age 6–23 months. Moran’s I values close to − 1 indicates poor consumption of iron rich foods which is dispersed, close to + 1 indicates clustered, and if Moran’s I value zero indicates randomly distributed [12]. A statistically significant Moran’s I value (p <  0.05) had a chance to reject the null hypothesis which indicates the presence of spatial autocorrelation. Hot spot analysis (Getis-Ord Gi* statistic) z-scores and significant p-values gave the features with either hot spot or cold spot values for the clusters spatially.

The spatial interpolation technique used to predict poor consumption of iron rich foods among children aged 6–23 months for un-sampled areas in the country. For the prediction of un sampled Enumeration Areas, we used geo-statistical Empirical Bayesian Kriging spatial interpolation techniques using ArcGIS 10.7(ESRI Inc., Redlands, CA, USA, version 10.7) software. Empirical Bayesian Kriging relaxes the assumption of the Gaussian distribution of the observed semi-variogram in the input data which rarely holds in practice. Empirical Bayesian Kriging interpolation works by generating a new simulated semi-variogram at each location from the estimated semi-variogram from the input data. The weight of the new simulated semi-variogram was calculated by Bayes’ rule (Krivoruchko, 2012 #1) [13].

We employed Bernoulli based model spatial scan statistics to determine the geographical locations of statistically significant clusters for poor consumption of iron rich foods among children aged 6–23 months using Kuldorff’s SaTScan version 9.6 software [14]. The scanning window that moves across the study area in which children aged 6–23 months with poor consumption of iron rich foods taken as cases and those good consumption taken as controls to fit the Bernoulli model. The default maximum spatial cluster size of less than 50% of the population used as an upper limit, allowing both small and large clusters to be detected, and ignored clusters that contained more than the maximum limit with the circular shape of the window. Most likely clusters identified using P-values and log-likelihood ratio tests based on the 999 Monte Carlo replications.

The nature of EDHS dataset was Hararichial which is children nested in household and household is nested in clusters. Therefore, the observations within the cluster correlated (dependant) that violates the assumption of independence. In correlated data, fitting models by ignoring these correlations will bias the parameter estimates. The correlation within the cluster diagnosed by intraclass correlation (ICC) value. If the ICC value greater than 0.25 [15] was reasonable to fit multilevel mixed-effects logistic regression model. The intraclass correlation (ICC) calculated by the variation between and within-cluster variation (ICC = ϭ²/ (ϭ² + π²/3). We fitted four models, the null model without independent variables, model I with only individual-level variables, model II with only community-level variables, and model III both the individual level and community level variables. The model was fitted by a STATA version 16/MP“xtmelogit” command for multilevel mixed-effects logistic regression analysis. We used the Log-Likelihood Ratio (LLR) for model comparison. The highest log-likelihood was taken as the best fit model. Median Odds Ratio (MOR = exp. [squrt (2ψΦ^− 1 (3/4))], where ψΦ^− 1 = variance of the cluster) was employed to observe the heterogeneity between clusters. The MOR quantifies the variation between clusters (the second-level variation) by comparing two persons from two randomly chosen different clusters. Considering the two persons with the same covariates, chosen randomly from two different clusters; If we select on child randomly from a cluster the median value of the odds ratio at the high-risk cluster compared to low risk cluster for the consumption of iron rich foods. The measure is always greater than or equal to 1. If the MOR is 1, there is no variation between clusters (no second-level variation). If there is considerable between-cluster variation, the MOR will be large. The measure is directly comparable with fixed-effects Odds Ratio [16]. Percentage Change Variation (PCV) measures the total variation attributed by individual and community level factors in the multilevel model as compared to the null model which is calculated by the difference from variance in the null model to variance in the full model divided by variance in the null model.

The authors, submitted concept note to DHS Program/ICF International Inc., letter of permission was waived from the International Review Board of Demographic and Health Surveys (DHS) program data archivists to download the dataset for this study.

From the EDHS dataset, a total of 3055 children aged 6–23 months were included in this study. More than half (53%) of the children were females. Eighteen and 36.6% of the children were between the age of 6–8 and 12–17 months. The mean ± SD of the age of the children was 14 ± 5 months. The majority (67.50%) of the children were born from mother’s age between 20 and 34 years. The mean ± SD age of the mothers was 28 ± 6 years. Approximately 61% of the mothers and 45% of fathers had no formal education. Around 44% of the children born from poor household wealth status (Table 1).

Overall, 21.41% (95% CI: 19.9–22.9) of children aged 6–23 months had good consumption of iron rich foods in Ethiopia. Consumption of egg was the most taken food items among children aged 6–23 months. Consumption of fish or shellfish foods were the least taken food (Table 2).

The incremental spatial autocorrelation evidenced that with 10 distance bands starting from 121.8 km (km) iron rich foods consumption clustering was detected at 152.4 km distance. Statistically, significant z-score indicates at 152.4 Km distance spatial clustering of iron rich foods consumption was most pronounced (Fig. 1). The spatial distribution of poor consumption of iron rich foods among children aged 6–23 months were non-random in Ethiopia (Global Moran’s I = 0.16, z-score = 10.45, P-value < 0.001). This evidenced that spatial hot spot and cold spot clustering identified in the regions of Ethiopia. Given the z-score of 10.48, there was less than 1% likelihood that this high-clustered pattern could be the result of random chance (Fig. 2).

The spatial hot spot analysis was predicted using incremental spatial autocorrelation maximum pick distance value 152.4 km. As shown in Fig. 3 below, the red color is intense clustering of high (hot spot) proportion with poor consumption of iron rich foods among children aged 6–23 months in Ethiopia. The prevalence of poor consumption of iron rich foods among children aged 6–23 months were clustered at Southeastern Amhara, Southern part of Afar, and Northern part of Somali Regional States of Ethiopia; whereas, Addis Ababa, Gamebela, Central Oromia, and Northern part of Southern Nations, Nationalities, and People’s Regional states of Ethiopia were less risk area for poor consumption of iron rich foods among children aged 6–23 months (Fig. 3).

As shown in Fig. 4 below, the red window indicates significant clusters. A total of 190 significant clusters identified. Among the significant clusters, 142 were most likely and the other 48 were secondary and tertiary clusters. The most likely (primary) clusters were located at 11.626646 N, 39.666950 E within 278.08 km radius in Afar, the Southeastern part of Tigray, and the Eastern part of Amhara Regional States of Ethiopia. Children aged 6–23 months living in the most likely cluster were 21% more likely vulnerable to poor consumption of iron rich foods than outside the window (RR = 1.21, LLR = 40.107, P-value < 0.001) (Table 3 and Fig. 4).

We employed an Empirical Bayesian Kriging interpolation method. The interpolation result revealed that children aged 6–23 months were venerable to poor consumption of iron rich foods in most part of Ethiopia. Relatively good consumption of iron rich foods among children aged 6–23 months were observed at Tigray, Gambela, Addis Ababa, and Central Oromia Regional States of Ethiopia (Fig. 5).

In the multilevel mixed effects logistic regression model, the full model contains both individual and community-level factors that were the best fit model (LLR = − 1223.36). From the null model, 27% of the variability of the odds of iron rich foods consumption accounted for intra class correlation (ICC =0.27). As a rule of thumb, the intra class correlation (ICC) variation greater than 0.25 indicates that intra-cluster correlation. The dependency within the cluster accounted for multilevel mixed effects model which is reliable by adjusting individual and community level factors. The experience of iron rich foods consumption was heterogeneous between clusters which were explained by Median Odds Ratio (MOR). In the full model, the MOR was 2.38 which means if we select one child in different cluster live in less risk cluster had 2.38 times more likely to consume iron rich foods than children live in high risk cluster (MOR = 2.38,95% CI: 1.91–4.34) (Table 2).

In multivariable multilevel mixed-effects logistic regression analysis, child age, educational status of the mother, wealth index, and region were statistically significant factors for good consumption of iron rich foods in Ethiopia.

After adjusting other variables, children aged between 12 and 17 months was 90% more likely to consume iron rich foods than children aged between 6 and 11 months (AOR = 1.90, 95% CI: 1.45–2.49). Besides, the odds of good consumption of iron rich foods among children aged between 18 and 23 months were 2.05 times higher than children aged between 6 and 11 months (AOR = 2.05, 95% CI: 1.55–2.73).

The odds of having good consumption of iron rich foods were 2.26 times higher than children who born from mother with secondary and above education level as compared to children born from non-educated mother’s (AOR = 2.26, 95% CI: 1.47–3.46). Similarly, children who born from mother with primary education was 42% times more likely to have good consumption of iron rich foods than children born from non-educated mother (AOR = 1.42, 95% CI:1.06–1.87).

Children aged 6–23 months from Afar and Amhara regional state of Ethiopia were 86% (AOR = 0.24, 95% CI: 0.09, 0.60) and 62% (AOR = 0.38, 95% CI: 0.17, 0.84) less likely to have good consumption of iron rich foods as compared to children live in Addis Ababa, Ethiopia respectively. Whereas, children live in Harari regional state of Ethiopia were two times more likely to consume iron rich foods than children living in Addis Ababa, Ethiopia (AOR = 2.11, 95CI: 1.02–4.39) (Table 4).

First, this study was conducted by using a nationally representative demographic and health survey dataset which may give better generalization. Secondly, accounting spatial heterogeneity was a strength for this study. Thirdly, accounting clustering effect using a mixed-effects model was also another strength for this study.

As well this study had some limitations. The first limitation was 2016 Ethiopian Demographic Health Survey dataset lacks some spatial coordinates in the Somali region which may mislead to spatial statistics estimate. Since the data were collected using face to face interview it might prone to recall and social desirability bias. The drawback of the secondary nature of data was inevitable.

A significant number of children aged 6–23 months hadn’t good consumption of iron rich foods. The spatial distribution of iron rich foods in Ethiopia was non-random. Spatial clustering of low consumption of iron rich foods was identified at Southeastern Amhara region, Eastern part of Tigray, the Southern part of Afar region, and some part of Northern Somali regional states of Ethiopia. Child age, educational status of the mother, wealth index and regions of Ethiopia were significantly associated with iron rich foods consumption among children age 6–23 months. Responsible stakeholders should give primary attention to children living in the hotspot areas of Ethiopia. Besides, complementary iron rich foods consumption should be promoted particularly in the age group of 6–11 months of children.