Predictors of stunting among children age 6–59 months in Ethiopia using Bayesian multi-level analysis

In developing countries including Ethiopia stunting remained a major public health burden. It is associated with adverse health consequences, thus, investigating predictors of childhood stunting is crucial to design appropriate strategies to intervene the problem stunting. The study uses data from the Ethiopian Demographic and Health Survey (EDHS) conducted from January 18 to June 27, 2016 in Ethiopia. A total of 8117 children aged 6–59 months were included in the study with a stratified two stage cluster sampling technique. A Bayesian multilevel logistic regression was fitted using Win BUGS version 1.4.3 software to identify predictors of stunting among children age 6–59 months. Adjusted odds ratio (AOR) with 95% credible intervals was used to ascertain the strength and direction of association. In this study, increasing child’s age (AOR = 1.022; 95% CrI 1.018–1.026), being a male child (AOR = 1.16; 95%CrI 1.05–1.29), a twin (AOR = 2.55; 95% CrI 1.78–3.56), having fever (AOR = 1.23; 95%CrI 1.02–1.46), having no formal education (AOR = 1.99; 95%CrI 1.28–2.96) and primary education (AOR = 83; 95%CrI 1.19–2.73), birth interval less than 24 months (AOR = 1.40; 95% CrI 1.20–1.61), increasing maternal BMI (AOR = 0.95; 95% CrI 0.93–0.97), and poorest household wealth status (AOR = 1.78; 95% CrI 1.35–2.30) were predictors of childhood stunting at individual level. Similarly, region and type of toilet facility were predictors of childhood stunting at community level. The current study revealed that both individual and community level factors were predictors of childhood stunting in Ethiopia. Thus, more emphasize should be given by the concerned bodies to intervene the problem stunting by improving maternal education, promotion of girl education, improving the economic status of households, promotion of context-specific child feeding practices, improving maternal nutrition education and counseling, and improving sanitation and hygiene practices.


Results
Socio-demographic characteristics of respondents. A total of 8117 children age 6-59 months were included in the analysis. The median age of children were 31 months (IQR = . The mean maternal BMI was 20.70 kg/m 2 with (SD ± 3.40). In addition, the mean height of mothers was 158 cm with (SD ± 6.7). Most of the study participants about, 7242 (89.22%) were rural residents.
Regarding family wealth index about 1901(23.42%) of respondents were from the poorest house.
Majority of children about, 7020 (86.49%) belongs to male household head. Regarding maternal education, nearly two third, 5388 (66.38%) of children had mothers who have no formal education.
In this study about, 3427 (42.96) of respondents were not anemic. Of the children, 7109 (87.59%) had no history of diarrhea and about, 4429 (54.57%) were not received vitamin A supplementation in the last 6 months.
Most of the households about, 7353 (90.59%) had not used improved toilet facilities, and about, 4582 (56.45%) of the households obtained water from improved sources ( Table 1).
Childhood stunting was more prevalent in males 1838 (44%) than females 1505 (38%). Also, the likelihood of stunting was highest in children from mothers who had no formal education 2388 (41%). Similarly the magnitude of stunting was higher among children from the poorest households 925 (49%) and who had mild anemia 130 (52.5%) ( Table 2). Predictors of stunting among children age 6-59 months in Ethiopia. In the Bayesian multi-level multivariate logistic regression analyses model both individual and community level factors were included.
Among individual level factors sex of child, child age, type of birth, anaemia status of the child, history of fever in the last 2 weeks, education level of the mother, maternal BMI, maternal height, birth interval, and family wealth index were identified as significant predictors of childhood stunting. Among community level factors included in the study only type of toilet facility and region were significant predictors of childhood stunting.
Individual level factors. The Bayesian multi-level multivariable logistic regression analysis result showed that being male children were 16% more (AOR = 1.16, 95% CrI 1.05-1.29) likely to be stunted as compared to female children.
Children of multiple birth (twin) was 2.55 times (AOR = 2.55, 95% CrI 1.78-3.56) more likely to be stunted as compared to their counterpart. Children who had fever were 23% (AOR = 1.23, 95% CrI 1.02-1.46) more likely to be stunted compared to the counterparts.
Children from poorest house hold wealth status were 78% (AOR =  Study design. A Community-based cross-sectional study design was conducted among children aged 6-59 months.

Methods
Population and sample. All children aged 6-59 months resided in Ethiopia (nine regions and two city administrations) were the source population and the study population was all children aged 6-59 months who were living in the selected households of Ethiopia. A total of 8117 children age 6-59 months which fulfil the inclusion criteria were included in the analysis.
A stratified two-stage cluster sampling procedure was employed where enumeration areas (EA) were the sampling units for the first stage and households for the second stage. In the 2016 EDHS, a total of 645 EAs (202 urban and 443 rural) were selected with a probability proportional to EA size (based on the 2007 housing and population census) and independent selection in each sampling stratum. Of these, 18,008 households and 16,583 eligible women were included. The detailed sampling procedure was presented in the full EDHS report 7 indicated in the Fig. 1 below.
Data collection procedure and variables. Stunting among children 6-59 months age was the dependent variable of the study. The independent variables were community level factors (region, clusters, residence, source of drinking water, and type of toilet facility), household factors (wealth index, household head family size, and number of under five children in the house hold), maternal factors(educational level, marital status, body mass index (BMI), maternal height, ANC visit, and birth interval), and child factors (age, sex, birth type, birth order, birth weight, diarrhea, acute respiratory tract infection (ARI), fever, vitamin A supplementation, and anemia status). www.nature.com/scientificreports/ Stunting is defined as children who have low height/length-for-age-Z score < − 2 SDs of the median value of the WHO Child Growth Standards median aged 6-59 months 1,31 .
Improved drinking water source includes piped water, public tap, standpipes, tube wells, boreholes, protected dug wells and spring, rain water, and bottled water 32 . Improved toilet facility includes any non-shared toilet of those types: flush/pour flush toilets to piped water systems, septic tanks and, and pit latrines; ventilated improved pit (VIP) latrines, pit latrines with slabs, and composting toilets 32 . Unimproved toilet facility includes flush to somewhere else, flush do not know where, pit latrine without slab/open pit, no facility/bush/field, bucket toilet, hanging toilet/latrine, others 32 .
Anemia level was determined based on hemoglobin levels adjusted for altitude in enumeration areas that are above 1000 m. They are classified as not anemic if the hemoglobin level is ≥ 11.0 mg/dl, mild anemic if hemoglobin level is from 10.0-10.9 mg/dl, moderate anemic if the level of hemoglobin is from 7.0-9.9 mg/dl, and severe anemic if the level of hemoglobin is < 7.0 mg/dl) 7 .
Body mass index (BMI) for women aged 15-49 years who are not pregnant and who have not had a birth in the last 2 months before the survey was determined by divided weight in kilograms by height in meters square (kg/m 2 ). Accordingly classified as under-weight/thin (BMI < 18.5 kg/m 2 ), normal (BMI in the range of 18.5-24.9 kg/m 2 ) and over-weight (BMI ≥ 25.0 kg/m 2 ) 7 . www.nature.com/scientificreports/ Anthropometric measurements. The length of children aged < 24 months was measured during the EDHS in a recumbent position to the nearest 0.1 cm using a locally made measuring board (Shorr Board) with an upright wooden base and moveable headpieces. Children ≥ 24 months were measured while standing upright. The length/height-for-age Z-score, an indicator of nutritional status, was compared with reference data from the WHO Multicenter Growth Reference Study Group, 2006 33 . Children whose height-for-age Z-score is < − 2 SD from the median of the WHO reference population are considered stunted (short for their age). Weight measurement was taken after children were undressed (no shoe, dresses and wet hat). For a child who stands on the weighing scale calmly, the measurement was taken in the nearest 0.1 kg. In the time of refuse to be scaled, children's mother carried and stood on the scale. Finally, the child actual weight was registered by subtracting mother's weight from mother and child weight 7 .
Before the actual data collection, interviewers were trained and a pre-test was performed. Interviews were also performed using local languages. A structured and pre-tested questionnaire was used as a tool for data collection. The 2016 EDHS interviewers used tablet computers to record responses during interviews. The tablets were equipped with Bluetooth technology to enable remote electronic transfer of files (transfer assignment sheets from team supervisors to interviewers and transfer of completed copies from interviewers to supervisors) 7 .
Data management and analysis. The data set was available at the DHS website https ://www.dhspr ogram .com/data/avail able-datas ets.cfm. After registering for the permission it was accessed through the DHS website; https ://dhspr ogram .com/data/datas et_admin /login _main.cfm. The kid recode (KR) data set in STATA file contains the outcome and predictor variables of the study. The data was explored, cleaned, coded, re-categorized and recoded. Then, the data was prepared in text file for analysis using publicly available software Win BUGS. Finally, in the Win BUGS program the distributional form of the data and the parameters were specified. Model specification in terms of the distributional relationships between observables and parameters (likelihood) and prior distributions of parameter, auxiliary files containing the data and initial values for unknowns were input to Win BUGS.
This study was based on secondary data analysis of 2016 EDHS by adjusting sample weights. Categorical characteristics and outcome of the study was described in terms of percentage and frequencies. Tables were used to present the result. A Bi-variable multi-level logistic regression analysis was carried out to see the crude Table 3. Multi-level logistic regression analysis of factors associated with childhood stunting in Ethiopia by using Bayesian statistical inference from January 18, 2016 to June 27, 2016 (n = 8117). Crl*: is Bayesian Credible Interval, σ 2 u : is residual variance between clusters , σ 2 e : residual variance between individuals in the same cluster , VPC a : is variance partition coefficient, Deviance b : is − 2log (L), DIC d : is deviance information criterion (model fit statistic).  www.nature.com/scientificreports/ stunting. WinBUGS is a windows-based computer program designed to conduct Bayesian Analyses of complex statistical models using Markov Chain Monte Carlo (MCMC) methods. The parameter estimation was done with MCMC simulation method by using Gibbs sampling technique. Non informative priors for fixed and random effect parameters were selected and three different initial values for these parameters were specified. Model specification with binary multi-level logistic regression was checked, data were loaded that were presented with rectangular text file format. Initial values compiled with the model and initialized. The deviance information criterion (DIC) statistic 35,36 was calculated for the different models (individual level, community level and both individual and community level) fitted with logit, probit and cloglog link functions. The DIC was used to evaluate and compare model performance of the full model and the reduced model. A model with lower DIC was considered as one with a better fit.
Convergence of parameters was checked visually using history plots, kernel density, auto correlation plots and Gulman-Rubin convergence diagnostic. Summary statistic (mean, SD, AOR and 95% credible intervals for parameters) from the posterior distribution of the parameters was calculated from simulated samples.
Variance partition coefficient (VPC) statistic was calculated to measure the variation between clusters (the random effect variable) 37 . It represents the percentage variance explained by higher level clusters. Hence, it was calculated as below: where σ 2 u is the between cluster variance, σ 2 e = 3.2.
Ethical issues. Ethical clearance was gained from the ethical review board of Institute of Public Health, College of Medicine and Health Sciences, University of Gondar. Written consent was obtained from Measure DHS International Program which approved the data-sets. All the data used in this study are publicly available, aggregated secondary data with not having any personal identifying information that can be linked to particular individuals, communities, or study participants. Confidentiality of data was maintained anonymously.

Confirmation of methods.
Author(s) confirm that all methods were carried out in accordance with relevant guidelines and regulations in the manuscript.

Discussion
The study demonstrated that child's age was significant predictors of childhood stunting. That is the older the child the higher the risk of being stunted. This finding is in line with studies done in Ethiopia and other developing countries [38][39][40][41][42] . This could be due to the fact that stunting is a chronic malnutrition and is commonly noticed long term nutritional deprivation. The older children are being weaned off breast feeding and also they are becoming mobile and get contaminated materials such as water, food, and soil then take into their mouth. These conditions expose children to infection which decreases food intake 43,44 .
Our study showed that male children were more stunted compared to female children. It is similar with reports from different regions 38,40,41,45,46 . The justification may be related that in early life childhood morbidity is higher among males than females. In addition, it is related with the higher number of male preterm births as compared to female preterm births that lead to increase stunting in under-five children 47 . In contrast to this finding studies demonstrated female children were more stunted compared to male children 48 . But this discrepancy could not be important to design interventions, because sex is not modifiable factor. Educational level of the mother showed a significant negative association with stunting among under-five children. This finding is similar with previous reports that maternal education has a positive outcome in reducing the child stunting 40,41,45,48,49 . One possible explanation is that knowledge that mothers get from their formal education could capable them to practice nutritional and other related behaviors that prevent chronic malnutrition/stunting. In addition to this, educated mothers have better health seeking behavior for childhood illnesses as compared to uneducated mothers 50 . A higher maternal education leads to better health care practice, acceptance of modern health practices and higher female autonomy, that affects health-related decisions again increases nutritional effects 51 . Further, education is one of an important tool for improvement of income which helps them to satisfy nutritional requirements. Furthermore, education increases skills and is highly linked with different socio-economic factors including life style, income, and fertility at individual and community level 8 .
In this study, maternal nutritional status was significant predictors of childhood stunting. Studies done in Ethiopia and Nigeria reported similar findings 39,40,44,52 . Maternal BMI is an essential predictor of childhood stunting and affected by maternal nutrition, hence proper nutrition for mothers during prenatal and postnatal period is highly demanded to enhance the growth of children. Maternal under-nutrition which leads poor growth of fetus leading to intrauterine growth retardation, this in turn strongly related with childhood stunting 39,53 .
Likewise, maternal height was found significant predictor of childhood stunting. The finding was consistent with studies done in India and Cambodia 54,56 . This could be explained by short maternal stature is associated with intrauterine growth retardation and low birth weight which are in turn predictors of impaired child growth 57 .
In addition, having birth interval < 24 months increases the likelihood of stunting. The finding is consistent with the previous studies 39,40,55 . The possible explanation could be those who had a short birth interval between births could be followed by adverse consequence on child nutrition because of it undermines intrauterine growth and caring quality of the child 58 .
The economic status of the households is related with nutritious foods at the household level which could determine the growth and the development of the children in early life 59  www.nature.com/scientificreports/ children with households in the poorest wealth quintiles were more likely to be stunted than children with richest wealth index, which is consistent with the previous studies 40,42,45 . It is clear that, increased income improves dietary diversity, which in turn improves nutrient intake and nutritional status of the children and the mother, and it will result appropriate growth and development 60,61 . From this study finding, children being twin were positively associated with stunting. This finding is in line with studies done in Cambodia 55 . However, the result is contradicting to studies done in India which showed being twin and stunting had negative association 62,63 . This might be due to socio-cultural differences in child care practices of the countries.
Children having fever were more likely to being stunted as compared to those children who had no fever. This may be due to the fact that during illness children have reduced intake of food and in turn affects their body nutrients requirement. Also, this study finding was opposite to studies done in Kenya, Nepal and Colombia that showed negative association between childhood stunting and fever [64][65][66] .
This study also showed that the likelihood of stunting was higher among children from households who had unimproved toilet facility compared to children from households who had improved toilet facility. This finding is consistent with findings from studies in Sub-Saharan Africa and Cambodia 39,55 . It is explained that children become more affected by environmental contamination when they start crawling, walking, exploring and taking objects to their mouth, that increases the risk of infection. This leads to diarrheal illness which in turn deteriorates nutritional status of children 67 . In this study region was found to be significant predictors of childhood stunting. Accordingly, children living in Dire-dawa, Amhara, Afar, Tigray, SNNPR, Oromia, and Benishangul Gumuz regions were more likely to be stunted compared to children living in Addis-Ababa. This finding is supported by studies conducted in Ethiopia 40 . In Ethiopia there are socio-cultural and socio-economic differences among regions, which influence the household food security, which in turn affect the risk of childhood stunting. In addition, the amount and distribution of rainfall and temperature influences household food availability, thus increasing the risk of a child to be stunted 68 .
Using Bayesian statistical analysis, which has high power and computes the posterior distribution which precludes the prior information with the current data and inference is based on the posterior distribution. Moreover, the use of multilevel logistic regression analysis, which was able to identify other factors beyond individual level factor that would not be identified by using standard logistic regression analysis, and a nationally representative population based study that could be made inference for all children in Ethiopia were the strength of study.
As a limitation the study didn't include important variables such as quantitative dietary consumption behavioral factors to substantiate the findings. In addition the study did not address the causal effects, since it was based on data from cross-sectional.
The findings of the study contributes knowledge to the new statistical approach (Bayesian inference) for researchers and guide public health planners and policy makers and other concerned bodies to design appropriate intervention programs including improving maternal education, promotion of girl education, improving the economic status of households, promotion of context-specific child feeding practices, improving maternal nutrition education and counseling, improving sanitation and hygiene practices to intervene the problem of stunting. In general, these findings are of supreme importance for the Ministry of Health, health bureaus, and partners to develop nutritional programs to reduce childhood stunting.

Conclusion
In this study, both individual and community level factors were significant predictors of childhood stunting. Accordingly, increased child's age, being a male child, a twin, children with anemia, having fever, have no formal education and primary education of the mother, short birth interval (< 24 months) and lived in middle or lowest household wealth status were predictors increased the likelihood of childhood stunting at the individual level. On the other hand, increased maternal BMI and maternal long stature were factors that reduced the odds of childhood stunting at the individual level. Using unimproved toilet facility and living in Dire-dawa, Amhara, Afar, Tigray, SNNPR, Oromia, and Benishangul Gumuz regions increases the likelihood of childhood stunting at community level.

Data availability
Data is available online from http://www.measu redhs .com. A letter of approval for the use of the data was obtained from the Measure DHS and the data set was downloaded from the website (https ://dhspr ogram .com/ data/avail able-datas ets.cfm). This study used EDHS 2016 child data set and extracted the outcome and explanatory variables.