Spatiotemporal Patterns and Determinants of Undernutrition among Late 1 Adolescent Girls in Ethiopia using Ethiopian Demographic and Health Surveys 2 2000 to 2016: Spatiotemporal and Multilevel approach.

Introduction : Under-nutrition of late adolescent girls in Ethiopia is the highest among Southern 2 and Eastern African countries. However, the spatial and temporal variations of under-nutrition as 3 a national context is not well understood. This study aimed the spatiotemporal patterns and 4 determinants of under-nutrition among Late Adolescent Girls in Ethiopia. 5 Methods : An in-depth secondary data analysis was conducted from women’s data of four 6 consecutive Ethiopian Demographic and Health Surveys (EDHS) 2000 to 2016. A total of 12,056 7 late adolescent girls were included in this study. The global spatial autocorrelation was assessed 8 using the Global Moran’s I statistic to evaluate the presence of geographical clustering and 9 variability of undernutrition. The significant cluster scan statistics using Bernoulli model to detect 10 local clusters of significant high rate and low rates of under-nutrition was explored. Multilevel 11 binary logistic regression model with cluster level random effects was fitted to determine factors 12 associated with under-nutrition among Late Adolescent girls in Ethiopia. 13 Results: undernutrition was clustered nationally during each survey (Global Moran’s I=0.009- 14 0.045, Z-score= 5.55-27.24, p value < 0.001). In the final model, individual and community level 15 factors accounted about 31.67% of the variations for under-nutrition. The odds of being under- 16 nourished girls in the age groups of 18 -19 years were 57 % (AOR = 0.43; 95 % CI: 0.35 - 0.53) 17 less likely than those from 15-17 years old. Being in higher educational status was 4.50 times 18 (AOR= 4.50; 95% CI: 2.33–8.69) more likely to be under-nourished compared with no educational 19 status. Undernutrition with occupation of sales was 40% (AOR=0.60; 95% CI: 0.43 – 0.84) lower 20 than those with not working. The odds of being undernourished adolescents were 1.77 times 21 (AOR=1.77; 95% CI: 1.24 - 2.53) higher than participants with unimproved latrine type. Rural residents were 2. 35 times (AOR=2.35; 95% CI: 1.41 - 3.92) more likely to be under nourish 23 compared with urban residents. Conclusion: undernutrition among late adolescent girls was spatially clustered in Ethiopia. The 25 significant high rate of undernutrition was observed in Northern and Eastern Ethiopia. Those 26 regions with high rates of under-nutrition should design interventions to combat under-nutrition. This also revealed that the highest prediction of under-nutrition prevalence was observed in different regions in different time periods. In 2000 survey, the highest prediction of undernutrition prevalence among late adolescent girls from unsampled areas were observed in regions of Tigray, Eastern Amhara reigion, Southeast and Southwest Oromia, Somali region, Western NNPR and parts of Beninshangul Gumuz regions. Likewise, EDHS 2005 revealed that the highest prediction of under-nutrition from unsampled enumeration areas were seen in Tigray, Northern Northern and Eastern Amhara reigion. Similarly, EDHS 2011 showed that the highest prediction was restricted in three regions ( Tigray, Afar and Northern & Eastern Amhara). During EDHS 2016, the highest prediction of under-nutrition among late adolescent girls from unsampled enumeration areas were observed in Tigray, Eastern Amhara reigion, Southern region and Gambell region.


Introduction
Page | 6 Data source, sample size and Sampling techniques of each EDHS data 1 We accessed the datasets using the website www.measuredhs.com after requested from the DHS techniques for each EDHS data was used and details of methodology was presented from each 8 EDHS report. In the first stage; a stratified sample of census enumeration areas (EAs) in the urban 9 and rural areas were selected with complete household listing using systematic probability 10 sampling based on sampling frame with population and household information from the 1994 and 11 2007 PHC. In the second-stage: the selection of households was carried out by equal probability 12 systematic sampling in the selected EAs. In each selected household, late adolescents were 13 interviewed with an individual questionnaire (5,7,21,22). Community level factors (level Two): 1 Region, residence, un-protected water source in the community, community poverty status, 2 community adolescent's educational achievement, unimproved latrine type in the community. Under-nutrition: is defined as nutritional status for which Body Mass Index is less than 18.5 5 kg/m 2 which is either stunting or underweight (thinness) (23) 6 Stunting: Is chronic malnutrition status that is measured by height-for-age z-score of less than -7 2SD of the World Health Organization or height <145 cm (2). 8 Total underweight (thinness): is defined as both acute and chronic malnutrition status that is 9 measured by BMI-for-age z-score less than -2SDs from median or BMI <18.5 10 kg/m 2 (2).

Wealth index:
A composite measure of a household's cumulative living standard divided into 14 5 quintiles using the wealth quintiles data, are derived using principal 15 component analysis (7). Kriging interpolation method was used to predict the spatial distribution of undernutrition from 14 unsampled areas in Ethiopia.

15
We explored spatio-temporal scan statistics using Bernoulli probability model to detect local 16 clusters of significant high rate and low rates of undernutrition using SaTScan 9.6 software. A 17 cluster is reported to be statistically significant when its log likelihood ratio (LLR) is greater than  We considered to use multilevel models, because each interviewed unit (household and 25 individual) are hierarchical and nested to EAs (7). Therefore, a two-level model was adopted, by In this model, logit (Yij) = ln (Yij / (1-Yij)) is log-odds for undernutrition called 'the logit link'.

6
The symbol, 'Yij' is a probability of under-nutrition for an adolescent girl i in any EA, rural/urban 7 region, 'j'.' β0j' is the cluster random intercept. 'j' is the residual for each cluster 'j'. '' is fixed (community level) in cluster j. 10 We considered four models to be fitted for multilevel analysis:   The fixed effects model has only one source of variability (εj, with its variance σ 2 µ), while the 22 random effects model has two components of variabilities (εj and ε0 with variances σ 2 µ and σ 2 ε 23 respectively). These two sources of variability showed the variability between predictors that are 24 in the same group, measured by the within-group variance σ 2 µ, and the variability between 25 observations that are in different groups, measured by the between-group variance σ 2 ε. The Page | 10 The ICC quantifies the variation of the under-nutrition within clusters. The ICC may range from  (Table 1).            The current study identified that the regional variation of under nutrition among late adolescent  under-nutrition (12). This was supported by a study in Northwestern Ethiopia (29), the age 9 groups of 15-17 years were 2 times more higher for being under-nourished than 18-19 years old 10 adolescents. On the other hand, adolescent girls after 18 years may be engaged into marriage and 11 may be better access to eating patterns in economically limited families. In contrast with above, 12 study from India (30) showed late adolescents were less likely for undernutrition compared with 13 early adolescents.
14 The current study identified that late adolescent girls with higher educational status were 4.50 15 times more likely to be under-nourished compared with no educational status. This may be 16 attributed to girls are more vulnerable for the influences of cultural and gender norms, which often 17 discriminate against frequent feeding and when dietary intakes are suboptimal, anemia and 18 micronutrient deficiencies are high among adolescent girls (12). On the other hand, during higher 19 education, those adolescents may go to place far from their parents, so that they are limited to get 20 timely feeding and food varieties due to economic barriers. A study stated that eating patterns and 21 behaviors are influenced by peer pressure, food availability, food preferences and cost, personal and cultural beliefs (1). A study from low and middle-income countries also remarked that about 23 40% of adolescent girls reported skipping their breakfast (31).

24
This study identified that late adolescent girls with occupation of sales were 40% lower for being 25 under-nourished than those with not working. This might be because of they have their own money 26 who easily to purchase varieties of food items.

27
The current study stated that the odds of under-nutrition was 1.77 times higher than participants 28 with un-improved latrine type. This may be attributed to poor sanitation could expose infestation of intestinal parasites that leads to illness, poor appetite and micronutrient deficiencies that leads 1 to undernutrition. This was consistent with SRMA study in Ethiopia (6).

2
Participants with frequency of listening to radio at least once a week were 28% less likely to be 3 under-nourished than those without listening radio. This could be because of better awareness and 4 information gain regarding importance of variety of food items and frequency of feeding patterns 5 among those with listening radio.

6
In the current study, community level factors were significantly associated with under-nutrition. 7 The odds of late adolescent girls' under-nutrition in the regions of Amhara, Oromia, Benishangul Conclusion: 5 The current study found that under-nutrition among late adolescent girls was clustered across 6 regions in Ethiopia in each survey. The spatio-temporal patterns of this study showed that there 7 was high spatial dependency across regions. The spatial scan statistics revealed that the significant  Competing interests: The authors declare that they have no competing interests.

18
Availability of data and materials: 19 The availability of data for this particular study was from the DHS program datasets using the