Geographical variations and associated factors of defaulting from immunization among children aged 12 to 23 months in Ethiopia: using spatial and multilevel analysis of 2016 Ethiopian Demographic and Health Survey

Ethiopia is located in the Horn of Africa and shares a border with Eritrea, Djibouti, Somalia, Sudan, South Sudan, and Kenya. The country covers an area of 1.1 million km² (square kilometer) with geographical diversity, ranging from 4550 meters (m) above sea level down to the Afar depression 110m below sea level, which is comprised of over 80 ethnicities and speaking over 80 different languages [31]. Administratively, Ethiopia is divided into nine regional states, and two city administrations subdivided into 68 zones, 817 districts, and 16,253 kebeles (lowest local administrative units of the country) in the administrative structure of the country [30]. Based on the 2018 world bank report, Ethiopia had a total population of 109 million with a gross national income per capita of US$ 790 [32]. Nearly one seventh (14.9%) of the total population is under the age of 5 years, with an infant mortality rate of 41 per 1000 live births and an under-five mortality rate of 59 per 1000 births [33]. Ethiopia’s health system comprises three tiers: a primary health care unit, a general hospital, and a specialized hospital [34, 35]. Vaccine-preventable diseases are a significant cause of childhood mortality, with under-five mortality of 25,970 deaths due to lower respiratory infections and 14,662 deaths due to diarrheal diseases in 2015 [36].

The data sources were extracted from EDHS 2016, particularly from the under-five children’s file (KR) (http://www.​measuredhs.​com). The shapefiles for Ethiopia’s administrative boundaries were obtained from the open African website (https://​africaopendata.​org/​dataset/​ethiopia-shapefiles).

The EDHS used a two-stage cluster sampling design with rural-urban and regions as strata. A community cluster was defined as a randomly selected area, which contained 150 to 200 households. In EDHS 2016, the actual size of each cluster on average contains 181 households (HH). After that, 28 HH per cluster were selected from 645 clusters drawn by the Ethiopian Central Statistical Agency from its master sampling frame of census 2007 using proportional allocation technique which provided 18,008 HH, and 16,583 reproductive women age 15 to 49 were sampled using a random selection from these clusters [37] (Fig. 1).

Moreover, the EDHS 2016 data set contains only 2036 children aged between 12 and 23 months from 645 sampled clusters that obtained an average of 3 children per cluster. Finally, 1638 children who received at least one dose of the recommended vaccine during their first year of age preceding 2016 with 552 clusters were included for community characteristics. Sixteen children whose geographical locations were not available at the global positioning system (GPS) were excluded from spatial analysis. Additionally, children younger than 12 months, older than 23 months, unvaccinated children, and children who have a missing value for a dependent variable were excluded from this analysis (Fig. 1).

The outcome variable for this study was defaulting from immunization among children aged 12 to 23 months. According to WHO 2017 guideline, children who received at least one dose and missed one or more doses of ten vaccines (one dose of BCG, three doses of pentavalent DPT–HepB–Hib, three doses of PCV, three doses of OPV, two doses of ROTA virus, and one dose of MCV1) at the age of 12 months were considered as defaulter [38].

First, we recoded each variable (recommended vaccinations) as “0” and “1” for children who did not take the recommended doses and those who took, respectively, based on vaccination cards, health facility records, and the mother’s report. Then we added all “0” and “1” and labeled the total as “immunization status”.

The immunization status was recorded as “1” if the child had received all the recommended doses of all vaccinations, categorized as “complete immunization” or “0” if the child had missed one or more doses of vaccinations, and categorized as “default from immunization”.

Individual-level factors were socioeconomic and demographic characteristics (age of mother, birth order, marital status, child’s sex, number of under-five children, mother’s educational level, partner/husband’s educational level, mother’s occupation, wealth index, exposure to mass media), reproductive health history (antenatal care (ANC) utilization, place of delivery, baby postnatal checkup within 2 months, distance to a health facility, availability of immunization card).

Community-level factors like residence, region, community poverty status, community unemployment, community-women education, community ANC utilization, community distance to a health facility, and community media exposure were the independent variables.

The EDHS did not collect data that can directly describe the clusters’ characteristics except the place of residence and region. Therefore, other common community-level data were generated by aggregating the individual characteristics with our interest in a cluster. The aggregates were computed using the proportion of a given variables’ subcategory in a given cluster since the aggregate value for all generated variables was not normally distributed. It was categorized into two groups (low and high proportion) based on their median values.

The aggregated count data of children whose default immunization were joined to the geographic coordinates based on each EA unique identification code. Location data (latitude and longitude coordinates) were also taken from selected EAs (clusters) in GPS data.

Exposure to mass media: a frequency of listening to the radio and watching television were considered exposure to mass media in this study by excluding exposure to magazines and newspapers. So, women exposed to either television or radio at least once per week considered exposed, if not exposed at all, taken as not exposed [30].

Antenatal care utilization was defined as mothers who had at least four antenatal care visits [39].

The EDHS 2016 data were pre-tested before the actual data collection. Data collectors had received training in interviewing techniques, field procedures, the content of the questionnaires, and how to administer both paper and electronic questionnaires; after all, questionnaires were finalized in English, then translated into Amarigna, Tigrigna, and Oromiffa [30]. Since this was secondary data, the data were maintained by processing, editing, and raw coding data, and re-coding, checking its completeness, and cleaning the missing values by running frequencies based on this research’s interest. Sample weights were applied to compensate for the disproportional probability of sampling and non-response rate between the strata that have been geographically defined. A detailed explanation of the weighting procedure can be found in the EDHS final report [30]. Cross tabulations and summary statistics were used to describe the study population. Descriptive and summary statistics were done using STATA version 14 software.

The data for this analysis included 1638 children nested within 552 EAs. Considering this hierarchical nature of the data and the assumption of independence among individuals within the same community and the assumption of equal variance across the community is violated in the nested data. Therefore, flat models could underestimate the effect sizes’ standard errors and lead to bias (loss of power or type I error), affecting the null hypothesis [40].

Multilevel logistic regression models were fitted to identify community and individual-level factors associated with defaulting from immunization.

Four models were developed; Model I (Empty model) was fitted without explanatory variables to test random variability in the intercept and estimate the intraclass correlation coefficient (ICC). Model II examined the effects of individual-level characteristics. Model III examined the impact of community-level variables, and Model IV examined the effects of both individual and community-level characteristics simultaneously. For association (fixed effect) measures, adjusted odds ratio with 95% confidence intervals at p value < 0.05 were used to declare statistical significance. For measurements of variation (random effects), ICC, median odds ratio (MOR), and proportional change in variance (PCV) statistics were computed. Model comparison was made based on Akaike information criteria (AIC) and deviance information criteria (DIC). The model with the lowest information criterion was considered to be the best fit model [40].

The aggregated defaulter and completed count data were joined to the geographic coordinates based on each EA unique identification code. Global spatial autocorrelations were assessed with ArcGIS version 10.5 using the Global Moran’s I statistic (Moran’s I) to evaluate whether the pattern expressed is clustered, dispersed, or random across the study areas. Moran’s I values close to − 1 indicated defaulters were dispersed, whereas I values close to + 1 indicated defaulters were clustered, and distributed randomly if I value was zero. A statistically significant Moran’s I (p < 0.05) led to the rejection of the null hypothesis and indicated the presence of spatial autocorrelation, and detect the existence of at least one cluster, but not the specific location of the cluster(s) [41].

Hot spot analysis was carried out to identify spatial clusters of defaulting from immunization. Since geographic coordinates were collected at the cluster level, the unit of spatial analyses was DHS clusters. Due to positive global spatial autocorrelation, local spatial association indicators were used to assess clusters and outliers by comparing the values in each specific location with values in neighboring locations. It allows for decomposing the pattern of spatial association into four categories (quadrants) [42]. Finally, we employed Kulldorff’s purely spatial scan statistic method using the Bernoulli probability model in SaTScan version 9.6 software to detect the local spatial clusters of areas with high default of immunization. We declare where spatially significant higher rates of aggregates were founded. Its output presents the hotspot areas in circular windows, indicating the areas inside the windows are higher than expected distributions compared with the areas outside of the cluster windows [43]. We used a maximum 50% of the population at risk for the spatial cluster size. A cluster was statistically significant if a p value < 0.05.

For the interpolation, we run the empirical Kriging technique to predict values for areas where data points were not taken.

For this study, ethical clearance was taken from the ethical review committee of Debre Markos University College of Health Sciences (CHS) with approval and a supporting letter. The EDHS 2016 data were then obtained and used with the Central Statistical Agency of Ethiopia’s prior permission. We registered for dataset access and wrote the study’s title and significance on the website after completing a short registration form. Downloading of datasets was done using the accessed website at http://​www.​measuredhs.​com on request with the help of ICF international. Downloading data were used only for this study. The dataset was not passed on to other researchers without the consent of EDHS. All EDHS data were treated as confidential, no need to identify any household or individual respondent interviewed in the survey.

This study focused on a sample of 1638 unweighted data of children and 1686 weighted children who received at least one specific vaccine during their first year of age before the survey. Three fifth (60.4%) of children had missed one or more doses of recommended vaccine. Two thirds (60.4%) of mothers had no formal education; of them, 40.1% of their children have defaulted from immunization, and nearly half (45.7%) of husbands had no formal education. More than half (53%) of mothers had no jobs, and 33.7% of their children have defaulted from immunization. Nearly one fourth (22.5%) of the household were in the poorest wealth quintile, with 16.8% of children default from immunization.

Regarding media exposure, 80% of women were not exposed to media in a week, and of them, 55.8% of their children have defaulted from immunization. The majority (53%) of the women practice home delivery, with 42.7% of their children default from immunization. Immunization cards were not available for 53.5% of children, and 36% of them were defaulters. In Table 1, the remaining variables were presented.

The study included 552 clusters, in which all the children among the age group of 12 to 23 months had lived. In the rest of the 93 clusters, child data for immunization were not taken. Three fourths (76%) of the clusters were in rural with 55.6% default areas, and more than half (57.6%) of clusters were a higher proportion of community poverty status. Of them, 38.4% of children have defaulted from immunization. Nearly half (48.9%) of clusters were a higher proportion of community women unemployed, with 31.2% of children defaulting from vaccination. Three fifths (58.9%) of the clusters was a higher proportion of community of non-educated women, and 39.3% of children have defaulted from immunization. Nearly two thirds (69.5%) of the clusters had a lower proportion of institutional community delivery, with 48% of defaulted children. Additionally, Table 2 illustrated immunization status for the remaining data.

The spatial cluster of autocorrelation on defaulting immunization indicates that statically negative autocorrelation (with high cluster, low outlier) was observed in the Afar region (zone 3, zone 4), Amhara region (South Gonder), Somalia (Fafan, Korahe, Jarar), Oromiya (Jimma), SNNPR, and Gambela (Nuer). Statistically negative autocorrelation (with high outlier, low cluster) was observed in the Tigray region (northwestern, central, and eastern), Dire Dawa, Harari, Addis Ababa (zone 4), Benishangul-Gumuz, and Oromiya (West Wallega, North Shewa, East Shewa, and West Shewa), while positive autocorrelation with high cluster and high outlier was observed in Oromiya and Amhara regions as well as that with low outlier and low cluster was observed in Gambela (Agnuak) (Fig. 4).

We estimated the spatial autocorrelation from defaulting immunization for areas where data points were not taken across Ethiopia by using the kriging method. More than 82% of children who live in an area with red color have defaulted from immunization, whereas less than 22% of children who resided in an area with green color have defaulted from immunization (Fig. 5).

The SaTScan spatial analysis detected four groups of statistically significant SaTScan cluster areas with high immunization defaulters. This means that the prevalence of immunization default was higher inside the SaTScan circular window compared with that outside the SaTScan window. The most likely primary SaTScan cluster of high default areas was detected (LLR = 28.35, p < 0.001 in Eastern Ethiopia, specifically in Afder, Shabelle, Korah, Doolo, Nogob, Jarar, and Fatan administrative zones of the Somali region. The first secondary most likely spatial SaTScan cluster was in the Afar region (LLR = 17.03, p < 0.001), specifically in Zone 1, Zone 2, and Zone 4 administrative zones (Fig. 6).

The primary cluster centered at 6.559519 N, 46.154797 E with 531.31-km radius, RR of 1.58, and LLR of 28.35 at p < .001 showed that children within the area had a 58% higher risk of defaulting from immunization than children outside the area. The first secondary clusters were centered at 12.53287 N, 41.1649 E with 165.63-km radius, RR of 1.63, and LLR of 17.02, p < 0.001 showed that children within the area had a 63% higher risk of defaulting from immunization than children outside the area (Table 3).

As shown in the results of multilevel logistic regression analysis in Table 4, the empty model (Model I) revealed that there was a considerable variation in the odds of defaulting immunization between communities (ICC = 34%), which implies 34% of the total variance in the defaulting immunization was attributed to differences between communities.

In Model II, only individual-level variables were added. The results showed that women’s educational level, antenatal care utilization, and delivery place were significantly associated with defaulting of immunization (Table 4).

In Model III, only community-level variables were added to assess how much the variation in defaulting of immunization is explained by community variation. The result revealed that children residing in Afar, Somalia, SNNPR, Oromiya, Amhara, Harari, Gambela, and place of delivery were significantly associated with defaulting of immunization (Table 4).

The final model (Model IV) included both the individual and community-level characteristics simultaneously depicted that women’s educational level and region were significantly associated with defaulting immunization (Table 4).

Children from mothers who had no formal education have 4.23 times (AOR = 4.23; 95% CI: 1.17, 15.78) higher likelihood of defaulting for immunization than children whose mothers had higher education. Those children who live in the Afar region have 7.53 times (AOR = 7.53; 95% CI: 2.06, 27.45), in the Oromiya region 3 times (AOR 3.070; 95% CI: 1.52, 9.03), in the Somali region 4.4 times (AOR = 4.39; 95% CI: 1.31, 14.69) more likely to default from immunization as compared with children who were residing in the Tigray region. Additionally, children who live in SNNPR (AOR = 3.44; 95% CI: 1.37, 8.67), of the Gambella people (AOR = 3.32; 95% CI: 1.11, 8.92) and in the Harari region (AOR = 2.94; 95% CI: 1.02, 8.48) had more likely to default immunization as compared with children who were residing in the Tigray region (Table 4).

As the multilevel logistic regression analysis results described in Table 5, Model I revealed statistically significant variation in defaulting of immunization across communities. One third (34%) of the variation in the odds of defaulting immunization is attributed to community-level factors (ICC = 34%).

After adjusting the model for individual-level factors (Model II), about 37.3% of the variation in the odds of defaulting immunization was attributed to the individual level factors (PCV = 37.3%), and 24.4% of the variance in defaulting of immunization was attributed to community-level factors (ICC = 24.4%) (Table 5).

Model III, which was adjusted for community-level factors, revealed that the community-level factors explained 49% of the variability in the odds of defaulting of immunization (PCV = 49%), and 21% of the variation among the clusters was attributed to community-level factors (ICC = 21%) (Table 5).

The final best-fit model (model IV) was adjusted for both individual and community-level factors simultaneously. In this final model, about 18.5% of the variability among communities in the odds of defaulting immunization was due to the community-level factors (ICC = 18.7%) and 55.6% of the variance in the odds of defaulting immunization (PCV = 55.6%) across communities was attributed to both individual and community-level factors (Table 5).

Including both individual- and community-level factors reduced the unexplained heterogeneity in defaulting immunization between communities from MOR of 3.5 in the null model to the MOR of 2.3 in the final model, which equals 1.2. This showed that the likelihood of having a default for immunization increased by 20% when children moved from low- to high-risk neighborhoods (Table 5).

As shown in Table 5 below, a small number of AIC and DIC and a large number of LLR are in Model IV, indicating that the explanatory value of the model increases for Model IV. In other words, Model IV explained the determinants better than Models II and III; this makes the final model the best-fitted model than others (Table 5).

This study cannot determine causality because of the cross-sectional study design. The other limitation of this study was that some random effect values not explained by the predictors including beliefs, cultural tradition, knowledge, attitude, and practices (KAP) were found to be barriers to child immunization from a previous study conducted in Southern Ethiopia in 2013. Due to the irregular shape of the earth, a circular SatScan window may miss intervention areas. There is no data in 2016 EDHS, that measure the distance of the household to the health facility. It was asked whether the distance to the health facility was a big problem or not. Lastly, the sample size may not be adequate to generalize at the cluster level.

Regarding the strength of the methodology to consider the hierarchical nature of the EDHS data, a two-level mixed-effects logistic regression was used to handle both the fixed effects of individual and community factors and random effects to explain the between-cluster variations simultaneously. Furthermore, it is noted that EDHS data often collects individual data; it does not collect data that describe the clusters directly except region and place of residence. As a result, this study endeavored to generate variables that can characterize communities by aggregating individual data into cluster values. This enabled the study to test whether community-level factors could influence the defaulting of child immunization, in addition to individual-level factors. The other strength of this study is with estimation adjustments for representativeness of EDHS data, such as applying weighting of data during the description of background characteristics of the study population and considering sample designs during the analysis of cross tabulation, results obtained based on EDHS data are assumed to be representative of the Ethiopian population.