Geographically weighted regression analysis of anemia and its associated factors among reproductive age women in Ethiopia using 2016 demographic and health survey

Introduction Anemia in reproductive age women is defined as the hemoglobin level <11g/dl for lactating or pregnant mothers and hemoglobin level <12 g/dl for none pregnant or non-lactating women. Anemia is a global public health problem affecting both developing and developed countries. Therefore this study aims to determine geographically weighted regression analysis of anemia and its associated factors among reproductive age women in Ethiopia using the 2016 Demographic and Health Survey. Method In this study, a total of 14,570 women of reproductive age were included. Multi-level binary logistic regression models were employed using STATA version 14. Odds ratio with a 95% confidence interval and p-values less than 0.05 was used to identify significant factors. Spatial scan statistics were used to identify the presence of anemia clusters using Kulldorf's SaTScan version 9.6 software. ArcGIS 10.7 software was used to visualize the spatial distribution and geographically weighted regression of anemia among reproductive age women. Result Overall 23.8% of reproductive-age women were anemic. The SaTScan spatial analysis identified the primary clusters' spatial window in Southeastern Oromia and the entire Somali region. The GWR analysis shows that having a formal education, using pills/injectables/ implant decreases the risks of anemia. However, women who have more than one child within five years have an increased risk of anemia in Ethiopia. In addition to these, in multilevel analysis women who were married and women who have >5 family members were more likely to have anemia. Conclusion In Ethiopia, anemia among reproductive age women was relatively high and had spatial variations across the regions. Policymakers should give attention to mothers who have a low birth interval, married women, and large family size. Women’s education and family planning usage especially pills, implants, or injectable should be strengthened. The study used population-based cross-sectional survey data from 2016 Demographic Health Surveys conducted in Ethiopia. Ethiopia (30–140 N and 330–480E) is located in the horn of Africa. The country covers 1.1 million Sq. Kilometers, with huge geographic diversity: from 4550m above sea level to 110m below sea level in Afar depression. There are nine regional states(Amhara, Afar, south nation nationality and peoples, Gambela, Benshangul Gumuz, Harari, Oromia, Somalia, and Tigry) and two city administrations (Addis Ababa and Dire Dawa). These areas are divided into 68 zones, 817 districts, and 16,253 kebeles (lowest local administrative units of the country) in the administrative structure of Ethiopia [12]. The source population was all women aged 15 to 49 within five years before the survey in Ethiopia, while all reproductive-age women in the selected enumeration areas were the study population. EDHS uses a two-stage stratified cluster sampling method, using the 2007 Population and Housing Census as the sampling frame. First, 645 enumeration areas (EA) were chosen with a probability proportionate to their size, and an independent sample was drawn at each sample level. And then 28 households were systematically selected on average. Hemoglobin level was done for 14,489 women and of them, 14,171 women were usually live in the surveyed households (de juries) and included in the study. Therefore, the final analysis in “Fig 1” uses a total weighted sample of 14,570 women. The data collection took place from 18 January 2016 to 27 June 2016. The current study is based on the altitude adjusted hemoglobin levels which were already reported in 2016 EDHS data. Anemia is defined as the hemoglobin level <11 g/dl for lactating or pregnant mothers and hemoglobin level <12 g/dl for none pregnant or non-lactating women [1]. Individual-level and community-level factors were used. The variables were selected based on the literature review for factors affecting anemia, then sociodemographic, maternal, as well as community-level factors, were identified as important factors for the occurrence of anemia. Individual factors included age, women education, religion, marital status, mass media exposure, alcohol consumption, khat chewing (stimulant plant), current pregnancy, lactating mother, history of abortion, contraceptive method, number of birth in last 5 years, wealth index, family size, cooking fuel, toilet facility, and drinking water source. Community-level factors such as place of residence, region, community poverty, community mass media exposure, and community women education were used. The recoding of community aggregate factors has been taken from national report percentages. For community poverty, according to the world bank (WB), in 2019/2020 report around 24% of the population is under poverty [13]. For community mass media exposure we have used 13.8% and also for community women’s education level we used 7.7% [6]. The normal distribution of aggregated community factors was assessed by histogram and Shapiro Wilks test but, they didn’t fulfill the normality assumption then we recode them based on the median value. We accessed the data sets using the website www.measuredhs.com after the rational request of the Demographic and health survey (DHS). The geographic coordinate data (latitude and longitude coordinates) were also taken from selected enumeration areas through the web page of the international DHS program. The required data treatment and cleaning process was made using Stata version 14 statistical software. Descriptive analyses were used to explain the prevalence of anemia among WRA groups. Before performing spatial analysis, the weighted proportion (using sample weight) of anemia among WRA and candidate explanatory variables data were exported to ArcGIS. Due to the hierarchical nature of the 2016 EDHS data, where individuals are nested within the community, the assumptions such as independent of observations and equality of variance have been violated. Therefore multilevel binary logistic regression was fitted for the study of determinants of anemia among reproductive age women. Four models were used in the multi-level analysis. The first model contained only the outcome variable which was used to check the proportion of anemia among WRA variability in the community. The second models contain only individual-level variables and the third model contains only community-level variables, whereas, in the fourth model, both the individual and community-level variables were adjusted simultaneously with the outcome variables. Model comparison was done using the loglikelihood ratio test and the fourth model, which has the highest log-likelihood ratio was selected as the best fit model. Both random effect and fixed effect model parameters were included in the model. Random-effects estimates the variation of prevalence of anemia among reproductive age women between clusters. We used the cluster number variable (v001) for random effect estimates. We estimated the intraclass correlation coefficient (ICC), the median odds ratio (MOR), and Proportional Change in Variance (PCV). The intraclass correlation coefficient (ICC) reveals that, the variation of anemia among reproductive age women due to the cluster difference. ICC=VAVA+3.29*100%, where; VA = area/cluster level variance [14–16]. The MOR can be understood as the increased risk (in median) that would have if moving to another area with a higher risk [16]. MOR = exp.[√(2 × VA) × 0.6745], or MOR=e0.95VA where; VA is the area level variance [14, 16]. The PCV reveals the variation in anemia among reproductive age women which is explained by all factors. The PCV is calculated as; PCV=Vnull-VAVnull*100% where; Vnull = variance of the first model, and VA = variance of the model with more terms [14, 16]. The fixed effect assesses the relationship between the possibilities of anemia among women of reproductive age and predictors. For the final model, factors with a p-value ≤ of 0.2 in crude odds ratio (COR) were selected. Associations between outcome and explanatory variables were assessed and its strength was presented using adjusted odds ratios with 95% confidence intervals with a P-value of <0.05 cut point. For spatial analysis, Arc GIS 10.7 and SaTScan version 9.6 software were used. A statistical measurement of spatial autocorrelation (Global Moran’s I) is used for the assessment of the spatial distribution of anemia among WRA in Ethiopia [17]. Hot Spot Analysis (Getis- Ord Gi* statistic) represents the cluster characteristics with hot or cold spot values spatially. Whereas the ordinary Kriging spatial interpolation technique is used to predict the proportion of anemia among WRA for unsampled areas in the country based on sampled EAs. Bernoulli-based model spatial scan statistics were employed to determine the geographical locations of statistically significant clusters for the prevalence of anemia among WRA. To fit the Bernoulli model, cases were taken from the scanning window that moves across the study area in which women had anemia, and controls were taken from those women who had no anemia. The default maximum spatial cluster size of < 50% of the population was used as an upper limit, allowing both small and large clusters to be detected. The primary, secondary, and other significant clusters were identified and ranked based on the likelihood ratio test (LLR) test using 999 replications of Monte Carlo. The circle with the highest statistic in the LLR test is defined as the most likely (primary) clusters, that is, the group with the least random occurrence. The ordinary least square analysis was done using variables that were found to be significant at the final multilevel model. The Ordinary Least Square regression (OLS) model is a global model that predicts only one coefficient per independent variable over the entire research area. Then, the model performance, as well as the model significance such as VIF, R-square, Koenker, and Jarque-Bera statistics, expected sign for coefficients, and spatial autocorrelation of residuals were checked. The model structure of ordinary least square analysis equation [18] is written as, where i = 1,2,…n; β0, β1, β2, …βp are the model parameters, yi is the outcome variable for observation i, xik are explanatory variables and ε1, ε2, … εn are the error term/residuals with zero mean and homogenous variance σ2 Unlike OLS that fits a single linear regression equation to all of the data in the study area, GWR creates an equation for each coefficient. The model structure of geographically weighted regression equation [19] is written as, where yi is observations of response y, (uivi) are geographical points (longitude, latitude), βk(ui,vi) (k = 0, 1, … p) are p unknown functions of geographic locations (uivi), xik are explanatory variables at the location (ui, vi), i = 1,2,…n and εi are error terms/residuals with zero mean and homogenous variance σ2. The OLS and GWR models were compared using different parameters. Finally, the coefficients which were created using GWR were mapped. The permission for access to the data was obtained from ICF International by registering and stating the purposes of the study. The data set has no household addresses or individual names. The data were used for the registered research topic only and were not shared with other subjects. All the data were fully anonymized before we accessed them and/or the ICF International waived the requirement for informed consent. There were no medical records used in the research since it was a demographic and health survey.