Background Alcohol consumption during pregnancy is a known contributor to teratogen and causes a range of effects on pregnancy and birth outcomes. This study aimed to investigate the spatial variation and determinants of alcohol consumption among pregnant women in Ethiopia. Methods A secondary data analysis was conducted using the 2016 Ethiopian Demographic and Health Survey data. A total of 1,135 pregnant women were included in the analysis. ArcGIS version 10.7 software was used to explore the spatial distribution of alcohol consumption, and SaTScan version 9.6 was employed to identify the significant spatial clusters of alcohol consumption. A mixed multi-level logistic regression analysis was employed to identify the determinant factors of alcohol consumption during pregnancy. Results The result showed that the prevalence of alcohol consumption during pregnancy was 22.49% (with a 95% CI: 18.18 to 26.17). The spatial analysis showed that the spatial distribution of alcohol consumption significantly varied across the country [Global Moran’s I value = 0.30 (P<0.001)]. The SaTScan analysis identified two most likely clusters with high rates of alcohol consumption such as northwest Ethiopia (Log-Likelihood Ratio (LLR) = 155.56, p<0.001) and central Ethiopia (LLR = 19.27, p<0.01). Never in union, divorced and/ widowed [Adjusted odds ratio (AOR) = 2.56; 95% CI: 1.07, 10.14], attended primary school [AOR = 0.45; 95% CI: 0.27, 0.95], having two or more lifetime sexual partners [AOR = 2.59; 95% CI: 1.11, 6.18], living in rural [AOR = 1.52; 95% CI: 1.12, 2.93] and higher community media exposure [AOR = 0.54; 95% CI: 0.28, 0.97] were the factors associated with alcohol consumption. Conclusion Alcohol consumption during pregnancy in Ethiopia was high. The spatial distribution of alcohol consumption was significantly varied across the country. Therefore, public health interventions targeting areas with high alcohol consumption are needed for drinking cessation and to prevent poor pregnancy outcomes related to alcohol use.
The study used the Ethiopian Demographic and Health Survey (EDHS) data of Ethiopia. Ethiopia is classified into nine regional states, two administrative cities, 611 Woredas, and 15,000 Kebeles. Administratively each region is divided into zones and zones into Woredas which is the third administrative division of the country. Finally, at the fourth level, Woredas are further subdivided into Kebeles which are the lowest administrative unit [24]. A population-based cross-sectional study was employed in Ethiopia in which the data was extracted from the EDHS 2016 dataset; this was collected from January 18 to June 27, 2016. All pregnant women aged 15–49 years in the selected enumeration areas of the survey were considered the study participants. The Ethiopian Demographic and Health Statistics 2016 survey used a two-stage stratified cluster sampling technique. The sampling frame was selected from the 2007 Population and Housing Census [24]. The regions were stratified into urban and rural, producing 21 strata. In each stratum, sample Enumeration Areas (EA) were selected independently in two stages by using proportional allocation and implicit stratification. In the first stage, a total of 645 EA (202 in urban areas and 443 in rural areas) were selected out of 84,915 EA. In the second stage of selection, a mean number of 28 households per cluster were systematically selected supported by the newly created household listing. EAs with “0” longitude and latitude data were dropped. Among 645 EAs two of them were not included initially in the DHS coordinate file. Of 643 EAs 487 were included in our analysis, the rest EAs were excluded due to dropped the zero GPS cells. Further detailed information about the sampling procedures and household selection has existed in the 2016 EDHS report [24]. For this study, the 2016 EDHS of the women dataset were used. A weighted sample of 1,135 pregnant women was used for the final analysis. The outcome variable for this study was alcohol consumption responses from the two survey questions. The first question was “have you ever taken a drink that contains alcohol?” and the second: “during the last 30 days, how many days did you have a drink that contains alcohol?” Current alcohol consumption was defined as those pregnant women who drank daily or had drunk in the past 30 days that contain alcohol based on these questions. The individual and community-level variables were considered independent variables in the study. Individual-level parameters comprised age, marital status, maternal education level, household wealth index, individual’s media exposure, current employment, tobacco smoking, Khat chewing, pregnancy term, number of sex partners/ husbands, wanted pregnancy, parity, and religion. The household wealth index was a categorized variable by the DHS as poorest, poor, middle, rich, and richest, and we used it as is for analysis; we have used it as it is. Some of these factors were re-categorized for the simplicity of analysis. Only half of the pregnant women in the data were screened for intimate partner violence and partner/husband alcohol consumption, hence these parameters were not included in the analysis. The variables considered as community-level factors were the place of residence and community-level media exposure. In the EDHS, participants’ media exposure was ascertained by 3 survey questions: “how often do you have read newspaper or magazine; how often do you listen to the radio, and how often do you have watching television? The responses were “not at all”, “at least once a week” and “more than once a week” for each question. Based on these questions, the individual level of media exposure was obtained by aggregating the specified ways of getting information such as reading news or magazine, listening to the radio, and watching television which gives a sum-total score ranging from zero to six. Then, the total score of media exposure was categorized as “yes” if the total score was greater than zero and “no” if the sum score was zero. Therefore, in this study, individuals’ media exposure was defined as those individuals who have a chance to get information through at least one of a specified mass media such as reading news or magazines, listening to the radio, and/or watching television at least once per a week. The community-level media exposure was obtained by aggregating the individual-level media exposure into clusters by using the proportion of those who had media exposure. This community-level media exposure shows the general media exposure within the community. Since the aggregated variable had a skewed distribution, and therefore median values were used to categorize as higher and lower. Data extraction, recoding, and descriptive statistics such as frequencies and percentages of variables were done using STATA-version 14. Sampling weights were performed before the analysis to restore the representativeness and to adjust the non-proportional allocation of the sample to strata and regions during the survey process. After the data adjustment and description, three statistical analysis models were preformed such as spatial autocorrelation and interpolation, spatial Scan distribution, and multilevel logistic regression analysis. Spatial distribution and a mixed multi-level logistic regression model were employed to identify the spatial variation and determinant factors of alcohol consumption during pregnancy, respectively. Spatial data analysis was performed using ArcGIS version 10.7 and Spatial Scan Statistics (SaTScanTM version 9.6) software. ArcGIS 10.7 was used for doing Moran’s I Analysis. Global Moran’s I statistics was used to determine the presence of spatial autocorrelation and whether alcohol consumption was dispersed, clustered, or randomly distributed across the country. Moran’s I value close to -1 indicated dispersed, Moran’s I value close to + 1 indicated clustered, or Moran’s I value was zero indicated randomly distributed. Moran’s I p-value < 0.05 indicated the presence of spatial autocorrelation. Hot spot analysis was done using Getis-Ord Gi* statistics to measure how spatial autocorrelation varies over the study location by calculating GI Bin for each area. High GIBin* in the statistical output indicated "hotspot" whereas low GI* indicated "cold spot. The ordinary Kriging spatial interpolation analysis was used to predict alcohol consumption for un-sampled areas based on sampled EAs. Spatial SaTscan analysis. It was conducted using Kuldorff’s SaTscan version 9.6 software. This helps to identify the geographical locations of statistically significant spatial clusters of alcohol consumption among pregnant women. Pregnant women who were not drinking alcohol were considered controls, and those who were drinking alcohol were taken as cases represented by a 0/1 variable and fitted in the Bernoulli model. The number of cases in each location had a Bernoulli distribution and the model required data with or without alcohol consumption. The default maximum spatial cluster size of < 50% of the population was used as an upper limit, which allowed both small and large clusters to be detected and ignored clusters that contained more than the maximum limit. The result was reported using both table and figure. Areas with a high log likely hood ratio (LLR) and p-value < 0.05 were considered to high risk of alcohol consumption as compared to areas outside the window. Finally significant and most likely clusters with LLR, RR, and P-values were reported. First bi-variable multilevel logistic regression analysis was performed using STATA-14 and those variables with a p-value <0.20 were selected for multivariable analysis. After selecting variables for multivariable analysis, four models; the null model (without explanatory variables), model II (containing only individual-level factors), model III (examined the effect of community-level factors), and model IV (which incorporates both individual and community level factors) were fitted. In the multivariable analysis, variables with a p-value of <0.05 were considered statistically significant and the factors associated with alcohol consumption were reported by an Adjusted Odds Ratio (AOR) at a 95% confidence interval. Model comparison and fitness was assessed using the log-likelihood and deviance and the model with a lower result of log-likelihood and deviance (Model IV) was considered the best–fitted model. The final model (model IV) was the best-fitted model and was selected for reporting of the results of the study. In addition, the measures of community variation (random effects) such as the Intra-Class Correlation (ICC), median odds ratio (MOR), and proportional change in variance (PCV) [25–27] were computed. These parameters were calculated to quantify; the degree of homogeneity of substance use within clusters, the degree of variation of substance use across clusters in terms of the odds ratio scale, and the proportion of variance explained by consecutive models, respectively. Ethics approval was not required since this study is a secondary analysis based on the 2016 EDHS data. Before conducting our study, we registered and requested the dataset from DHS online archive and received approval to access and download the data files from the DHS website: https://dhsprogram.com/data/dataset_admin/index.cfm All DHS data should be treated as confidential, and no effort should be made to identify any household or individual respondent interviewed in the survey. The data could be used only for statistical reporting and analysis, and only for our registered research. According to the EDHS 2016 report, all respondents’ data were anonymized during the collection of the survey data [24].