Trends and Spatio-temporal variation of female genital mutilation among reproductive-age women in Ethiopia: a Spatio-temporal and multivariate decomposition analysis of Ethiopian demographic and health surveys

We used 2000, 2005, and 2016 Ethiopian Demographic and Health surveys (EDHSs). These EDHSs are nationally representative cross-sectional surveys performed in 9 regions and 2 country city administrations every five years (Fig. 1). In each of the surveys, stratified two-stage sampling of clusters was carried out. Stratification was achieved by separating each region into urban and rural areas. Accordingly, a total of 21 sampling strata have been created. In the first stage, a total of 539 Enumeration Areas (EAs) for EDHS 2000, 540 EAs for EDHS 2005, and 645 EAs for EDHS 2016 were randomly selected proportional to the EA size. At the second stage, on average 27 to 32 households per EA were selected. A total weighted sample of 36,685 (15,367 in EDHS 2000, 14,070 in EDHS 2005 and 7248 in EDHS 2016) reproductive-age women used for this study. The comprehensive procedure for sampling was described in the complete EDHS report [10, 11, 23].

The outcome variable for this study was experienced FGM and coded as “Yes = 1” or “No = 0”. The EDHS asked women to answer the question “have you ever been circumcised?”. So, the response variable of the ith mother Yi was measured as a dichotomous variable with possible values Yi = Yes if ith mother had experienced circumcision and Yi = No if mother did not experience circumcision. The independent variables included in this study were: residence, religion, geographic region, responded age, maternal education, women occupation, media exposure, and wealth index.

The data were extracted from the Individual Record (IR) data sets. Before any statistical analysis, the data were weighted using sampling weight, primary sampling unit, and strata, to restore the representativeness of the survey and get reliable statistical estimates.

Trend analysis of FGM and decomposition of the decrease in the prevalence of FGM over time was done. The trend analysis has been done in three phases, phase 1 (2000–2005), phase 2 (2005–2016) and the overall trend (2000–2016), the trend and determinants was examined separately.

For the trend analysis multivariate decomposition analysis for non-linear response outcome was employed to identify the factors contributed to the decrease in FGM practice across the surveys. For our study, Logit based decomposition analysis was employed.

The Logit based multivariate decomposition analysis utilizes the output from the logistic regression model to parcel out the observed decrease in FGM over time into components. The main aim of multivariate decomposition is to identify the factors contributing to the decrease in FGM practice for the last 16 years. The decrease in FGM practice can be explained by the compositional difference between surveys (i.e. differences in characteristics) and/or the difference in effects of explanatory variables (i.e. differences in the coefficients) between the surveys. Hence, the observed decrease in FGM over time is additively decomposed into a characteristics (or endowments) component and a coefficient (or effects of characteristics) component.

For logistic regression, the Logit or log-odd of FGM is taken as:

E C

The E component refers to the part of the differential owing to differences in endowments or characteristics. The C component refers to that part of the differential attributable to differences in coefficients or effects.

The recently developed multivariate decomposition for the non-linear model was used for the decomposition analysis of female genital mutilation using the mvdcmp STATA command [24].

ArcGIS version 10.6 and SaTScan version 9.6 software were used for spatial analysis. The spatial autocorrelation (Global Moran’s I) statistic was used to assess whether there was significant clustering of FGM [25]. Moran’s I has a value ranging from-1 to 1. Positive Moran’s I value shows that FGM is clustered while negative Moran’s I indicates that FGM is dispersed [26]. The value of Moran’s I near zero has revealed that FGM is randomly distributed. Both Z-score and P-value are generated to assess the significance of the Moran index.

In spatial scan statistical analysis, Bernoulli based model was employed to identify significant spatial high FGM clusters using Kuldorff’s SaTScan version 9.6 software. The SaTScan uses a circular scanning window that moves across the study area. Women who were circumcised were taken as cases whereas those who were not circumcised were taken as controls to fit the Bernoulli model. 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. For each potential cluster, a likelihood ratio test statistic and the p-value were used to determine significant clusters. The scanning window with maximum likelihood was the most likely performing cluster. The primary and secondary clusters were identified and ranked based on their likelihood ratio test, based on 999 Monte Carlo replications [27].

The Ordinary Kriging spatial interpolation method was used to predict the un-sampled/unmeasured values from the sampled measurements.

Since the study was a secondary data analysis of publically available survey data from MEASURE DHS program, ethical approval and participant consent were not necessary for this particular study. We requested DHS Program and permission was granted to download and use the data for this study from http://​www.​dhsprogram.​com. There are no names of individuals or household addresses in the data files. The geographic identifiers only go down to the regional level (where regions are typically very large geographical areas encompassing several states/provinces. In surveys that collect GIS coordinates in the field, the coordinates are only for the enumeration area (EA) as a whole, and not for individual households, and the measured coordinates are randomly displaced within a large geographic area so that specific enumeration areas cannot be identified.

About one-third of the respondents in all three surveys were found in the Oromia region. There was a slight increment in urban residents from 18.2% in 2000 to 23.0% in 2016. In 2000 EDHS, 75.2% of women were not educated, while this figure decreased to 47.0% in 2016. The proportion of women who had media exposure has been increased from 34% in 2000 to 43.3% in 2016. Except for religion, maternal age, region, and wealth index, all other variables showed changes in the composition of women across the surveys (Table 1).

The overall FGM has been decreased from 79.9% (95% CI: 79.3, 80.5) in 2000 to 74.3% (95% CI: 73.5, 74.9) in 2005 to 70.4% (95% CI: 69.3, 71.4) in 2016 with annual reduction rate of 0.8%. The trends in female genital mutilation have shown variation by women’s characteristics over time. Major decreases in FGM were observed in some of the selected variables.

Tigray, Harari, Addis Ababa, and Dire-Dawa regions showed the greatest decline in FGM practice over the study period but the prevalence remains high in the Somali region (Fig. 2). Among urban residents, the largest decline in FGM practice has been observed during the third phase of the study period (2000–2016), at 24.3 percentage point decrease as compared with 11.3 percentage point decrease during the first phase (2000–2005) and a 13.0 percentage point decrease during the second phase (2005–2016). Young age women (15–19 years) showed a larger decrease in FGM by 19.1 percentage point change during the overall study period (2000–2016). Women with secondary and higher education showed a 34.8 percentage point decrease during the third phase of the study period (2000–2016) (Table 2).

The overall decomposition analysis revealed that about 95% of the overall decrease in FGM among reproductive-age women was due to difference in coefficient (effects of characteristics) across the surveys but the change due to the difference in composition (endowment) was not significant (Table 3). After controlling for the roles of compositional changes, a significant decrease in FGM practice was due to the difference in the effects of residence, region, educational status, religion, occupation, and media exposure. About 20% of the decrease in FGM practice over the last 16 years was attributable to the difference in the effect among the urban population (B = 0.02, 95% CI: 0.01, 0.03). As compared to the Somali region, the change in the effect of FGM practice among the Tigray regions (B = 0.01, 95% CI: 0.0008, 0.02) contributes about 12.2% of the overall decrease in FGM over the study period. Other factors such as a decrease in the effects of women higher education (B = -0.0008, 95% CI: − 0.001, − 0.00006), Muslim religion (B = − 0.05, 95% CI: − 0.08, − 0.01), Orthodox religion (B = -0.06, 95% CI: − 0.11, − 0.02), media exposure (B = -0.012, 95% CI: − 0.022, − 0.003) and occupation (B = -0.02, 95% CI: − 0.03, − 0.009) were significant predictors of the decrease in FGM practice over the last 16 years (Table 4).

The spatial distribution of FGM showed significant spatial variations across the country (Fig. 3). The highest prevalence of FGM was identified in the Somali, Harari, east Benishangul Gumuz, Amhara, Oromia, east SNNPRs, and southern Afar regions consistently over time. Whereas, Gambella, Tigray, northeast Amhara, west SNNPRs and northeast Benishangul Gumuz regions showed the lowest FGM. In EDHS 2000, the spatial scan statistics identified 427 significant primary and secondary clusters. Of these, 141 clusters were the most likely clusters (primary clusters), which was located in north Somali, Harari, Dire Dawa, south Afar, and southeast Amhara region centered at 9.867651 N, 43.086403 E with 348.25 km radius, a Relative Risk (RR) of 1.35, and Log-Likelihood Ratio (LRR) of 711.95, at p-value< 0.001. It showed that women within the spatial window had 1.35 times higher odds of having FGM as compared to women outside the spatial window, whereas the secondary clusters were located in Amhara, Tigray, Benishangul Gumuz, SNNPRs, and Oromia regions (Fig. 4).

In EDHS 2005, a spatial scan statistics identified a total of 468 significant primary and secondary clusters. Of these, 175 clusters were the most likely clusters (primary clusters), which was located in the entire Somali region and border areas of Somali regions, centered at 9.774395 N, 43.208576 E with a radius of 484.13 km, a Relative Risk (RR) of 1.42, and Log-likelihood (LLR = 576.40, p-value< 0.0001). It showed that women within the spatial window had 1.42 times higher odds of FGM as compared to women outside the spatial window and the secondary clusters were located in Benishangul, Amhara, Oromia, and Gambella regions (Fig. 5). In EDHS 2016, a spatial scan statistics identified a total of 581 significant primary and secondary clusters. Of these, 220 clusters were the most likely clusters (primary clusters) and the spatial window was located in the entire Somali, Afar and border areas of Somali, centered at 7.717178 N, 46.991580 E with a radius of 900.49 km, a Relative Risk (RR) of 1.42, and Log-likelihood (LLR = 248.65, p-value< 0.0001). It showed that women within the spatial window had 1.42 times higher odds of having circumcised as compared to women outside the spatial window (Fig. 6).

Based on EDHS 2000 sampled data, Kriging interpolation predict that the highest prevalence of FGM was detected in the entire Somali, Oromia, Harari, Dire Dawa, Addis Ababa, Benishangul Gumuz, southeast Afar, southeast Amhara, and east SNNPRs. In contrast, relatively low prevalence of FGM was predicted in the entire Tigray, Gambella, southwest Oromia, north Benishangul, northwest Amhara, and north Afar (Fig. 7). Based on EDHS 2005, the highest prevalence of FGM was detected in east SNNPRs, Oromia, west Benishangul, southeast Amhara, and most of the Afar regions. In contrast, predicted low prevalence of FGM was detected in Tigray, northwest Gambella, north Afar, and west SNNPRs (Fig. 8). Based on EDHS 2016 data, Kriging interpolation predicted that most parts of Oromia, entire Somali, Dire Dawa, Harari, Afar, and west Benishangul contained the highest FGM prevalence while most parts of Amhara, Tigray, Gambella and west SNNPRs contained relatively low FGM practice (Fig. 9).

This study had several strengths. First, the study was based on nationally representative large datasets, and thus it had adequate statistical power. Second, the estimates of the study were done after the data were weighted for probability sampling and non-response. Therefore, it can be generalized to all women in Ethiopia. Third, multivariate decomposition analysis was applied to identify the factors contributing to the decrease in FGM practice over time. Fourth, the use of GIS and SaTScan statistical tests helped to detect statistically significant hotspot areas of FGM practice across the surveys and to design effective public health programs.

As a limitation, some variables were not consistently collected in all EDHS surveys like the wealth index. Therefore, these variables were not used for the decomposition analysis. The other limitation was the SaTScan detect only circular clusters not irregularly shaped clusters. Furthermore, the EDHS survey did not incorporate community-level variables like community norm, culture, and beliefs rather it relied on mothers or caregivers report and might have the possibility of social desirability and recall bias since FGM is not socially acceptable though CSA claim that strong effort was made to minimize it mainly through extensive training of data collectors, recruiting experienced data collectors and supervisors this might underestimate our finding. As with other cross-sectional studies, the temporal relationship between exposure and outcome can’t be established. Besides, since we used repeated cross-sectional data (time series cross-sectional study) for the trend and Spatio-temporal analysis, the enumeration areas selected in each survey were different, and the data for each survey were collected in different population, this could result in under/overestimation of the prevalence of FGM practice hence there may be the migration of the population to different enumeration areas during the study period, this may dilute the difference in FGM practice across the surveys since the enumeration areas selected across the survey was not the same in EDHSs even if we have done Kriging Interpolation to predict the unobserved areas of the country over the surveys.