Study Design and Data Sources
This is an ecological study. We used a multilevel observational cross-sectional study design based on data from two main sources: Demographic and Health Surveys (DHS) and Afrobarometer. We used DHS for our outcome of interest, i.e., MMR, and Afrobarometer for our main independent variable, i.e., GBD. DHS are large-scale household surveys with an average sample size of 5,000 to 30,000 households, conducted across many low- and middle-income countries (LMICs) using standardized questionnaires. (29) Since DHS phase VII, the inclusion of two questions to exclude deaths due to an act of violence or an accident have improved maternal mortality estimates. (30) Therefore, this paper utilized data from DHS phase VII, conducted between 2015 and 2018 (see Appendix 1). Afrobarometer surveys measure social, political, and economic conditions in more than 30 African countries. (31) Ordinarily, these are measured with the help of face-to-face interviews in applicable local dialects with each survey round containing randomly selected samples of 1,200 or 2,400 people in each country. (31) Working with trained and skilled national partners ensures the quality of data collection. (31) From round seven (R7), Afrobarometer started to include a question to explore self-reported GBD of women. (32) Therefore, we used data from Afrobarometer R7, conducted between 2016 and 2018. We actively decided against using the DHS gender-based violence module to construct our GBD variable since we aimed to investigate and capture GBD in general, not only in its overt manifestation as violence. (29) Both data sources provide geocoded data at the region-level. We picked all SSA countries for which we had data from DHS and Afrobarometer for the mentioned time periods and matched the datasets at the corresponding geocoded regional level. That led to the following country selection: Benin, Malawi, Mali, Nigeria, Senegal, South Africa, Uganda, Zambia, and Zimbabwe. We then modeled the analysis as a random intercept two-level model to assess further information at the region- and at the country-level.
Variables and their measurement
Our outcome variable of interest is the MMR, which captures the number of maternal deaths per 100,000 live births. (33) To estimate maternal mortality, we built a variable from DHS using the sibling history approach. (34) Using a catalogue of detailed questions, this method asks women to name and count their sisters who died from maternal causes. (34) We coded this variable as continuous, differing by region of residence, and calculated it as follows (34):
$$MMR=\frac{MMRate}{FR} x \text{100,000},$$
1
$$MMRate=\frac{Number of maternal deaths}{Womenyears of exposure of sisters }x\text{1,000},$$
2
$$FR=\frac{Number of births}{Womenyears of exposure of female respondents}x\text{1,000},$$
3
where MMR is the maternal mortality rate, and FR is the fertility rate. We used the DHS guide to compute and estimate all relevant statistics with respect to DHS-7. We calculated the number of births and the number of maternal deaths for a period of zero to six years preceding the survey. (35) Women-years of exposure are the sum of women living in a given preceding time period, in our case zero to six years. (34) We generated the number of births out of the DHS Birth’s Recode and the number of maternal deaths as well as the women-years of exposure out of the DHS Individual Recode. (36)
Our main independent variable of interest is the proportion of women who reported having experienced gender-based discrimination in the past year. Afrobarometer asks: “In the past year, how often, if at all, have you personally been discriminated against or harassed based on […] your gender?” (32) with answer categories distinguishing never, once or twice, several times, or many times. We defined this variable as continuous, capturing the proportion of women who reported having experienced GBD several or many times out of the full sample of women interviewed in a given region.
At both levels, the region- and the country-level, we controlled for several covariates that have been shown to be associated with maternal mortality in previous analyses. (37–44) At the region-level we integrated the proportion of women who (1) have any school education, (2) are assigned with a high lived poverty index (LPI), which measures how frequently a person goes without basic needs such as food, water, or medicine (45), (3) have difficulties in obtaining medical treatments in general (not only relating to maternity), (4) never had to pay a bribe to obtain medical treatments, and who (5) have access to a piped water system close to their place of residence. At the country-level, we controlled for the (6) UNAIDS’ estimated HIV prevalence among the total population of ages 15–49, (7) the WHO’s calculated current health expenditure per capita in US$, and (8) the adolescents’ fertility rate measured by the United Nations Population Division, which means the average number of births per 1,000 women ages 15–19 (see Tables 1, 2). We controlled all independent variables for multicollinearity and found that correlation values are consistently below ± 0·73 (> 95% are below ± 0·7) (see Appendix 2). We did not control for median age or fertility rates, as these variables are captured by the outcome (dependent) variable, MMR.
Table 1
Variables, their measurements, and their data sources
Variable | Definition | Measurement | Data source and Date |
---|
Outcome Variable | | | |
MMR | Number of maternal deaths per 100,000 live births during the six years preceding the survey per region | Continuous Variable | DHS-VII, 2015-2018 Individual and Birth Recode (46) |
Main Independent Variable | | | |
Self-reported experiencing GBD | Proportion of women reporting experiencing gender-based discrimination in the year preceding the survey per region | Continuous Variable, Unit of Measure in % | Afrobarometer R7, 2016–2018 (32) |
Covariates region-level | | | |
School education | Proportion of women who have any kind of school education | Continuous Variable, Unit of Measure in % | DHS-VII, 2015-2018 Individual and Birth Recode (46) |
High LPI | Proportion of women who are assigned with a high lived poverty index | Continuous Variable, Unit of Measure in % | Afrobarometer R7, 2016–2018 (32) |
Difficulties in obtaining medical treatment | Proportion of women who have difficulties in obtaining medical treatments | Continuous Variable, Unit of Measure in % | Afrobarometer R7, 2016–2018 (32) |
Never pay bribes for medical treatment | Proportion of women who never had to pay a bribe to obtain medical treatment | Continuous Variable, Unit of Measure in % | Afrobarometer R7, 2016–2018 (32) |
Access to water | Proportion of women who have access to a piped water system | Continuous Variable, Unit of Measure in % | Afrobarometer R7, 2016–2018 (32) |
Covariates country-level | | | |
HIV Prevalence | HIV Prevalence among the population of ages 15–49 | Continuous Variable, Unit of Measure in % | UNAIDS, 2020 (47) |
Health Expenditure | Current health expenditure per capita | Continuous Variable, Unit of Measure in US$ | WHO, 2018 (48) |
Adolescents’ Fertility Rate | Average number of births per 1,000 women ages 15–19 | Continuous Variable, Unit of Measure in % | UNPD, 2019 (49) |
Table 2
Descriptive statistics of regional independent variables
| Study Population |
---|
N | % |
---|
Experienced gender discrimination - total | 5925 | 100 |
Yes | 427 | 7 |
No | 5489 | 93 |
Have school education - total | 160275 | 100 |
Yes | 44893 | 28 |
No | 115382 | 72 |
High Lived Poverty Index (LPI) - total | 5884 | 100 |
Yes | 1024 | 17 |
No | 4860 | 83 |
Obtaining medical treatment - total | 5927 | 100 |
Difficult | 1772 | 30 |
Easy | 2080 | 35 |
No contact | 2069 | 35 |
Having to pay a bribe to obtain medical treatment - total | 5927 | 100 |
Never | 3455 | 58 |
At least once | 397 | 7 |
No contact | 2069 | 35 |
Have a piped water system in PSU/EA - total | 5928 | 100 |
Yes | 2674 | 45 |
No | 3225 | 54 |
*Sum of counts for each variable may not add up to total due to missing data |
Analytical Approach
To generate an initial overview, we first mapped the levels of self-reported GBD and MMR in the selected 78 regions across the nine countries at a region-level. We classified four subgroups for each of the two continuous variables, from low, over medium-low, and medium-high, to high levels. For self-reported GBD the subgroups range from ‘<3%’, over ‘3–5·6%’, ‘5·7–10%’ to ‘>10%’. We formed these groups tailored to the distribution of the values, so that every group would contain approximately the same number of regions. We contrasted both variables with the help of a figure showing the levels in colors, ranging from dark green over light green, and light red, to dark red (see Fig. 1). For instance, in Mopti, Mali, we found, based on the Afrobarometer data, that 1·5% of the women reported experiencing GBD. Mopti belongs to the lowest level of < 3% reported GBD and that is why Mopti is colored in dark green in Fig. 1.
To examine the association between MMR and self-reported GBD and to assess the relevance of other region- and country-level factors in explaining MMR, we conducted a random intercept two-level model using the continuous variable measures In a linear regression. Level 1 is the region of residence, which is nested in level 2, the country. For this analysis, we used data from 160,275 women from the DHS survey and from approximately 5,900 women from the Afrobarometer survey, living in \({n}_{i}=78 \text{r}\text{e}\text{g}\text{i}\text{o}\text{n}\text{s}\) (level 1), nested in \({n}_{j}=9 \text{c}\text{o}\text{u}\text{n}\text{t}\text{r}\text{i}\text{e}\text{s}\) (level 2). Choosing a random-intercept model, we interpreted the slope, i.e., the coefficients of the independent variables, as fixed across all 78 regions. On the other hand, the intercept is random across all 78 regions due to region- and country-specific residuals. Our analysis consists of four different models, each adjusted for more covariates. First, we conducted the Null Model, which only concentrates on our outcome variable and gives out the overall mean of the MMR across all regions. Then we integrated step by step our main independent variable of interest (Model 1), all region-level covariates (Model 2), and all country-level covariates (Model 3). By adjusting the model for independent variables, the model produces coefficients that show the association between the independent variables and the MMR, and show the independent association of self-reported GBD with MMR. We conducted all analyses of our random intercept two-level model in the statistics software SPSS (PASW Statistics 18.0). The equation of the full model (Model 3) is:
\({y}_{ij}={\beta }_{0}+{\beta }_{1}{x}_{1ij}+{\beta }_{2}{x}_{2ij}+{\beta }_{3}{x}_{3ij}+{\beta }_{4}{x}_{4ij}+{\beta }_{5}{x}_{5ij}+{\beta }_{6}{x}_{6ij}\) \(+{\beta }_{7}{x}_{7j}+{\beta }_{8}{x}_{8j}+{\beta }_{9}{x}_{9j}+{u}_{j}+{e}_{ij},\) (4)
where \({y}_{ij}\) represents the MMR outcome variable for region \(i\) in country \(j\). Every \(x\) represents a predictor variable, \({\beta }_{0}\) is the overall mean of the MMR across all 78 regions, \({\beta }_{1-9}\) are the described coefficients or fixed effects of the predictor variables on the MMR, and \({e}_{ij}\) and \({u}_{j}\) are the region- and country-specific residuals. Besides the estimates of fixed parameters as well as individual residuals, each model produces the Akaike’s Information Criterion (AIC). The comparison of the AIC for each model shows whether the inclusion of the covariates improves the fit of the model or not, i.e., if the AIC is smaller than the AIC of the model conducted before, it is a fit improvement. (50)