Background: Gender is a crucial consideration of human rights that impacts many priority maternal health outcomes. However, gender is often only reported in relation to sex-disaggregated data in health coverage surveys. Few coverage surveys to date have integrated a more expansive set of gender-related questions and indicators, especially in low- to middle-income countries that have high levels of reported gender inequality. Using various gender-sensitive indicators, we investigated the role of gender power relations within households on women’s health outcomes in Simiyu region, Tanzania. Methods: We assessed 34 questions around gender dynamics reported by men and women against 18 women’s health outcomes. We created directed acyclic graphs (DAGs) to theorize the relationship between indicators, outcomes, and sociodemographic covariates. We grouped gender variables into four categories using an established gender framework: (1) women’s decision-making, (2) household labor-sharing, (3) women’s resource access, and (4) norms/beliefs. Gender indicators that were most proximate to the health outcomes in the DAG were tested using multivariate logistic regression, adjusting for sociodemographic factors. Results: The overall percent agreement of gender-related indicators within couples was 68.6%. The lowest couple concordance was a woman’s autonomy to decide to see family/friends without permission from her husband/partner (40.1%). A number of relationships between gender-related indicators and health outcomes emerged: questions from the decision-making domain were found to play a large role in women’s health outcomes, and condoms and contraceptive outcomes had the most robust relationship with gender indicators. Women who reported being able to make their own health decisions were 1.57 times (95% CI: 1.12, 2.20) more likely to use condoms. Women who reported that they decide how many children they had also reported high contraception use (OR: 1.79, 95% CI: 1.34, 2.39). Seeking care at the health facility was also associated with women’s autonomy for making major household purchases (OR: 1.35, 95% CI: 1.13, 1.62). Conclusions: The association between decision-making and other gender domains with women’s health outcomes highlights the need for heightened attention to gender dimensions of intervention coverage in maternal health. Future studies should integrate and analyze gender-sensitive questions within coverage surveys.
The Simiyu region lies in north-western Tanzania in the Lake zone, made up of 5 districts and Bariadi Town is the administrative headquarters. It has an overall population of 1,584,157 people, and 93% of the area is rural [33]. The main occupation is subsistence farming and pastoralism [34]. About half of all residents are under 14 years of age (51.3%) [33]. The Lake zone has an under-5 child mortality rate of 88 deaths per 1000 live births (compared to 67 deaths per 1000 live births in Tanzania overall), and only 50% of births were delivered in a health facility (compared to 63% in all Tanzania) [35]. In the Lake zone, 51% of births were attended by skilled health personnel, which is 13% less than the average of Tanzania. In Simiyu specifically, there is lower coverage of interventions compared to national averages: for instance, 68% of children ages 12 to 23 months received all their vaccinations, while the average in Tanzania is about 75% [35]. We used multi-stage cluster sampling, stratified by area of residence (urban, rural, mixed) to measure coverage within the Simiyu region. For the first stage, we used systematic random sampling with probability proportional to population size to sample 67 clusters. The 2012 Tanzania census served as sampling frame, and enumeration areas (EAs) from the census were used as clusters. Within each of these clusters, we listed and enumerated all the households, then sampled 30 households using systematic random sampling. In total, 2010 households were selected. Trained interviewers visited each sampled household and, after obtaining verbal consent, drew up a roster of all household members. Interviewers used this roster to identify all women aged 15–49 in all the sampled households and all men aged 15–49 in a randomly selected sub-sample of 1005 households (50% of the sample). The household response rate was 98.8% with 1915 household surveys completed. The women’s overall response rate was 94.1% with 2528 women included in our sample; the women’s questionnaire included questions on women’s sociodemographic information, fertility, antenatal care and childbirth, postnatal care, family planning, and gender modules. The men’s overall response rate was 87.2% with 1000 total men in our sample; the men’s questionnaire included questions about men’s sociodemographic information, family planning, and gender modules. All interviews were carried out February–April 2018 primarily in Swahili and in a few cases in Sukuma, a local language. We conceptualized gender power relations using Morgan et al.’s gender framework [11], which spans four major domains: (1) decision-making and autonomy, (2) labor sharing and partner involvement, (3) access to resources, and (4) norms and beliefs. Next, we developed a gender analysis matrix to identify key gender-related considerations and questions across relevant coverage survey domains (i.e. access to and utilization of services, quality of care, facility/infrastructure). Examples of these tools are accessible on the RADAR project website (www.radar-project.org) [36]. We reviewed the existing coverage survey to identify questions which could be used as proxies for gender analysis. We reviewed validated indicators (for example final reports and survey instruments used in published Demographic and Health Surveys [37] and Multiple Indicator Cluster Surveys [38]) to fill any existing gaps, prioritized questions for inclusion within the coverage survey (based on feasibility and appropriateness), and incorporated these new gender analysis indicators into the coverage survey. Questions span across the four gender analysis domains identified above. These additional variables necessitated the development of a standalone work and decision-making module within the women’s survey. We also developed a men’s questionnaire with relevant modules to complement the questions and modules within the women’s survey. Many, though not all, questions were asked in both the men’s and women’s questionnaires. The decision-making and norms/beliefs domains were most heavily represented within the survey. This was due both to the coverage survey tool’s goal of keeping the questionnaires as light as possible, and to the limited availability of validated indicators that could be used as proxies. We were unable to include additional labor sharing questions, for example, as these would require intensive time-use questions. Similarly, more comprehensive gender-based violence questions were not able to be included due to time intensiveness for training and interviewing, and the ethical requirements in the context of a household coverage survey. Further details of the questionnaire, including the specific questions used in the coverage survey, are available on the RADAR website [36]. In order to map our multi-domain gender indicators to our outcomes of interest, while controlling for potential confounding due to sociodemographic variables, we used a series of conceptual models to guide our thinking (Figs. 1 and and2).2). Drawing upon the broader conceptual models of the previously described gender analytic framework and the socioecological framework, we positioned our sociodemographic confounders as the foundation upon which many gender is a foundation. We grouped our indicators based on the four domains and drew relationships initially based on these overall groupings. Finally, we applied directed acyclic graph (DAG) notation from our measured indicators to one another, in order to formalize the regression model we intended to run [11] (Fig. 1). Conceptual framework with directed acyclic graph notation of gender indicator variables and their associations between one another and with health outcomes Conceptual framework of social power and equality in relationships and the impact of concordance or discordance in the relationship on health outcomes. Note: “Women +” refers to women’s endorsement of these gender variables as measured in our survey, while “Women -” refers to lack of endorsement. Similarly, “Men +” refers to men’s endorsement while “Men -” refers to their lack of endorsement of gender power dynamic variables. We hypothesize that men carry more social power in patriarchal societies than women, and that endorsing positive gender variables relates to greater equality within the relationship/marriage. We highlight here that it is the overlap and lack thereof of their gender variable endorsement (concordance or discordance of responses) that is of interest in this analysis of health outcomes These conceptual models examined women’s gender norms and beliefs, decision-making ability, access to resources, and paid labor and responsibility-sharing in the household. We also applied an extended framework to incorporate complimentary measures in the men’s survey. Utilizing a paired analytic approach, we limited our analyses to men’s and women’s responses who could be paired from the survey’s relational data. We hypothesize that social power and equality exist on two axes that men’s and women’s concordant or discordant responses to gender variable questions can then influence health outcomes (Fig. 2). To simplify the discussion of the discordance analyses, “men discordant” refers to when the men endorsed an item and their wife/partner did not; whereas, “women discordant” refers to when the women endorse an item and their husband/partner did not. The gender variables were assessed against 18 existing women’s health and health system access outcome indicators, spanning three levels: individual, family, and health system. Individual outcomes included whether a woman breastfed her last child. Family outcomes included those requiring consent or input from both partners, such as whether a woman used condoms or contraception during her last sexual encounter with their partner/husband. Health system access outcomes included outcomes that were at least partially dependent on health facility capacity and availability, such as whether a woman received Antenatal Care (ANC) from a skilled health care provider. We removed health outcomes from our analyses that had poor performance in prior RADAR coverage surveys, such as whether a skilled provider was present at a woman’s most recent delivery. We conducted the following statistical analyses: We undertook initial descriptive analyses to understand the distribution of gender variables within the sample population. For this assessment, we identified men and women residing in the same household. In addition, multiple men or women from the same household could be included in the main analysis, however additional people beyond the head of household and their partner were removed in the paired analysis. We also undertook descriptive analysis among paired men and women who were married to one another. We restricted our analysis to men who were heads of their households and their wives/partners. This is due to data availability; in the household roster, each household member’s relationship was recorded with respect to the head of household. We calculated the percent agreement and Pearson’s correlation coefficient for each of the responses between couples. Polygamous unions were not delineated from other unions, and only the primary husband-wife pair were included in the analysis. In order to conserve power, we generated 2 × 2 contingency tables for each gender variable and health outcome. If a contingency table had fewer than 10 cases in any given cell, we excluded the corresponding dependent/independent variable model from analysis. We then calculated logistic or linear regressions for each health outcome and the gender variables of interest, as applicable. We used a false discovery rate (FDR) method to adjust for multiple testing while maintaining power in items with fewer responses due to the survey skip pattern. Our intended significance threshold was defined as P < 0.05. We also generated an overarching DAG to conceptualize the relationship between gender variables and our health outcomes of interest. For health outcomes that showed consistent unadjusted associations with 2 or more gender variables, we generated more detailed DAGs that used individual gender variables in order to inform adjustment in subsequent regression models. For all unadjusted models that had statistical significance, we used models adjusting for covariates identified in the DAGs, including women’s and men’s education, household wealth quintile, and women’s and men’s ages. Sampling weights were accounted for in the survey design and for household and individual non-response. All analyses were weighted and used the Taylor linearization method to adjust standard errors for the effects of clustering and stratification. Data cleaning was conducted in Stata 14 [39]. All statistical analyses were performed using the R Statistical Software version 4.0.2 [40].
N/A