Household factors and under-five mortality in Bankass, Mali: results from a cross-sectional survey

We analysed cross-sectionally collected baseline household survey data from a three-year cluster randomized controlled trial being conducted in 137 villages distributed across seven of the 22 health catchment areas in Bankass (Kanibonzon, Ende, Dimbal, Doundé, Soubala, Koulongon, and Lessagou). The study area has a population of approximately 100,000 people. The trial primarily aims to determine whether door-to-door proactive case detection by community health workers reduces U5 mortality compared to passive, site-based care offered under the standard Integrated Community Case Management protocol [25]. Further details on the trial protocol are available elsewhere [19].

As a part of the trial, a survey is administered to all households in the study area at baseline and every 12 months thereafter during the study period. All households were censused between December 2016 and January 2017, just before the launch of the intervention, to enumerate all permanent residents (present more than 50% of the time in the past year). The census included a household roster to collect the age, date of birth, and sex of permanent residents, as well as information about deaths in the household in the past five years. All women in the household aged 15–49 years (i.e., women of reproductive age) who provided written informed consent were then administered a baseline survey (see DEMOGRAPHICS SURVEY – Baseline in Supplementary Material). The survey was adapted from the Demographic and Health Surveys (DHS) [26], and included detailed information on household factors. Based on standard DHS modules, women were given a reading test to assess literacy and asked if they contribute to household decision-making, if their husband has more than one wife or partner, and if they felt their husband hitting or beating them was justified under certain circumstances. Geographical co-ordinates were collected using global positioning technology. We supplemented the baseline survey data with information from the year-one follow-up survey (see DEMOGRAPHICS SURVEY – Year One in Supplementary Material), administered from February to March 2018, which used the same structure as the baseline survey but added details on women’s birth histories (i.e., probes to distinguish between live and still births, clarification of multiple births, and greater precision on birth dates); this enabled us to correct any misreporting in the birth histories recorded at baseline. The survey instrument was designed in Open Data Kit, which permits real-time quality and completeness control on data collection. The final dataset was de-identified, cleaned and compiled by the data management team at Muso, a non-government organization based in Bamako, and the Malaria Research and Training Center, University of Sciences, Techniques and Technologies of Bamako.

Household wealth was defined in quintiles using a principal components analysis of household possession of durable goods [27]. Women were considered to contribute to household decisions if they indicated participation in decision-making. Literacy was categorized based on the highest level of reading ability among women surveyed within the household. Schooling was categorized as having had any formal schooling. Polygamy was defined as a survey respondent in the household indicating that her husband/partner had more than one wife/partner. Women were considered to have tolerant views towards spousal violence if they indicated that their husband hitting or beating them was justified under any of the circumstances evaluated in the survey. We used the World Health Organization definitions of improved/unimproved water supply and sanitation [28]. Households were also asked if they treated their water to make it safer for drinking. Treatment included boiling, adding sterilizing chemicals, filtering, and solar disinfection. Roofing, wall, and flooring materials were defined as finished, rudimentary, or natural as per DHS definitions [26]. Distance to the nearest primary health center was defined as the Euclidean distance from the household village to the closest primary health center. All distance calculations were conducted in QGIS Version 3.4.6-Madeira (QGIS Development Team (2019), QGIS Geographic Information System, Open Source Geospatial Foundation Project, http://​qgis.​osgeo.​org).

U5 mortality rates for the five years immediately preceding the baseline survey were calculated using full birth history data. Women reported all live births using a birth history module modelled on the DHS [26]. We used a synthetic cohort life table approach to estimate the number of deaths per 1000 live births [29]. Briefly, this method combines mortality probabilities for small age segments into the standard age segments (neonatal, infant, child, and U5). We adopted the age segments 0, 1–2, 3–5, 6–11, 12–23, 24–35, 36–47, 48–59 months. Where children were missing information about their date of birth, we used available data to narrow the range of possible values. First, we determined an unconstrained range from available information about the month and year of birth. Next, we narrowed the possible range of dates based on other available information from the child’s mother, such as age at marriage and sibling dates of birth. Then, if a range of dates still existed, we randomly assigned a date within that range. For children with missing age or date of death, we imputed age at death based on a hot deck technique, assigning data from the preceding child of the same birth order and form of reporting (day, month, and/or year). No children were missing information on the date of interview or vital status.

The primary outcome was U5 mortality. Household-level factors analyzed in both boys and girls were ethnicity, wealth quintile, decision making contribution of women in household, highest level of reading ability among women in household, schooling among women in household, polygamy, domestic violence tolerance, water source, sanitation method, roofing material, wall material, flooring material, electricity access, primary cooking fuel, recent food shortage, livestock ownership, access to motorized transport, and distance of household from nearest healthcare center. Cox regression was used to evaluate factors associated with U5 mortality. Children were considered at risk of death from their date of birth. Follow-up was censored on the day children turned five-years-old or at the time their household completed the baseline survey, whichever came first. Children who died on their day of birth were considered to have one day of follow up so that they were not excluded from the analyses.

All regression analyses were adjusted for household clustering. Covariates were subject to univariate analysis and all were included in our multivariate analyses. Significance for hazard ratios (HRs) was defined as the lower bound of the 95% confidence interval (95%CI) being greater than 1 or the upper bound of the 95%CI being less than 1. Multicollinearity was evaluated by calculating variance inflation factors for each covariate. Children with missing covariate data were included in analyses but HRs for missing categories are not reported. Analyses were conducted with Stata 16 (Stata Corp., College Station, Texas).