Integrating nutrition into health systems at community level: Impact evaluation of the community-based maternal and neonatal health and nutrition projects in Ethiopia, Kenya, and Senegal

Maternal undernutrition and mortality remain high in several African countries. Key nutrition and health interventions improve maternal and birth outcomes. Evidence is scarce on how to strengthen health systems to ensure pregnant women and newborns are reached with these interventions. We conducted three quasi-experimental nonrandomized Community Based Maternal and Neonatal Health and Nutrition projects in regions of Ethiopia, Senegal, and Kenya to demonstrate how proven nutrition interventions could be integrated into health programs to improve knowledge and practices during pregnancy, birth, and postpartum. We evaluated impact on knowledge and practices related to maternal and neonatal care using logistic regression and repeated-measures models with districts as a fixed variable and adjusted for covariates. Combined country analyses show significant positive effects of the intervention on women receiving first antenatal care visit (ANC) during first trimester (OR = 1.44; p <.001), those consuming any iron and folic acid supplement during their latest pregnancy (OR = 1.60; p =.005), those whose <6 months infants were exclusively breastfed (OR = 2.01; p=.003), those whose delivery was facility based (OR = 1.48; p=.031), and those whose postnatal care was facility based (OR = 2.15; p<.001). There was no significant differences between intervention and control groups regarding one or more and four or more ANC visits, women consuming iron and folic acid for ≥90 days, and early initiation of breastfeeding. We conclude that integrating proven nutrition interventions into health programs at community level improved components of access to and use of ANC, delivery services, and postnatal care by women in three African countries. Kung'u et al. (2018), in this supplement, describe the health system context in the countries where CBMNH‐N was implemented, the components of the health system that were strengthened, and the overall logic model and the country‐specific logic models. These logic models illustrate how components of the interventions were linked with each other and with the ultimate outcome of reducing maternal morbidity and mortality. The CBMNH‐N interventions were intended to be delivered using primary health care platforms. We have therefore also described the different cadres of health workers in the health care platforms of the countries where we implemented the programmes. In summary, the key inputs of the CBMNH‐N interventions in the three countries included the following: Ethiopia—behaviour change interventions, integrating nutrition to the frontline health workers training guideline and family education and counselling take action booklet, routine community health management information systems and strengthening programme review process through collaborative quality improvement approach, and stop gap procurement and logistic support of maternal and newborn health and nutrition commodities such as iron and misoprostol; Kenya—community health worker training, behaviour change interventions at the community level, traditional birth attendant orientation, individual and collective incentive package, and team work and simulation training of health and nonhealth workers present at the health facility; Senegal—formation and training of facilitators to serve as influencers and lead pregnant women peer groups, training of community‐based health workers and providing an effective communication strategy designed for them, enlisting and equipping health huts, and addressing health posts supply and distribution issues for essential maternal and newborn health and nutrition commodities. We used quasi‐experimental, pre‐post nonrandomized intervention designs with intervention and control groups to collect repeated cross‐sectional data at baseline in 2013 and at endline in 2015. This study reports quantitative data from mothers with children 0–11 months. This cohort was represented in all three countries and was selected to ensure impact would represent the full pregnancy period but minimize recall bias. Mothers with children 0–11 months in Ethiopia, Kenya, and Senegal in the project regions were eligible to participate in the baseline and endline surveys. The primary sampling unit (PSU) was the district in Senegal, subcounty in Kenya and Woreda in Ethiopia. We calculated the required sample size using a two‐tailed alpha of 5%, a power of 90% (1‐β) in Kenya and Senegal and 80% in Ethiopia, and the minimum detectable effect size for the difference in change in two proportions of key outcome indicators of antenatal care, essential nutrition actions, delivery, and postnatal care at baseline and endline. The estimated sample sizes per group were adjusted for 10% attrition rate and a design effect of 2, a default value which should adequately compensate for the use of cluster sampling in most cases. Multistage cluster sampling was used in all three countries. In the first stage, we randomly sampled the departments and Woredas in Senegal and Ethiopia and purposively selected all primary health care facilities in Kenya within each subcounty that met the criteria of at least 10 deliveries per year. The second stage involved random selection of a village within each Woreda in Ethiopia, department in Senegal, or catchment area of the health facility in Kenya based on probability proportion to size. Lastly, within each village, we randomly selected households that met the inclusion criteria of having women who gave birth within 1 year preceding the survey and who are usual residents of the household. Random selection of households was from census lists if available by systematic sampling (where every kth household was selected, where k, the sampling interval = N, total number of households in a village ÷ n, number of households to be selected within each village). The first household visited was randomly selected. The method of selection of the first household depended on whether there was a list or not. Census lists were available in Senegal and Kenya. Community‐based personnel in Senegal and Kenya conducted a census of all eligible households with mothers with children 0–11 months in the village. Using this list, a sampling interval was calculated as indicated above. The households on the list were numbered and a random number selected from one to the highest numbered household on the list. The household on the numbered list where the number corresponded to the random number selected would be the first household to be visited. In Ethiopia, where there was no list and no census was conducted, a central location in the village was selected. Each team of two started out in different directions. The first household was randomly selected between the first and ninth households. From then onwards, data collection teams would call at every other household. If the household did not contain anyone eligible, the data collection team would move to the household next door and resume the data collection once they identify an eligible household. In households in which there was more than one person who met the criteria, the enumerators would randomly choose one person to interview. We collected data on the same PSU at baseline and endline. Combined across countries, there were five PSU, 1,670 women at baseline and 1,620 women at endline for the control group, and 10 PSU, 2,905 women at baseline and 2,570 women at endline for the intervention group, for a total sample size of 8,765 women (Figure 1). Selection of participating mothers with children 0–11 months. PSU = primary sampling unit 1Some mothers in Ethiopia from both the intervention and control groups were absent at the time of data collection for the endline survey because they had left their households due to drought. In all three countries, data were collected using face‐to‐face interviews at the respondent's home. The enumerators were comprehensively trained to standardize the data collection method and were supervised by trained supervisors and the research team. Training lasted 3 days in Kenya, 5 days in Ethiopia, and 2.5 days in Senegal. The difference in number of training days was because of the differences in scope of the surveys and the target groups in the three countries. This manuscript focuses on data collected from mothers with children 0–11 months in Ethiopia, Kenya, and Senegal in the project regions at baseline and endline. The questionnaires were finalized after pretesting in similar populations in neighbouring areas. Validation and verification of data were done at the end of each day of data collection. Supervisors were expected to check completed questionnaires on a daily basis and sign off each time they supervised the enumerators in the field. The evaluation team also made supervisory visits to the data collection sites. All surveys received ethical approval from nationally recognized ethical review committees. Permission to conduct the surveys was obtained from national and regional relevant bodies in each country. Before enlisting participants into the study, informed consent by signature or thumbprint was obtained from each participant. All individual identifiers of the respondents were removed prior to analysis. To develop the indicators for impact analysis, we reviewed the CBMNH‐N logic model, survey instruments, and project approach used in each country. For all three countries, description of the indicators and variables and the methodology used to create them are reported in the performance measurement framework of each project. We categorized the indicators into four major constructs: (a) quality and uptake of antenatal care; (b) essential nutrition actions in ANC, delivery, and postnatal care; (c) delivery with skilled and trained birth attendant; and (d) postnatal care. Because programme approaches and tools slightly differed across countries and surveys, some of the variables were not available for all three countries. Computer and manual coding of variables in baseline and endline surveys in three countries were done to create comparable variables where possible. All indicators used in the impact analysis were binary. For all indicators reported here except one, the denominator was all women. For the indicator “% whose < 6 months infants were exclusively breast‐fed”, the denominator was the total number of women with infants <6 months old. With two exceptions, all indicators had no or few (less than 4%) missing values. In the Ethiopia and Senegal baseline surveys, for percent of mothers who delivered with the assistance by Traditional Birth Attendants, 20% and 47% were missing, respectively. In the Senegal baseline and endline surveys, for the percent of mothers whose infants were put to the breast early, 20% and 27% were missing, respectively, because of a questionnaire error that did not bias the results. For each country separately, we used a fixed‐effects logistic regression repeated‐measures model that compared between control and intervention differences between baseline and endline by regressing each outcome indicator on binary indicators for intervention (vs. control), endline (vs. baseline), and their interaction; PSU was included in the model as a fixed effect, meaning that all differences among the PSU were explained in the analyses. Analyses were also adjusted for covariates: child age (months), mother age (years), mother completion of primary school (yes, no), mother in income generating activity (yes, no), and walking time from home to nearest health facility (≤1 hr, >1 hr). The estimates of interest for intervention effects came from the interaction terms that estimated the difference between intervention and control in the differences from baseline to endline, that is, difference in differences (Gertler and World Bank, 2016). Unadjusted effects were expressed as both difference‐in‐differences and odds ratios; adjusted effects were expressed as odds ratios. The estimator that combined effects from multiple countries considered each country equally by averaging the country‐specific differences and log‐odds. The p values for combined effects were calculated by averaging the Z‐statistics using the inverse normal method (Hedges & Olkin, 1985). The p values for adjusted combined effects were calculated without and with incorporating heterogeneity among country effects. The former reflects the evidence for effectiveness of the interventions in the three countries combined, whereas the latter reflects expectations if implementing the intervention in a new country. The p values <.05 were considered significant.