Assessing program coverage of two approaches to distributing a complementary feeding supplement to infants and young children in Ghana

The work reported here assesses the coverage achieved by two sales-based approaches to distributing a complementary food supplement (KOKO Plus™) to infants and young children in Ghana. Delivery Model 1 was conducted in the Northern Region of Ghana and used a mixture of health extension workers (delivering behavior change communications and demand creation activities at primary healthcare centers and in the community) and petty traders recruited from among beneficiaries of a local microfinance initiative (responsible for the sale of the complementary food supplement at market stalls and house to house). Delivery Model 2 was conducted in the Eastern Region of Ghana and used a market-based approach, with the product being sold through micro-retail routes (i.e., small shops and roadside stalls) in three districts supported by behavior change communications and demand creation activities led by a local social marketing company. Both delivery models were implemented sub-nationally as 1-year pilot programs, with the aim of informing the design of a scaled-up program. A series of cross-sectional coverage surveys was implemented in each program area. Results from these surveys show that Delivery Model 1 was successful in achieving and sustaining high (i.e., 86%) effective coverage (i.e., the child had been given the product at least once in the previous 7 days) during implementation. Effective coverage fell to 62% within 3 months of the behavior change communications and demand creation activities stopping. Delivery Model 2 was successful in raising awareness of the product (i.e., 90% message coverage), but effective coverage was low (i.e., 9.4%). Future programming efforts should use the health extension / microfinance / petty trader approach in rural settings and consider adapting this approach for use in urban and periurban settings. Ongoing behavior change communications and demand creation activities is likely to be essential to the continued success of such programming. Program implementation for Delivery Model 1 was carried out by a nongovernmental organization (NGO), CARE International in Ghana, in 13 neighboring rural communities in the East Mamprusi District of Ghana’s Northern Region. This region is much drier than southern areas of the country; the region experiences 3 months of rainfall annually between June and September, with a dry season extending from November to April [10]. Subsistence farming is the main source of income and is limited to staple grains and legumes [10]. Because of climate and distance from commercial centers, the Northern Region is one of the poorest and most food insecure regions of Ghana [10]. The target consumer age group for program delivery was children aged 6 months and over, with particular emphasis on the complementary feeding period from 6 to 24 months. The survey population consisted of children aged between 6 and 24 months and their principal caregivers (defined as the person who provides most care for the child and gives the child most meals on most days). A pilot survey was implemented before the start of the program to train survey staff and to test questionnaires and indicators (results not reported here). Three coverage surveys were implemented during the program delivery period. Survey rounds one and two were implemented at 3 and 10 months into the program, respectively. Survey round three was implemented 14 months after the start of the program. This was 3 months after BCC and demand creation activities had ceased. Survey samples were independent of each other. The surveys employed a two-stage sampling procedure. The first stage consisted of all 13 intervention communities as primary sampling units (PSUs). The second stage consisted of households sampled from each PSU using the quarter (QTR) method (i.e., division of a community into four areas of approximately equal population size) and a random walk (EPI3) sampling method in each quarter [11]. The EPI3 method selects the first household in a quarter to be sampled using the EPI strategy, with subsequent households selected by choosing a random direction and selecting the third nearest house in that direction [11]. This method has been shown to yield results comparable to simple random samples and to be better than the unmodified EPI sampling strategy when a wide range of indicators is being assessed [11]. Sample sizes were calculated for estimating a proportion with a finite population correction [12]. Assuming an expected coverage of 50%, a desired precision (i.e., half-width of the 95% confidence interval) of ± 8%, a maximum expected survey design effect of 2.0, and a population of N ≈ 12,000, the sample size required was calculated to be n = 300 households per survey. This was increased to n = 312 households per survey (giving n = 24 households from each village) to simplify partitioning of the within-community sample into quarters as required by the QTR sampling method. Delivery Model 2 was implemented by the local not-for-profit arm of a pan-African social marketing company (Exp Social Marketing—ESM) in three neighboring districts (Nsawam, Suhum, and Asamankese) in the Eastern Region of Ghana. These districts were selected because of their large urban and peri-urban populations, their proximity to Accra, and to avoid interfering with other regions in Ghana where known nutritional trials were ongoing. Income levels vary across the three districts depending on the degree of urbanization. Subsistence farming and petty trade dominate in rural areas. Commerce, manufacturing, building, and service sector activities dominate in the urban and peri-urban areas [10]. The Eastern Region benefits from substantial annual rainfall and a second growing season, supporting a more productive and varied agriculture than is possible in the Northern Region [10]. Poverty and food insecurity, while present, are both less severe and less prevalent than in the Northern Region [10]. The target consumer age group for the program was the same as in Delivery Model 1 (i.e., children aged between 6 and 24 months). The coverage surveys reported here sampled children aged between 0 and 24 months. The rationale for including younger children was to pilot a simple structured IYCF indicator set for use elsewhere. Data for children aged under 6 months are not presented here. Two coverage surveys were implemented during the program delivery period. Survey round one was implemented at 2 months into the program, and round two at 11 months into the program. Survey samples were independent of each other. The surveys were designed to be spatially representative, that is, the sample was distributed evenly across the survey area, using a spatial sample design that selected communities located closest to the centroids of a hexagonal grid laid over the survey area. The resulting sample is a triangular irregular network [13, 14]. A variable intensity sampling design was used [13]. In rural areas, the sample density was such that no person lived more than about 8 kilometers from a sampling point. Sampling density increased with increasing population density. Each survey used a sample of n = 18 caregiver-child pairs from m = 58 PSUs (villages or city blocks). The within-community sample in villages used systematic sampling of dwellings in the villages (or parts of villages) organized as a ribbon (or ribbons) of dwellings, and a random walk EPI3 sampling strategy in villages (or parts of the villages) organized as clusters of dwellings. Sampling in urban communities used systematic household sampling with a sampling interval calculated in the field. This sample design provides implicit stratification, selecting a sample that is distributed across both the entire survey area and within sampled communities [15]. This type of sample tends to spread the sample among important subgroups of the population, (e.g., rural, urban, and peri-urban; different administrative areas; ethnic / religious subpopulations; and various socioeconomic groups) and often improves the precision of estimates made from survey data [15, 16]. Ethical clearance to conduct the coverage surveys for both delivery models was obtained from the Ghana Health Services Ethical Review Committee (protocol ID number GHS-ERC-05092012). Oral consent to participate was obtained from the child’s principal caregiver on the basis that participation in the survey was voluntary. Written consent was not sought due to concerns regarding the adult female literacy rate in Ghana. Consent was recorded in survey supervisors’ logbooks. The Ghana Health Services Ethical Review Committee approved this consent process. Trained interviewers under the supervision of experienced field supervisors collected data. For Delivery Model 1, data were collected using mobile devices (Open Data Kit version 1.3) with pre-coded logical responses to ensure data quality. For Delivery Model 2, data were collected using paper forms, with data entry and interactive checking (for consistency, ranges, and legal values during data entry) and batch checking (double-entry and validation, as well as a batch application for consistency, range, and legal value checks) performed using EpiData (version 3.1) [17]. Data were collected on demographics and socioeconomic status; education levels within the household; housing conditions; recent infant and child mortality; water, sanitation, and hygiene (WASH) practices; food security; child health; IYCF practices; maternal dietary diversity; coverage of fortified staples; product coverage; and maternal and child anthropometry. The same survey instrument was used in both sets of delivery model assessments. All survey modules (i.e., question and indicator sets) were taken from validated guidelines with language, wording, and layout finalized through pilot testing in the field. All case-definitions (e.g., for maternal and child undernutrition, hunger, poor sanitation, and suboptimal IYCF practices) adhered to internationally recognized standards. Product coverage question sets and indicators were adapted from those used in semi-quantitative evaluation of access and coverage (SQUEAC) and simplified lot quality assurance evaluation of access and coverage (SLEAC) coverage assessments [18]. Three key indicators of risk (or need) were used to investigate the targeting efficiency of the two delivery models. These were poverty, poor maternal dietary diversity, and suboptimal IYCF practices. Poverty was assessed using an adapted multidimensional poverty index (MPI) [19]. Adaptations followed published guidelines. The MPI score is constructed as a weighted sum of indicators in three dimensions (health, education, and living standards) and ranges between 0 and 1. Fig 1 shows the component indicators and weightings used to calculate the MPI score used in the assessments reported here. A household was classified as being in poverty if the MPI score was greater than or equal to one third. HH = Household; HHS = Household Hunger Score; JMP = WHO/UNICEF Joint Monitoring Program for Water Supply and Sanitation; MUAC = Mid-upper arm circumference; PBH = Previous birth history; BCG = Bacillus Calmette–Guérin vaccine; WAZ = Weight-for-age z-score (WHO Growth Standards); Edema = the presence of bilateral pitting edema. Maternal dietary diversity was assessed using the Women’s Dietary Diversity Score (WDDS), which is a count of food groups (from a list of nine food groups) consumed in the previous 24 hours [20]. Poor maternal dietary diversity was defined as having a WDDS below the sample median WDDS. Suboptimal IYCF practices were assessed using an Infant and Child Feeding Index (ICFI) [21, 22]. All children aged between 6 and 24 months received an ICFI score between 0 and 6. The ICFI score is a measure of age-appropriate child feeding practices using age-appropriate scoring for breastfeeding, dietary diversity, and meal frequency (Table 1). Children with a total score less than 6 were classified as having suboptimal IYCF practices. Children with a total score less than 6 are classified as having suboptimal IYCF practices. Three measures of coverage were assessed following the model of Tanahashi [23]: “message coverage” (i.e., has the caregiver ever heard of the product?), “contact coverage” (i.e., has the child ever been fed the product?), and “effective coverage” (i.e., has the child been fed the product at least once in the previous 7 days?). Three summary statistics were estimated for each of the three coverage measures: A program may be classified as well functioning based on observing a high RC or efficient targeting of need (i.e., high MN with a CR above 1). Fig 3 shows how the three summary statistics are calculated from a two-by-two table. Data were analyzed using the R language for data analysis and graphics (version 3.1.2) and the R-AnalyticFlow scientific workflow system (version 3.01). A blocked weighted bootstrap estimation technique was used [24]. Bootstrap replicates consisted of a set of within-PSU survey samples that were sampled with replacement and with a probability proportion to PSU population size using a roulette wheel (also known as stochastic sampling with replacement) algorithm [25]. For each bootstrap replicate, a total of m PSUs were sampled with replacement (where m is the number of PSUs in the survey sample). Observations within selected PSUs were also sampled with replacement with the same within-PSU sample size that was achieved in the survey. A total of r = 400 bootstrap replicates were used. The required summary statistic was calculated from each replicate. The resulting estimate consisted of the 2.5th (lower 95% confidence limit), 50th (point estimate), and 97.5th (upper 95% confidence limit) percentiles of the distribution of the statistic across all replicates [26]. This procedure accounts for unequal selection probabilities in the sample design (by applying posterior weighting), as well as for any variance lost due to the clustered nature of the sample [24]. Coverage observed at each survey round was compared with the coverage observed at the previous survey round using a two-sample z-test. Individual standard errors were calculated as: where UCL and LCL are the upper and lower 95% confidence limits on the coverage proportion. The resulting standard errors were pooled: and the test-statistic calculated as: A two sided p-value was calculated. The design of the sample used for the assessment of coverage of Delivery Model 2 allowed for results to be mapped. RC for message, contact, and effective coverage were mapped. Interpolation between sampling points was performed using inverse distance weighting, using a global neighborhood with the weighting power that minimized errors in a twofold “holdout” cross-validation [27, 28].