Bottlenecks and predictors of coverage and adherence outcomes for a micronutrient powder program in Ethiopia.

Abstract A theory‐driven evaluation was conducted to assess performance of a trial to deliver micronutrient powder (MNP) through the Ethiopian Ministry of Health. We adapted an approach to coverage assessment, originally developed to identify bottlenecks in health service delivery, to examine sequential program outcomes and their correlates using cross‐sectional survey data of caregivers of children 6–23 months (N = 1915). Separate multivariable Poisson regression models were used to estimate adjusted risk ratios of conceptually relevant determinants of coverage and adherence. Caregivers of children >11 months were more likely to have received MNP than caregivers of younger infants, yet children 12–17 months were 32% (P < 0.001) and children 18–23 months 38% (P < 0.001) less likely to have been fed MNP in the 14 days preceding the survey than children 6–11 months. Among caregivers who initiated feeding MNP, the most frequently reported reasons for discontinuing use were not obtaining additional supply (36.1%) and perceived child rejection of food with MNP (22.9%). For each additional time a caregiver met with frontline workers in the 3 months preceding the survey, they were 13% more likely to have recently fed MNP (P < 0.001). Caregivers' perception that MNP produced positive changes in children was associated with a 14% increase in the likelihood of having recently fed it (P < 0.001). These results emphasize the importance of counselling for MNP and infant and young child feeding for initial use and the importance of multiple contacts with frontline workers for continued use.

biological pathway to impact for MNP is well articulated (De-Regil, Suchdev, Vist, Walleser, & Peña-Rosas, 2013), but program implementers are often constrained by insufficient understanding of the factors and processes underlying successful delivery of the intervention .
Only 7% of Ethiopian infants and young children 6-23 months of age consume the minimum number of food groups associated with better nutrient adequacy and growth (Arimond & Ruel, 2004;Central Statistical Agency Ethiopia & ICF International, 2007;Marriott, White, Hadden, Davies, & Wallingford, 2012). A recent national micronutrient survey found that 34% of children under 5 years of age experience anaemia, 18% are iron deficient, 14% vitamin A deficient and 35% zinc deficient (Ethiopian Public Health Institute, 2016). Point-of-use fortification of complementary foods with MNP has been demonstrated to improve growth and haemoglobin status among Ethiopian children 6-23 months of age (Samuel et al., 2018).  (Tumilowicz, Schnefke, Neufeld, & Pelto, 2017). The PIP described two sequential systems: (1) the program delivery system, which includes the processes by which the intervention is delivered to the household, and (2) the household delivery system, which consists of all steps required for a child to consume a biologically impactful dose (Carroll et al., 2007).
For MNP interventions, the processes underlying the household delivery system are based on adherence-the extent to which a caregiver's trial and adoption of behaviour and the child's consumption are congruent with the MNP recommendations offered by the delivery system Vrijens et al., 2012). Figure 1 presents the flow of the PIP in the form of program component exposures and actions by the caregiver. The PIP describes the transmission of MNP and behaviour change communication (BCC) activities from the program delivery system to the household delivery system, as well as the sequence of activities to prepare food with MNP and feed it to the child.
The extent to which a nutrition program is delivered as intended substantially influences its potential for impact (Mbuya, Menon, Habicht, Pelto, & Ruel, 2013). This challenge is exacerbated in interventions and programs with either multiple or complex delivery systems (Pearce & Merletti, 2006). The key challenges in studying the handover processes within and between delivery systems are to conceptually and analytically elucidate them in order to ascertain their determinants and identify how to create conditions that improve the effectiveness of nutrition interventions (Mbuya et al., 2013). In this analysis, we adapt the Tanahashi approach to coverage assessment (Tanahashi, 1978), originally developed to identify bottlenecks in health service delivery, to describe changes in MNP coverage and adherence outcomes. We add to the Tanahashi method an examination of determinants for the sequential bottlenecks. Our intention is to contribute to the expanding evidence base on MNP programs and apply a novel analytic approach to examine the determinants of each outcome separately.

| MNP integration into MOH Health Extension Program
The trial aimed to deliver the MNP locally branded as Desta and infant and young child (IYC) feeding BCC activities to approximately 71,000 children 6-23 months in six woredas in Amhara Region (Kobo, Gubalafto, Delanta, Dessie Zuria, Kalu and Bati) and five woredas in

Key messages
• Continued MNP use was positively associated with living in rural Tigray Region, child age <12 months, female child, increased frequency of contact with frontline workers and perceived positive outcomes from feeding MNP.
• Caregivers face challenges to continue feeding MNP through their children's second year of life as they experience periods of illness and poor appetite, progress through developmental stages that affect feeding behaviours and stop routine attendance at health services.
• Frequent contacts with frontline workers foster continued use of MNP, likely through ensuring caregivers have supply of MNP, alerting caretakers to positive changes and potential negative side effects in the child, and helping address MNP and IYC feeding difficulties.
Tigray Region (Saesi Tsaedaemba, Mereb Lehe, Werie Lehe, Gulomahda and Ahferom). The distribution of Desta was integrated into the IYC feeding BCC strategy developed by Alive & Thrive and implemented through the MOH's Health Extension Program (Kim et al., 2015(Kim et al., , 2016. The MOH, Concern Worldwide and Global Alliance for Improved Nutrition created guidelines, a training manual, job aids and materials for caregivers, which included information on the composition and purpose of Desta, instruction on food preparation with Desta and contraindications and potential side effects of Desta (see Table SS1 in the supporting information for key messages regarding Desta that HEW and health centre workers were trained to provide to caregivers). Starting in May 2016 and until May 2017, HEW or health centre staff provided a box containing 30 Desta sachets every 2 months to caregivers at the health post, health centre or during home visits. Caregivers were instructed to feed food mixed with one sachet of Desta to the child 6-23 months of age every other day and return to the HEW or health centre for a new supply after 2 months. At the initial distribution, caregivers were provided with a "Desta child card" that could be used to keep track of distribution dates. The card included appointment dates for caregivers to return to receive a new box of Desta, which was scheduled 2 months after the date of the current distribution.

| Study design and population
We identified two main questions to be addressed in the evaluation: (1) what proportion of caregivers received the intervention and adhered to Desta recommendations and (2) how did the delivery and utilization processes that characterized the trial affect program outcomes, including coverage and adherence to recommendations regarding Desta. Examination of these two questions required a complex study design and mixed methods. Therefore, we created two intersecting streams of inquiry: (1) an examination of sequential program outcomes and their correlates using cross-sectional survey data and (2) a focused ethnographic study (FES) to investigate delivery experiences and Desta use from the perspectives of caregivers. We report on the results of the FES in a separate paper of this supplement (Pelto et al., 2019). This paper presents key findings from a population-based cross-sectional survey conducted at the same time as the FES, approximately 8 months after initiation of MNP

| Sampling
The implementation area was divided into three strata: rural Amhara, rural Tigray and urban Amhara and Tigray combined. As it was not budgetarily feasible to collect data for the number of enumeration areas (EA) required to create separate urban strata for the regions, urban areas in Amhara and Tigray were combined into a single stratum, which permitted comparisons across the two rural areas and the combined urban area. The sampling frame used for the first stage  Program impact pathway for micronutrient powder delivery system in Ethiopia. HEW: health extension worker; HC: health centre staff; IYC: infant and young child; Desta, brand of micronutrient powder used in the trial to conducting the survey, six of the urban EA were excluded from the survey because a recent administrative reorganization allocated these EA to areas outside the trial's implementation woredas.
Prior to the second stage of sampling, a household listing was carried out in each selected EA to identify all potentially eligible children 6-23 months of age and to update the population size estimate. The resulting child lists in each EA were used in the second stage of sampling, which consisted of a simple random selection of 18 children 6-23 months of age in each EA. If the caregiver provided written informed consent, a questionnaire was used to collect data on Desta coverage and adherence outcomes, caregiver exposure to the elements of the trial design and perception-of-use factors, variables that could have affected program delivery exposure or adherence outcomes and covariates such as demographic factors.
The questionnaire response rate of the 2,052 selected mothers was 95.6% overall (N = 1,963), 96.8% (N = 697 of 720) in Tigray, 94.4% (N = 680 of 720) in Amhara and 95.8% (N = 586 of 612) in the urban stratum. After further elimination of 47 cases with implausible responses (e.g. reporting to have fed Desta despite not having heard of it or received it), the total sample of caregivers used in the estimation of outcomes was N = 1916 (93%). The sample size for the multivariable analysis was 1915 because one caregiver's age was missing and therefore excluded from the analyses.

| Outcome variables
We adapted the Tanahashi (1978) approach to health program coverage assessment (Tanahashi, 1978), to the PIP of the Ethiopia trial and indicators collected by the survey. We specified four main sequential outcomes to analyse bottlenecks. The first two outcomes relate to the trial's program coverage and latter two to the household adherence to Desta recommendations: a. "Ever heard," the proportion of all caregivers who ever heard of Desta; analogous to "availability coverage" as defined by Tanahashi b. "Ever received," the proportion of all caregivers who ever received a box of Desta; analogous to "accessibility coverage" as defined by Tanahashi c. "Ever fed," the proportion of all caregivers who ever fed Desta to the child; analogous to "contact coverage" as defined by Tanahashi d. "Recently fed," the proportion of all caregivers who fed Desta to the child at least one time in the 14 days preceding the survey; analogous to "effectiveness coverage" as defined by Tanahashi The data did not permit construction of a more accurate indicator of effectiveness coverage or adherence to recommendations (e.g. number of sachets used over the last 7 days). Although imperfect, we considered "used at least once within the 14 days preceding the survey" to be indicative of continued use and effectiveness coverage.
Each outcome is conditional on having accomplished the previous one; for example, mothers can only have "ever fed" if they "ever received" Desta. For "ever heard," the study population provides the comparator.
A large difference in the proportions of the target population accomplishing adjacent outcomes implies the existence of a bottleneck in the PIP. The magnitude of the bottleneck is the attrition in proportions between one outcome and the next.
To examine the determinants of each step through which Desta is transferred from program implementers to caregivers and ultimately to their children, we computed the following outcomes' proportions, first relative to diminishing populations according to progression along the PIP, and then relative to the total population of caregivers ( Figure 1

| Independent variables
Prior to this study, we conducted an analysis of the literature to identify factors that affect MNP adherence , which provided a theoretical basis for the selection of individual factors (independent variables) that may have a role in either facilitating or limiting MNP intervention outcomes. These potential determinants related to socio-demographic factors, caregiver and child characteristics, IYC feeding practices, exposure to the intervention and perception-of-use factors. Table S2 lists all variables assessed in the analysis.
A socio-economic status (SES) score was derived using principal component analysis (PCA) as per (Chasekwa et al., 2018) from variables covering household characteristics, asset ownership, water and sanitation. All variables were recoded as binary indicator variables, and those with frequencies <5% or >95% or >5% missing were excluded. This cutoff was used to offer discriminating power by excluding particularly uncommon and very common characteristics.
Of the 60 variables initially considered, 39 met the inclusion criteria.
To compute a SES score we first carried out polychoric PCA (with orthogonal varimax rotation) on the set of 39 variables and identified the first principal component, which explained 21% of the total variation. We then computed the SES score as the sum of standardized factor loadings from this first principal component. The resulting score was then validated internally and externally by examining trends in the proportion of ownership or presence of some assets across five quantiles of the score. Several variables were examined, some used in the construction of the index (internal validity) and others which were excluded (external validity).
IYC feeding practices were assessed using questions adapted from the WHO/United Nations International Children's Emergency Fund document on "Indicators of Infant and Young Child Feeding" (World Health Organization (WHO), 2008). Caregivers reported all foods the child consumed in the previous day and the frequency of feeding, and data collectors coded responses based on standard food groups.
In the questionnaire the responses were recorded in 11 food groups that were subsequently recoded during data analysis into the seven WHO/United Nations International Children's Emergency Fund food groups of (1) grains/roots/tubers, (2) legumes/nuts, (3) dairy, (4) flesh foods, (5) eggs, (6) vitamin-A-rich fruits/vegetables and (7) other fruits/vegetables. From these scores, a child dietary diversity score was calculated and minimum dietary diversity defined as ≥4 of seven food groups, minimum meal frequency as solid/semi-solid foods ≥2 times per day for breastfed infants 6-8 months, ≥3 times per day for breastfed children 9-23 months and ≥4 times per day for nonbreastfed children 6-23 months.
Household hunger was assessed using the Household Hunger Score (HHS; Ballard, Coates, Swindale, & Deitcher, 2011). The HHS is an indicator of household hunger and focuses on the food quantity dimension of food access. HHS consists of three occurrence questions and three frequency-of-occurrence questions. The HHS occurrence questions ask whether a specific condition associated with the experience of food insecurity ever occurred during the previous 30 days.
The HHS frequency-of-occurrence questions ask how often a reported condition occurred during the previous 4 weeks: rarely, sometimes or often. Households with HHS > 1 are categorized as experiencing moderate or severe hunger.
We used polychoric PCA with orthogonal varimax rotation to compute perceptions-of-use latent factors and reduced the dimensions of the 25 perception-of-use questions (Chasekwa et al., 2018). Caregivers were asked without prompting to list all the positive changes, negative side effects and challenges that they perceived related to feeding Desta to their child. All 25 perception-ofuse variables (Table S2) were recoded as binary indicator variables, and those with frequencies <5% or >95% or >5% missing responses were excluded as they were either too rare or too common to offer discriminating power. This resulted in 14 variables being included in the analysis (Table S3). Based on a scree-plot of eigenvalues and examination of the factor loadings, we identified three principal components, which were conceptually meaningful. The three-factor solution explained 45.3% of the total variance, 18.7% for factor 1, 14.4% for factor 2 and 12.2% for factor 3. A principal component score for each caregiver was then calculated by adding the standardized loadings for all variables in the set. Items that had loadings >0.15 on the first principal component (factor 1), which we refer to as "perceived positive outcomes," included the number of positive changes caregivers reported, perceived physical growth, perceived increased strength, perceived increased activity, perceived increased immunity and perceived increased appetite. Items that significantly loaded on the second principal component (factor 2), referred to as "perceived negative side effects," included the number of negative side effects reported by the caregiver, perceived loose stool/diarrhoea and perceived nausea/vomiting. Items that loaded on the third principal component (factor 3), referred to as "perceived challenges to feeding Desta," included the number of challenges to feed Desta reported by caregiver, child rejection of food with Desta and perceived black stool. For ease of interpretation and comparison across factors, factor scores were standardized so that each has a mean of zero and standard deviation of one.

| Data entry and statistical analysis
Double data entry was conducted by dedicated data entry clerks using Epi Info™ 7.2.1.0 (Dean et al., 2011). Subsequently, thorough consistency checks of the merged databases were conducted. Statistical weights applied during all analyses accounted for the differences between estimated EA populations used in first-stage sampling and the actual EA population counted by the survey teams. Additional statistical weights were used in analyses of the entire sample to correct for different sampling fractions in different strata. We did not use a finite population correction factor in the design because the sampling fraction of the population was much lower than 5%. As a result, the N are the same as those reported in the actual survey. All analyses were carried out using Stata 14 (StataCorp, 2015) and accounted for survey design and weights. For continuous and categorical variables, means and proportions were compared across strata using a Wald or Pearson chi-square test, respectively.
The association between each of the four conditional outcomes (P1, P2, P3, and P4) and outcome among all caregivers (PT) with independent variables was estimated using generalized linear models with a Poisson log-link function and cluster-robust standard error estimates. The coefficients for each step (P1-P4) are the effects as one moves from P1 (ever heard) through P4 (fed Desta at least once within the last 14 days). Thus, changes in a coefficient over the P reflect changes in its association as one moves along the PIP. A final model for PT was estimated, which gave the adjusted relative risk of continued feeding among all caregivers. To construct multivariable models, first bivariate analyses were explored with each independent predictor variable (Table S4). A full multivariable model was then constructed using all predictors with P < 0.25 in bivariate analyses or with a strong theoretical basis for inclusion. In sensitivity analyses, parsimonious models were identified using forward step-wise selection permitting all covariates to enter at P < 0.25. The adequacy of the parsimonious models was assessed against the full models using Akaike information criterion and Bayesian information criterion. The full models were found to adequately explain the data for each outcome.

| Ethical considerations
The survey protocol was approved by the Ethiopia National Research Ethics Committee, clearance number IEHD/250/12/08. Written informed consent was sought from the caregiver of each selected child. Children with a mid-upper arm circumference <110 mm were referred to the nearest health facility following MOH protocol for the treatment of severe acute malnutrition.

| RESULTS
Descriptive data of the survey sample show differences among strata, with the largest contrasts being between the urban stratum and the two rural strata of Amhara and Tigray (Table 1). Although most caregivers from the urban stratum were classified in the highest SES tercile, most caregivers from rural strata were distributed between the middle and lower SES terciles. Compared with caregivers from the rural strata, caregivers in the urban stratum were more educated, and a lower proportion perceived the distance to access HEW and health services as "very long." Children in the urban stratum had  18.5, 27.9) of caregivers who had stopped feeding Desta did so because they perceived the child rejected food mixed with Desta, with no significant differences in proportions of caregivers reporting this problem across strata or age group.
Examining the changes of a single independent variable across levels of coverage and adherence outcomes from P1 to P4 gives insights about their associations with bottlenecks as one moves along the PIP (Table 4). In other words, one can identify variables that are associated with increased or decreased outcomes as one moves across the PIP. Sometimes, the direction of variables' association may go in opposite directions at different steps and thus can cancel each other out in PT. Below we discuss independent variables with significance P < 0.05 as one moves along the PIP from P1 to P4, as well as their overall effects reflected in PT (Table 4)

| Caregiver education
Among caregivers who heard of Desta, those with <5 years of education were less likely to have received Desta than caregivers with ≥5 years of education (P2, ARR (95% CI): 0.96 (0.92, 0.99) P = 0.026). Caregiver education was not found to be associated with any other step in the PIP or PT.  Perceptions of use and outcomes scores c Perceived positive outcomes score, mean

| Child dietary diversity
Increasing child dietary diversity was slightly negatively associated with having received Desta among those who heard about it (P2, FIGURE 2 Bottlenecks in coverage and adherence outcomes among all caregivers (N = 1916) reflective of transfer of information and Desta from the program to the caregiver and household delivery system. Desta, brand of micronutrient powder used in the trial. "Ever heard of Desta," proportion of all caregivers with children 6-23 months of age who ever heard of Desta. "Ever received Desta," proportion all caregivers with children 6-23 months of age who ever received Desta. "Ever fed Desta," proportion of all caregivers with children 6-23 months of age who ever fed Desta to the child. "Recently fed Desta," proportion of all caregivers with children 6-23 months of age who fed Desta to the child at least one time in the 14 days preceding the survey

| Perception of use and outcomes scores
Among caregivers who initiated feeding Desta, one Z-score increase in the caregivers' perception that Desta produced positive changes in children was associated with a 14% increase in the likelihood of having recently fed it (ARR [95% CI]: 1.14 [1.10, 1.19] P < 0.001). The addition of perceptions of use and outcomes scores derived from factor analysis to the multivariable model did not substantially change the other determinants' effects (Table 5).

| DISCUSSION
The process of adherence to health interventions has been articulated in a model by Vrijens et al. (Vrijens et al., 2012). In this model, the process starts with "prescription" of the intervention, which occurs when individuals learn about the intervention, receive it and agree to try it.
The process continues with "initiation," consisting of trying the intervention and then continuing to practice the new behaviour. The final phase is "continuation" or "persistence," defined as the period between initiation and the last dose. The study presented here was designed to permit us to differentiate the steps through which messages, and Desta were transferred from program implementers to TABLE 3 Reasons for not feeding Desta among caregivers who did not feed Desta to child in 14 days previous the survey (N = 556)    caregivers and from caregivers to their children, so that we could examine the determinants of each step in the process of adherence.
Using this approach, we discovered the central role and  (Table 3). Our findings from the FES, which was conducted in a subsample of caregivers who participated in the trial (Pelto et al., 2019), provide insights for the interpretation of this result.
In the FES sample, nine caregivers spontaneously said that rejection of food mixed with Desta was the reason for stopping it, but all except one reported that initially, and even for periods of up to 9 months before they began to refuse it, the child had accepted it. These findings may be attributable to a natural evolution of taste preferences. Many children in the second and third years of life enter a "neophobic" phase during which some previously liked foods are no longer accepted, and the introduction of new foods becomes difficult (Nicklaus, 2009  and no correlations between perceived negative side effects and discontinuation of MNP (Jefferds et al., 2015;Mirkovic et al., 2016;Tumilowicz et al., 2017). The influence of negative side effects on adherence appears to depend on the counselling received by caregivers (Loechl et al., 2009;Tripp et al., 2011). When caregivers are forewarned about the possible negative side effects of MNP (i.e. changes in stool including darker than usual, mild diarrhoea or a mild form of constipation), experiencing them does not appear to deter continued MNP use, and, paradoxically, their appearance may even encourage continued use (Loechl et al., 2009;Mirkovic et al., 2016;Tripp et al., 2011). Caregivers in Ethiopia were advised that in the first few days of taking Desta, they may observe changes in the child's stool, which will usually disappear after a period of 4-5 days (Table SS1).
Shifting attention to the effect of sex, the positive association of being a female child on recent feeding of Desta is not explainable based on available evidence. The prevalence of stunting among Ethiopian girls 0-23 months of age has been reported to be lower than that of boys, but the difference disappears after the second year of life (Woodruff et al., 2017). Sex and gender differences in IYC feeding practices and growth are evident across very different national and cultural contexts, including Senegal (Bork & Diallo, 2017), Guatemala (Tumilowicz, Habicht, Pelto, & Pelletier, 2015) and the Philippines (Adair & Guilkey, 1997;Popkin et al., 1990). Indigenous mothers in Guatemala were found to perceive boys as hungrier and requiring earlier and more frequent complementary feeding compared with girls. In contrast to the findings of the current study, we expect Guatemalan indigenous boys would be fed food with MNP more often than girls.
We encourage future studies in Ethiopia and other contexts to explore the interactions of biological-cultural-ecological factors that produce differences between sexes in IYC feeding and growth patterns. A model for this type of research has been previously proposed (Tumilowicz et al., 2015).
The main limitation of this study is its cross-sectional design. A stronger case for attributing the relative longitudinal influence of specific factors to caregivers' interaction with the program delivery system and adherence to MNP recommendations requires a longitudinal cohort study design. Following children over a period of months would allow for deeper exploration and testing of hypotheses regarding why the use of MNP declines with age. It is also required to construct a more accurate indicator of effectiveness coverage or adherence to recommendations. Despite the design limitation, examining the steps in the PIP helps to interpret the results of the combined model (PT) and pinpoint at which steps in the PIP opportunities lie for program implementers to improve coverage and adherence outcomes. The analytic methodology presented in this study, as well as the emergent, substantive findings of the factors associated with key bottlenecks, can be used to inform the design of epidemiological studies and evaluations of MNP programs.
Our findings indicate that caregivers face challenges to continue feeding MNP through their children's second year of life as they experience periods of illness and poor appetite, progress through developmental stages that affect feeding behaviours and stop routine attendance at health services. Notwithstanding these challenges, frequent contacts with frontline workers fostered continued use of MNP, likely through ensuring caregivers have supply of MNP, alerting caretakers to positive changes and potential negative side effects in the child, and helping address MNP and IYC feeding difficulties. This study contributes to growing evidence of the critical role of frontline workers in MNP adherence (Reerink et al., 2017;Tumilowicz et al., 2017) and more broadly the critical role of community-based programs to improve maternal and child health and nutrition (Lewin et al., 2010;Menon et al., 2016;Singh & Sachs, 2013). The Ethiopia MOH Health Extension program and platforms like it provide opportunities to scale-up MNP delivery but will require strengthening so that frontline workers have the time and ability to effectively support caregivers.