Animal source food consumption in young children from four regions of ethiopia: Association with religion, livelihood, and participation in the productive safety net program

Introduction: Child undernutrition remains a challenge globally and in the geographically diverse country of Ethiopia. Improving dietary diversity and consumption of animal source foods are important for improving child nutrition and corresponding health outcomes. Objective: The objective of the study was to identify household and community factors associated with consumption of animal source foods among 6 to 36-month-old children from four regions of Ethiopia. Methods: A cross-sectional survey using multistage probability sampling in eight geographic zones and four regions of Ethiopia took place in 2015 with parents/caretakers of 6 to 36-month-old children. Data was collected on demographic information, proxy indicators of socioeconomic status, and food consumed by the child the day before the survey. Results: Increased child age, pastoral livelihood, Muslim religion, and participation in the Productive Safety Net Program were associated with increased consumption of animal source foods. Odds of animal source foods consumption increased by 8% with each 3-month age increase. Children from pastoralist households were the most likely to have consumed animal source foods in the preceding 24 hours as compared with those in agro-pastoralist households (0.21 times as likely) or those in agriculturalist/farming households (0.15 times as likely). The odds of consumption of animal source foods for families with food aid or safety net support was 1.7 times greater among those receiving traditional support from the Productive Safety Net Program and 4.5 times greater for those in the direct support arm of the program. Conclusions: The findings illustrate the importance of accounting for local context and community characteristics, such as livelihood and religion, when undertaking programming designed to improve diversity of children’s diets through increasing animal source foods. In addition, the Productive Safety Net Program may be a critical determinant of dietary diversity for young children in these regions. The study was conducted in eight zones and four regions in Ethiopia: Afar, Amhara, Benishangul-Gumuz, and Tigray from October through December 2015. Geographic remoteness and resource limitations necessitated limiting the survey to two zones per region. The zones were purposively selected based on IYCF factors identified from previous research to capture a range of experiences within the region [39]. The eight zones surveyed were: Zone 1 and 4 in Afar, South Wollo and West Gojjam in Amhara, Asosa and Metekel in Benishangul-Gumuz, and Eastern and North Western Zones in Tigray. The Afar region in the northeastern part of Ethiopia is a predominantly pastoralist region. Afar is a region where the highest burden of underweight (36%) and second highest burden of wasting (18%) in under-five children was reported [7]. Amhara, Benishangul-Gumuz, and Tigray are largely agrarian regions. The Amhara region is characterized by the highest prevalence of stunting in Ethiopia where 46% of the under-five children in the region are stunted [7]. The second highest prevalence of stunting among under-five children is reported from Benishangul-Gumuz [7]. About 39%, 23%, and 11% of the under-five children from Tigray are stunted, underweight, and wasted, respectively [7]. The heavy burden of child malnutrition, low levels of child dietary diversity, and population diversity of these regions present an optimal area for investigating the research question. The data in the present study were collected for operation purposes by an international non-governmental organization and analyzed separately by the study authors as strictly de-identified. A cross-sectional study design was employed to conduct the study on ASF consumption in 6 to 36- month-old children from the four regions of Ethiopia. The inclusion criteria included children of 0 to 36 months of age living in the households of interviewed mothers/caretakers in rural areas. Because the outcome of interest (ASF) should not be given to children younger than 6 months of age based on international recommendations, this analysis excluded children in the sample younger than 6 months of age, but the remaining children were still the youngest in the household since only one child per household was included. A multi-stage sampling technique was used to select study participants. The sample size was calculated to detect a 10% change at household level in receipt of nutrition services across the sampled zones with the anticipation of a follow up survey. One hundred simulations were conducted to determine the sample size using the R package clusterPower [40] with 80% power and type-1 error rate of 5%. Between-cluster variance was assumed to be 0.1 in all survey rounds and sample size simulations. The results indicated that sampling of ten clusters per zone with fifteen households per cluster would allow equal probability of selection for all households that met the inclusion criteria per zone. Minimum samples of 143 households per zone were indicated to detect the identified change. In order to preserve internal validity, 12 clusters per zone were used so that the minimum sample sizes could be met without the need for replacement of non-responsive households or unreachable clusters. These methods were employed to eliminate sampling bias as much as possible. Although the initial sample size was calculated for a different outcome, the sample of 1009 was amply large to detect the hypothesized differences in the present study. Using G*Power [41], a 15 percentage point difference in consumption of animal source foods across three groups, similar to the religion variable, is detectible using a contingency table chi square test for independence at 80% power and 5% type-1 error rate with a sample of 531. As required by the local government partner, the Ethiopian Central Statistical Authority (CSA) used probability proportion to size random sampling to select twelve clusters in each zone from the 2007 census listing of rural enumeration areas. These were sampled without replacement if the cluster was deemed unreachable during data collection due to unfavorable field conditions. Upon arrival in each cluster, the survey team consulted community leaders (e.g., health extension workers) to define the geographical confines of the cluster, which corresponded to a local village. The last stage of sampling began with the full enumeration of the cluster to develop a listing of all eligible households (those with 0–36-month-old children) in the area. The survey team leader then used a random number table to randomly sample fifteen households. Enumerators were instructed to revisit sampled households three times before it was considered non-responsive. Each household completed only one household level survey, based on the youngest child living in the house, so the household survey gives data on one child/caregiver dyad. The household survey questionnaire included sections on demographics, water and sanitation, reproduction, breastfeeding, complementary feeding, nutrition services, maternal and child health, and food security. The questionnaires were developed in English, translated to Amharic for Amhara, Afar, and Benishangul-Gumuz regions and Tigrigna for Tigray, back translated to English, and pre-tested. Experienced enumerators who spoke the languages of the surveyed areas implemented the study and were trained on the survey protocol and tools during a four-day training session held in Addis Ababa prior to data collection. The enumerators conducted the household surveys in the homes of caregiver participants. Paper questionnaires were used for data collection. The completed questionnaires were then entered into a database by separate data entry persons and converted into excel spreadsheets. Verbal informed consent was obtained from all participants prior to data collection during fieldwork. The authors received approval from the Tulane University Health Sciences Institutional Review Board for secondary analysis of data collected by a non-governmental agency. The primary outcome measure for the present analysis was consumption of ASF by 6–36-month-old children. The questionnaire asked parents/caretakers what types of foods the child had eaten in the 24 hours prior to the survey. These questions were matched to those used in the 2011 EDHS [25], and the respondent stated if the child consumed any amount of foods from a predefined list. The animal source foods in the list included eggs, fish, yogurt, cheese, milk, meat (including beef, poultry, pork, lamb, and any other meat not mentioned), and organ meats (e.g., liver). As a secondary outcome of interest, minimum dietary diversity (MDD) was calculated to identify if a positive relationship between ASF and MDD was present in this sample. MDD was defined according to the WHO definitions of IYCF indicators [19], requiring a minimum consumption of four out of seven pre-defined food groups in the previous 24 hours for the child’s diet to be considered adequately diverse. The seven food groups include: grains, roots, and tubers; legumes and nuts; dairy products (milk, yogurt, cheese); flesh foods (meat, fish, poultry, and liver/organ meat); eggs; vitamin-A rich fruits and vegetables; and other fruits and vegetables [19]. Since three of the seven food groups are ASF, it was anticipated that ASF consumption contributed to dietary diversity in the sample. A group of cases from the region of Benishangul-Gumuz were identified to have answered the dietary recall questions in a way that did not meet the requirements of the study protocol. This resulted in the need to exclude 50 participants to eliminate introducing bias in the measurement of the outcome of interest. Measured independent variables were selected based on their hypothesized influence on consumption of ASF. Child and household characteristics included in this analysis were child sex and age; respondent age, education, and occupation; and household total number of children, assets, livelihood, religion, and livestock ownership. Education is a dichotomous variable identifying caretakers reporting any versus no formal or informal (e.g., religious schooling that includes learning to read) education. Respondent occupation was reduced to a three-level variable to include farmer, housewife (defined as having no employment outside the home), and other occupations. Household asset ownership was determined from a set of twelve assets (equivalent to those asked on the EDHS) including electricity, a watch or clock, a radio, a television, a mobile telephone, a non-mobile telephone, a refrigerator, a table, a chair, a bed with a mattress (cotton/sponge/spring), an electric mitad (a grill or cooktop used for preparing injera or bread), and a kerosene lamp/pressure lamp. Ownership of these assets were summed to assign a score of asset ownership to households. Family livelihood was determined from three questions regarding activities of anyone in the household: ownership of agricultural land, use of land to grow crops to consume or sell, and movement for the purpose of obtaining food or water for livestock. Households were categorized as agriculturalists if they owned land and used it to grow crops for consumption or to sell but did not move in search of food or water for livestock. Participants that grew crops for sale or consumption on their land and moved in search of food and water for livestock were categorized as agro-pastoralists. Households moving in search of food or water for livestock, but not owning land used to grow crops for sale or consumption, were classified as pastoralist households. Religion was constructed as a three-level variable to identify Ethiopian Orthodox Christians, Muslims, and other religions (including other Christian faiths). Livestock ownership was assessed as the total number of animals owned. Occupation and livelihood were highly correlated (Pearson correlation coefficient = 0.650) since the farmer occupation and agriculturalist livelihood captured almost the same group. Thus, these were combined to eliminate issues of multi-collinearity in the multivariable logistic regression model, and allowed for fewer missing cases. Participants/caregivers who reported their occupation as housewife or other but were assigned to pastoralist or agro-pastoralist livelihoods were analyzed as pastoralist or agro-pastoralist, not as housewives or other occupation. The resulting variable was considered as the household-level livelihood/occupation. Food security was assessed using a portion of the Household Food Insecurity Access Scale (HFIAS) developed by the Food and Nutrition Technical Assistance Project in 2007 [42]. Time limitations prevented the full set of HFIAS questions from being included on the questionnaire. The questions used were intended to measure insufficient food intake and its physical consequences, defining severe food insecurity by the HFIAS [43]. These asked about the occurrence (yes/no) and frequency (rarely, sometimes, often) during the previous month of the following food insecure events: was there ever no food in the household, did they ever go to bed hungry because there was not enough food to eat, and did they ever go a whole day and night without eating because they did not have enough food. A severe food insecurity score was calculated per the full HFIAS questionnaire [42], by summing the frequency of occurrence (0 for never, 1 for rarely, 2 for sometimes, and 3 for often) of each event, giving a score ranging from 0, indicating the household had not experienced any severe form of food insecurity, to 9, indicating all three severe food insecurity events happened often. Participants were asked about receipt of food aid and participation in the PSNP during the previous year. The PSNP provides food, cash, or a combination in exchange for labor on public works projects supported by the Agriculture Extension Program [32]. Direct support from the PSNP where no labor is required was also evaluated. Pregnant and lactating women, and those with children suffering from acute malnutrition, qualify for temporary direct support from the PSNP with some conditionalities of antenatal care and nutrition education. Data was transferred into SPSS Statistics, version 24 (IBM Corp, Armonk, NY, USA) for analysis. Descriptive characteristics are given as proportions (%) or means for categorical and continuous variables, respectively. Bivariate analysis of ASF consumption with independent variables was conducted using Pearson’s Chi Square test of independence for categorical variables, and independent samples t-test and Wilcoxan rank-sum test for continuous variables. Multivariable logistic regression was used to calculate adjusted odds ratios (OR) and 95% confidence intervals (CI). Two adjusted models are presented. The first is adjusted for age to remove the effect of increased consumption of foods, including ASF, resulting from potential age differences between groups. The second adjusted model includes age, livelihood, religion, and the food aid/safety net variable. The model building strategy was similar to that outlined by Hosmer and Lemeshow [44]. The initial selection of variables was based on a priori thought of their relation to the outcome and their availability in the data. The initial multivariable model included all effects that were associated with the outcome in unadjusted analysis at p < 0.25, then those that were not significantly improving the model (p > 0.05 based on likelihood ratio tests) were removed until a final parsimonious model was reached with stable parameter estimates. Variables that had unadjusted significant associations with ASF but had been removed were re-entered into the final model to check again for significant contribution in the smaller model. None were retained at this point. Child sex was not included in the final model since it was not associated with ASF consumption, but when it was included, it did not alter the results. Interactions among the variables in the final model were not tested due to the small cell sizes that would result. The final model did not show indication of poor fit, based on Hosmer and Lemeshow’s test for goodness of fit using deciles (p = 0.6736) [44]. All adjusted ORs from the final model are presented regardless of statistical significance. Missing values were limited for the majority of variables used and were excluded from analyses that used those variables but were not excluded across all other analyses in order to preserve sample size. Imputation was not used for multiple logistic regression.