Data for this study were extracted from the longitudinal ENGINE Birth cohort study that was implemented from 2014 to 2016 in three woredas of Oromiya region in Ethiopia, two of which were part of the USAID ENGINE program while the third was not. The sample size for the Birth Cohort study was estimated at 4680 with 1560 women recruited in each woreda to allow for a comparison between woredas. Due to missing data for a few independent variables, 4600 women were included in our analysis sample. For the main study, the sample size calculation was based on the outcome of height for age Z-score. For this analysis, where the outcome is anemia, this sample allows us to detect an effect of 0.018 change in odds of anemia with 80% power at the 0.05 level of significance.
Pregnant women ages 14 to 50 years old were recruited from the three woredas from a total of 78 kebeles (N = 4680). Data were collected twice during pregnancy, at birth, and then every 3 months until the child reached 12 months of age. The data used in this study were obtained through surveys and assessments administered to the pregnant women during the time of recruitment. In addition, the present study used data collected from the survey administered to the household head at recruitment. Data was collected electronically through a tablet using Open Data Kit.
Data calculations and variable definitions
Hemoglobin levels were measured with the HemoCue® system for mobile screening. Hemoglobin cutoff values were adjusted for altitude and trimester according to method described by Cohen and Hass . The average kebele altitude was substituted for missing values for altitude. Women with hemoglobin levels below the adjusted cutoff point were classified as anemic. A binary variable for anemia status was used as the outcome variable in the analysis.
Marital status was categorized into 3 groups: married and monogamous, married and polygamous, and not married (single, cohabitating, separated, divorced, widowed). Most of the households recruited were Muslim or Orthodox (90%), thus religion was categorized as Muslim, Orthodox, or other. Women were classified as literate if they could read specific sentences in Oromifa and numerate if they correctly answered a simple math problem (i.e. "If you sell eggs for 30 Birr and chicks for 50 Birr, how many Birr do you have?"). A wealth index was constructed using polychoric principal component analysis to represent a composite measure of a household’s cumulative living conditions and then separated into quintiles. This method was the same as described by the Demographic and Health surveys for Ethiopia .
Mid-upper arm circumference (MUAC) was used as measure of nutritional status. The average of three MUAC measurements was calculated and then categorized as normal or low MUAC. A MUAC measurement less than 23 cm was classified as low MUAC . The variable for antenatal care visits was coded as a binary variable for whether they have sought antenatal care (ANC). Alternatively, or in addition to clinic visits, some women may have been visited at home by a health extension worker. A binary variable was created for whether they received any home visits from health workers in the past year. Iron supplementation was coded as a binary variable, which is defined as the receipt or purchase of iron supplements during the current pregnancy. Likewise, a binary variable was created for receiving treatment for intestinal worms during the current pregnancy. Whether a woman has had previous pregnancies was coded as a binary variable (0 = first pregnancy, 1 = previous pregnancies). A handwashing score was computed using seven self-reported questions about the critical times for hand washing (when dirt is visible, after toilet use, after cleaning a child following defecation, before preparing food, before serving a meal, before eating, before feeding a child).
Minimum dietary diversity scores (MDD-W) were constructed from a 24-h qualitative recall as a proxy indicator for nutrient adequacy of the diet . Foods were grouped into the following categories: all starchy staple foods, beans and peas, nuts and seeds, dairy, flesh foods, eggs, vitamin A-rich dark green leafy vegetables, other vitamin A-rich vegetables and fruits, other vegetables, and other fruits for a maximum score of 10. Household food insecurity access scale (HFIAS) score was constructed using the method described by Coates et al. .
Crop production diversity was calculated as a simple count of the crop groups produced annually by the household. Crops were grouped as cereals, roots and tubers, legumes, cash crops, vegetables, fruits, oil seeds, and spices for a maximum score of 8. Livestock production diversity was created as a count of products from livestock. The score was constructed from the following products: beef, milk, butter, cheese, cattle hides, cattle manure, yogurt, sheep meat, wool, sheep hides, sheep manure, goat meat, goat milk, goat hides, goat manure, eggs, bird manure, honey, wax, and propolis for a maximum score of 20.
Because recruitment occurred on a rolling basis, it was necessary to control for lean season. Months of adequate household food provisioning (MAHFP) were used to define the lean season. In our sample, the highest MAHFP scores, which signify the highest levels of food insecurity, occurred between June–September. A binary variable for lean season was created based on recruitment during those months. The variable for market access was defined as minutes to the nearest local or major market.
Data were analyzed using Stata Corp 2013, StataSE 14 software. Descriptive statistics included generation of means and standard deviations along with bivariate analysis before variables were included in the model. A multivariate logistic regression analysis was conducted to ascertain the factors associated with being anemic in this population. The dependent variable was a binary variable of presence or absence of anemia (prevalence) while the independent variables included in the model were lean season, presence of low MUAC, previous pregnancy, trimester, number of antenatal visits to the clinic (by the pregnant woman), number of health worker home visits, use of iron supplementation, use of deworming, handwashing score, age, market access, HFIAS, wealth quintile, minimum dietary diversity score, crop and livestock production diversity, woman’s literacy and numeracy. The model was adjusted for clustering at the kebele level, woreda (this also controlled for presence or absence of ENGINE as an intervention) and religion and includes robust standard errors using the vce command in Stata. A p-value of less than 0.05 was considered as a statistically significant result.
After the preliminary model, multiple iterations were tested including the removal of insignificant variables and addition of interaction terms. However, neither the inclusion of interaction terms nor the removal of insignificant variables improved the model. Furthermore, the presence of interaction terms worsened the model as determined by Akaike’s and Schwarz’s Bayesian information criteria (AIC/BIC) using the estat ic command in Stata for assessing the model fit. Adjusted odds ratios and 95% confidence intervals are reported.