Study design and sample size
The study was further analysis of the 2011 Ethiopian Demographic and Health survey (EDHS) data. EDHS is periodical survey with five years interval, sometimes different from five with special cases. The 2011 EDHS was conducted on a nationally representative sample of nine regions and two city administrations of the country. It was conducted from September 2010 to January 2011 to provide current and reliable data on fertility and family planning behavior, child mortality, adult and maternal mortality, children’s nutritional status, use of maternal and child health services, knowledge of HIV/AIDS, and prevalence of HIV/AIDS and anemia. Samples were selected using stratified two-stage cluster design technique taking census enumeration areas as the sampling units. In the first stage, 624 clusters of enumeration areas were selected from the list of the 2007 Population and Housing Census sample frame. A total of 17,817 representative households were selected for the 2011 EDHS. Hemoglobin was measured onsite using battery operated portable Hemacue analyzer from finger prick for all children aged 6–59 months, and women aged 15–49 years. The detail of the methodology is available in the EDHS 2011 report [9]. Only households having children aged 6–59 months were considered for this study. Data of 7636 children aged 6 to 59 months with complete information on the selected predictors of childhood anemia status were used.
Variables and source of data
The 2011 EDHS data were obtained from Central Statistical Agency (CSA), Addis Ababa, Ethiopia. The outcome variable in this study was anemia status of children aged 6 to 59 months categorized into three: severe or moderate, mild, and non-anemic. Anemia status was determined based on hemoglobin concentration in blood adjusted to altitude. Adjusted concentration 10.0–10.9 g/dl was considered as mild anemia, 7.0-9.9 g/dl as moderate anemia and less than 7.0 g/dl as severe anemia.
This study tried to include the most important expected determinants of anemia from various literature reviews [3, 5, 10–19], and their theoretical justification from the source of data [9]. The explanatory variables at individual and household levels included were child’s size at birth, sex of child, child’s age, stunting status of child, wasting status of child, mother’s educational level, husband/partner’s educational level, mothers’ anemia status, mothers’ age, mothers’ marital status, mothers’ current employment status, place of residence, religion of child, source of drinking water, number of under five years old children in the household and child’s birth order.
Method of analysis
Ordinal logistic regression model was employed because of child anemia status is ordered. Specifically, proportional odds model (POM) was employed because of the following appealing features: (a) it is invariant under several categories as only the signs of the regression coefficients change when the coding of the response variable are inverted [20, 21]; (b) it is invariant under collapsibility of the ordered categories as the regression coefficients do not change when response categories are collapsed or the category definitions are changed [22]; and (c) it produces the most easily interpretable regression coefficients as exp(−β) is the homogenous odds ratio (OR) over all cut-off points summarizing the effects of the explanatory variables on the response variable in a single frequently used measure [20].
The POM for the categorical variable Y with C ordered categories and a collection of P explanatory variables for the l
th subject \( {\boldsymbol{X}}_{\boldsymbol{l}}^{\mathbf{\prime}}=\left({x}_{1\boldsymbol{l}},\;{x}_{2\boldsymbol{l}}, \dots,\;{x}_{p\boldsymbol{l}}\right),\;\boldsymbol{l}=1,\;2, \dots,\;n \) is given as:
$$ \begin{array}{c} logit\left[{Y}_{\boldsymbol{l}}\le\ i\Big|{x}_{\boldsymbol{l}}\right]= \log \left[\frac{\pi_i\left({X}_{\boldsymbol{l}}\right)}{1-{\pi}_i\left({X}_{\boldsymbol{l}}\right)}\right]\ \\ {}={\alpha}_i-{\beta}_1{x}_{1\boldsymbol{l}} - \dots -{\beta}_p{x}_{p\boldsymbol{l}}={\alpha}_i-{\boldsymbol{X}}_{\boldsymbol{l}}^{\hbox{'}}\beta \\ {}for\ i=1,2,\dots, c-1;l=1,\ 2, \dots,\ n\end{array} $$
where π
i
(X
l
) = Pr (Y
l
≤ i|X
l
) and β is a column vector of P regression coefficients and α
i
is i
th intercept coefficient.
After the best model has been chosen, test of parallelism was assessed. A non-significant chi-square test of parallelism was taken as evidence that the logit surfaces are parallel and that the odds ratios can be interpreted as constant across all possible cut-off points of the outcome variable.