Factors associated with anaemia among women of reproductive age in Ethiopia: Multilevel ordinal logistic regression analysis

Abstract Anaemia has prevailed as a mild to severe public health problem in Ethiopian women of reproductive age. Many studies carried out on anaemia have been limited to subnational assessments and subgroups of women. The effects of potential factors thought to affect anaemia and severity levels of anaemia have not been well considered. Therefore, this study identifies individual, household and community level factors associated with anaemia among women of reproductive age in Ethiopia applying multilevel ordinal logistic regression models. Proportional odds assumption was tested by likelihood ratio test. About 35.6% of the variation on anaemia was due to between household and community level differences. Pregnancy (adjusted odds ratio [AOR] = 2.30, 95% confidence interval [CI]: 1.82, 2.91), HIV (AOR = 2.40, 95% CI: 1.76, 3.25), giving birth once (AOR = 1.2, 95% CI: 1.05, 1.40), giving birth more than once (AOR = 1.4, 95% CI: 1.19, 1.71), living with five or more family members (AOR = 1.24, 95% CI: 1.05, 1.47), living in poorest households (AOR = 1.34, 95% CI: 1.2, 1.61) and rural area (AOR = 1.57, 95% CI: 1.28, 1.92) were associated with greater odds of more severe anaemia compared with their respective counter parts. Secondary and above education (AOR = 0.83, 95% CI: 0.70, 0.97) and use of pills, implants or injectable (AOR = 0.67, 95% CI: 0.59, 0.77) were associated with lower odds of more severe anaemia. Anaemia prevention and control programmes need to be strengthened for women living with HIV/AIDS and during pregnancy. Household poverty reduction and social protection services need to be strengthened and integrated in anaemia prevention and management activities in women.


| INTRODUCTION
Anaemia is a global public health problem affecting about one-third (32.8%) of all women of reproductive age (WRA), 40% of pregnant and 32.5% of nonpregnant women in 2016 (The World Bank, 2018). It is a severe public health problem in many countries of the world.
Defined as a condition characterized by a lower than normal concentration of haemoglobin (Hb) (World Health Organization, 2015), anaemia contributes to maternal and perinatal mortality. It doubles the risk of maternal death and accounts for 20% of all maternal deaths (Balarajan, Ramakrishnan, Özaltin, Shankar, & Subramanian, 2011;Black et al., 2008;Daru et al., 2018).
Iron deficiency is the most common underlying cause of anaemia contributing to more than half of all cases anaemia (Hay et al., 2017).
Other micronutrient deficiencies (e.g., folate and vitamin B12), acute and chronic diseases (e.g. malaria, hook worm) and inherited blood disorders that affect red blood cells (RBC) production or survival are other common causes of anaemia (Balarajan et al., 2011;Kassebaum et al., 2014). Risk factors of anaemia are many, contextual and complex. These factors operate at different levels of the hierarchy, i.e.
Both the risk and consequences of anaemia are more prominent in low-and middle-income countries (Balarajan et al., 2011). Africa was ranked the second highest in terms of anaemia (38.6%) and severe anaemia (1. 8%;World Health Organization, 2015). It is a moderate and severe public health problem in all WRA (38.6%) and pregnant women (54%) of the Sub-Saharan Africa (SSA), respectively (The World Bank, 2018).
The Ethiopian Federal Ministry of Health (FMoH) has been struggling to prevent anaemia focusing on pregnant women by providing iron (Fe) and folic acid, nutrition education, providing drugs for deworming, promoting sanitation and preventing and treating malaria (Federal Ministry Of Health, 2008). However, in the last 15 years, the trend of anaemia has remained inconsistent. Although the prevalence of all form anaemia decreased from 27% in 2005 (Central Statistical Agency [Ethiopia] and ORC Macro, 2006) to 17% in 2011 (Central Statistical Agency [Ethiopia] and ICF International, 2012), it increased to 23.63% in 2016. The prevalence of anaemia was higher among pregnant, breast feeding, rural women and women who have had six or more births. It was also more prevalent with decreasing educational level and household wealth (Central Statistical Agency (CSA) [Ethiopia] and ICF, 2016).
Recognizing it as a worldwide public health problem, the WHO target is set to reduce anaemia in WRA by 50% in 2025 (World Health Organization, 2014). Maternal nutrition is one of the top priorities of Health Sector Transformation Plans (HSTP). The prevalence of anaemia in WRA is among the outcome indicators of HSTP targets (Federal Ministry of Health, 2015). In settings with limited resource like Ethiopia, up-to-date and empirical evidence on severity levels of anaemia and their associated factors is essential to achieve such targets and give priority to those who are at risk of higher levels of anaemia (Neha, Moutish, Ramanan, & Purnima, 2015).
Many studies focusing on anaemia in women have been limited to subnational assessments and subgroups of the women, with most studies being carried out with pregnant women. Therefore, the results of these studies are not generalizable to all WRA. Furthermore, the effects of some potential factors like cooking fuel, khat chewing, alcohol consumption, lactation, community education and smoking have been given little attention and are not well understood. Although factors affecting anaemia are considered to operate in different levels, existing national studies used single level analysis techniques which may lead to incorrect estimation of parameters, standard errors and false conclusion of the association. Also, single level analysis cannot measure the contribution of the household and community clustering to the total variation of anaemia (Diez-Roux, 2000). To the investigators knowledge, the effect of the household level variability on anaemia has not been examined in the context of Ethiopia. The effect of anaemia differs depending on the severity levels (mild, moderate or severe; Riva et al., 2009). Accounting for the ordered nature of

Key messages
• Anaemia remains a public health concern in women of Ethiopia.
• Anaemia was found to be influenced by factors evolving at individual, household and community levels.
• There was a sizable heterogeneity in likelihood of anaemia across households and clusters.
• Pregnancy, HIV, giving birth once or more than once, living with five or more family members, living in poorest households and rural area were associated with greater odds of worse anaemia. Secondary and above education and use of pills, implants or injectable were associated with lesser odds of worse anaemia.
• The Ethiopia FMoH should strengthen anaemia prevention and control programmes for vulnerable WRA.
anaemia may reduce bias in parameter estimation, increase precision and power and consider as an efficient use of information available (Hedeker, 2015). Multilevel ordinal logistic regression (OLR) examines the effects of explanatory variables at different levels simultaneously.
It produces more accurate estimates of regression coefficients, standard errors, confidence intervals and significance tests compared with single-level logistic regression. Furthermore, it quantifies within and between groups variation on outcome which might be due to unobserved factors at different levels (Diez-Roux, 2000).
The aim of this study, therefore, was to identify individual, household and community level factors of anaemia in WRA simultaneously using multilevel OLR analysis technique. Additionally, the study aimed to quantify the amount of unexplained variation in odds of anaemia which is attributable to community and household clustering.

| Analytic samples
All WRA who had data on anaemia status were included in this study.
A total of 14,489 WRA who were tested for anaemia were included in the analysis.

| Anaemia
Anaemia is an ordered categorical variable categorized as none, mild, moderate and severe anaemia based on Hb level. Blood samples were taken from a finger prick of the voluntarily consented women and collected in a micro cuvette. Hb analysis was carried out on-site using a battery-operated portable HemoCue analyser (Central Statistical Agency [CSA; Ethiopia] and ICF, 2016). Hb levels were adjusted for pregnancy because during pregnancy the increase in maternal blood volume and the iron needs of the fetus decrease the blood Hb level (World Health Organization, 1989). It was also adjusted for smoking and altitude. The World Health Organization (WHO) Hb cut off points for diagnosis of anaemia are given in Table 1

| Predictors of anaemia
After reviewing recent literature, potential risk factors of anaemia were extracted from the data set. Due to the hierarchical nature of the 2016 EDHS data, the extracted variables were classified as individual, household and community level variables. Individual level variables were characteristics of the women which were specific to each woman (Table 2).  Table 3.
Community level variables were characteristics which are common for all women residing in the same community (cluster) and include place of residence, region, community (cluster) women education, community poverty, community women unemployment and community mass media exposure. Variables like community women education, community poverty, community women unemployment and community mass media exposure were generated by aggregating individual characteristics within the cluster. The generated variables were further categorized as low or high based on the national median values of the generated variables. These variables are measured as shown in Table 4. OLR is a well-suited technique to this study because of the ordered nature of outcome variable (none, mild, moderate and severe anaemia; Hedeker, 2015).
Stata software version 14 was used for analysis of the data. A P-value ≤ 0.25 in bivariate analysis was used to consider candidate variables for multivariable analysis (Stoltzfus, 2011). In a multivariable analysis, a P value < 0.05 was used to identify variables significantly associated with anaemia. Adjusted odds ratios with 95% confidence intervals were estimated and interpreted (Raman & Hedeker, 2005). The proportion of variations in odds of anaemia between households and communities was expressed using variance partition coefficients (VPC). The VPC measures the proportion of outcome (anaemia) variation unexplained by the predictor variables that lies at each level of the model hierarchy. It measures the relative importance of clusters, households and individual (women) as sources of variation on anaemia status (Leckie & French, 2013).

| Model specifications
The mixed-effects OLR (proportional odds) model can be written in terms of the cumulative logits as below in the box: c-is a model threshold or intercept for C-1 level of anaemia, and it is a fixed parameter.
It represents the cumulative logits of being at or below C-1 level of anaemia when the covariates and random effects equal to zero. It is strictly increasing (i.e., γ1 < γ2 < Á Á Á < γC − 1).

C = number of categories of anaemia which equals to 4.
β-is a coefficient (fixed effect of explanatory variable).
X ijk -is a covariate vector for kth individual in jth household and ith cluster.
u ij -is level-2 (household) random effect, and it is assumed to be normally distributed with variance σ 2 (v2).
Violation of the proportional odds assumption is common. In such occasions, a model which relaxes the assumption is nonproportional or partial-proportional odds model in which covariates are allowed to have different effects on the C − 1 cumulative logits. It is given as below in the box: u ijk -is a covariate vector for set of variables for which proportional odds is not assumed.
α c -is a vector of regression coefficients associated with these covariates for C-1 levels of outcome.
Because α c carries the c subscript, the effects of these covariates are allowed to vary across the C − 1 cumulative logits.
T A B L E 1 The WHO Hb cut off points for diagnosis of anaemia

Anaemia status
Hb cutoffs for pregnant women (g/dl) Hb cutoffs for NP-NL (g/dl)
VPC (3) is a proportion of total variation on anaemia attributable to cluster random effect.
It is given as in the box: ð Þ+ π 2 =3 , where π 2 /3 is individual level variance which equals to 3.29.

of 15
bs_bs_banner VPC for level-2 and 3 clustering effects (VPC (2 + 3) ) is a proportion of total variation on anaemia attributable to both household and cluster level random effect. It is given as below (Leckie & French, 2013).

| Proportional odds assumption
The proportional odds assumption states that the effects of all covariates are constant across categories of outcome variable. After fitting both proportional and nonproportional odds models, the proportional odds assumption was tested using likelihood ratio test. It tests the null hypothesis that there is no difference in the effects of explanatory variables across the levels of anaemia. The P value ≥ 0.05 is desirable to retain null hypothesis (Bauer & Sterba, 2011). The likelihood ratio test supported the nonproportional odds assumption. Furthermore, each variable in the model was tested to identify the variables for which the proportional odds assumption was violated.

| Model fit statistics
An Akaike information criterion (AIC) was used to select the final model which fits the data best compared with other fitted models.
The AIC of the all models were compared, and the model with the lowest AIC was considered as the best model fits the data (Hox, Moerbeek, & Van De Schoot, 2010; Table 6).

| Significance test of random effects
For a substantial number of clusters and households, the 95% confidence intervals of random intercepts do not overlap zero. This implied that the random effects of the many households and clusters on anaemia were significantly different from zero (above or below zero; Leckie & French, 2013).
2.6.6 | Model adequacy (normality of random effects distribution) The normality assumptions were tested graphically using quantilequantile plots (Leckie & French, 2013

| Ethical considerations
The permission for access to the data was obtained from ICF International by registering and stating the objective of the study. The data set has no individual names or house hold addresses. The data were 3 | RESULTS Table 5 shows individual, household and community level characteristics of the study population. The mean ± standard deviation (SD) age of the respondents was 28 ± 9 years. More than half (62.20%) and 1,055 (7.30%) of the women were neither pregnant nor lactating and pregnant, respectively. More than a quarter (25.70%) of the respondents was current contraceptive users. The prevalence of HIV/AIDS was 1.21% (175). A total of 3,425 (23.63%) women were anaemic.

Model 1 (null model)
It is the intercept only model without independent variable. It provided the variance components of random effects which helped to determine variation between households and clusters to favour multilevel analysis over a standard ordered logistic regression.
This model showed that more than one-third (VPC (2 + 3) = 35.64%) of the total variation in anaemia was due to unobserved household and community level factors. About a quarter (VPC (3) = 24.09%) and 11.56% (VPC (2) ) of the variation was attributable to community and household level unobserved factors, respectively. This quantity of VPC (2 + 3) was a suggestive of using multilevel model rather than single-level model. Furthermore, the intrahousehold clustering (VPC = 11.6%) was a suggestive of using three-level model rather than two-level which ignores the household level variability (controlling for intrahousehold and intracluster variability). Models 2, 3 and 4 are models adjusted for individual, household and community level variables, respectively (Table 6).

Model 5: Adjusted for individual, household and community level factors together
This model showed that 18% (VPC (2 + 3) ), 9% (VPC (3) ) and 9% (VPC (2) ) of the unexplained variation in anaemia could be attributable to unobserved community and household level factors together, community level factors alone and household level factors alone, respectively.
The PCV 3 showed that 70.73% of the cluster level variation was explained by the variables in the model. This suggested that, after controlling for the variables in the model, there are still significant factors of anaemia at household and community levels which are not accounted for. Model 5 was found to be the best fit model for the data (Table 6).  (Table 7).

| DISCUSSION
The main aim of this study was identifying multilevel (individual, household and community level) factors associated with anaemia among women in Ethiopia. Accordingly, the study found that anaemia was influenced by factors evolving not only at individual level but also at household and community (cluster) levels factors too. There was a sizable heterogeneity in a likelihood of anaemia across households and clusters. This heterogeneity implied that the entire variation in anaemia might not be fully explained by individual level factors alone.
This could be due to blood loss, decreased and ineffective RBC production and increased RBC destruction associated with comorbidities, medications and micronutrient deficiencies in HIV positive women (Volberding et al., 2004). This is in contrast to another study in Ethiopia (Bekele, Tilahun, & Mekuria, 2016) where HIV status was not associated with anaemia. This might be due to the possibility that HIV positive women in the later study might be in the early stage of the disease or properly screened and treated for anaemia, and they might had high adherence to antiretroviral drugs and other interventions.
Pregnant women had higher odds of anaemia than pregnant or lactating women. Furthermore, its effect was stronger for moderate and severe anaemia than for mild and above anaemia. Findings from Tanzania (Wilunda et al., 2013), Malawi (Adamu et al., 2017) and Ethiopia (Ejigu et al., 2018) also supported the observed association.
The likely explanations to this association could be the increased risk of infections and obstetric complications during pregnancy that leads to blood loss and under nutrition (Balarajan et al., 2011;Ebuy, Alemayehu, Mitiku, & Goba, 2017). Use of pills, implants or injectable was negatively associated with anaemia. This is consistent to the findings from Tanzania (Wilunda et al., 2013), Bangladesh (Kamruzzaman et al., 2015) and Ethiopia (Asres, Yemane, & Gedefaw, 2014;Lakew, Biadgilign, & Haile, 2015). The noncontraceptive benefits of the hormonal contraceptives like iron content of the pills, prevention of the heavy menstrual bleeding and regulating menses could explain the low prevalence of anaemia among contraceptive users (Brant, Ye, Teng, & Lotke, 2017).
Women with one or more births in the last 5 years before survey had greater odds of being at higher levels of anaemia than women with no births in the last 5 years. This finding is in agreement with the results from studies in India (Perumal, 2014) and Ethiopia (Bekele et al., 2016). This could be due to the possibility that narrow birth Women with secondary and above education were less likely to be at higher levels of anaemia as compared with noneducated women. This is in agreement to the findings from studies in China (Ma et al., 2017), Bangladesh (Kamruzzaman et al., 2015) and Ethiopia (Ejigu et al., 2018). There is evidence that women with secondary and higher education have higher health seeking behaviour and service utilization than noneducated women so that they get preventive and curative services for conditions that contribute for anaemia (Bobo, Yesuf, & Woldie, 2017;Wonde & Tadele, 2015). However, this finding is in conflict with another study in Ethiopia where education was not associated with anaemia (Kefiyalew, Zemene, Asres, & Gedefaw, 2014). This might be due to the reason that the latter study has controlled for factors which might be more proximal factors of anaemia than education.
Wealth index of the household was positively associated with anaemia in women. This is supported by the findings in China (Siddiqui et al., 2017) and Bangladesh (Kamruzzaman et al., 2015).
Inability to afford micronutrient-rich food which is expensive in Ethiopia and lack of priority to women in poor households could make them more vulnerable to anaemia than women of richest households.
There is also evidence that poor households are poor in seeking and timing health care so that they might not be treated for diseases that may result in anaemia (Bobo et al., 2017;FMOH, 2008;Mebratie et al., 2014). This finding disagrees with the finding in another study done in Ethiopia (Ejigu et al., 2018) where wealth index was not associated with anaemia. This might be due to the increase in the prices of foods and health services over time that widens the difference between richest and poorest in terms of affording foods and health services in Ethiopia.
Women who were living with five or more family members had greater odds of higher levels of anaemia as compared with those living with two or less family members. This is in agreement with findings in Ethiopia (Bekele et al., 2016). This could be due to the possibility that large family size might results in household food insecurity which compromises women's access to balanced diet. In addition, traditionally women might be expected to serve their family first and eat the leftover if any at the end.
Rural women had greater odds of higher levels of anaemia as compared with urban women. The same evidence was reported in studies from Ethiopia (Ebuy et al., 2017;Ejigu et al., 2018;Kefiyalew et al., 2014). Due to poor health seeking behaviour and service utilization associated with inadequate infrastructure, rural women might not get care for maternal services and disease which may contribute to anaemia (Begashaw, Tessema, & Gesesew, 2016;Bobo et al., 2017).
This finding disagrees with a study conducted in Malawi (Adamu et al., 2017) in which place of residence was not related to anaemia.
The difference in health service utilization and food access and sociocultural differences between the countries might explain the disagreement between studies. Unlike Ethiopia, national estimates of the maternal health service utilization were higher, and there was little difference between urban and rural women in terms of maternal health service utilization in Malawi (Malawi, 2015). Being a moderate public health problem also suggests the coverage and consistency issue of the current intervention methods need to be considered. Monitoring and evaluation of programmes is inadequate in many countries. Anaemia prevention programmes have challenges which include poor attendance at antenatal clinics or insufficient emphasis on behavioural aspects of taking supplements on a regular basis and consuming diverse diets which have limited their effectiveness (World Health Organization, 2016). Revision of existing programmes would be important to provide evidence on challenges and methods of improving their effectiveness.

| Limitation of the study
Cross-sectional nature of the study which cannot show the temporal relationship between risk factors and anaemia. Use of secondary data and inability to include variables like hookworm infection, malaria, schistosomiasis, menorrhagia, chronic diseases, dietary intake, cultural taboos and body mass index might made difficult to identify the independent effects of the considered variables. Recall bias might be introduced during measuring some events in the past.

| CONCLUSIONS
Anaemia was found to be influenced by factors evolving at individual, household and community levels. At individual level, HIV infection, pregnancy, increased number of births and living with husband were positively associated with anaemia. However, secondary and above education and use of contraceptive pills, implants or injectable were negatively associated with anaemia. At household level, living in households with large family members and with poorest, poorer and middle wealth index was positively associated with anaemia. Living in rural areas, in Afar, Somali and Harari regions and Dire Dawa city was positively associated with anaemia. There was a sizable heterogeneity in likelihood of anaemia across households and clusters which is unaccounted for by the predictor variables.
Daily iron supplementation during pregnancy is recommended as part of the standard care by WHO (World Health Organization, 2012).
A daily iron supplementation to women during pregnancy was found to reduce anaemia at term by 70% (Bhutta et al., 2013;Peña-Rosas, De-Regil, Garcia-Casal, & Dowswell, 2015). Countries like Indonesia and Sweden have effectively controlled anaemia in pregnant women (Suharno, Karyadi, West, & Hautvast, 1993;World Health Organization, 2001). The Ethiopia FMoH should strengthen anaemia prevention and control programmes for WRA living with HIV/AIDS and during pregnancy. To tackle the anaemia among poor, household poverty reduction and social protection services should be strengthened and integrated in anaemia prevention and controlling programmes in women. Information on dietary intake of women and indicators of infections need to be incorporated during DHS, and further study is needed to determine additional factors.