Multilevel multinomial logistic regression model for identifying factors associated with anemia in children 6–59 months in northeastern states of India

Public Interest Statement

A multilevel multinomial logistic regression model was considered to identify potential risk factors of anemia in children aged 6–59 months in northeastern India. Level of anemia was considered as the outcome variable with four ordinal categories (severe, moderate, mild, and non-anemic) based on hemoglobin concentration in blood. A two-level random intercept model was used with state of residence as the level-2 variable. Considering the hierarchical nature of the data, we did not find place of residence (urban vs. rural), sex of child, or household living standard (low, medium, and high SES) to be significant. Rather, mother’s age at marriage and the number of children ever born were significantly associated with anemia. Our findings suggest that as the child becomes 48 months or older, likelihood of severe anemia decreases. This study throws light on the fact that anemia intervention needs to focus more at state level followed by individual level.

1. Introduction

Anemia is considered the most prevalent nutritional deficiency globally (McLean, Cogswell, Egli, Wojdyla, & de Benoist, Citation2009). About one-third of the global population (over two billion) is anemic (de Benoist, McLean, Egll, & Cogswell, Citation2008) while the prevalence of anemia among preschool children is 47.4% (IIPS, Citation2000). According to WHO estimates, India has the highest prevalence of anemia among the South Asian countries in all age groups. The Third National Family Health Survey (IIPS, Citation2007) of India revealed that 70% of the children are anemic in the age group of 6–59 months, including 3% severely anemic, 40% moderately anemic, and 26% mildly anemic (IIPS, Citation2000). As per five major surveys (National Family Health Survey (NFHS Assunção, Santos, de Barros, Gigante, & Victora, Citation2007; de Benoist et al., Citation2008), District Level Household Survey 2 (DLHS), Indian Council of Medical Research (ICMR) Micronutrient Survey (Togeja & Singh, Citation2004) and Micronutrient Survey (NNMB, Citation2004), over 70% of preschool children are anemic in the country.

Iron deficiency (ID) is listed as one of the “top ten risk factors contributing to death” (Dubey, Citation1994). Iron deficiency anemia (IDA) is prevalent in South Asia, predominantly India, Bangladesh, and Pakistan. However, the prevalence of IDA has declined substantially in neighboring countries such as Bangladesh and Pakistan (Gillespie, Citation1997), whereas it has plummeted from 20 to 8% within a decade in China (Lokeshwar, Manglani, Rao, Patel, & Kulkarni, Citation1990). Studies carried out in Egypt (Seshadri, Gopaldas, Walter, & Heywood, Citation1989), India (Soemantri, Gopaldas, Seshadri, & Pollitt, Citation1989), Thailand (Pollitt, Hathirat, Kotchabhakdi, Missell, & Valyasevi, Citation1989), and the USA (Pollitt, Citation1991) revealed that iron and folate deficiency anemia reduce the learning capacity, attentiveness, and intelligence of children aged below five years. Several studies have been carried out on anemia in India since 1980. However, very few studies focused on child anemia at the national and regional levels (Bharati, Pal, Chakrabarty, & Bharati, Citation2015; Dey, Goswami, & Dey, Citation2013; Singh & Patra, Citation2014). The risk factors of anemia most often cited in the literature are: low family income and low maternal level of education, lack of access to health care services, inadequate sanitary conditions, and a diet with poor quantities of iron (Oliveira, Osório, & Raposo, Citation2007; Osório, Lira, & Ashworth, Citation2004) .

Nutritional problem is very common among children below five years in all the states of India, particularly, in those states whose performances are very poor in respect of demographic and socioeconomic indicators. There are several factors responsible for IDA, for example: low dietary intake, inadequate iron (less than 20 mg/day) and folic acid intake (less than 70 mg/day), and chronic blood loss due to infection such as malaria and hookworm infestations, among others (NNMB, Citation2004; Togeja & Singh, Citation2004). A recent study on the determination of socioeconomic and demographic determinants of anemia among Indian children aged 6–59 months found an increasing trend of prevalence of anemia up to two years of age and then decreased thereafter (Goswmai & Das, Citation2015). Therefore, it is imperative to find out the factors that contributed to anemia and to examine the contribution of existing programs in combating child anemia, especially in the less developed areas such as northeastern states of India. In view of the above, our objective is to determine the prevalence of anemia among children (6–59 months) from the northeastern states of India. To comprehend the prevalence of anemia and to identify the significant predictors, socioeconomic differentials have been taken into consideration in our study.

2. Methods

The present paper uses the data-set from the 2005–2006 National Family Health Survey (released in 2008) for the northeastern states of India. Respondents were selected through a systematic multistage stratified sample survey conducted in all 29 states of India (IIPS, Citation2007). In each state, populations were stratified by urban and rural area of residence, and the sample size at the state level was proportional to the size of the state’s urban and rural population. A uniform sample design was adopted in all states. In each state, rural sample was selected in two stages: at the first stage, primary sampling units (PSUs), which are villages, were selected with probability proportional to population size (PPS). At the second stage, households within each PSU were selected randomly. In urban areas, a three-stage sampling procedure was adopted. At first, wards were selected with PPS sampling. In the second stage, one census enumeration block (CEB) was randomly selected from each sample ward. In the final stage, households were randomly selected within each sample CEB. The study population constitutes a nationally representative cross-sectional sample of singleton children aged 0–59 months and born after January 2000 or January 2001 (n = 50,750) to mothers aged 15–49 years from all 29 states of India. Information on children was obtained by a face-to-face interview with mothers, with a response rate of 94.5%. The survey also provided district-level information on the prevalence of under-nutrition [weight-for-age, using the standard deviation (SD) classification] among children in the age group of 0–59 month(s); prevalence of anemia (Hb estimation by HemoCue Hb201 + analyzer, Angelholm, Sweden) for the children aged 6–59 months, adolescent girls aged 10–19 years, and pregnant women; household availability of iodized salt; and the coverage of vitamin A program, with appropriate dosage. To meet the objectives of the study, we have produced a data-set that pertains to the northeastern states of India. Our study comprises 10,136 children within 6–59 months of age.

3. Variables

The variables of the study are briefly described in the following.

4. Objective of the study

Our objective is to study the relationship between some household-level predictors and some maternal variables on the probability of being at or below a hemoglobin level. Based on our classification of anemic children, we calculate odds ratio and the predictive probability of being at or below a hemoglobin level based on a polytomous logistic regression model.

In particular, we like to answer the following research questions:

(1)	What is the likelihood of being at or below each level of hemoglobin in the blood (g/dL) for children at a typical household?
(2)	Does the likelihood of being at or below each level of hemoglobin (i.e. anemia level) vary across state of residence?
(3)	What are the relationships between the household and maternal variables, and the likelihood of a child being at or below a given hemoglobin level?

5. Statistical analysis

In the present study, we employ bivariate as well as multivariate techniques to identify the factors that are associated with anemia in children of 6–59 months of age. In the bivariate analysis, cross-tabulation was made between the potential risk factors and the presence of anemia. Pearson’s chi-square test and p-values were used to test for the significance of each of the potential risk factor in bivariate analysis. Further analyses were carried out to study the relationship of the potential risk factors with the hemoglobin level in a multivariate setup. We use multinomial multilevel logistic regression model to predict the level of anemia as a function of mother’s age at marriage, number of children ever born to mother, religion, literacy of mother, household living standard, place of residence, sex of the child, and age of the child (in months). The wealth index was constructed using household asset data. Each household asset was assigned a score generated through principal components analysis and the resulting scores were standardized and summed up for each household (IIPS, Citation2007).

Due to the stratified nature of data in NFHS, the children are naturally nested into mothers, mothers are nested into households, households into PSUs, and PSUs into the states. Hence, keeping in view of the hierarchically clustered nature of the survey data, we use multilevel regression model to avoid possible under-estimation of parameters from a single-level model (Griffiths, Matthews, & Hinde, Citation2002). The advantage of multilevel model is that they properly account for the correlation structure of the data that frequently occurs in social sciences and in multistage survey sampling.

To properly account for the hierarchical nature of the data, we consider state of residence as the level-2 variable under which the respondents are nested. Thus, respondent-level variables are the level-1 variables used in this study to predict the level of anemia in children. Our outcome variable is ordinal polytomous with four levels for anemia. The model for such data uses a multinomial distribution and cumulative logit link function to compute the cumulative odds for each category of the response (O’Connnell, Goldstein, Rogers, & Peng, Citation2008).

For the polytomous response variable, let us denote ${\mathit{\eta }}_{kij}$ as the log odds of a child in the ith household in the jth state being at or below kth level of hemoglobin in blood. The model for level-1 may be written as:(1) $${\mathit{\eta }}_{kij}=log\left(\frac{P(R_{ij}\le k)}{1-P\left(R_{ij}\le k\right)}\right)={\mathit{\beta }}_{0j}+{\mathit{\beta }}_{1j}{X}_{ij}+{\mathit{\delta }}_k$$(1)

Here, ${\mathit{\beta }}_{0j}$ is the intercept, assumed to be random, represents the average log odds of being at or below severely low hemoglobin level; $P\left(R_{ij}\le k\right)$ represents the probability of responding at or below kth level of the outcome variable; X_ij represents level-1 predictor ${\mathit{\beta }}_{1j}$, the slope coefficient corresponding to X_ij which measures the change in the probability of being at a given hemoglobin level per unit change in the level-1 predictor; and δ_k epresents the difference between the kth category and the preceding one. Essentially, for k = 1, we have δ₁ = 0.

We consider the intercept to be random having the following model: $${\mathit{\beta }}_{0j}={\mathit{\gamma }}_{00}+u_{0j},$$

where ${\mathit{\gamma }}_{00}$ is the log odds of severe anemia relative to no anemia when the predictor variables in the model are evaluated at zero, and u_0j is the random error term with $u_{0j}\sim N\left(0,{\mathit{\tau }}_{00}\right)$ where $u{\mathit{\tau }}_{00}$ is the level-2 error variance.

Now, substituting the expression for β_0j in equation (McLean et al., Citation2009), we get(2) $${\mathit{\eta }}_{kij}=log\left(\frac{P(R_{ij}\le k)}{1-P\left(R_{ij}\le k\right)}\right)={\mathit{\gamma }}_{00}+{\mathit{\beta }}_{1j}{X}_{ij}+{\mathit{\delta }}_k+u_{0j}$$(2)

This model represents a scenario with no level-2 predictor (as in our case), and one level-1 predictor. However, in our study, we have more than one level-1 predictor, and so the model is meant to be augmented by those additional covariates as they are considered in the analysis.

Model (2) gives log odds when fitted to data. Alternatively, predicted probability of the event of interest (e.g. being at or below a hemoglobin level) can be calculated using the following formula as discussed in (Ene, Leighton, Blue, & Bell, Citation2014):(3) $$\text{Predicted}\text{probability}\text{(PP)}{\mathit{\theta }}_{kij}=\frac{e^{{\mathit{\eta }}_{kij}}}{1+e^{{\mathit{\eta }}_{kij}}}$$(3)

This is simply a conversion of log odds of the event of interest to probability of the event of interest in order to provide a meaningful interpretation of the numbers. In this expression, θ_kij is the probability of the event (e.g. being at or below a hemoglobin level), 1 − θ_kij is the corresponding probability of being above a given hemoglobin level, η_kij represents the log odds of the event of interest, and e is approximately 2.72.

We used PROC GLIMMIX in SAS 9.4 for multinomial distribution, and CLOGIT link function. To the best of our knowledge, there is no model diagnostic procedure available under PROC GLIMMIX for multinomial model. We considered several candidate models, and the best model was selected based on −2 log likelihood as a criterion. Details are discussed later in the paper.

6. Results

7. Discussion

Detection of the risk factors is detrimental to planning and implementation of programs to eradicate child anemia, especially in those groups where prevalence is very high. Our analysis suggests that childhood anemia is highly prevalent in the northeastern states of India. The overall prevalence of anemia in children in the northeastern states is 53%. There are important and significant relationships between anemia and some of the selected potential risk factors. Bivariate analysis suggests that there is no significant association between anemia level and household living standard, sex of child, literacy of mother, total children ever born to mother, and age of the child (months), whereas results from multivariate analysis show that age of the child, mother’s age at marriage, and number of children ever born have significant effects on child anemia. We also found statistically significant variation across states for the chance of being at or below a certain hemoglobin level. The intra-class correlation coefficient, ICC = 0.0577, indicates approximately 6% of the total variation in the response variable explained by the state of residence. This is a sizeable contribution in explaining the total variability in the response variable.

In contrary to our findings, several other studies (le Cessie et al., Citation2002; Unsal, Bor, Tozun, Dinleyici, & Erenturk, Citation2007) indicated that male children were at greater risk of anemia than female children. Also, children living in rural areas were at greater risk of anemia as compared to their urban counterparts. Results also show that children belonging to households with low and medium standards of living index are more susceptible to anemia compared to their counterparts which corroborates the findings in Brazil and other countries (Assunção et al., Citation2007; Mamiro et al., Citation2005). We also notice that there is a declining tendency of the prevalence of anemia after two years of age. This may be due to the iron intake which improves with age as a result of a more varied diet, including the introduction of meat and other hemoglobin-containing diets.

Interestingly, the prevalence of any form of anemia is low among illiterate mothers. Results also indicate that the highest prevalence rate of any form of anemia was among the Muslim children (62%) followed by Hindus (56%) and Christians (48%). The reason might be that the children from Muslim and Hindu religions are from lowest quintiles and illiterate mothers. The fertility of the mother impacted anemia in children i.e. the higher number of children of the mother increases the requirement for childcare, demand for food, and inadequate supply of nutritional diet to all the children which ultimately make the children more vulnerable to the risk of anemia (Singh & Patra, Citation2014)

Mother’s age at marriage had also a considerable effect on anemia in their children. Children of women who got married between 18 and 26 years were at greater risk of anemia (Agho, Dibley, D’Este, & Gibberd, Citation2008; Unsal et al., Citation2007).

The results from this study indicate that ignoring the hierarchical nature of the data could result in over statement of the significance of some of the variables included in the model. Most importantly, the standard errors would be biased downward. Anemia is a widespread health problem in India, especially in less developed regions like northeastern states of India. The strength of the paper is that it provides important information for policy-makers, program planners, and implementers that seek to reduce childhood anemia in northeastern states of India. The study findings have some important and relevant policy messages. High prevalence of mild and moderate anemia demands multiple interventions and strategies to tackle the burden of anemia. Special policies must be formulated and executed for prevention of anemia.

The present study demonstrates the endemicity of childhood anemia in northeastern states of India which may be due biological, social, and cultural factors. However, in this study, mother’s age at marriage and the number of children ever born are the factors significantly associated with anemia in children. Results also suggest that age of the child is also a significant predictor, as the child ages to 48 months or more, the likelihood of severe anemia decreases. This study throws light on the fact that anemia intervention needs to focus more at state level followed by individual level. The limitations of the study are that indicators such as mother’s anemia level, serum ferritin, dietary intake, worm infestations, malaria, and infectious diseases (Sinha, Deshmukh, & Garg, Citation2008) have not been considered.