Looking upstream: enhancers of child nutritional status in post-flood rural settings

Design and sample

Jagatsinghpur is a coastal district of the state of Odisha, located within the Bay of Bengal, India. The district’s population is around one million, 90% of which inhabit rural areas (The Census of India, 2011). The region is located in a large and fertile flood plain, crossed by large rivers, which makes it very attractive for fishing activities, livestock rearing and crop farming. However, this land is also regularly affected by heavy monsoons which often produce flooding. In the last decade, Jagatsinghpur has been hardly hit by five major floods, including coastal flooding associated to cyclone Paradip (05B) in 1999, followed by heavy rain floods in 2001, 2003, 2006 and the latest flood occurring prior our investigation, which took place in mid-September 2008 (Government of Orissa, 2008).

We conducted a population-based survey one year after the 2008 flooding in rural Jagatsinghpur district, Odisha, India. A two-stage cluster survey was necessary to obtain a probability sample of children 6–59 months of age in 265 villages from four severely flood-affected blocks of the district (Kujang, Biridi, Balikuda and Tirtol). Overall, 122 of these were flooded and 143 non-flooded in September 2008. The percentage of households flooded in each village was obtained for 2006 and 2008 events from OSDMA data (Odisha State Disaster Mitigation Authority, 2009; Fig. 1).

We used a two-stage cluster design to select our sample as our population of interest was clustered in villages and the information on the population was scant (Groves et al., 2009). In this study, the Primary Sampling Units (PSUs) were the villages and the children were the Secondary Sampling Unit (SSU). A 30 (PSU) by 30 (SSU) design, which should provide a probability sample of 900 children, was fixed. We initially enumerated the 265 villages along with each village population size projected for 2009 from census data (The Census of India, 2001), and subsequently 30 clusters from 29 villages were selected. This first selection was done using Probability Proportionate to Size (PPS) sampling with replacement (i.e., one large village was selected twice). This method picks up villages randomly, but the chances of selection are proportional to the size of the village, with selection probabilities favoring larger villages (Groves et al., 2009). At the second stage a fixed number of children (i.e., 30) are selected by cluster, independently of village size. The result is that chances are compensated with equal probabilities of selection of any eligible child listed in the sampling frame (Groves et al., 2009). Importantly, an updated list of eligible children was obtained from the ICDS centers (1 ICDS center for every 1,000 inhabitants) and validated with the ward members of each village. A total 3,671 eligible children were listed from 29 villages within a month prior training and piloting of the survey instrument. Once the lists compiled, it was detected that three flooded villages (Korana, Jamphar and Raghunathpur) and two non-flooded (Muthapada, Sureilo) did not have enough eligible children (i.e., less than 30). We created 5 new clusters by merging the list of children in each of these 5 villages (n = 280) with those of the closest non-selected flooded or non-flooded village, respectively, of our list (Fig. 1). Finally, thirty children per cluster were randomly selected.

Ethical approval

This study was approved by the Community Health Ethics Committee, Voluntary Health Association of India, New Delhi. Persons eligible to participate in the study were not offered any monetary incentive. Written informed consent was obtained for every head of household visited. In case the respondent was an illiterate, we asked a literate person from the community to read out the consent form and explain it to the head of the family. We then obtained the thumb impression of the respondent. In those cases, the person who read out the consent form also signed as a witness. Research procedures were consistent with the Declaration of Helsinki (1997). Interviews were administered after obtaining informed consent. The protocol was reviewed by a small group of scientists who had experience working with survivors of natural disasters and amended based on their recommendations.

Data collection: instruments and measures

Our survey instrument was adapted from the core one developed and approved by a multidisciplinary consortium of researchers from the MICRODIS project. The instrument development was based on interim literature reviews, and follows the UNICEF conceptual framework on child malnutrition (United Nations Children’s Fund, 1997). The questionnaire was validated prior to our study in research sites in India, Indonesia, the Philippines, Vietnam and the UK. We collected background information at the household level, more specifically from mothers and fathers, covering basic socio-demographic characteristics, wealth, child caring practices, healthcare access, maternal and paternal education, income and credit practices, water and sanitation, food consumption patterns; demographics, nutrition and health status data at the child level.

To assess nutritional status using anthropometric indicators, weight and height/length were recorded. Children were weighed without clothes. Weight was measured to the nearest 100 g by trained research assistants using a beam balance (<10 kg; Raman Surgical Co., Delhi-33, India) and an electronic balance for those children heavier than 10 kg. For children younger than 2 years of age, length measurements were taken to the nearest millimeter in recumbent position with an infantometer (Narang Medical, Delhi-110 028, India). If children were older than 2 years, they were measured standing up with an adjustable board calibrated in millimeters. Each research assistant measured and weighted each child twice to minimize measurement errors and use the average value of both measurements to gain precision. These instruments were calibrated daily.

Study questionnaires were administered by twelve experienced research assistants (Rodriguez-Llanes et al., 2011) from the Voluntary Health Association of India (VHAI) who received specific training on anthropometry and interview procedures for this study in late August 2009. The questionnaire was piloted in 12 households (6 in flooded villages and 6 in non-flooded) and improved based on the inputs of the pilot exercise. The study questionnaire was translated to the local language in Odisha (Oriya) and subsequently back translated into English by different professional translators, and a researcher checked the level of agreement between both versions. Duration of interviews ranged from 45 to 60 min, and all field work was completed between 6 and 24 September 2009.

Study variables

As our aim was to estimate the association effect of formal maternal and paternal education on child undernutrition, we excluded variables in our dataset which may have mediated this effect (Schisterman, Cole & Platt, 2009). We supported our choice by examining these variables (e.g., caregiving practices, food security, and access to health care, water and sanitation) framed within the UNICEF framework for child malnutrition (United Nations Children’s Fund, 1997). Overall 51 variables were assessed in a previous study (Rodriguez-Llanes et al., 2016) but for the purpose of this study we only analyzed distal determinants of child undernutrition.

Two outcomes were used in our study: stunting (height-for-age) and wasting (weight-for-height). Stunting is an indicator of chronic malnutrition, whereas wasting often evaluates acute nutritional stress at individual and population levels. The new WHO standard was used to calculate the z scores for these indicators. Malnutrition was a binary variable indicating whether a children is malnourished, z score ≤ 2 (1) or not (0) at the time of the interview.

Overall, 17 variables were examined as potential predictors. The two fundamental variables of this study were the level of formal education attained by both children’s parents, mother and father. Formal education was the only assessed in this study, as we excluded parents with technical training or professional studies (Fig. 2). We studied education using the same categories as recorded in the original questionnaire, which follow benchmarks in the Indian educational system: never attended school (0 years of schooling), completed elementary school (5 years of schooling), middle school (8 years), high school (12 years) and completed university studies (15 years or more). For mothers, age at marriage, age at first delivery and age at birth of the selected child were reported and analyzed in years; the later calculated by subtracting the age of each child (converted to years) to the mother’s age. The father’s age at birth of the selected child was obtained by same calculations. Child’s sex was a binary variable. The age of each child in months was obtained from birth certificates and vaccination cards. If these were not available, local calendars were used. The birthweight of each child was recorded in grams from birth certificates and for analytical purpose we expressed birthweight per 100 g. The count of children younger than 5 years living in a household were used in our analyses. The principal means of livelihood were dichotomized as agriculture (taken as reference), grouping households dedicated to fishing, livestock rearing and crop farming versus non-agricultural for any other reported activity. Two religions were present in the study area: Hinduism, taken as the reference group, and Islam. The caste of the household was based on the household head and was grouped as a scheduled caste, other backward class or general class (reference category). The general class is the higher caste status. The scheduled caste is the social group historically subject to the higher deprivation levels in the country. Owned land was originally collected in acres but was expressed in hectares and analyzed as a continuous variable. Annual household income was recalculated per capita using data on household size and expressed per 1,000 Indian rupees (INR) and modeled as a continuous predictor. A household owning any livestock, including chicken, were modelled against those households not owning any. Finally, we recorded the exact number of persons residing in each household (household size) but were dichotomized as more than four or otherwise to account for overcrowding as done by a previous study (Semba et al., 2008). The levels of occupation (employed, unemployed or working as housewife) of mother and father were examined to better comprehend the level of gender empowerment within the household.

Statistical methods

Two expert data managers entered the data. An external researcher (JMR) run exploratory data analysis to identify errors in data entry and other implausible values (Day, Fayers & Harvey, 1998). ENA for SMART software (version November 2008) was used to calculate nutritional indicators with the 2006 World Health Organization Standard (ENA, 2008). ENA software is built up with specific functions for detection of outliers and impossible values. These were used to identify further problematic values among the calculated nutritional indicators. Each potential error discovered was discussed and investigated to determine where they originated and subsequently corrected.

The main predictor in this study was mother and father education. The original sample size was determined based on requirements of a previous study (Rodriguez-Llanes et al., 2016). Regarding the main research question addressed by this report, overall, available sample sizes were sufficient for subgroup analyses to detect prevalence ratios of two or more with a 80% power (Sullivan, Dean & Soe, 2009).

We examined the relationship between parental education and child nutritional status separately in repeatedly flooded and non-flooded cohorts. A first consideration was not to model maternal and paternal education jointly, but in separate analyses (Fig. 2). Intuitively, individuals with similar education tend to get married more often and this was reflected by our data: variance inflation factor (VIF) did show these two variables being highly collinear. As such, eight models were fitted on this data (2 datasets × 2 outcomes × 2 predictors each). These models are summarized in Fig. 2.

Bivariate and multiple adjusted logistic regression models with a quasi-binomial distribution to control for overdispersion were fitted (Lumley, 2010). For multivariate models, we included only variables with p < 0.1 in bivariate associations with undernutrition. We first ran a full model with all variables having a p-value lower than 0.1, a bit more conservative that some authors recommend (Vittinghoff et al., 2005). On each full model, the VIF on each predictor was calculated and predictors with the highest VIF were sequentially eliminated until all remaining predictors had a VIF lower than four. Backward selection was then applied to the remaining variables. At each step the non-significant variable with the largest p-value was excluded to obtain a final model including only significant variables at alpha level of 5% (i.e., p-value < 0.05). As proposed in recent literature (Siri, 2014; Muttarak & Lutz, 2014), we examined the impact of household annual per capita income on the effects on paternal education to ensure that education and income are independent contributors of child nutrition outcomes. We did that in all eight models.

Results were provided as prevalence ratios, crude [PR] or adjusted [aPR] with 95% confidence intervals. Given the limits of our sample size, interactions were not examined. All tests were two-tailed with α = 0.05. All analyses were weighted. Weights were calculated as the inverse of the selection probabilities. Statistical analyses were conducted in R (version 3.0.2) (R Development Core Team, 2008) with the survey package (Lumley, 2010).

Study sample

A sample representative of the children population 6–59 months of age was obtained in 265 villages of Jagatsinghpur district. Our analyses excluded children with missing observations on relevant variables and from villages inundated in September 2008 but not in 2006. Figure 2 provides further details on each stage of the sampling and exclusions before statistical analysis.

Table 1 presents descriptive information for the variables analyzed in recurrently flooded communities (n = 299) and those non-flooded (n = 385). The illiteracy of mothers was less prevalent in flooded areas compared to non-flooded. Whereas more mothers completed middle school in flooded communities, more completed university studies in non-flooded ones. The percentage of fathers completing high school was 10% higher in flooded communities compared to non-flooded, while those having university degrees was similar. In general the flooded population was more educated relative to the non-flooded, and fathers were more educated than mothers. Almost 75% of the mothers attained at least middle school amongst the flooded villages relative to 67% in the non-flooded. For a similar comparison amongst fathers, the percentages were higher, 89% and 81.2%, respectively.

Flooded households showed higher reliance on agriculture activities as per their higher percentages of land and livestock owned, which corresponded to higher proportions dedicated to agricultural activities (41.8% in flooded vs. 35.8% in non-flooded). Households of the highest cast were also more common in the sample of flooded villages whereas schedule and other backward castes were found more often in non-flooded communities. Unemployment rates among fathers were 2.1% in non-flooded compared to 2.3% in recurrently-flooded populations. As many as 97.7% of all interviewed mothers worked as housewives in both flooded and non-flooded communities.

Correlates of child stunting

In univariate analyses, paternal and maternal education played an important role in reducing child stunting (Fig. 3). Overall, the protective effect of education on child stunting were substantial for most educated groups, amongst flooded communities and in men. Compared to mothers never attending school, a 3-fold (200%) lower prevalence of stunting was observed in most educated ones (i.e., those completing university studies). But generally the effects observed on paternal education were larger and stronger compared to maternal education (Fig. 3).

An additional hectare of land owned in flooded communities had a substantial protective effect, although non-significant (p = 0.07; detailed results in Table S1), but the effect was lower in non-flooded. A per capita household annual income improvement of 1,000 rupees was also associated with a decrease of about 6% in child stunting, but again, only in flooded households (Fig. 3, Table S1). Belonging to a backward caste was detrimental to child stunting but only in flooded communities, and especially deleterious in the most deprived caste in India, the scheduled caste, with more than a twofold associated increase in the prevalence of child stunting.

Multivariate analyses on child stunting

In adjusted analyses, the positive effect of maternal and paternal education on child nutrition weakened and remained statistically significant only for models on paternal education in flooded communities, and marginally significant in non-flooded (Table 2). In repeatedly-flooded communities, fathers with completed middle, high school or university studies were consistently associated with 2.5-, 2.1- and 2.7-fold lower prevalence of child stunting, respectively. In those non-flooded communities, completed middle or high school by fathers conferred lower protection to their children, with associated 1.9- (p = 0.055) and 1.8-fold (p = 0.085) lower prevalence of stunting but none reached statistical significance. Having completed college was comparably as protective in non-flooded areas as in flooded (2.8-fold reduction), and the effect statistically significant (p = 0.015).

Per capita annual income was a consistent effect in flooded studied areas and across models on maternal and paternal education with 4.7%–4.9% lower prevalence of stunting for each yearly increase in 1,000 rupees per capita at household level. Schedule caste remained a relevant factor in multivariate analysis modeling the effect of maternal education but only in flooded communities (Table 2). The two final models for non-flooded communities were also adjusted for the effect of income but no substantial change in the coefficients were noted.

Correlates of child wasting

Crude associations showed an important role of parental education to reduce child wasting, with a preponderance of paternal over maternal effects of formal education and the largest effects observed in non-flooded communities (Fig. 4). An association between means of livelihood (non-agriculture vs. agriculture) and child wasting favoring households not relying on agriculture activities was observed, with a 47% lower prevalence of child wasting compared to those dedicated to agricultural activities (Fig. 4, detailed results in Table S2).

Multivariate analyses on child wasting

The results for paternal education were different in adjusted models compared to those showed by bivariate associations (Table 3, Table S2 and Fig. 4).

In flooded communities, the protective effects of different levels of paternal education were all significant (p < 0.001) except for that of primary education (p = 0.117). However, larger effects were observed in the non-flooded communities surveyed in our study starting from primary education (2.9-fold lower prevalence, p = 0.019), middle (2.9 times lower, p = 0.012) or high school (2.7-fold difference, p = 0.008) up to those with university degrees who were associated with two times lower prevalence of wasting among their children, though the association remained insignificant (p = 0.081). In flooded households dedicated to activities other than agriculture, a 50–51% lower prevalence of child wasting was estimated (Table 3).

In modeling the effect of maternal education in flooded communities, we found that the mother’s age at birth of the studied child was associated with an improved nutritional outcome of their children. Each additional year was associated with a 3.4% lower prevalence of wasting (p = 0.005; Table 3). In flooded communities only and for the model on maternal education, belonging to other backward castes showed to be deleterious on child wasting (Table 3). All four final models in this section were also adjusted for annual per head income but none of those inclusions impacted our results.

Strengths and limitations of the study

There a few methodological considerations, strengths and limitations that need to be addressed here. Firstly, all the reported results were adjusted, as an additional robustness check, by annual per capita household income. As such, all effects reported were independent of income, and thus our results can be methodologically comparable to those recently published (Muttarak & Lutz, 2014). Second, we often found in the literature maternal and paternal education jointly modelled (Semba et al., 2008). However, we found that these two variables were highly correlated in our data and enough evidence using variance inflation factor analysis, which supported that they should be modelled separately. Future studies should be aware of this. Third, father education was highly correlated to caste and mother’s age at birth of a child, and this is the reason why models modeling father education do not include these variables. Strengths of this study include carefully consideration of income, as well as having analyzed data on a non-flooded group which allowed us to observe that some effects such as a mother’s age at birth, income or caste become important determinants only in the extreme circumstances of flooding. Additional positive features of this study were its population-based design, the consideration of recurrently flooded communities only, coupled with the analysis of mid-term and more long-term child undernutrition. On the minus side, our study was not based on a very large sample size, which could have been missed the detection of smaller effects and provided tighter confidence intervals. Similarly, a larger sample size might have allowed us to analyze most severe forms of malnutrition, such as severe wasting and stunting. In the aftermath of flooding in Bangladesh, Choudhury & Bhuiya (1993) have shown maternal education to play a role in increasing severe forms of underweight. Also, as the findings presented are based on cross-sectional data, care is needed to not attribute causal relationships to the associations showcased here. Importantly, it is difficult to generalize our results to other settings such as urban slums, as vulnerabilities might be different there. However, we are confident that there might be similarities between our results and other rural areas around the world in which subsistence farming is the norm. With its limitations, this is one of the first studies to look at child undernutrition in flood settings and study potential preventive factors in short and long-term. More studies are needed to solidify the evidence base for global action.