Factors associated with child hunger among food insecure households in Bangladesh

Data was derived from FSNSP, a surveillance system that operated to track nutrition and food security. FSNSP followed a repeated cross-sectional design for collecting data countrywide at the household level. The data was collected after the two major harvest seasons, the post-aman harvest season (January-April) and the post-aus harvest season (September-December) and also during the monsoon season (May-August). The study area, the whole country, was divided into 13 strata consisting of six vulnerable zones (coastal belt, eastern hills, haor region, padma chars, northern chars and the northwest region) and the seven administrative divisions (Dhaka, Chittagong, Rajshahi, Barisal, Khulna, Sylhet and Rangpur) which contain all the upazilas not included in a surveillance zone, correspond to the seven divisions of Bangladesh. A map of FSNSP zones is illustrated in Fig. 1. From each strata, a set number of upazilas were selected with replacement for further sampling based on the number of upazilas in the strata. From each of the surveillance zones, upazilas were selected by rotation into the sampling frame in order to reduce random variation in estimates between rounds, as has been recommended for surveillance systems by the United Nations (UN) and is commonly done in labour participation surveillance [18]. FSNSP’s rotational pattern ensures that 50% of all upazilas in zones are identical between the same season in subsequent years and between two consecutive rounds. In each round, three new upazilas were selected for sampling and the remaining nine upazilas are drawn from past rounds in each surveillance zone. Each selected upazila remains in the system for four rounds of data collection. In the second stage of sample selection, three rural villages or urban mohalla were chosen at random and without replacement from all the villages/mohalla in each selected upazila. There was no stratification of rural and urban areas during the second stage of selection. The third stage of sample selection was done in the field. In each community, 32 households were selected systematically and interviewed. The starting point for interviews in each village was the first eligible house from a randomly assigned approach road (north, south, east, or west) determined by a random number generator. The next and subsequent households for interview were chosen systematically by counting five or ten households from the previously interviewed household (depending on the size of the village) and, in a “zigzag” fashion, selecting households from both sides of the road. In situations where the identified household was not eligible for inclusion or refused participation, the next household that met the inclusion criteria was selected [19]. Sample size was calculated to obtain representative prevalence estimates for indicators of food insecurity and children’s and women’s undernutrition by surveillance zone. Sample size calculations were based on the estimated prevalence of seven key indicators such as Round-wise estimation of acute childhood malnutrition, child underweight, chronic childhood malnutrition, proportion of women with chronic energy deficiency, proportion of households with food insecurity, proportion of households with “household food deficit”, and proportion of households with poor or borderline food consumption patterns. A trial profile is shown as Fig. 2. The primary respondent was the mother. Data quality was ensured through multiple procedures of review and cross-checking. Monitoring officers reviewed all questionnaires on the day of completion by the data collectors so that any errors or inconsistencies identified could be corrected in the field. Quality control officers revisited a randomly selected sub-sample (around 10%) of interviewed households within 48 h of the initial visit by the data collection team to verify the quality of data collected. For all three rounds conducted in the second field year of FSNSP, internal FSNSP quality control operations were supplemented by Bangladesh Bureau of Statistics (BBS) staff performing a 10% post-enumeration check using a shortened questionnaire. Quality control data were compared to the surveillance data collected by data collectors. Inconsistencies were reviewed by the project manager, project coordinator, training officer, and field managers to identify possible reasons for the discrepancy and to implement appropriate solutions, such as a review session on selected indicators during the refresher training or a revision of the questionnaire.

Data of 14,712 Bangladeshi children aged 6–59 months belonging to food insecure households as per Household Food Insecurity Access Scale (HFIAS) collected between June 2011 and November 2013 was analyzed for this paper. Information on food insecurity at the household level was collected for the month preceding the survey. FSNSP measures food security through HFIAS which defines food insecurity based on lack of access originated due to poverty rather than shortage of supply [20]. The scale contains 9 questions (Worry about food/unable to eat preferred foods/eat just a few kinds of foods/eat foods they really do not want eat/eat a smaller meal/eat fewer meals in a day/no food of any kind in the household/go to sleep hungry/go a whole day and night without eating) to assess the level of anxiety and uncertainty of the participants about household food supply, insufficient quality of food and insufficient food intake [21]. Data has been collected about child hunger separately in this project so as to determine its relation with food insecurity; to ascertain whether child hunger increases with food insecurity.

Child hunger was defined in this study as a household having at least one child who consumed only rice or an amount of food less than required or fewer meals than required or went to bed hungry or a whole day and night without food on the day preceding the survey. The household with such a child was defined as a household having at least one hungry child. Cited literature suggested the selection of variables for this study, the selected variables were seasonality, residence type, sex of household head, educational level of household head, educational status of household women, degree of household food insecurity, amount of rice consumed by the household members and asset index. Seasonality, for this paper, has been defined as the segregation of the year into the post-aman (January-April), monsoon (May-August) and post-aus (September-December) periods, type of residence has been dichotomized into rural and urban. Education level of household head has been categorized into ‘no formal education’, ‘below Secondary School Certificate (SSC)’ and ‘SSC complete and above’, educational status of household women was determined on the basis of households having or not having at least one woman, with at least, one year of formal schooling. The asset index used in this study is a composite indicator of household wealth calculated using principal component analysis, following a method similar to that used in the Bangladesh Demographic and Health Survey (BDHS) [11, 22]. The calculation was based on ownership of household electrical appliances, furniture, livestock and vehicles, the type of household construction materials, kitchen fuel and latrine used and the source of drinking water. Food insecurity was categorized as mildly food insecure, moderately food insecure, and severely food insecure.

All analyses were conducted in STATA v10 (StataCorp; College Station, Texas, USA) using the svyset command to adjust strata and cluster for complex survey data. Descriptive statistics served to illustrate the general characteristics and simple logistic regression was used to assess the strength of association (unadjusted) between child hunger and other variables. Multivariate logistic regression analysis was conducted to identify the statistically significant risk factors for child hunger. Logistic regression is the most suitable mode of analysis since the outcome variable is binary and both categorical and continuous variables can be fitted into the regression model. All variables which had p-value significant at 0.25 were included in the logistic regression [23]. In the final multiple logistic regression model, variables were considered statistically significant only if the p-value was less than 0.05. Seasonality, residence type, educational status of household head were adjusted in all the steps of the model.

Our analysis suggests that 14,712 households were food insecure among which 11,428 households had data on hunger and had at least one child aged 6–59 months. Around 94% households were from rural area. The average number of household member was five and the average consumption of rice by each household was 13 kg in the week preceding the survey. Among households who experienced hunger, 81% were severely food insecure. Overall, 10% of the children were hungry and additionally, around 1% went a whole day and night (24 h) without eating anything or slept at night being hungry. On the basis of seasonality, the proportion of hungry children was around 8% during post-aman, 10% during monsoon and 10% during post-aus season. Ninety three percent of the households were headed by male and 51% of the household heads had no formal education. Around 59% of the primary earner of the households was day laborer, whereas around 16% households had no women with at least one-year of formal education. All the descriptive findings are tabulated in Table 1.

Our bivariate and multivariate analyses (Table 2) revealed that seasonality, type of residence as well as education level of household head was not significantly associated with child hunger. The odds of child hunger was 1.87 times [95% CI: 1.43–2.45; p < 0.001] in female headed households compared to male headed households. Severe food insecurity in the households appeared to be significantly associated with child hunger [OR: 10.5 (1.43–76.6); p < 0.05] in comparison to mildly food insecure households. It was also found that child hunger was 1.56 times [95% CI: 1.27–1.92; p < 0.05] more in a household having no educated woman. Odds of a child being hungry in the poorest household was 1.54 times higher [95% CI: 1.11–2.15; p < 0.05] compared to the richest quintile. The amount of rice consumed by the household members in the past week was negatively associated with child hunger [OR: 0.94 (0.92–0.96); p < 0.001].

While the explanatory variables indicate risk factors for child hunger, causal inferences cannot be established due to the cross-sectional nature of the data. A possibility of recall bias remains regarding HFIAS data, as information of 1 month preceding the survey was gathered through maternal response. While, the strength lies in the large sample size and adjustment of seasonality along with separate measurements for food insecurity and hunger. Nevertheless, the present work contributes greatly to our understanding of the socio-economic characteristics related to child hunger among food insecure households.