Stagnating trends in complementary feeding practices in Bangladesh: An analysis of national surveys from 2004‐2014

Abstract Bangladesh has experienced steady socio‐economic development. However, improvements in child growth have not kept pace. It is important to document complementary feeding (CF) practices—a key determinant of children's growth—and their trends over time. The study aims to examine trends in CF practices in children aged 6–23 months using data from Bangladesh Demographic and Health Surveys conducted in 2004, 2007, 2011, and 2014. Multilevel logistic regression models were applied to identify independent predictors of four CF practice indicators among children 6–23 months, namely, timely introduction of complementary foods, minimum meal frequency, minimum dietary diversity, and minimum acceptable diet. Introduction of complementary foods was achieved among 64–71% of children between 2004 and 2014. The proportion meeting minimum meal frequency increased from 2004 to 2007 (71–81%) and declined and held steady at 65% from 2011 to 2014. The proportion meeting minimum dietary diversity in 2011 and 2014 was low (25% and 28%), and so was minimum acceptable diet (19% and 20%). From 2007 to 2014, child dietary diversity decreased and the most decline was in the consumption of legumes and nuts (29% to 8%), vitamin A‐rich fruits and vegetables (54% to 41%), and other fruits and vegetables (47% to 20%). Young child age (6–11 months), poor parental education, household poverty, and residence in the Chittagong and Sylhet independently predicted poorer feeding practices. Dietary diversity and overall diet in Bangladeshi children are strikingly poor. Stagnation or worsening of feeding practices in the past decade are concerning and call for decisive policy and programme action to address inappropriate child feeding practices.

has declined significantly, the annual rate of reduction has slowed down in more recent years. Under optimistic estimation, if the current rate of decline persists at~3% point per year, more than 25% of children under five will still be stunted or underweight in 2025 (Government of the People's Republic of Bangladesh, 2017).
In recent years, Bangladesh has shown impressive economic growth. According to the World Bank Databank, between 2004 and, the gross domestic product per capita in Bangladesh increased at an average annual rate of almost 5% (World Bank, 2017; see Figure   S1). This strong economic growth has lifted many people out of poverty in both rural and urban areas ( Figure S1a). Other socio-economic and child health care indicators have also improved ( Figure S1a), including primary education completion rate, coverage of improved water source and sanitation facilities, and women's empowerment.
However, socio-economic development and health policy commitments do not seem to have solved the challenge of child undernutrition in Bangladesh. The current prevalence of stunting and underweight in 2014 remained unacceptably high at 36% and 33%, respectively ( Figure S1c). The prevalence of child wasting (~14%), low birth weight (~22%), and anaemia in women (~44%) and children (~51%) have remained high, showing no significant change over the past decade ( Figure S1c).
Though child malnutrition has multiple causes, it is widely agreed that inadequate infant and young child feeding (IYCF) is one of the most immediate determinants (Black et al., 2013;Stewart, Iannotti, Dewey, Michaelsen & Onyango, 2013). Rapid growth and critical development during infancy and early childhood require diverse, energy-, and nutrient-dense complementary foods that are introduced at the right time and fed with the right frequency and in right quantities (Arimond & Ruel, 2004;Black et al., 2013;Dewey, 2003;Frongillo et al., 2016;Horta & Victora, 2013). Child growth faltering is most evident from 3 to 24 months (Victora, de Onis, Hallal, Blössner, & Shrimpton, 2010), a window during which suboptimal feeding likely takes place, including disrupted exclusive breastfeeding in the first 6 months, early or delayed introduction of complementary foods, discontinued breastfeeding before age two, and inadequate quantity and/or quality of foods and feeding frequency. Because of the critical role of IYCF, several major nutrition programmes and policies in Bangladesh that aim to reduce child undernutrition have emphasized IYCF as a key component. These include the National Nutrition Programme The objective of this analysis is to document trends and predictors of CF practices in Bangladesh using repeated national survey data over a 10-year period (2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014) In all four BDHS, respondents were asked to recall if the child ate any solid, semi-solid, or soft foods in the previous 24 hr, during the day and at night. Introduction of solid, semi-solid, or soft foods (Intro) was calculated as the proportion of infants 6-8 months of age who received complementary foods in the previous 24 hr.
Respondents were also asked to recall how often the child was fed solid, semi-solid, or soft foods, and how many times the child consumed dairy products, including tinned, powdered or fresh animal milk, infant formula, and yogurt. For breastfed children, Minimum meal frequency (MMF) was calculated as a frequency ≥ 2 for children aged  Table   S1. The food items were grouped into seven food groups, including grains, roots, and tubers; legumes and nuts; flesh foods; eggs; vitamin A-rich fruits and vegetables; other fruits and vegetables; and dairy products. There were some differences in the IYCF module Finally, if a breastfed child met MMF and MDD or a nonbreastfed child met MMF and was fed ≥2 milk feedings plus ≥4 food groups excluding milk products, we determined that the child had met the overall minimum requirement. We then calculated the proportion of children 6-23 months of age who met the overall minimum requirement in the previous day or night as minimum acceptable diet (MAD). MAD was only calculated in 2011 and in 2014 with incomplete food group data.

| Individual and household level
At individual level, we selected characteristics of the child, the mother, and/or the father that could be associated with child feeding and nutritional status (Black et al., 2013;Stewart et al., 2013). For children, child sex, breastfeeding status, age, birth order, birth interval, perceived birth weight, vitamin A supplementation in the previous 6 months, iron supplementation in children in the previous 7 days (not available in 2004 and 2007), vaccination status, and reported morbidity in the previous 2 weeks (diarrhoea, fever, cough) were considered. For mothers, we included age, body mass index (defined as weight in kilograms/squared height in meters), use of reproductive health care including place of delivery, type of delivery assistance, caesarean delivery, number of antenatal clinic visits, timing of postnatal check-ups on women and child (not available in 2004), highest education level completed, marital status, exposure to media (newspaper, radio, TV), involvement in decision making (regarding large purchases for the households, freedom to visit family and friends, woman's own health care), and attitude towards domestic violence (whether beating wife is justified if she goes out without telling husband, neglects the children, argues with husband, refuses to have sex with husband, or burns the food). A composite women's empowerment score was calculated using an established coding scheme of available items under the "decision making" and "domestic violence" module (Jennings et al., 2014;Na, Jennings, Talegawkar, & Ahmed, 2015).
For fathers, we included age, highest education level, and occupation.
At the household level, we selected household head sex, number of household members, number of children under 5 years of age, types of cooking fuel, water (source of drinking water, location of water source, time to get water) and sanitation (toilet facility and whether toilet facility was shared) characteristics, and a precomposed wealth index using available socio-economic variables and principal components analysis (Rutstein, & Johnson, 2004).

| Community level
The community level was defined at the PSU or cluster level. Place of residence, region, proportion of women completing primary or higher education, and mean women's empowerment score were used to describe the distribution of attributes among the eligible clusters, in which the mother-child dyads lived. In addition, we created a summary index of community-level access to health care using all available data. The detailed algorithm to compose the index and methods to categorize five quintiles has been described previously (Na, Aguayo, Arimond, & Stewart, 2017). In sum, eight variables were available in

| Statistical analysis
STATA/SE 14.1 (STATA Corporation, College Station, TX, USA) was used to analyse data. The "svy" command was used to adjust for sampling weights in calculating mean, median, and proportions at the population level. The average annual rate of increase was calculated to measure the geometric progression ratio, at which proportion changes constantly over the period between first and latest observed year.
Nonparametric tests were performed to test trends in ordinal variables over year. To test if rates of change in CF indicators differed in subgroups of the sample, an interaction term was created between the group and year and was added to the logistic regression model for each feeding indicator. The slopes representing the linearized rates were estimated, plotted, and compared against each other by contrasting the marginal effects. Delta-methods were used to determine statistical significance (Cameron & Trivedi, 2005).
We constructed multivariable models to identify independent predictors following several steps to determine the number of levels in the model variance structure and final variable selection. First, we constructed the year-specific null model for each CF indicator with no predictors under a one-(individuals only) and a two-level (individuals and clusters) variance structure. Likelihood ratio tests were used to compare between the two nested models to determine whether the two-level model significantly increased the proportion of explained variance for each CF indicator. With the exception of the Intro indicator in 2011, all tests resulted in a p value < .05 and we decided to apply the two-level models in subsequent analysis. Second, we constructed year-specific two-level univariable models for each pair of predictor variable and CF indicator (Tables S2-S5). Third, to construct year-specific multivariable models for each CF indicator, we included all predictive variables with p < .1 in the univariable models as the initial step. We then removed variables with variance inflation factor (VIF) greater than 5, starting from the variable with the largest VIF, one by one until all VIF < 5. Finally, we pooled the data across years , 2007, 20112011 and2014 for MDD andMAD) and included the combined set of predictors remaining in the final year-specific multivariable analyses, and reran the analysis with the pooled (all years) data. After removing variables with VIF > 5, we finalized the two-level multivariable models for Intro, MMF, MDD, and MAD, respectively, in the pooled analysis. Trends and odds ratios were considered statistically significant at p < .05. Sensitivity analyses were performed to check robustness of multilevel models by comparing results from (a) multilevel models using stepwise selection at p = .1 level; (b) multilevel models with child sex, maternal, and paternal age fixed as covariates; and (c) multilevel models with and without the year variable and models including year as a continuous or a categorical variable.
The proportion of mothers watching TV increased slightly from 44% to 49% (p trend < .05). Indicators of women's empowerment improved by 10-24% points (all p trends < .05). The proportion of fathers completing primary or higher education increased slightly (27% to 30%, p trend < .001). At the household level, the proportion with unimproved sources of drinking water remained low (3.6% to 2.8%, p trend = .39), whereas there was a significant reduction in the prevalence of households with unimproved toilet facilities (44% to 31%) and shared toilets (45% in 2007 to 35%, both p trend < .001).
The distributions of community characteristics are presented for the 361, 356, 584, and 582 eligible clusters (Table 1). Around two-thirds of the clusters were in rural areas (63-66%). The rate of child completed age-appropriate vaccination increased from 71% to 82% and the utilization of reproductive health services and nutritional supplementation, including health facility delivery, professional assistance at delivery, Caesarean delivery, antenatal care, postnatal care, child vitamin A supplementation, and child and maternal iron supplementation, also increased significantly between 2004 and 2014 (all p trend < .001).

| Trends in CF indicators
The proportion of children meeting the different CF indicators is presented by child age and year ( Figure 1). Intro was achieved by 68%, 71%, 64%, and 67% of children aged 6-8 months in 2004, 2007, 2011, and 2014, respectively (p trend = .31). In the same years, the proportion of children aged 6-23 months who met MMF was 71%, 81%, 65%, and 65%, showing a decreasing trend regardless of child age group (all p trend < .001). The proportion of children meeting MDD was 25% and 28% in 2011 and 2014, respectively (p trend = .10), whereas the proportion of children meeting MAD was 19% and 20% in 2011 and 2014, respectively (p trend = .15). Among the three age groups, the proportion meeting MMF, MDD, and MAD was the lowest in children aged 6-11 months, among whom only 52%, 13%, and 11% met each minimum required criterion, respectively. This general secular trend in CF indicators held true in other population subgroups with some differential changes over time ( Figures S2-S4). For example, the proportion of children meeting the MMF criterion decreased over time but more so among children whose mothers had no education ( Figure S2, difference in slope   associated with differences in community-level access to health care narrowed over time ( Figure S4).

| Trends in food group consumption
The distribution of child dietary diversity scores and the proportion of children consuming individual food groups are presented by child age (Figure 2). Among children aged 6-11 months, the weighted mean (95%CI) dietary diversity score was 2.2 (2.0, 2.3), 1.6 (1.5, 1.7), and whereas the corresponding values among children aged 18-23 months were 3.7 (3.6, 3.9), 2.9 (2.7, 3.0), and 3.0 (2.9, 3.1), respectively. In all three age groups, there was a significant declining trend in child ing eggs was also similar across the two survey years (23% and 26%, respectively; Figure S6).

| Independent predictors of appropriate CF practices
The predictors included in the analysis of pooled data (all years) are presented for the four CF indicators in Table 2 (the year-specific results are available in Tables S2-S5 There were other predictors that were consistently associated with MDD and MAD but not Intro or MMF, including 33-36% lower odds of meeting MDD and MAD among children who were firstborn (comparing with second to fourth birth), 24-26% lower odds among children whose fathers had no education (comparing with secondary or higher education), and 5-8%, 28-33%, 42-43%, 45-48% lower odds among children living in the progressively poorer household wealth quintiles (comparing with the richest quintile).   (Jain & Zeller, 2015). However, this relationship may have changed in the past 20 years, during which many women shifted from working on farms to off-farm employment, such as in manufacturing garment factories (Kabeer & Mahmud, 2004) and in microfinance groups (Bangladesh Bank, 2002). A recent systematic review has revealed mixed findings linking microfinance participation to child nutritional outcomes (Orton et al., 2016). Programmes and initiatives to improve CF practices in Bangladesh need to pay attention to promoting and supporting increased meal feeding frequency. Particularly-but not exclusively-among disadvantaged subgroups of children whose mothers are less educated, children who live in poverty, and children who live in communities with poor access to health care and nutrition services.
Similarly, the apparent disparities in MDD and MAD call for policy and programme attention in all groups, particularly the most disadvan-   increasing the intake of protein and improving linear growth in young children (Iannotti et al., 2017).
Third, future nutrition behaviour and social change education interventions should be continuously tailored to audiences with limited or no formal education. Successful nutrition education trials that have improved feeding practices and child growth, share characteristics of good design (e.g., visual tools including charts and posters), actionable messaging (e.g., home-prepared recipes), standardized training and quality control, and frequent visits Lassi, Das, Zahid, Imdad, & Bhutta, 2013   and CARE (Owais et al., 2017) in Bangladesh that were delivered by trained community health and nutrition workers have been effective at improving CF practices in the intervention areas.
Finally, strategies to improve access to food need to be coupled with communication interventions for behaviour and social change, as financial, physical, and social constraints to optimal CF may coexist in Bangladeshi families (Manikam et al., 2017).
Our analysis has strengths and weaknesses. It was limited by the cross-sectional nature of the data, with different sampling frames between recent and more remote survey years including some missing values. Therefore, our analysis can only identify associations and cannot determine causal relationships. The data available in the BDHS also limits our ability to examine all potential risk factors for poor CF at the individual, household, and community levels. The lack of information regarding access to infant and young child feeding-specific services also prevented us from examining more direct health access indicators at individual and community level. However, our analysis has important strengths, including the use of nationally representative data over multiple years that were collected using a common and globally agreed upon methodology. The trend analysis over the past 10 years used nationally representative data and likely represented true distributions in the general population. Multivariable multilevel models were applied in the predictor analysis, the results of which were robust and were supported by a series of sensitivity analysis (data not shown). The standard methodology used for data analysis and presentation provides for potential comparability with previous and ongoing analyses across different countries in South Asia.
In sum, our analysis indicates stagnation or worsening trends in some of the key CF indicators in Bangladesh over the past decade.
These worrying findings call for advocacy, policy, and programmatic .40 Note. Intro = introduction of solid, semi-solid, and soft foods; MMF = minimum meal frequency; MDD = minimum dietary diversity; MAD = minimum acceptable diet; OR = odds ratio; CI = confidence interval. *p < .05; **p < .01; ***p < .001. efforts to reprioritize the improvement of complementary foods and feeding for infants and young children in the context of the larger national development agenda in Bangladesh.