Menu
Menu
Menu
Menu
Childhood linear growth faltering remains a major public health concern in Nepal. Nevertheless, over the past 20 years, Nepal sustained one of the most rapid reductions in the prevalence of stunting worldwide. First, our study analysed the trends in height-for-age z-score (HAZ), stunting prevalence, and available nutrition-sensitive and nutrition-specific determinants of linear growth faltering in under-three children across Nepal’s Family Health Survey 1996 and Nepal’s Demographic and Health Surveys 2001, 2006, 2011, and 2016. Second, we constructed pooled multivariable linear regression models and decomposed the contributions of our time-variant determinants on the predicted changes in HAZ and stunting over the past two decades. Our findings indicate substantial improvements in HAZ (38.5%) and reductions in stunting (−42.6%) and severe stunting prevalence (−63.9%) in Nepalese children aged 0–35 months. We also report that the increment in HAZ, across the 1996–2016 period, was significantly associated (confounder-adjusted p <.05) with household asset index, maternal and paternal years of education, maternal body mass index and height, basic child vaccinations, preceding birth interval, childbirth in a medical facility, and prenatal doctor visits. Furthermore, our quantitative decomposition of HAZ identified advances in utilisation of health care and related services (31.7% of predicted change), household wealth accumulation (25%), parental education (21.7%), and maternal nutrition (8.3%) as key drivers of the long-term and sustained progress against child linear growth deficits. Our research reiterates the multifactorial nature of chronic child undernutrition and the need for coherent multisectoral nutrition-sensitive and nutrition-specific strategies at national scale to further improve linear growth in Nepal. [Correction added on 6 November 2020, after first online publication: in abstract, the citation year in the fourth sentence has been changed from ‘2001’ to ‘2011’.].
This study is reported according to the STROBE checklist for cross‐sectional studies (von Elm et al., 2007). To examine changes in child linear growth outcomes over time, we analysed Nepal's Family Health Survey (NFHS) 1996 (n = 3,703) and four rounds of Nepal's Demographic and Health Surveys (NDHS) 2001 (n = 3,729), 2006 (n = 3,003), 2011 (n = 1,424), and 2016 (n = 1,403). For consistency across all five survey rounds, our analysis was restricted to individual‐level recode data from children aged 0–35 months (Corsi, Neuman, Finlay, & Subramanian, 2012). This age range covers the crucial postnatal window in which most population growth faltering occurs in LMIC (Victora et al., 2010). Furthermore, NFHS and NDHS multicluster cross‐sectional surveys of ever‐married women of reproductive age (15–49 years) are well suited to our purpose, insofar as they are high quality, nationally representative, and standardised across rounds in their coverage of a wide, albeit nonexhaustive, range of hypothesised nutrition‐sensitive and nutrition‐specific drivers of child linear growth faltering and anthropometric outcomes. Further details of these data sets are reported in Angdembe, Dulal, Bhattarai, and Karn (2019), and country‐specific surveys are found in ICF International (2019). Roth et al. (2017) recently reported child linear growth faltering as a whole‐population condition, thus affecting the entire height‐for‐age z‐score (HAZ) distribution. Therefore, our research paper focused on HAZ, measured against the median of the WHO, 2006 Child Growth Standard (WHO, 2006; WHO & UNICEF, 2009), as the main dependent variable. In addition, due to the large pooled sample size (n = 10,880), we analysed the prevalence of stunting (HAZ ≤ 2 SD) and severe stunting (HAZ ≤ 3 SD). At present, stunting is regarded as the standard metric to monitor commitments and progress towards global (and national) chronic child undernutrition targets (de Onis et al., 2013; Devkota, Adhikari, & Upreti, 2016). Our time‐variant independent variables at child‐, parental‐, and household‐level were selected based on the Black et al. (2013) framework and a review of previous regression–decomposition analyses of HAZ (Cunningham, Headey, et al., 2017; Headey et al., 2015, 2016, 2017; Headey & Hoddinott, 2015; Menon et al., 2018). These covariates, representing nutrition‐sensitive and nutrition‐specific domains hypothesised to affect child linear growth outcomes over time, are straightforward inclusions in nutrition models (Table 1). The strengths and weaknesses of various indicators are discussed later. Nevertheless, we do note: Following Filmer and Pritchett (2001), we used a DHS asset index to proxy for household wealth. These and other authors have shown that such DHS asset indices are fair proxies for household socioeconomic status in terms of sharing strong correlations with other welfare indicators, including child linear growth outcomes (Headey & Hoddinott, 2015). To construct a common household asset index across all five data rounds, we conducted a pooled principal components analysis using five consistently measured durables. These five indicators and their respective factor loadings were bicycle ownership (0.31), television ownership (0.58), radio ownership (0.12), non‐natural flooring material (0.51), and household access to electricity (0.55). After applying these loading weights, we rescaled the household asset index to vary between a minimum score of 1 and a maximum score of 10. Variable definitions Source: Authors' construction. Abbreviations: BCG, Bacillus Calmette‐Guerin vaccine against tuberculosis; DPT, diphtheria–pertussis–tetanus vaccine; MCV, measles antigen‐containing vaccine. In addition, we adopted a flexible specification of time‐invariant control variables to adjust the associations between time‐variant independent variables and child linear growth outcomes, including month‐specific child age dummy variables (capturing the progressive growth‐faltering process that chronically malnourished populations undergo until around 24 months of age; Victora et al., 2010), religion and ethnicity variables, regional and agroecological zone variables, maternal age (in five‐year intervals), child sex, stratum, and NFHS and NDHS survey round variables. Data management and statistical analysis were conducted using Stata version 15.1 (StataCorp, 2017). The weighted prevalence and average values of child linear growth outcomes and time‐variant independent variables in each survey round were calculated considering the DHS sampling weight factor using the svyset command. Our analysis excluded all extreme HAZ values beyond the range of ±6 SD from the median. We followed a two‐step regression–decomposition approach to evaluate the important drivers of the change in child linear growth faltering from 1996 to 2016. First, to identify the key nutrition‐sensitive and nutrition‐specific determinants of child linear growth outcomes, we fitted multivariable ordinary least squares (OLS) regression models for the continuous HAZ outcome and multivariable linear probability models with a robust variance estimator (LPM) for the binary stunting outcomes on pooled data from all available survey rounds. The use of LPM for binary outcomes is well established in econometrics and allows for a straightforward interpretation of the average marginal effect of an explanatory variable, expressed as a probability difference using percentage points (p.p.; Hellevik, 2009; Wooldridge, 2002). The functional form (linearity assumption) of the relationships between HAZ and the time‐variant continuous variables were examined using kernel‐weighted local polynomial smoothing graphs. Our multivariable regression models are represented in Equation (1) below, assessing the associations between linear growth outcomes (N) for a child i at time t and vectors of time‐variant nutrition‐sensitive and nutrition‐specific determinants (X), vectors of mainly time‐invariant control variables (μ i ), trend effects represented by a vector of year dummy variables (T), and a standard error term (ε i,t ). In Equation (1), the vectors of coefficients (β) on X constitute the set of parameters of principal interest, which are used to answer the first of our two questions about the determinants of child linear growth faltering, that is, which nutrition‐sensitive and nutrition‐specific determinants best explain variations in child linear growth outcomes among children aged 0–35 months in Nepal from 1996 to 2016? Second, we used the estimated parameters from Equation (1) to conduct a simple statistical decomposition at means described in Equation (2) below (under the assumption that the β coefficients are time invariant and the error term has a mean of zero). For our analysis, we selected the earliest NFHS 1996 round (t = 1) and the most recent NDHS 2016 round (t = k). To evaluate the contribution of important nutrition‐sensitive and nutrition‐specific determinants on the observed trends in child linear growth outcomes, our analysis entailed multiplying the β coefficient from Equation (1) by the change in the means of each explanatory variable across the 1996–2016 period. This gives the predicted change in child linear growth outcomes due to the change in an explanatory variable over the past 20 years and thus shows the estimated contribution of each time‐variant variable to changes in child linear growth outcomes. To illustrate, presume that the average years of paternal education increased by 2.5 years between the NFHS 1996 and NDHS 2016 rounds, thus X¯ t = k − X¯ t = 1 = 2.5, and that the estimated β coefficient of paternal education from the multivariable OLS model equalled 0.040 (p < .10). Multiplying the two components yields 0.10. This indicates that the hypothesised change in paternal education over the 1996–2016 period predicted a 0.10 SD increase in HAZ. We can perform equivalent calculations for other nutrition‐sensitive and nutrition‐specific drivers of chronic undernutrition to gauge the extent to which a determinant explains changes in child linear growth outcomes over time, as well as how all our time‐variant independent variables as a whole (i.e., the models) perform in explaining changes in HAZ and the prevalence of (severe) stunting over time. To check the robustness of our regression–decomposition results, we performed various additional statistical analyses. First, we tested the differences between our LPM β coefficients and average marginal effects estimated from multivariable logistic regression models for (severe) stunting. Second, to assess the assumption of time‐invariant β coefficients, we conducted an Oaxaca‐Blinder decomposition testing for systematic differences in β coefficient between the NFHS 1996 and NDHS 2016 rounds (Jann, 2008). Furthermore, we checked the interaction terms between our time‐variant covariates and five data rounds to test if associations between predictors and child linear growth outcomes were modified by survey year. Third, we used quantile regressions as an alternative method of exploring potential changes in the importance of our hypothesised nutrition‐sensitive and nutrition‐specific determinants across different levels of the HAZ distribution (Block, Masters, & Bhagowalia, 2012). Fourth, as a sensitivity analysis, we conducted separate regression–decompositions for rural and urban samples. Fifth, we estimated models that excluded potentially endogenous health care and demographic variables. Lastly, we tested potential multicollinearity among the time‐variant independent variables in the multiple regression models using variance inflation factors (≥4).
