Path analyses of risk factors for linear growth faltering in four prospective cohorts of young children in Ghana, Malawi and Burkina Faso

Stunting prevalence is an indicator of a country’s progress towards United Nations’ Sustainable Development Goal 2, which is to end hunger and achieve improved nutrition. Accelerating progress towards reducing stunting requires a deeper understanding of the factors that contribute to linear growth faltering. We conducted path analyses of factors associated with 18-month length-for-Age z-score (LAZ) in four prospective cohorts of children who participated in trials conducted as part of the International Lipid-Based Nutrient Supplements Project in Ghana (n=1039), Malawi (n=684 and 1504) and Burkina Faso (n=2619). In two cohorts, women were enrolled during pregnancy. In two other cohorts, infants were enrolled at 6 or 9 months. We examined the association of 42 indicators of environmental, maternal, caregiving and child factors with 18-month LAZ. Using structural equation modelling, we examined direct and indirect associations through hypothesised mediators in each cohort. Out of 42 indicators, 2 were associated with 18-month LAZ in three or four cohorts: maternal height and body mass index (BMI). Six factors were associated with 18-month LAZ in two cohorts: length for gestational age z-score (LGAZ) at birth, pregnancy duration, improved household water, child dietary diversity, diarrhoea incidence and 6-month or 9-month haemoglobin concentration. Direct associations were more prevalent than indirect associations, but 30%-62% of the associations of maternal height and BMI with 18-month LAZ were mediated by LGAZ at birth. Factors that were not associated with LAZ were maternal iron status, illness and inflammation during pregnancy, maternal stress and depression, exclusive breast feeding during 6 months post partum, feeding frequency and child fever, malaria and acute respiratory infections. These findings may help in identifying interventions to accelerate progress towards reducing stunting; however, much of the variance in linear growth status remained unaccounted for by these 42 individual-level factors, suggesting that community-level changes may be needed to achieve substantial progress. Based on several previous frameworks,5 6 14 we developed a conceptual path model of potential influences on 18-month linear growth status (figure 1). We tested the following pathways, which correspond to the labels of the arrows in figure 1. Conceptual model. At an individual level, maternal height may be related to the child’s genetic potential for adult height that can be attained. In populations with a high prevalence of stunting, maternal height is also partly a reflection of growth restriction experienced by the mother during early life. Therefore, inclusion of maternal height in the model served two purposes: first, to adjust for a proxy of genetic potential, and second, to test the pathway that intergenerational effects of maternal growth during early life, reflected by maternal adult height, on child linear growth may be mediated by (1.1) socioeconomic conditions of the current generation, (1.2) maternal adult factors (nutritional status, illness, stress, depression, cognition), (1.3) caregiving practices, (1.4) child factors or (1.5) may directly affect linear growth (arrows for each individual pathway not drawn in the figure). Socioeconomic disparities and other environmental effects on child linear growth may be mediated by (2.1) maternal factors, (2.2) caregiving practices, (2.3) child factors or (2.4) may directly affect linear growth. Effects of maternal factors on child growth may be mediated by (3.1) child factors, (3.2) caregiving practices or (3.3) may directly affect child growth. Effects of infant feeding practices on child growth may be mediated by child factors (4.1) or caregiving practices may directly affect child growth (4.2). Effects of preterm birth or intrauterine growth restriction on later child linear growth status may be mediated by (5.1) postnatal child factors (appetite, illness, haemoglobin (Hb)/iron status, physical activity, stress) or (5.2) may be direct effects. Effects of child factors (appetite, illness, Hb/iron status, physical activity, stress) on child growth (6.1) may be mediated by caregiving behaviour in response to these factors, or (6.2) may be direct effects. In the iLiNS-DYAD-G trial in Ghana (n=1320) and the iLiNS-DYAD-M trial in Malawi (n=869), pregnant women were enrolled at ≤20 weeks of gestation. In the iLiNS-DOSE trial in Malawi (n=1932) and iLiNS-ZINC trial in Burkina Faso (n=3220), infants were enrolled at age 6 and 9 months, respectively. All participants were assigned to receive various doses and formulations of lipid-based nutrient supplements (LNS), or to control groups until age 18 months, when length was measured.15–18 The effects of the interventions on 18-month child length-for LAZ differed across trials, with positive effects in Burkina Faso17 and Ghana,15 but not in Malawi.16 18 For further information, see supplemental methods. In the path analyses reported here, we included all children for whom LAZ at age 18 months was available, comprising 1039 children in iLiNS-DYAD-G, 684 in iLiNS-DYAD-M, 1504 in iLiNS-DOSE and 2619 children in iLiNS-ZINC. Detailed reports of the data collection procedures in each trial have been published elsewhere15–18; therefore, we summarise the procedures for collection of variables that were used in the analyses presented here. Online supplementary table 1 presents further details of the data collection procedures and variable definitions. Figure 2 shows the data collection schedule for each variable in each cohort. Data collection timeline in each cohort. bmjgh-2018-001155supp001.pdf Data on socio-demographic characteristics and maternal anthropometric status were gathered at enrolment. In the two DYAD trials, maternal pre-pregnancy body mass index (BMI) was estimated based on BMI and gestational age at enrolment. Capillary or venous blood samples were collected from mothers and/or children at multiple time points for the assessment of (1) malaria using a rapid diagnostic test, (2) Hb concentration (g/L), (3) biomarkers of iron status, including zinc protoporphyrin (ZPP) concentration (μmol/mol heme), and soluble transferrin receptor (sTfR, mg/L), and (4) biomarkers of inflammation, including alpha-1-acid glycoprotein (AGP; g/L) concentration. In ZINC, known HIV infection was an exclusion criterion, but HIV was not tested; therefore, HIV status of women who were enrolled was unknown. In DOSE, HIV status was also unknown. In DYAD-G, HIV infection was known from antenatal cards and HIV-positive women were excluded. In DYAD-M, women were tested for HIV at enrolment but were not excluded. Maternal and/or child saliva samples in DYAD-M and DYAD-G were collected at several time points for the measurement of cortisol concentration (nmol/L). Maternal self-reported stress was measured in DYAD-M at multiple time points using the Perceived Stress Scale.19 20 Mothers were interviewed regarding depressive symptoms at 6 months post partum in DYAD-M using a locally validated adaptation of the Self-Reporting Questionnaire and in DYAD-G with the Edinburgh Post-natal Depression Scale.21 In DYAD-M, at 6 months post partum, maternal cognition was assessed using digit span forward and backward, verbal fluency, mental rotation and functional health literacy tests.22 Children were visited weekly for morbidity surveillance. At these visits, caregivers were asked whether the child experienced any illness symptoms, including fever, diarrhoea, vomiting, cough, nasal discharge, respiratory distress or poor appetite during the past seven days and/or data collectors measured the child’s auricular temperature. Longitudinal prevalence and/or incidence of diarrhoea, fever, malaria, acute respiratory infection and/or poor appetite were calculated (online supplementary table 1). In DOSE and DYAD-M, physical activity at age 18 months was measured over 1 week with the hip-worn ActiGraph GT3X+accelerometer (Pensacola, Florida, USA).23 Infant feeding practices were assessed at multiple time points through qualitative 24 hours and/or 7-day dietary recall questionnaires.24 In all four trials, developmental stimulation was measured at age 18 months using the Family Care Indicators interview.25 The mother was interviewed with regard to the variety of play materials and activities that adults used to engage with the child in the past three days. In DYAD-G and DYAD-M, gestational age at enrolment was mainly determined by ultrasound and this was used to calculate gestational age at birth. In DYAD-G, infant weight and length were measured within 48 hours of birth or between 3 and 14 days after birth for 87 (9.4%) when the former was not possible. In DYAD-M, infant weight and length were measured within 6 weeks of birth. We estimated length and weight at birth based on LAZ and WAZ measured within 6 weeks of birth, assuming that LAZ and WAZ did not change from birth to the time of measurement. Length and weight for gestational age at birth z-scores were then calculated based on the INTERGROWTH-21st standards.26 In all four cohorts, length was measured at age 18 months. All length measurements were conducted to the nearest 1 mm by teams of two trained and standardised anthropometrists using length boards. We examined the distribution of each independent variable separately by cohort. We log-transformed skewed variables and winsorised outliers to the 1st and 99th percentile. If transformation did not result in a normal distribution, we created a binary variable. All continuous variables were standardised to SD units by subtracting the mean and dividing by SD. We performed exploratory mediation analyses according to the following steps. We refer to figure 3 to describe each step in the mediation analysis. Mediation analysis. The first condition for inclusion in our mediation analysis was that X is associated with Y, represented by c in figure 3, so we examined independent associations between each predictor and 18-month LAZ and dropped any that were not associated at p<0.05. The second condition was that X is associated with M, represented by a. The third condition was that M is associated with Y, represented by b. In this first step, we tested the significance of b and c. Second, we reduced the number of variables measuring the same construct by eliminating variables that were collinear. If two variables were highly collinear (r>0.6), we dropped the one that was less strongly associated with 18-month LAZ. Third, we examined four multivariate models with each category of factors together predicting 18-month LAZ and dropped any that were not associated at p<0.05. If, at this step, no variable was significantly associated when controlling for the others, we retained the one that was most strongly associated with 18-month LAZ. The four multivariate models were (1) all environmental factors together, (2) all maternal factors together, excluding maternal height, (3) all caregiving factors together and (4) all child factors together, excluding gestational age at birth and LGAZ at birth. For each pathway in figure 1, we examined the association between each pair of variables on the pathway to determine which variables were potential mediators. In this second step, we tested the significance of a in figure 3 and dropped any pathways for which p>0.05. In figure 1, unidirectional arrows represent pathways tested. Bidirectional arrows represent associations that were checked for collinearity, but were not otherwise modelled in the path analysis. All analyses up to this point were conducted using SAS V.9.4 (SAS Institute). Next, for each independent variable with potential mediators, we tested the mediation model using Stata V.14.1 (StataCorp) binary mediation program. We ran the multiple mediation model including all potential mediators together, rather than testing each mediator one by one in separate models. In the final path model, we included all pathways for which the indirect association of X with Y through M was significant. If the interaction between X and M was significant at p<0.05, we stratified the sample at the median of the independent variable and tested the indirect association of X through M at both high and low levels of X. If the indirect association was significant at both high and low levels of X, then we retained the pathway in the model; otherwise, we removed that pathway. Finally, we ran the final path model using the sem command in Stata with the mlmv option to estimate the model on the full data set using maximum likelihood estimation for missing values. All models controlled for three covariates: randomly assigned trial group (LNS vs no LNS), child sex and child age at 18-month LAZ assessment. In the final models, we corrected p values for multiple comparisons using the Benjamini-Hochberg correction,27 which is recommended for controlling the false discovery rate in structural equation models.28 We applied the correction separately for each model (ie, for each cohort). If any pathway was not significant at corrected p<0.05, then we did not draw that pathway in the path diagram. For objective 3, we examined the 16 variables that were available in all four cohorts: household asset index, household food insecurity access index, maternal and paternal education, household water and toilet, maternal age, height, and BMI, child diarrhoea and fever prevalence, child 6-month or 9-month Hb and ZPP, mean dietary diversity across time points, variety of play materials and activities with caregivers at age 18 months. We report the R2 in the models with these 16 predictors in each cohort separately and in the pooled model to determine whether the pooled analysis accounted for more variance in 18-month LAZ than the within-cohort models.