Why Do Children in Slums Suffer from Anemia, Iron, Zinc, and Vitamin A Deficiency? Results from a Birth Cohort Study in Dhaka

Considering the high burden of micronutrient deficiencies in Bangladeshi children, this analysis aimed to identify the factors associated with micronutrient deficiencies and association of plasma micronutrient concentration trajectories from 7 to 24 months with the concentrations at 60 months of age. Plasma samples were collected at 7, 15, 24, and 60 months of age, and hemoglobin, ferritin, zinc, and retinol concentrations of 155, 153, 154, and 155 children were measured, respectively. A generalized estimating equation was used to identify the factors associated with micronutrient deficiencies, while latent class growth modeling identified the trajectories of plasma micronutrients from 7 to 24 months and its association with the concentrations of micronutrients at 60 months was examined using multiple linear regression modeling. Early (AOR = 2.21, p < 0.05) and late convalescence (AOR = 1.65, p < 0.05) stage of an infection, low ferritin (AOR = 3.04, p < 0.05), and low retinol (AOR = 2.07, p < 0.05) were associated with increased anemia prevalence. Wasting at enrollment was associated with zinc deficiency (AOR = 1.8, p < 0.05) and birth weight was associated with ferritin deficiency (AOR = 0.58, p < 0.05). Treatment of drinking water was found protective against vitamin A deficiency (AOR = 0.57, p < 0.05). Higher trajectories for ferritin and retinol during 7–24 months were positively associated with plasma ferritin (β = 13.72, p < 0.05) and plasma retinol (β = 3.99, p < 0.05) at 60 months.


Latent class growth modeling (LCGM) and multiple linear regression models
We used latent class growth modeling (LCGM), also called group-based trajectory modeling, to identify distinct clusters or classes of children following similar trajectories with regard to the pattern of hemoglobin, ferritin, retinol and zinc during the age of 7 to 24 months. LCGM is a semi-parametric, finite mixture modeling technique which analyzes longitudinal data using maximum likelihood to identify meaningful and distinct groups of individuals who follow similar progression over time for a given variable. LCGM relaxes the assumption that all individuals are drawn from a single population and allows for differences in growth parameters across unobserved subpopulations. However, it assumes that the intercept and slope are fixed for all individuals within each distinct group [4][5].
We built separate trajectory models for hemoglobin, ferritin, retinol and zinc. LCGM requires at least three measurement time-points for each case to generate reliable parameter estimates of trajectories [6]. Therefore, the analyses were restricted to the children for whom data on the outcomes were available at 7, 15 and 24 months (three time-points). In addition, for ferritin, three very large and unusual values were dropped from the dataset. The sample sizes for LCGM were reduced to 155 for hemoglobin, 153 for ferritin, 154 for retinol and 155 for zinc.
We built and compared several models for each outcome with 1 to 4 trajectories. A censored normal distribution approach was used for modeling all three outcomes. To identify the optimal number of trajectories, we fit models of increasing complexity starting from a single trajectory and finalized the model that best fit the data. In identifying the distinct trajectories of the outcomes, linear and quadratic functions of time (age in months) were examined. Non-significant quadratic terms were removed from a model for a given trajectory. Models generating trajectories with insufficient cluster size (less than 5% of the study population) were not considered.
We selected the final models with optimal number and shape of trajectories based on Bayesian information criteria (BIC), log Bayes factor, the statistical significance of quadratic terms, whether 95% confidence intervals of trajectories overlapped, and the percentage of the population in each trajectory group. The smallest absolute value of the BIC indicated the best fit.
A value greater than 6 for the estimated log Bayes factor, which is equal to two times the difference in the BIC values calculated by subtracting the BIC of the simpler model from that of the more complex model, was interpreted as strong evidence for the more complex model [7].
After selecting the final model, we calculated the posterior probabilities for each individual of belonging to each of the trajectory groups, and individuals were assigned to a trajectory group based on the maximum-probability assignment rule [8].
We used the following criteria to assess the goodness of fit of the final models: whether the average posterior probability of assignment was greater than 0.8 for each of the subgroups, whether the odds of correct classification was greater than 5, and whether the estimated/modeled group probabilities were in good agreement with the proportions of group assignments [27]. We reported the findings of the LCGM following The GRoLTS-Checklist: Guidelines for Reporting on Latent Trajectory Studies [9].
In subsequent analyses, multiple linear regression models with robust standard errors were fitted to examine the association of levels of hemoglobin, ferritin, retinol and zinc at the age of 60 months with the trajectories identified through LCGM. As zinc was found to have a single trajectory (described in detail in the results section), level of zinc at 24 months was used instead of any trajectory to assess the predictive association with level of zinc at 60 months. We considered several covariates collected at the age of 60 months in building these models which include sum of energy in Kcal for the day of food recall, sum of protein in grams for the day of food recall, sum of fat in grams for the day of food recall, sum of carbohydrates in grams for the day of food recall, sum of iron in mg for the day of food recall, sum of Vitamin A in ug for the day of food recall, sum of zinc in mg for the day of food recall, phytate to iron ratio of the diet, phytate to zinc ratio of the diet, energy in Kcal from carbohydrates (minus fiber) as percent of total energy, energy in Kcal from protein as percent of total energy, energy in Kcal from fat as percent of total energy, socioeconomic and WASH variables including WAMI score, and child sex. However, the final regression models included only iron intake, the percentage of energy from protein, and WAMI score for hemoglobin, level of zinc at 24 months, total energy intake, protein intake and WAMI score for ferritin, vitamin A intake and WAMI score for retinol, and zinc intake and WAMI score for zinc. Due to missing values for the outcome variables, the sample sizes for the linear models were reduced to 142 for hemoglobin, 138 for ferritin and retinol and 140 for zinc.
The statistical analyses related to LCGM were performed using the "traj plugin" in Stata (StataCorp, College Station, Texas 77845 USA, version 14.1) [10], a Stata equivalent of the widely used "proc traj" in SAS [28]. Outputs of LCGM models were plotted using "traj" in Stata and the R packages "lcmm" and "ggplot2" in R (version 3.5.1). Generalisability 21 Discuss the generalisability (external validity) of the study results 17-20

Other information
Funding 22 Give the source of funding and the role of the funders for the present study and, if applicable, for the original study on which the present article is based 21 *Give information separately for exposed and unexposed groups.
Note: An Explanation and Elaboration article discusses each checklist item and gives methodological background and published examples of transparent reporting. The STROBE checklist is best used in conjunction with this article (freely available on the Web sites of PLoS Medicine at http://www.plosmedicine.org/, Annals of Internal Medicine at http://www.annals.org/, and Epidemiology at http://www.epidem.com/). Information on the STROBE Initiative is available at http://www.strobe-statement.org.