Early Life Risk Factors for Childhood Adiposity – a Comparison of Marginal Growth Models Using Hong Kong's “Children of 1997” Birth Cohort
Risk factors for childhood obesity identified were largely based on cross-sectional data analysis considering obesity at one single time point. Recent development of growth modeling techniques has facilitated the investigation from a longitudinal perspective, but no systematic comparison of marginal models has been performed. The main purpose is to determine risk factors for growth at a population average level and the influence of different statistical approaches.
METHODS:
Body mass index (BMI) from a large population-representative Chinese birth cohort, Hong Kong’s “Children of 1997” was transformed to sex- and age- specific z-scores. Maximum likelihood estimation (MLE) under the assumption of normally distributed data, generalized estimating equations (GEE), and quantile regression (QR) for estimation of the median growth in relation to early life risk factors for childhood adiposity were compared. The models were compared with and without multiple imputation (MI) for missing BMI and risk factors under the assumption of missing at random.
RESULTS:
The distribution of BMI z-scores seemed symmetric and did not strongly deviate from normality. However, the three methods differed in the estimates, standard errors, significance levels, and interpretation. Most notably MLE seemed to deviate from GEE and QR, while GEE and QR seemed more similar, irrespective of the use of imputation or not. For example, a subgroup of children was identified for which the MLE demonstrated a decline in BMI z-scores over time of 0.15, while GEE and QR showed an increase of 0.2.
CONCLUSIONS:
The results did not indicate one optimal method for a population-averaged interpretation. Given the lack of a “gold standard” statistical method, we suggest assessing the consistency of conclusions with all these methods when analyzing longitudinal data to eliminate possible discrepancies due to statistical methods. Further studies will be needed to determine the best method for analyzing this type of longitudinal growth data.