Modelling longitudinal child growth data in African settings

Abstract

Rationale: With more and more studies in human and biological sciences involving longitudinal, multi-level or hierarchical data, skills in manipulation and analysis of such data have become very essential in understanding public health problems and in guiding public health policy. Longitudinal child growth studies are particularly useful in monitoring child growth and understanding relationships between early childhood growth and later life health outcomes. However, one of the challenges of longitudinal studies is the inevitability of missing data due to missed visits or lost to follow up. Use of appropriate statistical methods that deal with missing data in longitudinal physical growth measurements and also take into account the correlations in measurements is thus very essential in understanding these relationships. Aims: The main aims of the thesis were to apply mixed effects modelling and various advanced statistical methodologies to longitudinal physical growth data from 2 African growth cohorts in order to: identify biological growth curves that best fit childhood physical growth measurements in these African settings, identify statistical methods that efficiently deal with missing data in physical growth measurements and, then explore the relationship between postnatal growth velocity and early adolescent obesity in the 2 cohorts. Methods: The study used physical growth measurements from the Birth to Twenty (BT20 - an urban South African cohort) and from the Lungwena Child Survival Study (a rural cohort from Malawi). There were differences in the intensity of the data collection waves in the 2 cohorts. Several parametric and non-parametric growth curves were fitted to height and weight measurements from birth to 10 years, using Linear Mixed Effects (LME) modelling. Both cohorts were modelled from birth to around 10/11 years. However, there were shorter intervals between data collection waves in the Lungwena than the BT20 cohort. Several goodness of fit statistics were used to compare how well the different curves fitted to the data. The parameter estimates of the Berkey-Reed model , which was found to fit better to the data than the other models, were then used to compare the efficiency of using Multiple Imputation, Regression Imputation or using available case analysis (ACA) methods to deal with missing growth data. The study used LME models as an ACA method. Lastly, the study further used LME models to derive growth velocity curves and then used logistic and multiple linear regression models to explore the relationship between postnatal growth and adolescent obesity. Results: Identification of growth curves: In comparing how the different human growth models fitted to the 2 cohorts, the study found that the Berkey-Reed 1st order model fitted well to both weight and height measurements in both cohorts compared to other growth models. Overall, the fitness of different models was affected by length of time between data collection waves, especially in the first year of life, as evidenced by smaller residuals in the Lungwena cohort, which had data points that were closer together than the BT20 cohort. There was improved model fit when there were more data points in early years (birth to 2 years), because this allowed for better capturing of the expected. The number of data points in early years also affected predictions of initial weight/length (birth weight/length) by the models, as evidenced by better prediction in the Lungwena cohort. There were also variations in precision of the estimated initial weight or height by the different models. In general, most models failed to pick out the pre-puberty rapid growth (at 7-9 years). Overall, there was better fit to height measurements than weight measurements due to the monotonic nature of height measurements. Human growth models are monotonic functions, primarily derived to model monotonic biological processes. However, individual weight fluctuates and is more sensitive to changes in ecological and environmental factors that affect growth. Dealing with missing physical growth measurements In comparing methods of dealing with missing data in longitudinal studies, the study found that there were no significant differences in the growth model parameter estimates derived after MI or using regression imputation or when using LME modelling, which uses all available information. However the efficiency of MI or LME was affected by the length of the period between data collection waves. Bias in the estimated parameters was consistently affected by the number of data points (amount of information from each child), with the Lungwena cohort parameters having reduced bias because of the larger number of data points. There was also more bias in MI values if imputation model used did not take into account the individual child’s growth profile (i.e. the longitudinal aspect of the data). The regression imputation method produced smaller standard errors than the ACA-based LME method, due to the increased number of observations created through the imputation process. Relationship between infant growth and early adolescent obesity Having found no significant gain in using Multiple Imputation or regression imputation in growth curve modelling, the study used LME modelling (which allows for missing data ) to examine the relationship between early child growth and adolescent obesity. LME is simple to use in growth curve modelling, especially in deriving other growth parameters such as peak weight/height velocities or time at peak velocity. The study found that there were significant differences in growth between the 2 cohorts, shown by the differences in the growth model parameters and weight/height growth velocities. BT20 boys and girls exhibited higher growth rates than their Lungwena counterparts. The differences between the 2 cohorts were also highlighted by the changes in relationships between growth parameters when models were adjusted for inherent cohort differences. However, no significant differences were observed between boys and girls within each cohort. No significant relationship was found between size at birth (birth weight) and adolescent obesity, even after taking into account inherent cohort differences. Rapid growth in infancy, independent of size at birth (birth weight) was highly associated with high BMI in early adolescent. In general, the risk of being an overweight adolescent increased with increase in growth velocity. The relationship between growth velocity and adolescent body mass index (BMI) was strongest for infant rather than childhood growth velocity. There was a general decrease in the strength of the relationship between weight velocity and adolescent BMIZ over time even after adjusting for birth weight, with the strongest relationship observed in infancy. Adolescent obesity was also associated with age at peak velocity, with infant that reached peak velocity early having higher risk of being obese in adolescence. Conclusions: Shorter intervals between data collection waves in the first 24 months of life (a period of general rapid growth) will lead to better fit of the growth models. Thus, for optimal study of infant and early childhood growth using these types of growth models, it is recommended to have measurements at least every 3 months. There is no gain in using MI or Regression Imputation in dealing with intermittent missing data in physical growth measurements in early childhood (birth to 10 years), especially if the time intervals between data collection waves are short. Available Case Analysis using LME method can produce sufficient and unbiased results. The method allows for analysis of unbalanced repeated measures that might arise by study design or due to missing data and is also simpler to use than MI. Regression Imputation, which also uses LME method to predict values, has the advantage of increasing the number of observations used, and thus helps in increasing the precision of the parameter estimates (reduced standard errors). Overall, rapid weight gain in infancy is highly associated with adolescent obesity. However the effect of rapid growth can have different health outcomes depending on what stage of nutritional transition the population is in. For a rural population that is still in the early stages of nutritional transition, rapid weight gain in infancy may have beneficial effects as it protects an adolescent child from the effects of under-nutrition, with children who experience rapid growth having reduced risk of stunting and under-weight. For an urban population in later stages of nutritional transition, rapid weight gain in infancy has detrimental effects, which exacerbates the effects of adolescent over-nutrition, thus increasing the risks of adolescent obesity. The study highlighted the diversity in nutritional problems that exist in Africa as a continent and the need to understand each country in terms of stage of nutritional transition, when designing public health interventions and also how other countries in the continent can learn from South Africa in mitigating the effects of over-nutrition.