Statistical modelling approaches for longitudinal multiple outcome data from immuno-epidemiological studies in Entebbe, Uganda

Abstract

Background Longitudinal studies are indispensable for studying changes in variables over time, and are increasingly common in many fields of research. However, in comparison with other study designs, longitudinal studies have some practical challenges including increased time requirements and associated costs. Many challenges arise when analysing data from longitudinal studies including multivariate measures with complex correlation and random error structures, time-varying covariates and missing data. Such challenges increase the complexity of longitudinal data analysis, and this is particularly the case for immuno-epidemiological studies. Immuno-epidemiological studies investigate the influence of immune responses on the epidemiology of conditions such as infectious diseases, cancer, hypersensitivity and autoimmunity. These studies have increased recently as a result of technological developments in sample processing and analyte measurement allowing for the simultaneous measurement of many immunological parameters, generating data for the exploration of both simple and complex associations between disease, immunity, social, environmental, genetic or other factors. Consequently, the complexity of associations between several immunological parameters poses a major challenge of how to select the most appropriate statistical approach to extract the maximum relevant information from such complex datasets whilst avoiding spurious findings. A small number of journal articles have given an overview of application of statistical techniques to immuno-epidemiological data, however, these focus mainly on cross-sectional data. Published guidance on the analysis of longitudinal immunological data remains limited, particularly for longitudinal data with multivariate outcomes. This PhD project therefore aimed to investigate and extend appropriate statistical modelling methods for longitudinal data, and to apply them to longitudinal multiple-outcome data from immuno-epidemiological studies conducted in Entebbe, Uganda. Methods

This PhD study utilised secondary data from two immuno-epidemiological cohort studies, carried out in Entebbe, Uganda. The Infant BCG study (IBS), an observational longitudinal study, aimed to examine the timing, magnitude and quality of the initial response to BCG immunisation and to determine the effect of prenatal exposure to maternal latent Mycobacterium tuberculosis infection (LTBI) on these outcomes. The Entebbe Mother and Baby Study (EMaBS), a birth cohort study, aimed to investigate the impact of worms and their treatment on responses to immunisation, illness and allergy-related diseases (ARDs). For the IBS, latent variable modelling approaches involving dimension reduction including

principal component analysis (PCA) and three-way component analysis, in combination with mixed effects models, were employed to study the evolution of multivariate cytokine responses over time and how that evolution depends on infant characteristics such as maternal LTBI status and age, sex, birthweight or other covariates. For the EMaBS, latent class analysis (LCA) was employed to characterise the infection experience of study participants and how this relates to susceptibility to ARDs at 9 years. To handle missing data, multiple imputation and direct likelihood approaches were incorporated in the analyses for both the IBS and EMaBS. Results The IBS showed that there was remarkably high early sensitisation to mycobacterial antigens in utero, in Uganda, but this sensitisation had no impact on the infant response to immunisation with BCG. PCA in cord blood showed IL-10 and TNF tending to group separately from the other cytokines, but there was no evidence that maternal LTBI was associated with a differing pattern of response, or that differences in these cord blood profiles impacted the subsequent infant response to BCG. The pairwise joint modelling approach was applied to the IBS data in a situation where fitting of a full multivariate mixed model had failed due to model complexity. Parameter estimates from the pairwise approach had better precision than those from the univariate mixed models. Initial patterns from PCA of cord blood responses were not reflected in the profile of responses that developed post-BCG as indicated by the PCA on the correlation matrix of random intercepts. This improved our interpretation of the data by suggesting that probably the effect of BCG immunisation was potent enough to over-ride patterns established in utero. LCA of data from the EMaBS identified two distinguishable groups of children, during each of the first five years of life, based on their early infection-exposure. In the first year of life, latent class (LC) 1 was characterised by higher probabilities of malaria, diarrhoea and LRTI, and membership of this class was associated with lower proportions of wheeze, eczema, rhinitis and SPT positivity for Dermatophagoides at 9 years. Associations between LC membership and ARDs were less consistent for infections in years two to five, suggesting that infancy may be a critical period during which the risk of ARDs in later childhood is established. Conclusions This PhD study adapts and applies recently developed robust statistical analysis methods in novel analyses of longitudinal immuno-epidemiological data with case studies from Entebbe, Uganda. Notably, this work shows that maternal latent M.tb infection is not the reason for reduced effectiveness of BCG in tropical countries and therefore, all infants are likely to benefit from BCG immunisation regardless of their maternal LTBI status. This work also shows that exposure to common infections, especially during the first year of life, was associated with lower proportions of ARDs in later childhood. The statistical analysis approaches used, including the pairwise joint modelling approach and latent class analysis, improve our understanding and interpretation of longitudinal immuno-epidemiological data.