Methods for modelling repeated measures in a lifecourse framework

Doctors are increasingly interested in long-term influences on health. For example, it is thought that people who were born small as babies, and grew faster during childhood, may be more likely to suffer cardiovascular disease in later life. Longitudinal studies ? studies in which individuals are followed over periods of many months or years ? are of great importance in understanding how aspects of people?s childhood lifestyle or environment influence their later health and wellbeing. When one measure (e.g. weight) is made several times on the same individual, the values at different ages are likely to be related. This raises difficult issues in the analysis of data from longitudinal studies, and failing to address these appropriately can lead to results that are biased (they differ from the results that would be observed if the analysis had been appropriate) or lead to inappropriate conclusions (the results of the analyses are interpreted incorrectly). Statistical methods that do address these issues have been proposed, and have the potential to decrease bias and increase ease of interpretation in analyses of longitudinal studies. However, these methods can be highly complex and difficult to apply. We will develop solutions to some of the problems with applying these methods, including developing strategies for modelling change over time (e.g. growth in childhood), and relating this change over time to later outcomes. We will incorporate our new methods into existing software, to maximise their future use, as well as publishing the results in scientific journals.

There is increasing emphasis in medical research on fetal and childhood antecedents of disease, and how these interact with other exposures throughout the lifecourse to influence later-life conditions. Answering questions about the relative importance of aspects and timing of growth, behaviour and health status for longer term outcomes requires longitudinal data. Analysis of such data must account for dependencies between repeated observations on the same person, including serial autocorrelation (greater correlation among measurements closer together in time). Where there are repeated measures of exposures related to a later-life outcome, standard regression models may be affected by multicollinearity. Measurement error may vary over time, and there will usually be dropout over time (due to death, illness, etc). The proposed research will focus on developing methods for more robust analysis of lifecourse data, by tackling these problems. First, we will improve the modelling of trajectories of change over time. We will develop methods for detecting and modelling autocorrelation (e.g., autocorrelation around a growth spurt or a period of illness). The impact of the assumption of Normality and constant variance in modelling growth trajectories will be examined, and methods for re-parameterising the variance function developed. We will then develop (and include in standard statistical software) methods for selecting the best-fitting parameterisation of growth curves. Second, we will tackle problems inherent in modelling the relationship between correlated exposures and later outcomes. The role of partial least squares in reducing the dimension of multiple correlated exposures will be examined. There are often several competing lifecourse hypotheses related to any given problem, and we will develop methods for distinguishing between these. We will develop (and include in standard statistical software) methods for including the uncertainty in the estimation of growth trajectories or latent classes of growth when using them as exposures in a model for later outcome. Finally, we will explore the joint modelling of exposures and outcomes using cross-lagged models, and the impact of modelling the initial conditions of such models. We will compare our methods to others (including simple lifecourse approaches, path analysis) to answer scientifically important questions pertaining to the ALSPAC study. Data from other longitudinal studies available to the co-applicants, and simulated data, will be used to test the generalisibility of our methods and conclusions. We will develop guidelines for deciding when simpler methods are likely to be useful and unbiased, and when more statistically sophisticated methods would be more appropriate.