Looking beyond the mean: what within-person variability can tell us about dementia, cardiovascular disease and cystic fibrosis

Longitudinal studies are studies where the same information from the same units of interest or the same individuals are collected over time. For example, in longitudinal studies of human beings, information from biological quantities of interest (for instance, cognitive function, blood pressure, lung capacity) is collected repeatedly over time from the same individuals. This permits an understanding of how these quantities change over time on average for each individual, and also how measurements vary and deviate over time from each person's average trajectory.

Traditionally, researchers interested in these studies have focused on the persons' average curve over time. For example, modelling the average trajectory of blood pressure or the average trajectory of cognitive functioning. However, emerging research suggests that large deviations from a person's average curve (intraindividual variability) may be indicative of an underlying pathological process. For example, individuals whose blood pressure is sometimes very high and some other times very low may be suffering from some underlying condition that may require attention. Similarly, those whose memory is sometimes very good and at other times very poor, may need to visit a doctor for a check-up. Hence, gaining a better understanding of intraindividual variability may offer a window of opportunity to better understand what is going on with these individuals and whether they are at higher risk of negative health events. We have identified three disease areas in which the study of intraindividual variability is highly relevant: blood pressure variability and the link with cardiovascular disease, variability in scores assessing cognitive function and their link with dementia, and the link between lung function variability and mortality in cystic fibrosis patients.

However, existing methods to measure intraindividual variability and to identify factors that may be associated with it are very limited. Furthermore, the effect of ignoring intraindividual variability in standard statistical methods, or using inappropriate statistical methods, is still unclear. An incomplete understanding of how intraindividual variability affects results generated from commonly used statistical methods in the three disease areas of interest may result in misleading findings. Therefore, we aim to generate knowledge about intraindividual variability, focusing on the three critical health research areas, cardiovascular disease, ageing and dementia and cystic fibrosis, that will help researchers improve existing knowledge. With this aim in mind, we propose to: (1) evaluate existing resources and statistical methods to understand intraindividual variability; (2) develop new methods to improve our understanding of intraindividual variability and apply them to our three areas of interest (cardiovascular disease, ageing and dementia and cystic fibrosis); (3) combine the new methods with statistical methods for predicting the probability of an event occurring, such as a cardiovascular event, a dementia diagnosis or death.

We will conduct the proposed research in leading institutions including the MRC Biostatistics Unit, University of Cambridge, and the Edinburgh Dementia Prevention Research group, University of Edinburgh. The work will be done collaboratively between the groups and supported by experts in the three disease areas. We will benefit from an experienced team and facilities in both institutions to maximise opportunities to disseminate our findings to researchers, practitioners and the wider research community.

The aim of this programme is to improve understanding of the role played by within-individual variability around mean trajectories of longitudinal outcomes in three key disease areas: cardiovascular disease, dementia and ageing, and cystic fibrosis. We use multi-outcome models with a mixed-effects location scale sub-model for longitudinal trajectories and a time-to-event sub-model for disease outcomes.

We will initially review existing methods and software for modelling within-individual variability and evaluate the performance of these methods using simulation studies. We will then develop new methodologies motivated by current clinical and epidemiological questions in our three key disease areas, including models for dealing with multiple longitudinal outcomes, models using standard Gaussian processes to distinguish between long-term variability, short-term variability and measurement error, change-point models, and multi-state models. Finally, we will embed the new methods into dynamic risk prediction models to allow statistical measures of intraindividual variability to be included as predictors. We will utilise (1) a landmarking approach and (2) a joint modelling approach, which both allow the risk prediction of the probability of some event occurring to be updated dynamically in response to additional longitudinal information.

All methodological developments will be accompanied by companion papers in clinical and epidemiological journals. We will make use of existing data resources, including long established cohort studies and registry data. Code with illustrative examples will be disseminated via a dedicated github repository to encourage uptake of our methods.