Recent advances in the application of compositional data analysis: Bayesian modelling for longitudinal compositional data and evolutionary algorithms for optimal time-use compositions

For our next talk, in the Digital Phenotyping seminar series, we will hear from Dr Dot Dumuid, Senior Research Fellow, Alliance for Research in Exercise, Nutrition and Activity, University of South Australia and Flora Le, PhD student, Sleep and Circadian Rhythm program, Monash University, on 12 June, 2:00 pm – 3:00 pm, at the Big Data Institute (BDI).

Title: Recent advances in the application of compositional data analysis: Bayesian modelling for longitudinal compositional data and evolutionary algorithms for optimal time-use compositions

Date: Wednesday 12 June
Time: 2:00 pm – 3:00 pm
Venue: BDI/OxPop Seminar Room 0; followed by refreshments in the atrium

Daily time must always sum to 24 hours, thus time-use data are compositional. Compositional data analysis has been applied to explore the relationships between reallocations of time and a large range of health measures. In our talk, we will present recent advances in the application of compositional data analysis to –
(1) a Bayesian framework: disentangling between- and within-person effects in longitudinal models, and
(2) quality diversity evolutionary algorithms: identifying optimal time-use compositions.

Multilevel compositional data, such as repeated measures of the 24-hour movement behaviours in intensive, longitudinal studies, are common, yet analytically challenging. We present a novel methodology for analysing multilevel compositional data using Bayesian inference. This method enables investigation of the longitudinal association between reallocation of movement behaviours and other phenomena (e.g., emotion, cognition) at both the between-person and the within-person levels. We explain the theoretical framework and software implementation of this novel method in our R package, multilevelcoda, with a user-friendly setup only requiring the data, model formula, and minimal analysis specification. We also demonstrate its application in a real data example and robust performance via a simulation study.

We demonstrate the application of evolutionary algorithms within a compositional log-ratio framework, considering a four-part time-use composition with four health objectives (adiposity, fitness, life satisfaction and cognition). Objective functions linking the time-use compositions with health outcomes are defined in the log-ratio space. To ensure solutions are feasible and to avoid extrapolating out of the space covered by the sampled data, we add a quadratic constraint to the feasible solution space using the distributional (multivariate Gaussian) quantiles of the sampled data. We employ a quality diversity approach (MAP-Elite algorithm) which randomly crosses two ‘parents’ (solutions), thus generating an offspring on which the mutation operator is applied. If the offspring solution for similar behaviour descriptors (time-use composition) is better in terms of meeting the objective, it replaces the existing solution. This process is iteratively repeated until termination criteria are met. A distinguishing feature of the MAP-Elite algorithm is that competition takes place among solutions with similar behaviour descriptors, increasing the diversity among solutions. We visualise the objective functions, feasible solution space and optimal time-use solutions in an interactive web application. Finally, we provide an example of how we have implemented time-use optimisation methods in a real-world intervention study, and demonstrate an interactive online research translation interface we have co-designed with the target population.

Flora Le is a final year PhD student in the Sleep and Circadian Rhythm program, Monash University. Her research interest lies in understanding the effects of balancing daily sleep-wake behaviours on daily psychological experiences (e.g., stress and affect) and long-term sleep and mental health (e.g., insomnia, depression, and anxiety). In collaboration with her supervisors, Flora has developed an R package for Bayesian multilevel compositional data analysis to enable a streamlined, efficient workflow in analysing longitudinal sleep-wake behaviours.

Dr Dot Dumuid is a Senior Research Fellow at the Alliance for Research in Exercise, Nutrition and Activity, University of South Australia. Her research seeks to identify the healthiest way to spend our time across daily activities such as sleeping, sedentary behaviours and physical activity. She brings together analytical expertise from diverse disciplines to explore how to achieve the best balance of these activities for health and wellbeing. Her funded projects seek to implement the new knowledge in behavioural interventions for children and young people, adults at risk of chronic disease and the elderly, including those in residential care.
Hybrid Option:
Please note that these meetings are closed meetings and only open to members of the University of Oxford. Please respect our speakers and do not share the link with anyone outside of the University. The purpose of these seminars is to foster more communication among employees throughout the University, so we strongly advise in-person attendance whenever feasible.
Microsoft Teams meeting
Click here to join the meeting
Meeting ID: 377 786 547 645
Passcode: qdc9ku
If you wish to know more or receive information related to training and events at BDI, please subscribe by emailing You’ll then receive an email from SYMPA and once you reply you’ll be on the list!