Efficient and effective use of longitudinal data

Information gathered in medical studies is often underused. Sometimes, this is because of problems with data collection, where participants decline to provide answers to certain questions, or decide to discontinue treatment before the end of a drug trial. Other times, a lack of simple tools to extract relevant findings can restrict a researcher‘s ability to delve deeply into the data.

Much existing data is longitudinal, where study subjects are followed over several weeks, months or years. Longitudinal data is particularly valuable because it is possible to observe changes in health conditions resulting from life events or clinical interventions. One convenient approach to understanding longitudinal data is to focus on these changes, using statistical models to separate out the predictable effects of past lifestyle and treatment decisions from the unpredictable aspects of human biology and behaviour.

This fellowship will provide statistical tools to make fuller use of longitudinal studies. These tools are designed to deal simply with the uncertainty surrounding incomplete patient records, allowing meaningful conclusions to be drawn about disease prognosis and progression. Re-examining existing data using these new tools has considerable potential to improve human health through better scientific understanding.

In order to reason causally from medical studies, a minimal requirement is that the temporal ordering of events is explicitly recognised. As this principal became established there was a corresponding increase in the number of research projects recording many variables over several waves of observation. Nevertheless, some statistical analyses of this type of data still fail to recognise the critical role of time. Consequently, existing data sources contain a wealth of information and a rich temporal structure that has not been fully exploited.

Introduced in my PhD thesis, the linear increments model focusses on observed changes in measured quantities, and hence offers a natural framework for acknowledging study chronology. This fellowship will develop and disseminate the linear increments approach, arguing for its routine use with longitudinal designs. An important feature of the incremental approach is that missing responses (due to dropouts, for example) pose no particular analytical problem, other than the unavoidable reduction in sample size.

Several facets of the work will serve to make this statistical technology more relevant to medical practitioners. One proposed development will more neatly distinguish aspects of the model that are of principle interest from those of secondary importance, just as in a proportional hazards model the effect of treatment on survival is emphasised over the precise baseline pattern of events. Another extension will allow inference to centre on surviving patients, rather than on a manifestly hypothetical immortal cohort. Often, multiple outcomes may be clinically pertinent, and so a third aspect of the fellowship will be to enable exploration of the dynamics of related variables.

An important goal of the fellowship will be to make the case that the linear increments model is the natural analogue of the popular Kaplan-Meier approach to event times. To this end, a unifying mathematical theory of dynamic observation of stochastic processes will be sought, incorporating at one extreme continuous survival, and discrete panel data at the other. Equally crucial will be the successful application of the methodology to substantive questions arising in randomised trials, observational studies and epidemiological research.