Prof. Peter Diggle, University of Lancaster
The choice of approach to the analysis of longitudinal data can, and arguably should, be informed by the purpose of the modelling. At the risk of over-simplifying, a broad distinction can be drawn between whether population-averaged or subject-specific effects are of primary scientific interest; see, for example, Diggle, Heagerty, Liang and Zeger (2002). In this talk, I will briefly review some widely used methods of longitudinal data analysis before describing a case-study in renal medicine. Progression towards end-stage renal failure is typified by an asymptomatic period that can extend over many years, but can be assessed indirectly by bio-chemical analysis of blood-samples. Also, early diagnosis and treatment can materially slow the rate of progression and so postpone the need for expensive and invasive renal replacement therapy (dialysis or transplantation). An important question is therefore: how can we use information that is easily obtained through routine blood-testing to provide early warning of end-stage renal failure? To address this question, I will formulate a dynamic linear model for a large data-set in which several thousand subjects in both primary and secondary care settings have had their renal function measured imperfectly, at irregular sequences of time-points, and use the model in conjunction with a Kalman filter algorithm to enable real-time updating of the predictive probability that an individual subject’s underlying renal profile has crossed a clinically defined intervention threshold. Finally, I will suggest that statistical modelling has an important role to play in so-called e-health research and give some other examples of related work-in-progress at Lancaster and Liverpool.
