UCL Partners Biostatistics Network Symposium "Contemporary Statistical Methods in Medical Research"
The Second Biostatistics Network Symposium on Contemporary Statistical Methods for Medical Research will take place on Thursday 18th October 2012 in the Kennedy Lecture Theatre, UCL Institute of Child Health, 30 Guilford Street, London WC1N 1EH.
All are welcome, but registration is required, with a £15 fee for non-UCL members to cover for coffee and lunch. As the number of places available is limited, please register at your earlier convenience
The symposium will have three themed sessions. The schedule is the following
9:30 - 9.45: Welcome and introduction, Prof Tim Cole (UCL)
9:45 - 11.15: Session 1 Nonparametric and Robust Methods in Medical Research
Chair: Prof Rebecca Hardy (UCL)
9.45-10.45: Prof. Adrian Bowman (University of Glasgow)
How to be flexible: an introduction to models based on smoothness, with applications to the face and the brain
Methods which allow the assumptions of statistical models to be relaxed from linear to simply 'smooth' are widely used and have proved to be very valuable modelling tools. Some popular approaches will be outlined and illustrated on a variety of examples. The thinking involved will be explored in greater detail in two particular application areas. One involves the analysis of three-dimensional surface images in anatomical settings, particularly of facial shape in the context of surgery. The other involves the analysis of spatiotemporal MEG brain images and the identification of dipole patterns associated with brain activity and response to stimulus.
- 10.45-11.15: Dr. Alessio Farcomeni (University of Rome La Sapienza, Italy)
- Longitudinal quantile regression in presence of informative drop-out through longitudinal-survival joint modelling
- In longitudinal studies subjects may be lost at follow-up due to events, like death, which are associated with the outcome of interest. Failure to model drop-out may lead to biased estimates in such cases. A particularly useful approach is represented by Joint Models (JM), which jointly estimate the longitudinal and drop-out parameters, avoiding possible bias due to the non-ignorability of the attrition. Participation to the study is described by a survival model of time-to-dropout.
JM may have a double aim: (i) remove bias in regression on a longitudinal response in presence of informative drop-out and (ii) deal with survival models in presence of measurement error and/or missing data in time-dependent covariates.
While effective, JM focus on the expected value of the longitudinal outcome, which may not always be the summary of interest or effectively remove bias in the survival model, for instance in presence of outliers. Modelling conditional quantiles is a well-known alternative. Quantile regression models are robust with respect to outliers, and are more useful than the mean in many biomedical applications (e.g., when interest lies in one of the tails). Informative missing data are ubiquitous in statistical applications, especially in longitudinal studies, but there are very few approaches to quantile regression with informative drop-out, which are mostly limited to weighting and assume that the drop-out can occur only at one of pre-specified time points. We propose a joint model for a time-to-event outcome and the quantiles of a continuous response repeatedly measured over time.
The quantile and survival processes are associated via shared latent variables. From the longitudinal quantile regression standpoint, our joint model provides a flexible approach to handle informative dropout. From the survival standpoint, the measurement error in time dependent covariates is avoided by considering the predicted quantile rather than the predicted mean. We discuss the meaning and implications of this feature of our model. Simulations and real data applications illustrate and evaluate our method. In particular, we illustrate how ignoring informative drop-out may lead to biased estimates of the quantile regression parameters, and how this can be avoided through our joint model approach.
11.15 am - 11.45 am: Coffee break
11:45-1:15 pm: Session 2 Infectious Disease Modelling
Chair: Dr Gianluca Baio (UCL)
11.45-12.45: Prof. Christl Donnelly (Imperial College London)
Badger culling and the science-led policy challenge
Bovine TB is a disease that is currently costing UK taxpayers £90million a year to control. Last year some 26,000 cattle were compulsorily slaughtered after testing positive for the disease. In 1997 a committee led by (now Lord) Prof John Krebs was asked to review the scientific evidence relating to badger and badger culling to control TB in cattle. The committee recommended a large-scale randomised trial of two badger culling policies. It was subsequently designed, overseen and analysed by the Independent Scientific Group on Cattle TB which issued its final report in 2007. Whether to cull badgers to control cattle disease remains controversial despite considerable agreement on the science base underpinning the discussions. The evidence will be reviewed along putting into context the various arguments being put forward. Finally, questions are posed about the role of science and scientists in the policy making sphere.
- 12.45-1.15pm: Dr. Anne Presanis (MRC Biostatistics Unit, Cambridge)
Key characteristics of infectious disease such as prevalence, incidence and severity are often difficult to measure directly or to estimate from a single study. However, there may be plenty of information indirectly informing these quantities through functions of the parameters of interest. Evidence synthesis methods are therefore increasingly being employed to combine multiple sources of evidence in a Bayesian framework. This requirement is accompanied by a need of systematically criticise such models to understand what parts of the evidence base and model are driving inference. The synthesis of multiple data sources informing common parameters may lead on the one hand to conflicting evidence and on the other to issues of identifiability when evidence is sparse. These aspects of model criticism are illustrated with examples such as estimating HIV prevalence or the severity of influenza.
1.15 pm - 2.30 pm: Lunch
2.30 pm - 4.00 pm: Session 3 Statistical Methods for Physical Activity Data
Chair: Dr Marco Geraci (UCL)
2.30pm-3.30pm: Prof. Murray Smith (University of Nottingham)
The switching regimes model (SRM) arises when a random variable Y is explained in different fashions across alternate regimes, where individuals choose their regime membership. The simplest of these model types has just two regimes, and is due originally to Roy (1951). While Roy's model concerned a binary switch, situations can arise in which multiple regimes are possible, as occurs in our present case. With three or more regimes, we examine the Ordered SRM (OSRM) in which regime-membership is determined by an ordered response connected by observation rules to an underlying latent economic utility variable, the latter can be associated with Y. The literature has discussed the OSRM in terms of the multivariate normal distribution. We extend the model to an arbitrary distributional setting by using a copula approach to model specification. The OSRM is illustrated by applications in health, in particular the modelling of sporting and physical activity data; for example, modelling spell-length duration of activity alongside of the intensity or vigour with such activities are undertaken.
3.30pm-4.00pm: Mr. Francesco Sera (UCL)
The use of accelerometers in population-based studies provides continuous, objective measurements of physical activity (PA) levels. The resulting data are temporal trajectories of PA levels that show marked daily, weekly and seasonal variations. Typically, daily summaries of PA mean level based on accelerometers counts, and time spent in physical activities defined by varying levels of intensity are used to model PA data; however, these summaries do not wholly capture the richness of information available. Functional data analysis (FDA) can be used to better characterise PA patterns in their entirety. For example, FDA can be used to filter signals (low or high pass frequencies) modelling individual physical activity trajectory; to identify daily and weekly periods when children are most or least active using the fitted individual trajectories; and to evaluate temporal effects of factors that explain variations of the physical activity profile via functional analysis of variance (fANOVA) or functional regression. Moreover, functional principal components analysis (fPCA) can be used to extract latent PA profiles that may explain observed correlations between individual trajectories. We have applied FDA to model accelerometer data from the UK Millennium Cohort Study, a population-based national survey. In this talk, we will discuss some computational and inferential problems that we faced whilst using FDA to model a large dataset of individual trajectories, and will present the results of our analyses.
For registration and further information, please send an email to: Dr Marco Geraci
This event is funded by the Education Theme of the UCLH/UCL Comprehensive Biomedical Research Centre.
Past editions: Biostatistics Network Symposium 2011
Page last modified on 26 oct 12 21:03