Seminars, 2019/20
Department of Statistics, University of Leeds

Seminars are normally held on Fridays during term time at 2:00. Here is the list so far for 2019/20. The abstracts have been collected at the end of the page.

List maintained by John Kent . Last updated 26 November 2019

Friday 27 September 2019 at 2:00

Speaker: Dr Jongrae Kim
Affiliation: School of Mechanical Engineering, University of Leeds
Title: Nonlinear projection filter (almost exact nonlinear estimator) with negative-free transformation
Location: RSLT 11
Click for Abstract

Friday 4 October 2019

No seminar scheduled

Friday 11 October 2019

No seminar scheduled

Friday 18 October 2019

No seminar scheduled

Friday 25 October 2019

No seminar scheduled

Friday 1 November 2019 at 2:00

Speaker: Dr Rebecca Killick
Affiliation: University of Lancaster
Title: A Bayesian Circular Changepoint Method to Identify Changes in Daily Activity Levels in the Elderly
Location: MALL 1
Click for Abstract

Friday 8 November 2019 at 2:00
(POSTPONED, due to illness)

Speaker: Dr Michael J. Crowther
Affiliation: Department of Health Sciences,University of Leicester
Title: Extended mixed effects regression models for linear and non-linear outcomes
Location: MALL 1
Click for Abstract

Friday 15 November 2019 at 2:00

Speaker: Dr Frank Dondelinger
Affiliation: Lancaster University
Title: Predictive models for progression in neurodegenerative diseases
Location: MALL 1
Click for Abstract

Friday 22 November 2019

No seminar scheduled

Friday 29 November 2019 at 2:00

Speaker: Dr Theo Damoulas
Affiliation: University of Warwick
Title: Multiresolution Multitask Gaussian Processes: Air quality in London
Location: MALL 1
Click for Abstract

Friday 6 December 2019 at 2:00

Speaker: Jacob Cancino-Romero and Kathryn Laing
Affiliation: PhD students, University of Leeds
Title: Jacob: Causal joint models for the relationship between frailty, recurrent falls and mortality in the elderly
Kathryn: Preprocessing for Efficient Dominance Testing in CP-nets
Location: MALL 1
Click for Abstract

Friday 13 December 2019 at 2:00

Speaker: Prof Janine Illian
Affiliation: University of Glasgow
Title: In a world of many processes — what’s the point?
Location: MALL 1
Click for Abstract

Abstracts for the seminars

27 September 2019; Jongrae Kim; Nonlinear projection filter (almost exact nonlinear estimator)

Abstract: The Fokker-Planck equation is the governing equation of the nonlinear estimation problem. The nonlinear projection filter presented in the late 90's provides almost exact solution of the nonlinear estimation problem, which calculates the probability density function (pdf) directly unlike the Kalman filter estimates only the first two moments. In this talk, I present computational methods to implement the filter and highlight a potential of the filter to some practical problems. There are two main computational obstacles in the filter implementation: the computational complexity and the non-negativeness of the pdf. i) The filter requires high-dimensional integrations and the complexity of the filter increases exponentially with the dimension of systems. Over the past few decades, the computational power has been increasing rapidly. With advances of the parallel computation architectures it provides new opportunities for calculating multi-dimensional integration in real-time. All integrations required in the filter implementation are naturally parallel and directly implemented on a parallel computing architecture. ii) The estimated pdf must satisfy the two essential properties of pdf: the integration of pdf over the whole sampling space is equal to 1 and the value of pdf in the sampling space is greater than or equal to zero. The first constraint can be easily achieved by the normalisation. On the other hand, it is very hard to impose the non-negativeness in the sampling space. In the pdf estimation, some areas in the sampling space might have negative pdf values. It produces unreasonable moment values such as negative probability or variance. A negative-free transformation is developed to guarantee the non-negativeness of the pdf over a chosen sampling space and the nonlinear projection filter is re-formulated with the transformation. The effectiveness of the filter is demonstrated using a simple example and a battery state-of-charge estimation problem.

Return to the main seminar list

1 November 2019; Rebecca Killick; A Bayesian Circular Changepoint Method to Identify Changes in Daily Activity Levels in the Elderly

Abstract: According to Age UK there are 11.6m over 65's, 3.64m who live alone and over 25% need help with at least one daily activity. A growing body of research indicates that changes in daily routine signal a change in health and well-being. Motivated by an industrial collaboration with Howz, we are using motion sensors in the home to automatically detect these changes as a key step in improving outcomes for elderly people and vulnerable members of society who live alone. Traditional changepoint methods to identify changes in activity levels view time as linear and thus are able to identify the day/night routine on a day-to-day basis. The typical assumption of independence of segments results in estimated changepoints and parameters that are unlikely to be consistent from day-to-day. As changes in routine happen on longer time scales, traditional methods make determining the normal daily patterns more challenging. We demonstrate a new changepoint method in the Bayesian framework that views time as circular in order to estimate the time-of-day changepoint events between different activity levels by pooling together information from across multiple days. These daily patterns can then be monitored for significant changes in daily changepoint locations and/or parameter estimates within segments.

Return to the main seminar list

8 November 2019; Michael Crowther; Extended mixed effects regression models for linear and non-linear outcomes

Abstract: Multivariate data occurs in a wide range of fields, with ever more flexible model specifications being proposed, often within a multivariate generalised linear mixed effects (MGLME) framework. In this talk, I’ll describe some current work developing an extended framework, encompassing multiple outcomes of any type, each of which could be repeatedly measured (longitudinal), with any number of levels, and with any number of random effects at each level. Many standard distributions are described, as well as non-standard, custom, non-linear models. The extension focuses on a complex linear predictor for each outcome model, allowing sharing and linking between outcome models in an extremely flexibly way, either by linking random effects directly, or the expected value of one outcome (or function of it) within the linear predictor of another. Non-linear and time-dependent effects are also seamlessly incorporated to the linear predictor through the use of splines or fractional polynomials. I’ll further discuss level-specific random effects distributions and numerical integration techniques to improve usability, relaxing the normally distributed random effects assumption to allow multivariate t-distributed random effects. I’ll consider some special cases of the general framework, describing some new models in the fields of clustered survival data, joint longitudinal-survival models, and discuss various applications in the health domain. User friendly, and easily extendable, software is provided in both Stata and R.

Return to the main seminar list

15 November 2019; Frank Dondelinger; Predictive models for progression in neurodegenerative diseases

Abstract: Identifying factors that influence the clinical progression of neurodegenerative diseases is of critical importance to both experimentalists trying to understand the disease mechanisms, and clinical researchers trying to develop improved therapies. While much effort has gone into the detection of risk factors for a given disease, most of these approaches ignore the inherent variability in the clinical phenotypes. We have developed two approaches for dealing with this variability: a penalised linear regression method using fusion penalties to share information across phenotypes, and a longitudinal high-dimensional mixture model approach for jointly solving the problem of data-driven estimation of clinical phenotypes and prediction of disease progression. We demonstrate the performance of both our methods by applying it to data from the PROACT database on amytrophic lateral sclerosis, as well as data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI, Mueller et al., 2005). We show that information sharing across phenotypes allows for improved prediction performance, and joint inference of the subtypes and predictors leads to the inference of meaningful subtypes, and hence produces potentially clinically useful insights.

Return to the main seminar list

29 November 2019; Theo Damoulas; Multiresolution Multitask Gaussian Processes: Air quality in London

Abstract: We consider evidence integration from potentially dependent observation processes under varying spatio-temporal sampling resolutions and noise levels. We offer a multi-resolution multi-task framework, termed MRGPs, while allowing for both inter-task and intra-task multi-resolution and multi-fidelity. We develop shallow Gaussian Process (GP) mixtures that approximate the difficult to estimate joint likelihood with a composite one and deep GP constructions that naturally handle scaling issues and biases. By doing so, we generalize and outperform state of the art GP compositions and offer information-theoretic corrections and efficient variational approximations for inference. We demonstrate the competitiveness of MRGPs on synthetic settings and on the challenging problem of hyper-local estimation of air pollution levels across London from multiple sensing modalities operating at disparate spatio-temporal resolutions.

Return to the main seminar list

6 December 2019: Jacob Cancino-Romero and Kathryn Laing

Abstract (Jacob Cancino-Romero): (joint work with Stuart Barber, Leonid V. Bogachev, Jeanne Houwing-Duistermaat) Joint modelling of longitudinal and time-to-event data is often used to explore the joint distribution of multiple outcomes. The complexity of joint models conveys challenges for statistical analysis and are computationally demanding. This strategy has been an area of active research for description and prediction, but causal inference has received less attention. In this work we investigate the causal relationships between three outcomes in the elderly: geriatric frailty, recurrent falls and mortality. We formulate two causal diagrams and their corresponding statistical joint model with the aim of de- termining if the effect of frailty on mortality differs between the two alternatives: (1) a shared frailty joint model where the three outcomes are linked by random effects, and (2) a joint model for frailty and mortality treating frailty as an endogenous time-varying covariate subject to measurement error and falls as an exogenous time-varying covariate for mortality. We explore with simulation studies the consequences for causal inference conclusions when the wrong model is used to analyze the data. This work is motivated by the Yorkshire & Humber Community Ageing Research 75+ study (n = 282) conducted in Northern England. The aim of the study is to understand the causal connections between frailty, recurrent falls and mortality in order to guide the early detection of treatable problems and intervene with the goals of improving quality of life and reducing the risk of mortality. Results of model (1) suggest effects of frailty on recurrent falls (p = 0.0004) and mortality (p = 0.03), but no effect of recurrent falls on mortality.

Abstract(Kathryn Laing): Conditional preference networks (CP-nets) are structures used for modelling a user's conditional preferences. These models are primarily of interest to the AI community for applications such as recommender systems and automated decision making as they allow us to reason about user preferences. A dominance query over a CP-net asks, given a pair of outcomes, which does the user prefer? Answering such queries is important for preference reasoning but has been shown to be a complex task. In this talk, I will give an introduction to CP-nets and the problem of dominance testing. In particular, I will illustrate our method of improving dominance testing efficiency by pre-processing the CP-net.

Return to the main seminar list

13 December 2019: Janine Illian

Abstract: In recent years point process methodology has become increasingly familiar to ecologists. Spatial point processes have been originally developed within mathematical statistics as stochastic processes that have spatial point patterns as realisations. Nowadays, there has been a shift towards using them as a tool for modelling the locations of objects or events in space (and time) in practical scientific applications. This shift implies a change in the aims of the statistical analysis and in the focus of the associated statistical research.

Within mathematical statistics, point processes form part of stochastic geometry and hence the aim is to define mathematically tractable processes that best mimic the geometric properties of a certain type of point pattern. In the context of applied statistics, however, the main aim of a modelling exercise is to answer scientific questions. Hence appropriate inference, model interpretation and model assessment are of primary concern. As a result, very different - and new - questions need to be answered throughout the modelling process. Statistical approaches that do this well still need to be developed.

This talk explores this shift in focus, reviews recent progress in making point process methodology more relevant in practice and highlights opportunities for research. In the light of this, we discuss issues concerning model assumptions, model construction and model assessment and draw on a number of concrete examples from ecology and beyond for illustration.

Return to the main seminar list