BELFAST

2-5 SEPTEMBER 2019

RSS International
Conference
ABSTRACTS BOOKLET
Headline sponsor

Visit:
rss.org.uk/conference2019
#RSS2019Conf
Abstracts are ordered in date, time and session order for oral
presentations followed by the poster presentations

1
1.1 Contributed - Medical: Data linkage
Tuesday 3 September 9am

Risk prediction models that use routinely collected electronic health data:
generalisable and useful in heterogeneous settings?

Yan Li, Matthew Sperrin, Miguel Belmonte, Alexander Pate, Darren Ashcroft, Tjeerd Jeerd
Farr Institute for Health Informatics Research University of Manchester

Objective: To assess the extent of variability between practices on individual patients’ risk of
cardiovascular disease (CVD) that is not taken into account by the risk prediction model
QRISK3.

Methods/Design: Longitudinal cohort study from 1st Jan 1998 to Jan 2015.Setting: 392
general practices (including 3.6 million patients) from the Clinical Practice Research Datalink
(CPRD)Methods: Shared frailty model to incorporate QRISK3 predictors, practice variability
and simulations to measure random variability.

Results: There was considerable variation in data recording between general practices.
Practices on 5th percentile of missingness of Body mass index have 18.7% patients with
missing values and 60.1% on the 95th percentile (for ethnicity, these were 19.6% and
93.9%, respectively). The crude incidence rates also varied considerably between practices
(from 0.4 to 1.3 CVD events per 100 patient-years, respectively). The estimates of individual
CVD risks with the random effect model were inconsistent with the estimated QRISK3 risk.
For patients with a QRISK3 CVD risk of 10%, the 95% range of predicted risks were
between 7.2% and 13.7% with the random effects model. Random variability only explained
a small part of this. The random effects model was similar to QRISK3 for discrimination (C-
statistic: 0.852 (95% CI: 0.850, 0.854)) and calibration (Brier score: 0.067 (95% CI: 0.067,
0.068)).

Conclusions: Risk prediction models that use routinely collected electronic health data can
have limited generalisability and accuracy in predicting individual patient risks in
heterogeneous settings. They need to be based on more robust evidence on causal risk
factors.

2
1.1 Contributed - Medical: Data linkage
Tuesday 3 September 9am

Education and health outcomes of children treated for chronic conditions

Michael Fleming1, James McLay2, Danel Mackay1, Jill Pell1

1
University of Glasgow, 2 University of Aberdeen

Objectives: This retrospective cohort study linked Scotland-wide education data to national
health data to explore associations between childhood chronic conditions and subsequent
educational and health outcomes. Conditions studied were diabetes, asthma, epilepsy,
attention deficit hyperactivity disorder (ADHD) and depression. We also explored
neurodevelopmental multimorbidity (comorbid autism, learning disability, ADHD or
depression). Educational outcomes were school absenteeism and exclusion, special
educational need, academic attainment and subsequent unemployment. Health outcomes
were hospital admissions and all-cause mortality.

Methods: Pupil census data and associated education records for all children attending
primary and secondary schools in Scotland between 2009 and 2013 were linked to national
prescribing data, acute and psychiatric hospital admissions, death records and retrospective
maternity records enabling conditions to be studied whilst adjusting for sociodemographic
and maternity factors and comorbid conditions. Conditions were ascertained from
prescribing data and school records.

Results: All conditions were associated with increased school absenteeism, special
educational need, and hospitalisation. All, excluding diabetes, were associated with poorer
academic attainment and all, excluding ADHD were associated with increased mortality.
ADHD and depression were associated with increased exclusion from school whilst epilepsy,
ADHD and depression were associated with subsequent unemployment. Children
experiencing neurodevelopmental multimorbidity had poorer outcomes across all
educational domains. Depression was the biggest driver of absenteeism and ADHD was the
biggest driver of exclusion.

Conclusion: In addition to poorer health outcomes, schoolchildren with these chronic

conditions appear to experience significant educational disadvantage compared to their
peers. Therefore interventions and further understanding of the intricate relationships
between health and education among children with these conditions is required.

3
1.1 Contributed - Medical: Data linkage
Tuesday 3 September 9am

Evaluating the effects of an admission avoidance home visiting programme for frail,
elderly patients in NHS Forth Valley

Maria Cristina Martin1, Matt-Mouley Bouamrane2, Kim Kavanagh1, Paul Woolman3

1
University of Strathclyde, 2 University of Edinburgh, 3 NHS Forth Valley

An increasing proportion of hospital inpatients are elderly, leading to the development of

alternatives avoiding admission. Such interventions are often complex with several
interconnecting parts. Their evaluation can prove challenging, with randomisation often
being unfeasible. Routinely collected healthcare data provide an opportunity for evaluation in
real-world settings, but carry several limitations. This research uses routinely collected
retrospective data on elderly patients receiving admission avoidance home visits in NHS
Forth Valley (FV). This submission presents the results of a matched cohort analysis and the
difficulties of using real-world data for evaluation in this setting. Deterministic data linkage
was used to compile an analytical dataset, linking several local and national datasets
containing demographic and hospital activity data, for the entire elderly population in FV
through Structured Query Language (SQL). The “Matching” R package was used to match
intervention patients and controls. The outcome variable of whether an inpatient admission
occurred or not in the year following intervention was evaluated using multivariate logistic
regression.A cohort of 61,467 patients (566 of these intervention patients) were included.
Intervention patients were found to be significantly different to general elderly patients on
several variables, selected as matching variables for a control group (age, gender, Charlson
comorbidity group, care home stay, living alone, and number of inpatient admissions
experienced in the year prior to the intervention based on a dummy index date allocated to
control patients) with 2:1 allocation. 477 intervention patients were successfully matched to
954 controls. Intervention patients were found to be more likely to experience an admission
(OR 3.07, 95% CI 2.43-3.87) despite adjustment for residual differences. Despite matching,
baseline hospital admissions for the two groups differed; hence, the selected matching
variables may be insufficient for selecting an appropriate control group. Such issues in using
routinely collected data for the purpose of evaluation will be discussed.

4
1.2 Contributed - Official and Public Policy: Quality and value in official statistics
Tuesday 3 September 9am

Hospital Standardised Mortality Ratio - Improving quality and safety of Scottish

hospitals through the exploration and modelling of mortality

Robyn Munro, Lucinda Lawrie, David Caldwell

NHS NSS

OBJECTIVES:Information Services Division has produced quarterly Hospital Standardised

Mortality Ratios (HSMR) for Scottish hospitals since December 2009. Data on hospital
mortality have an important role to play in improving the quality of patient care. Over recent
years there has been considerable debate about the utility and limitations of these
measures. However, used wisely and as part of a wider suite of measures on quality and
safety, these data can help drive improvements in patient care. In order to maximise the
utility of HSMR, ISD have undertaken research to: Review the model methodology to ensure
it continues to be appropriate and relevant;Interpret and disseminate key patterns shown in
the data

METHODS:The HSMR is calculated by obtaining routinely collected death certificate data.

These crude mortality data are adjusted to take account of some of the factors known to
affect the underlying risk of death.The HSMR is calculated as: HSMR = Observed Deaths/
Predicted DeathsThe observed number of deaths is the total number of patients who died
within 30-days of admission to hospital. The predicted number of deaths is calculated from a
case-mix adjusted model.Using a three year dataset, logistic regression analyses were
performed in order to test the relationship between explanatory variables and the outcome
(whether the patient was alive or dead within 30 days). This involved fitting a model to data,
evaluating fit, and estimating parameters that are later used in the prediction equation.

CONCLUSIONS:The existing HSMR methodology is robust and helps drive improvements in

patient care; however to ensure comparisons which are made against the national average
continue to be appropriate and relevant for each point in time the following recommendations
were made:Regularly update the three year reference period used to model the relationship
between explanatory variables and 30-day mortality.Update specific variables used for risk
adjustment.

5
1.2 Contributed - Official and Public Policy: Quality and value in official statistics
Tuesday 3 September 9am

Reproducible Analytical Pipelines for Health and Social Care Publications

Jack Hannah, Anna Price

NHS National Services Scotland

Background: The Information Services Division (ISD) of the National Health Service
Scotland produces approximately 200 health and social care publications annually. Most
publications are produced using proprietary software such as SPSS, Business Objects and
Microsoft Excel. Production of data and reports is time and labour intensive, involving
extensive manual formatting and checking, and multiple movements of data between
software. The Transforming Publishing Programme aims to modernise the publication
process by creating Reproducible Analytical Pipelines (RAPs).

Methods: RAP combines the concept of reproducible research with data science best
practices. It improves the quality, auditability and speed of publication production, as well as
ensuring knowledge transfer in organisations with high turnover in staff. Our team focused
on one publication as a proof of concept, developing an R package to produce the Quarterly
Hospital Standardised Mortality Ratios (HSMR) publication. Git and GitHub were used for
version controlling and facilitating collaborative working. We then developed a buddy system
and toolkit to help other teams automate their reports, including an R style guide, project
templates and RMarkdown publication templates. We also developed levels of code maturity
and automation to help teams select an appropriate level to aim for, based on their skills and
available development time.

Results and Conclusions: A standard RAP project includes each step of the publication
production process: extraction from databases using SQL; data wrangling; documentation
and unit testing of functions; and production of the final report using RMarkdown. The
process of scaling RAP widely within the organisation has been facilitated by a buddy
system whereby a RAP champion, who has previously worked on a RAP project, supports
analysts from another team by providing expert advice, training and code reviews. This
system has enabled the successful development of further RAPs at ISD and the opportunity
for staff to develop code and new skills, with input and guidance from the Transforming
Publishing team.

6
1.2 Contributed - Official and Public Policy: Quality and value in official statistics
Tuesday 3 September 9am

Dependent on Dover? Estimating and visualising the value of EU trade by UK port of

entry for consumer goods

Jonathan Lewis
Civil Service

If a UK port were to suddenly become unusable, what is the most optimal route that a
container of Radio equipment from Germany would take instead to import its products. What
is the value of all Toys coming through Dover from Calais, and from which countries and EU
ports did they originate. These are the questions that this project is aiming to answer, by
developing a model that combines Economic and Statistical techniques to estimate the most
optimal routes for products from over 20 EU directives, from every European Country, into
the UK. We have then combined these probabilities with existing HMRC data to produce a
Transport Mapper, which estimates how much trade is likely to be arriving into each UK port
for each type of product.The model estimates which types of goods will prioritise time over
cost and vice-versa in deciding the most optimal route, and a beta value will be used to
adjust the sensitivity to using alternative routes on the final output. This evidence base does
not yet exist, and could be very useful in understanding our trade flows with the EU into each
UK port. It can for instance predict the next best alternative route to Dover in the expected
scenario that there is disruption at Dover port.An interactive Transport Mapper dashboard
has been developed which will allow policy makers to view those ports with the highest
estimation of trade by product, along with the country of origin and which EU port it arrived
from. This Dashboard will be live at the presentation to showcase the work.

7
1.3 Contributed - Environmental / Spatial Statistics:
Tuesday 3 September 9am

Combining citizen science and survey data in a log-Gaussian Cox process framework
to estimate the monthly space-use of Southern Resident Killer Whales

Joe Watson, Marie Auger-Methe

University of British Columbia

Species distribution models (SDMs) are useful tools to help ecologists quantify species-
environment relationships, and they are being increasingly used to help determine the
impacts of future climate and habitat changes on species. Estimating SDMs can be tricky
from a statistical point of view since the effects of spatial and temporal autocorrelations, land
cover and environmental covariates and detectability functions all need to be considered and
inherently modeled. Furthemore, such models often assume that data have been collected
from well-designed surveys and/or studies. In practice, data are often of the form of
presence-only sightings collected from 'citizen scientists' and/or industry, and their 'search
effort' can be difficult to quantify. Furthermore, search effort from such sources is often
concentrated in areas in which the expected count of the species under study is high, and/or
where population density is high. Ignoring the search effort can lead to severely biased
estimates of the species distribution.We look at data collected on Southern Resident Killer
Whales (SRKWs), an ecotype with designated 'species at risk' status found off the coast of
Vancouver Island. Data from from a variety of 'citizen science' sources and government
surveys are considered. We present a method to combine the different data sources and
estimate the monthly SRKW space-use in a statistically-rigorous manner using spatio-
temporal log-Gaussian Cox processes within the R-INLA and inlabru packages. Improved
(effort-corrected) estimates of the SRKW distribution will hopefully help ecologists and
policy-makers safeguard the future of the SRKW.

8
1.3 Contributed - Environmental / Spatial Statistics:
Tuesday 3 September 9am

Detecting and Communicating changes in Waste Water Treatment Plant performance

in Ireland

Jason Larkin
Environmental Protection Agency Ireland

A small team at the Environmental Protection Agency regulates Waste Water Treatment
Plants (WWTPs) in Ireland. We collect hundreds of thousands of environmental monitoring
data points each year relating to the performance of these plants. The Analytics team is
striving to turn this hard monitoring data into information about how each plant is doing, and
how it is impacting the wider environment.This presentation will show how we scaled from a
small investigation into the performance of a single WWTP, to generating a set of metrics
that measure the performance of each of ~500 WWTPs. These metrics use simple statistical
tools like t-tests and sign-tests to summarise the monitoring data.It will also show how we
communicate our metrics in an r-shiny dashboard.

9
1.3 Contributed - Environmental / Spatial Statistics:
Tuesday 3 September 9am

Understanding model fit for simulating species dispersal using alternative cost
metrics

Laura Merritt1, Justin Travis2, Steven White3, Tom Oliver4, Rob Salguero-Gomez5, James
Bullock3
1
University of Reading/Centre for Ecology and Hydrology, 2 University of Aberdeen, 3 Centre
for Ecology and Hydrology, 4 University of Reading, 5 University of Oxford

Calculating measures of dispersal such as ‘wave-speed’ through integrodifference equations

relies on the fitting of probability density functions called dispersal kernels. However, these
estimates are particularly sensitive to the underestimation of rare long-distance dispersal
events. Visual inspection of graphs showing fitted dispersal kernels compared to the original
count data seemed to suggest consistent underestimation of long distance dispersal events.
The objective of this study was to statistically quantify the difference in fit between these rare
events, and shorter dispersal distances. Due to the magnitude of difference in counts
between the first and last measured points, a novel cost metric was developed based on
proportional differences between observed and predicted values. We conclude that rare
long-distance dispersal events are being underestimated in dispersal kernels, while there is
little error in shorter distances. We therefore suggest methods for fitting that can limit this
underestimation and provide more accurate predictions of species dispersal speeds.

10
1.4 Contributed - Social Statistics: Neighbourhoods
Tuesday 3 September 9am

Life at the Frontier: Conceptualising the Causes and Consequences of Ambient Social
Frontier Propensity

Gwilym Pryce
University of Sheffield

Quantitative research on segregation has focussed primarily on measuring the overall

degree of separation within and across neighbourhoods, rather than on understanding the
nature of the spatial transition from one neighbourhood to another. Yet, it is the nature of the
residential interface between separate communities that might be of greatest importance.
This paper aims to extend the theory of “social frontiers” and their impacts. Social frontiers
are defined sharp spatial divisions in the residential make-up of adjacent communities (Dean
et al. 2018), as opposed to more gradual blending of groups. Such frontiers represent
invisible “cliff edges” in the socio-economic geography of neighbourhoods. We conceptualise
social frontier formation through the lens of residential sorting approaches and reflect on
their impacts on human wellbeing through a variety of complementary mechanisms. A series
of hypotheses are proposed regarding the potential impact of proximity to social frontiers on
educational attainment and life outcomes. We argue that, to properly account for such
effects, we need to consider a number of important overlooked aspects of social frontiers,
including the ambient propensity toward social frontiers in the surrounding areas and in
wider society. This leads us to propose the concept of “Ambient Frontier Propensity” and to
suggest a Bayesian statistical framework for estimating it.

11
1.4 Contributed - Social Statistics: Neighbourhoods
Tuesday 3 September 9am

Career Satisfaction, Work Resources and Health of Employees and of Their Children:
Evidence from 1,883 Chinese Dual-Earner and Only-Child Households

Chunyi Chen, Mengjie Xu, Liang Guo

Shandong University

Prior studies have paid considerable attention to the relationship between work and family
and there is a consensus that work can affect employee’s mental and physical health.
However, little attention has been devoted to analyzing how parents’ work affects their
children’s health. This study tends to fill in this gap. Our samples include 1,883 dual-earner
and only-child households in China. We follow the path analysis approach to estimate a
series of regression models. Our study reveals that perceiving to gain valuable resources at
work significantly promotes not only the health of working parents (0.077, p<0.01) but also
the health status of their children (0.058, p<0.05). We also find that career satisfaction was
positively related to better parental health (0.094, p<0.01), but it has a long-arm, negative
effect on children’s health (-0.070, p<0.01). These effects are partially mediated by parental
health. Our study contributes to the work-family literature in the following ways. Firstly, we
provide both theoretical and empirical justifications for the fact that parental career
satisfaction promotes parental health but jeopardizes the health of their children. Secondly,
we find that work resources promote the health of both parents and children. And finally, our
sample size is the largest of its kind, representing both urban and rural households working
in multiple industries. The findings are robust and may be generalized across different
social-economic status.

12
1.4 Contributed - Social Statistics: Neighbourhoods
Tuesday 3 September 9am

Neighbourhood change in Britain, 1971-2011

Chris Lloyd1, Gemma Catney1, Paul Norman2, Nick Bearman3

1
Queen's University Belfast, 2 University of Leeds, 3 University College London

The paper brings new perspectives on major demographic and socioeconomic trends
witnessed in Britain over the last 40 years – counterurbanisation and return to cities, a north-
south divide in population growth, the decline of social housing, increases in overcrowding,
the decline of manufacturing and mining, and the growth of the service sector. The analysis
makes use of 1km gridded counts derived using Census data for 1971, 1981, 1991, 2001
and 2011 for Britain. The data were produced as a part of the ESRC-funded ‘PopChange’
project. The project entailed transfer of counts from source geographies (enumeration
districts or output areas) to 1km grid cells to produce geographically-consistent data. The
analyses presented are part of a large programme of work which has explored changes in
population densities, deprivation, health, housing tenure, and housing spaces. The paper
outlines the approaches used to construct the dataset before moving on to assess major
geographical trends and local complexities in terms of age, country of birth, qualifications,
employment, industry, housing tenure, and overcrowding. The analyses reveal marked
changes in the geography of age profiles between 1971 and 2011, with a decline in the
diversity of age groups in many neighbourhoods, including much of London. Other major
temporal patterns include increased overcrowding in the south east of England, and
considerable declines in manufacturing and mining - illustrated in far more geographical
detail than in previous analyses. Common changes (variables which increase or decrease
together) are explored using multidimensional scaling of correlation matrices for percentage
point change variables. A final stage presents a classification schema developed using -k-
means, allowing exploration of how neighbourhoods change in multiple ways over the 40-
year period covered. The paper concludes by considering how the analyses can be
extended beyond 2011, and it also highlights ways in which analyses of neighbourhood
change can be vital in assessing the needs of local communities.

13
1.5 Contributed - Methods and Theory: Causal Inference
Tuesday 3 September 9am

Bayesian nonparametric estimation in longitudinal mediation: A Baron-Kenny based

estimator for cross-lagged models

Andrej Srakar
Institute for Economic Research (IER), Ljubljana and Faculty of Economics, University of
Ljubljana

There has been a significantly raised interest in causal inference in both statistics and
econometrics in recent decades. Mediation analysis has its roots in the literature of structural
equation models, both in the context of linear (MacKinnon, 2008; Glynn, 2012) as well as
nonlinear (Rubin, 1974; Holland, 1988; Halpern, 1998; Pearl, 2001) models. Limitations of
cross-sectional models to analyze mediation effects have been pointed in the literature and
can be overcome by longitudinal modelling (Gollob and Reichardt, 1987; Selig and Preacher,
2009; Bernal Turnes and Ernst, 2016) which is the particular reason for studying longitudinal
mediation, apart from it being a highly uncommon and underresearched methodology. Also,
existing models for longitudinal mediation are estimated under strong parametric
assumptions which imposes statistical problems (Bernal Turnes and Ernst, 2016).We derive
a nonparametric Bayesian estimator for cross-lagged longitudinal mediation models (based
on »classical« Baron-Kenny approach to mediation, see Baron and Kenny, 1986). As
longitudinal mediation for cross-lagged models demands a dynamic panel modelling having
lagged variables in the nonparametric part, we follow Su and Lu (2013) using iterative local
kernel-based approach with sieves as initial estimators. To map to a Bayesian
nonparametric "space" we use objective Dirichlet prior, recommended in the literature
(Alvares, Armero and Forte, 2018) while also exploring in Bayesian nonparametric literature
more commonly used Dirichlet Process Mixtures. We show the constructed estimator attains
optimal information rate (Alaa and van der Schaar, 2018). We explore the properties of the
approach in a simulation study, comparing the performance to the existing (parametric)
estimators for cross-sectional and longitudinal mediation. In a short application, we study the
mediated effects of cultural policy funding on the performance of nascent cultural firms using
Amadeus firm-level data in the period 2007-2016.The article is the first exploration of
Bayesian nonparametric approach in longitudinal mediation literature providing this field with
important statistical considerations for its future development.

14
1.5 Contributed - Methods and Theory: Causal Inference
Tuesday 3 September 9am

Interpreting estimates of mediated effects from studies with attrition: an example from
a study of maternal depression and child neurodevelopment.

Nicola Fitz-Simon1, Alina Rodriguez2

1
National University of Ireland Galway, 2 University of Lincoln and Imperial College London

Depression during pregnancy is common, affecting about 15% of pregnant women.

Depression has been linked to compromised foetal development, including the foetal brain,
via biological mechanisms, and is related to behavioural, emotional, and cognitive problems,
i.e. neurodevelopmental deficits, in childhood.We analysed a longitudinal study of births in
Sweden that recruited n=3,046 pregnant women at antenatal clinics in 1999 and 2000. We
found a positive association (adjusted for baseline covariates) between maternal depression
during pregnancy and child behavioural problems at age 5. By considering hypothetical
interventions on alternative mediating pathways via postnatal maternal mental health and
infant temperament, we estimated interventional direct and indirect effects, using Monte
Carlo simulation. We found the adjusted association between pregnancy depression and
child behaviour at age 5 was largely explained by postnatal maternal depression, thus
suggesting little evidence of a foetal effect. However, this study was affected by attrition,
with 5 year follow-up data available for only 55% of the cohort. Maternal depression, both
during pregnancy and the postnatal period, was associated with 5-year attrition. Other
variables associated with attrition were markers of socioeconomic position, parity, and
stressful life events, which are potential confounders of the mediator-outcome associations.
To investigate the impact of attrition on the mediated effect estimate, we did a series of
sensitivity analyses. We found that an unmeasured confounder of plausible magnitude, that
also predicts attrition, can give rise to a biased estimate of the mediated effect. Moreover,
this bias is increased with increasing attrition. Mediated effect estimates in this study with
high attrition must be interpreted with caution.

15
1.6 Contributed - Medical: Risk Factors
Tuesday 3 September 9am

Evaluation of two-part models for semi-continuous patient reported outcome

measures: an application to a clinical trial of lower back pain.

James Griffin1, Ranjit Lall2, Jane Hutton1

1
Department of Statistics, University of Warwick, 2 Warwick Clinical Trials Unit, University of
Warwick

Patient reported outcome measures (PROMs) are extensively used to assess patients’
perspective on impairment and pain in clinical studies and medical practice; and are often
the primary outcome in randomised controlled trials of new treatments. PROMs sometimes
have a point mass at one end of the scale, and skewed values over the remaining values
and lead to heteroscedastic errors and non-constant variance. Hence conventional linear
regression approaches; even with mathematical transformations applied, are invalid. In
econometrics such distributions are termed “semi-continuous” as they are neither exclusively
discrete nor continuous.We consider a two-part model approach to address these problems
in semi-continuous PROMs. Econometricians have used such models for decades but there
have been very limited application in medical statistics and clinical trials. Two-part models
require treating the observed PROM values as arising from two distinct stochastic
processes; a binary outcome taking value 0 if the score was zeroes and 1 if the score was
non zero (the first part); and another governing the non-zero values conditioning on a non-
zero response (the second part).We compare specifications of two-part models to predict
treatment effects using the Roland Morris Disability Questionnaire in the Back Skills Training
Study. Models studied included logistic and probit models for the first part, with transformed
regression and generalised linear models for the second part.We present an approach to
assess model fit, prediction errors and the validity of model assumptions which extends
traditional residual checking approaches for conventional one part regression models. We
also discuss how to derive and evaluate treatment effects in the complex setting of a two-
part model and demonstrate the inadequacies of single number summaries.Our results
demonstrate that two-part models are an interpretable and practical family of models and
with careful consideration they should be an attractive option to analyse semi-continuous
PROMs.

16
1.6 Contributed - Medical: Risk Factors
Tuesday 3 September 9am

Quantifying effects of some socio-demographic risk factors on Lyme disease

incidence in Scotland.

Jude Eze1, Roger Evans2, Harriet Auty1, Rita Ribeiro1, Sally Mavin2, Roger Humphry1, Gunn
George1
1
SRUC, 2 NHS Highland

Problem: Lyme disease (LD) is a bacterial infection transmitted to humans through a bite by
Ixodes ricinus tick infected with Borrelia burdgeferi. It is the most significant vector-borne
disease in Europe and America with increasing number of cases reported annually. If left
untreated, LD could lead to debilitating disease and serious morbidity with associated human
and health costs. Variations in the incidence of LD over time and space may depend,
among other factors, on the socioeconomic and demographic characteristics of the
population. These characteristics may be strongly associated with behaviors and activities
which expose individuals to risk of infection. LD prevalence studies have focused mainly on
environmental and climatic factors. Few studies have indicated differences in prevalence
rates among age groups and sex. This study aims at using advanced statistical models to
obtain adjusted estimate of effects of patient’s socioeconomic and demographic
characteristics after accounting for correlation, trend and seasonal effects.

Methodology: Laboratory confirmed cases of LD between 2006 and 2017, with patient
demographics, were supplied by National Lyme Borreliosis Testing Laboratory (NLBTL),
Scotland. Using Scottish Index of Multiple Deprivation (SIMD) quintiles (1 = most deprived to
5 = least deprived) as proxy for socioeconomic status, we developed a multivariable logistic
mixed effect model that accounted for within health board correlation.

Findings: Incidence of LD was strongly positively associated with patient’s age and
socioeconomic status. Increasing age by one year increased the odds of infection by 15% in
females and 9% in males. Odds of infection among the most affluent were 1.33 times that of
the most deprived. Being a male increased the odds of infection by 1.76 times relative to
being female.

Conclusion: Being older, or male, or on the higher rungs of socioeconomic status, means
greater odds of LD infection. This highlights the need to target individuals with these
characteristics in any campaign to reduce LD infection.

17
1.6 Contributed - Medical: Risk Factors
Tuesday 3 September 9am

Incorporating misclassification error from finite mixture models into generalised

linear models: an illustration from serological survey of RSV in England

Ania Zylbersztejn1, Lucy Pembrey2, Harvey Goldstein1, Guy Berbers3, Rutger Schepp3,
Fiona van der Klis3, Charles Sande4, Dan Mason5, Rosalind Smyth1, Pia Hardelid1
1
UCL Great Ormond Street Institute of Child Health, London, UK, 2 London School of
Hygiene and Tropical Medicine, London, UK, 3 National Institute of Public Health and the
Environment (RIVM), Bilthoven, The Netherlands, 4 Kemri-Wellcome Trust Research
Programme, Kilifi, Kenya, 5 Bradford Institute for Health Research, Bradford, UK,

Background: Bronchiolitis due to respiratory syncytial virus (RSV) is a leading cause of

hospitalisations in infants. Community-based serological surveys of RSV can inform vaccine
strategies but require costly, time-consuming and invasive repeated blood sampling. Instead,
we estimate burden of RSV through secondary analysis of stored blood samples, using a
two-stage modelling strategy based on finite mixture models (FMM).

Methods: We used stored blood samples, tested for RSV IgG antibodies, linked to
questionnaires, primary and secondary care records from the Born in Bradford cohort study,
collected at 1 and 2 years of age. We fitted FMM to antibody data and classified children as
infected according to their posterior probability of infection. We estimated risk ratios for
infection according to age, ethnicity, date of blood sample and indicators of population
mixing (e.g.: childcare attendance) using Poisson regression with robust error variances. To
account for possible misclassification of infection status, we simulated infection status
according to the posterior probability of infection from the FMM and re-estimated risk ratios
50 times. We pooled the estimates of risk ratios using Rubin’s rules.

Results: The study included 477 children. 249 (52%) children had serological evidence of
RSV infection in infancy; 94 of whom (38%) had been in contact with healthcare during RSV
season. Date of sampling, having older siblings, Pakistani ethnicity, and attending formal
childcare were predictive of RSV infection in infancy. The infection status was well defined –
462 (97%) children had posterior probability of infection <10% or >90%. On average, 248
children (standard deviation=2) were indicated as infected across simulated datasets.
Pooled risk ratios were comparable with the main analyses.

Conclusions: This is the first community-based serological survey of RSV in England.

Secondary use of stored serological data combined with appropriate statistical analyses
provide a time-efficient and cheap method for carrying out serological surveys to determine
the community-burden of RSV.

18
1.7 Contributed - Data Science: Words
Tuesday 3 September 9am

Feature2Vec: Distributional Semantic Modelling of Human Property Knowledge

Steven Derby1, Paul Miller2, Barry Devereux2

1
Queens University Belfast, ECIT, Department of Data Science and Scalable Computing, 2
Queens University Belfast, School of Electronics, Electrical Engineering and Computer
Science

In computational linguistics, the distributional hypothesis of word meaning has allowed us to

construct distributional semantic models by mining large corpora of text to extract co-
occurrence statistics of words. Neural approaches such as word2vec learn two sets of matrix
representations from mined word-context counts, by using dot product between target and
context vector with a sigmoid function to measure the probability of positive association. By
using negative sampling techniques and gradient descent optimization, we can learn an
approximation of word meaning. Our goal is to construct vector representations for a set of
human-derived properties by using a neural topology, similar to the skip-gram word2vec,
which uses the target word to predict the surrounding windowed context from mined
statistics. Surveying a number of concepts for human interpretable features is costly and
time-consuming, but unsupervised learning of vector space models from text data is cheap
and accessible. We learn feature meaning by sampling a tiny subset of a pretrained set of
word embeddings for which we know the properties. Negative sampling along with gradient
descent applied only to the matrix of representations allows us to learn feature meaning in
relation to the pretrained word vectors. In this case, a word and a feature have meaningful
association if their vectors are close together, which we can measure using cosine similarity.
Ranking these features for a given concept, we can extract salient features for the word.
Furthermore, since these come from a wider vector space model, we can sample unseen
words for features. The process allows us to extract possible feature of words, which could
make further surveying the concepts for properties much faster. Active learning would then
allow us to repeat this process with a larger lexicon which could be then surveyed again, this
time with a higher probability of correctly sampling features.

19
1.7 Contributed - Data Science: Words
Tuesday 3 September 9am

Clustering the citation network of a computer science conference

Clement Lee
Lancaster University

Social network analysis can be applied to a wide range of topics and data. One of its
applications is in bibliometrics, where networks concerning academic publications are being
constructed and then analysed in a quantitative way. While different networks could be
constructed from the same data, it is networks with publications as the nodes that received
less attention. In our work, we analysed these citation networks to gain insights into how
publications, rather than authors, are grouped .The data concerned are the Proceedings of
the SIGCHI Conference on Human Factors in Computing Systems. As the premier
conference in Human-Computer Interaction (HCI), it has amassed over 6000 publications
from 1981 to 2018. These publications are not only electronically available but also properly
indexed in the ACM Digital Library (dl.acm.org). The detailed information on the references
of these publications allows identification of citations between these ~6000 publications, and
therefore the construction of the full citation network with ease.After collecting and cleaning
the data, a stochastic block model is applied, which is a popular model in social network
analysis for clustering nodes. Bayesian inference, specifically Markov chain Monte Carlo
(MCMC), is carried out to estimate the group memberships of the publications, which are
treated as latent variables in the model. After modelling and inference, the texts of the
publications in each group are processed, to find out the over-represented words, which can
be seen as the topic words of the group. It is revealed that these topics found correspond
very well with the main themes in the field of HCI. This suggests that citation networks are a
good data source, alternative to collaboration networks, that can be explored to understand
the landscape of a field in the academic literature.

20
1.8 Contributed - Medical: Changing practices
Tuesday 3 September 9am

A fresh look at the James-Stein estimator shows that ‘dynamic borrowing’ of

historical data is an illusion

Nicholas Galwey
GlaxoSmithKline

There are strong reasons for seeking to use historical control data to supplement the results
from a clinical trial or other experiment, and it has been suggested that this approach is of
value even when the only information concerning the bias of the historical data comes from
the observed offset between historical and concurrent controls – an approach known as
‘dynamic borrowing’ (Viele et al., 2014). However, the properties of the James-Stein
estimator indicate that this may not be the case. This estimator of the mean of each of p
groups of observations can be obtained by shrinking the usual least-squares estimates
towards the grand mean, and always gives an increase in the average accuracy of the
estimates (i.e. a reduction in mean square error (MSE) = bias + variance), provided that p >=
3. It is intuitively plausible that no such gain in accuracy is possible when p = 2, but the
papers that introduced the James-Stein estimator (Stein, 1956; James and Stein, 1961) are
heavily mathematical, and for many readers will not provide an easily-understood, rigorous
basis for this limitation. This presentation will demonstrate the limitation in terms familiar to
statisticians, relating it to the balance between two sources of information on the among-
group variance: (mean(group of interest) – mean(other groups)) and var(means of other
groups). When p = 2 the second source is absent and the information on bias and variance
for the group of interest is confounded, confirming that nothing can be gained from the use of
historical control data in this way. A straightforward, transparent way to overcome this
limitation is explicitly to place an informative prior distribution on the bias.

References: Viele, K., et al. (2014) Pharmaceutical Statistics 13:41-54.Stein, C. (1956) Proc. Third
Berkeley Symp. Math. Statist. Prob. 1:197–206.James, W. and Stein, C. (1961) Proc. Fourth Berkeley
Symp. Math. Statist. Prob. 1:361-379.

21
1.8 Contributed - Medical: Changing practices
Tuesday 3 September 9am

Improving transparency about “power” and trade-offs in subgroup selection:

assessing criteria and statistical models for subgroup selection

Claudia Geue1, Neil Hawkins1, Jeremy Oakley2, Richard Grieve3, Monica Hernandez2,
James Carpenter3
1
University of Glasgow, 2 University of Sheffield, 3 London School of Hygiene & Tropical
Medicine

Introduction: Population health can be improved by identifying treatments with favourable

risk-benefit or cost-effectiveness ratios for identifiable sets of patients. This is described as
precision medicine and underpinned by subgroup analysis. Subgroup analysis requires
statistical models that address multiplicity to avoid inflation of type-1 error, which would lead
to inappropriate identification of differential treatment effects. However, this should be
balanced against the risk of missing true subgroups. We suggest metrics to assess the
performance of a given approach for subgroup identification in hypothetical scenarios where
true subgroup effects are known. These criteria could be used to describe the “power” of a
proposed subgroup analysis or in simulation studies that compare alternative statistical
models.

Methods: We reviewed performance metrics in published simulation studies of subgroup

analysis. Based on this we suggest a set of metrics that consider estimation and inference
and reflect trade-offs between avoiding false positives and false negatives in subgroup
analysis.

Results: Subgroup identification methods ranged from t-tests or likelihood-ratio tests

including and excluding the predictive term for individual models, to methods using step
functions, splines or simultaneous threshold interaction modelling. Performance metrics
included the rate of identifying correct treatment-covariate interactions, true-positive rate,
true-negative rate, the probability of detecting at least one false-negative subgroup and the
probability of detecting at least one false-positive subgroup. Most studies did not mention
assessing simulations in terms of power and associated trade-offs. One study assessed
model performance under different scenarios for sample size and power.

Conclusion: Misinterpreting results from subgroup analyses could potentially harm patients
(missing true subgroups/accepting false subgroups) or lead to futile and costly trials (falsely
identified subgroups). Decision-makers must balance the risk of accepting ‘spurious’
subgroup effects against the risk of rejecting true subgroup effects. Transparency is required
in particular about trade-offs associated with avoiding false negatives and false positives.

22
1.8 Contributed - Medical: Changing practices
Tuesday 3 September 9am

Epidemiological characterisation and classification of disease

Anthony Webster, Benjamin Cairns, Robert Clarke

NDPH, University of Oxford

Modern healthcare classifies diseases by anatomical systems. We used UK Biobank data

and their linked Hospital Episodes Statistics to explore whether a combination of Big Data
and machine learning methodologies can provide new insights into classification systems for
disease. We considered whether supervised learning methods can generate conventional
groupings, such as the International Classification of Diseases Version 10 (ICD-10), and
explored the potential for novel data-driven classification of diseases. Our aims were to: (i) to
provide new insights into the causes and consequences of disease, (ii) to develop new
methods to characterise the links between diseases and their underlying risk factors, and (iii)
to understand and improve systems for disease classification. By understanding the links
between multiple diseases and their underlying risk factors, we hope to transform the way
that we study and manage individual diseases and associated multi-morbidity patterns.

23
2.1 Medical: Investigation of and overcoming practical dilemmas within medical
research
Tuesday 3 September 10.10am

Under-representation in Clinical Trials: Participants with rare diseases, reporting and

awareness.

Helen McAneney
Queen's University Belfast

Introduction: Trial populations need to reflect those in the community who may benefit from
the treatment being tested if the results of the trial are to maximise patient health gains. The
European Union defines a rare disease as one affecting fewer than 5 in 10,000 of the
general population. With over 8,000 known rare diseases, almost 6% of the population will
be affected by a rare disease. This equates to approximately 3.5 million people in the UK
and 30 million people across Europe. Consequently one would expect 6% of those recruited
to clinical trials to have a rare disease; however this is at best unknown and likely to not be
the case. This study will investigate the representation of rare disease participants within
clinical trials and possible reasons for under-representation of these participants, including (i)
lack of reporting that a participant has a rare disease, (ii) potential exclusion due to the
recruitment criteria to the clinical trial and/or (iii) lack of awareness of rare diseases.

Methods: A sample of clinical trials for anti-hypertension medications, listed on

ClinicalTrials.gov, will be investigated. Representation of participants will be considered by
analysing the reporting of participants’ rare disease and the inclusion/exclusion criteria from
protocols, reports and publications of the clinical trials. Awareness of rare diseases will be
explored through a network analysis of the citation of rare disease literature using
CitNetExplorer, specifically citation of the European Union definition of a rare disease.

Potential relevance and impact: The aim of the UK Strategy for Rare Disease is to ‘ensure
no one gets left behind just because they have a rare disease’. Clinical trials are pivotal to
the improvement of patient health, and require representation of all patients, including and
inclusion of rare disease patients in clinical trials.

24
2.2 Official Statistics & Public Policy: International comparisons of health outcomes -
opportunities and challenges of using routinely collected administrative databases
Tuesday 3 September 10.10am

Comparing maternal and child outcomes across countries using prospectively

planned, pooled analyses of administrative data: the devil is in the detail

Katie Harron
UCL

Careful review and comparison of different systems can help provide insight into the types of
system that deliver the best health outcomes for mothers and their children. In particular,
international comparisons provide an important opportunity to help us focus on the specifics
of care that are likely to lead to improved outcomes. Provided careful consideration is given
to the comparability and quality of data collected in different contexts, international
comparisons can offer powerful external benchmarks and evidence on where and how
improvements to health and services might be made. In this talk I present methods for
bringing together data from different countries to facilitate cross country comparisons, based
on three international studies relating to infant admissions, outcomes of teenage
motherhood, and long term maternal mortality in mothers of infants with neonatal abstinence
syndrome. I discuss the strengths of using hospital data captured in different settings and
approaches used to address limitations in comparability. Results demonstrate that even
when countries appear to have similar healthcare systems, detailed understanding of
healthcare seeking behaviours and the policy context is vital in interpreting different patterns
of outcomes. Whilst cross country comparisons may not be able to provide definitive
answers, they can help us raise the right questions about the best policies to promote
maternal and child health.

25
2.2 Official Statistics & Public Policy: International comparisons of health outcomes -
opportunities and challenges of using routinely collected administrative databases
Tuesday 3 September 10.10am

International comparisons of primary care quality using admissions for ambulatory

care sensitive conditions: the example of asthma in children

Irina Lut1, Kate Lewis1, Linda Wijlaars1, Tiffany Fitzpatrick2, Hong Lu2, Astrid Guttman2,
Sharon Goldfeld3, Shaoke Lei3, Geir Gunnlaugsson4, S Hrafn Jonsson4, Reli Mechtler5, Mika
Gissler6, Anders Hjern7, Pia Hardelid1
1
UCL Great Ormond Street Institute of Child Health, London, UK, 2 ICES & Dalla Lana
School of Public Health, Toronto, Canada, 3 Murdoch Children’s Research Institute &
Department of Paediatrics, University of Melbourne, Australia, 4 School of Social Sciences,
University of Iceland, Iceland, 5 Johannes Kepler University, Austria, 6 National Institute for
Health and Welfare, Finland, 7 Centre for Health Equality Studies & Department of Clinical
Epidemiology, Karolinska Institutet, Sweden

Objectives: International comparisons of health indicators are a powerful tool in health

research. Differences in delivery of care mechanisms or how measures are recorded could
affect observed differences in health outcomes across countries that have similar health
systems and national income levels. We examine challenges in interpreting cross-country
comparisons of hospital admission rates for asthma as an indicator of quality of primary care
provided to children with long-term conditions.

Methods: Seven jurisdictions provided administrative data on asthma-related hospital

admissions and population denominators for children aged 6-15 years inclusive between
2008 and 2015. Five countries provided admissions with asthma as a secondary diagnosis
and three included asthma-related attendances to emergency departments. We compared
incidence rate ratios (IRRs) between admissions with asthma as the primary diagnosis,
admissions with asthma as a secondary diagnosis and asthma-related emergency
department attendance.

Results: 74,682 asthma admissions were recorded among 56 million children. Across all
jurisdictions, hospital admission rates were higher for boys aged ≤12 and higher for girls 13-
15 years. Over our study period, Victoria (Australia) had the highest primary admission rate
and Iceland the lowest, with an 8-fold variation between countries. The rate for admissions
with asthma recorded as a secondary diagnosis or within emergency departments was over
three times higher than for asthma primary diagnosis admissions (IRR 3.5, 95%CI: 3.5-3.6
and IRR 3.0, 95%CI: 3.0-3.1 respectively). England’s admission rate with asthma as a
primary diagnosis was 7.1 times higher than Sweden (95%CI: 6.7-7.6), but 10.8 times higher
when including asthma recorded as a secondary diagnosis (95%CI: 10.4-11.2).

ConclusionsFactors including prevalence and severity of asthma, hospital admission

thresholds and recording practices influence hospital admission rates. Clearly defining
outcome measures is a critical first step for international comparisons in the quality of care
for children with asthma.

26
2.2 Official Statistics & Public Policy: International comparisons of health outcomes -
opportunities and challenges of using routinely collected administrative databases
Tuesday 3 September 10.10am

Using administrative linked datasets to explain differences in child mortality between

England and Sweden: opportunities and challenges

Ania Zylbersztejn, Pia Hardelid

UCL Great Ormond Street Institute of Child Health, London, UK

Child mortality is considered an important indicator of health of a nation. England has one of
the highest rates of child mortality in Western Europe, nearly twice as high as that of
Sweden – a country with comparable levels of economic development and universal
healthcare. The differences in child mortality between these two countries have been
commonly attributed to differences in quality of healthcare for children and wider
socioeconomic inequalities in England. Any differences, however, are likely to be at least
partly explained by England’s high rates of adverse birth characteristics, such as low birth
weight, preterm birth or presence of congenital anomalies. Detailed information about child’s
health at birth is therefore needed to disentangle the origins of intercountry differences in
child mortality.In this presentation, we use national birth cohorts from administrative linked
databases in England and Sweden to determine which factors before and after birth
contribute most to the gap in mortality between England and Sweden. We used data on
almost 4 million singleton live births in England and over 1 million in Sweden in 2003-2012,
with detailed information about health of a child at birth (e.g.: length of gestation, birth
weight, presence of birth defects), socioeconomic factors and causes and timing of death.
We show what steps are needed to assess comparability of national birth cohorts developed
from different data sources (hospital admission and mortality records in England, birth, death
and hospital admission registers in Sweden), how to compare recording of diagnoses and
causes of death across different settings. Our findings and recommendations will be relevant
to future comparisons of health outcomes in countries with administrative linked databases.

27
2.2 Official Statistics & Public Policy: International comparisons of health outcomes -
opportunities and challenges of using routinely collected administrative databases
Tuesday 3 September 10.10am

The unique opportunities and critical limitations of using routinely collected national
data for international comparisons of maternal and newborn health

Jennifer Zeitlin
Inserm

International comparisons of perinatal health indicators provide performance benchmarks

and underpin maternal and child health policies. The Euro-Peristat project began in 2000
with the aim of providing high-quality, comparable health information on maternal and
newborn health and healthcare in Europe in order to enable these comparisons. The project
compiles data on 30 indicators from participating countries which now include all 28 EU
Member States, Iceland, Norway and Switzerland. Participants are clinicians, statisticians
and epidemiologists who work with routine data. Indicators are compiled from routine
national-level sources, including: vital statistics, birth registers, hospital discharge abstracts
and surveys. By conducting research using indicators of perinatal health derived from
national routine sources, Euro-Peristat has been able to identify the benefits as well as the
dangers of using these data to inform policy and practice. This presentation will illustrate key
lessons learned in two main domains (1) defining maternal and infant mortality and morbidity
for benchmarking between countries and (2) interpreting correlations between indicators and
time periods. These analyses require realistic assessments of data quality limitations and of
the appropriate scope of inference to avoid ecological fallacies. These assessments depend
on involving representatives from participating countries with knowledge of medical and data
collection practices and developing a common vocabulary through collaboration over time.
New data on births in 2015 will be used to illustrate potential errors in interpretation as well
as the unique insights that can be drawn from these comparisons.

28
2.3 Environmental / Spatial Statistics: Applications of hidden Markov models in
ecology
Tuesday 3 September 10.10am

A continuous-time Arnason-Schwarz model for the annual movement of bottlenose

dolphins

Sina Mews1, Roland Langrock1, Ruth King2, Nicola Quick3

1
Bielefeld University, 2 University of Edinburgh, 3 Duke University Marine Lab

Our modelling approach is motivated by individual sighting histories of bottlenose dolphins

off the east coast of Scotland. Due to ongoing offshore development, conservation
managers seek to better understand the temporal movement patterns of the dolphin
population between different sites. The Arnason-Schwarz (AS) model is often used to
analyse such transitions between different states, which here correspond to the location
(site) of a given individual. Within the AS model, transitions between states are modelled
using a discrete-time Markov chain. In our case, however, the capture occasions do not
follow a regular sampling protocol, which is why we develop a continuous-time analogue of
the AS model. In contrast to the AS model, the new modelling framework does not require
capture occasions to be regularly spaced in time.Statistical inference is carried out by
regarding the capture-recapture data as realisations from a (continuous-time) hidden Markov
model (HMM), where an individual's sighting history corresponds to the observed state-
dependent process and the individual’s (true) state corresponds to the unobserved state
process. Embedding the capture-recapture setting in the HMM framework allows the
associated efficient algorithms to be used in particular for (numerical) maximum likelihood
estimation and state decoding. For scenarios with time-varying covariates affecting the state-
switching rates, we develop an approximate maximum likelihood approach.In our present
analysis, we are particularly interested in seasonal effects on the migration rates of
bottlenose dolphins along the Scottish east coast. The results reveal seasonal migration
patterns between two main areas, information that can help to inform conservation
management. While motivated by a particular data set, our modelling framework is generally
applicable to irregularly sampled capture-recapture data where individuals traverse through
different states.

29
2.3 Environmental / Spatial Statistics: Applications of hidden Markov models in
ecology
Tuesday 3 September 10.10am

A test for the underlying state-structure of Hidden Markov models: A case study of
partially observed capture-recapture data

Rachel McCrea
University of Kent

Hidden Markov models (HMMs) are being widely used in the field of ecological modelling,
however determining the number of underlying states in an HMM remains a challenge. Here
we focus on a special case partially-observed capture-recapture models, where some
animals are observed but it is not possible to ascertain their state, whilst the other animals'
states are assigned without error. We propose a mixture test of the underlying state structure
generating the partial observations, which assesses whether they are compatible with the
set of states directly observed in the capture-recapture experiment. I demonstrate the
performance of the test using simulation and through application to a data set of Canada
Geese.

30
2.3 Environmental / Spatial Statistics: Applications of hidden Markov models in
ecology
Tuesday 3 September 10.10am

Modelling latent animal movement and behaviour in population abundance surveys

using hidden Markov models

Richard Glennie1, Stephen Buckland1, Roland Langrock2, David Borchers1

1
University of St Andrews, 2 Bielefeld University

Distance sampling and spatial capture-recapture are statistical methods to estimate the
number of animals in a wild population based on encounters between these animals and
scientific detectors. Both methods estimate the probability an animal is detected during a
survey, but do not explicitly model animal movement and behaviour. The primary challenge
is that animal movement in these surveys is unobserved; one must average over all possible
histories of each individual. In this talk, a general statistical model, with distance sampling
and spatial capture-recapture as special cases, is presented that explicitly incorporates
animal movement. An efficient algorithm to integrate over all possible movement paths,
based on quadrature and hidden Markov modelling, is given to overcome the computational
obstacles. For distance sampling, simulation studies and case studies show that
incorporating animal movement can reduce the bias in estimated abundance found in
conventional models and expand application of distance sampling to surveys that violate the
assumption of no animal movement. For spatial capture-recapture, continuous-time
encounter records are used to make detailed inference on where animals spend their time
during the survey. In surveys conducted in discrete occasions, maximum likelihood models
that allow for mobile activity centres are presented to account for transience, dispersal,
survival, and heterogeneous space use. These methods provide an alternative when animal
movement and behaviour causes bias in standard methods and the opportunity to gain
richer inference on how animals move, where they spend their time, and how they interact.

31
2.3 Environmental / Spatial Statistics: Applications of hidden Markov models in
ecology
Tuesday 3 September 10.10am

Modelling population dynamics using hidden Markov models

Takis Besbeas
Athens University Economics Business / University of Kent

Hidden Markov models have gained popularity within the statistical ecology community
thanks to their flexibility to accommodate various types of time series data. They have been
routinely applied to model animal movement, but their potential has also been demonstrated
in capture-recapture and occupancy modelling. In this talk I will discuss recent developments
for modelling population dynamics using hidden Markov models. In particular, we show how
hidden Markov model methodology provides a flexible and efficient framework for parameter
estimation and model selection from time series data of population abundances. We also
show how hidden Markov modelling machinery can be used to provide exact inference for
Integrated Population Modelling (IPM), where demographic data are also available. In both
of these cases, a single hidden Markov model is employed, corresponding to data from a
single species. We last show how multiple hidden Markov models can be engaged to
provide a Multispecies Biodiversity Indicator when indices of abundance from multiple
species are available. We illustrate the methodology using real data in each case.

32
2.4 Social Statistics: The better understanding of society - methodological innovation
on Understanding Society
Tuesday 3 September 10.10am

Understanding the impact of web mode on quantitative analysis of data from

Understanding Society

Paul Clarke, Yanchun Bao

University of Essex

Understanding Society has, until recently, mainly used interviewers to administer its
questionnaires. However, starting from Wave 8, it is now allowing individuals to participate
via the web. Survey mode is known to affect participants' responses so, to assess its impact,
a randomized sequential-design experiment was carried out during the first year of fieldwork
for Wave 8. This design, akin to encouragement designs in clinical research, permits the
identification of mode effects which would otherwise be confounded by non-random
selection. We set up a general framework based on extended structural mean models to
efficiently estimate the effects of mode not only on the mean but also the variance,
covariance and (for categorical variables) the entire distribution of mixed-mode survey
variables. We also show how this framework can be used by analysts to test whether the
introduction of web mode has made a difference to the results of their particular analysis.

33
2.4 Social Statistics: The better understanding of society - methodological innovation
on Understanding Society
Tuesday 3 September 10.10am

A latent class approach to inequity in health using biomarker data

Apostolos Davillas1, Vincenzo Carrieri2, Andrew Jones3

1
Office for Health Economics and University of Essex, 2 University of Catanzaro, 3 University
of York

We develop an empirical approach to analyse, measure and decompose Inequality of

Opportunity (IOp) in health, based on a latent class model. This addresses the limitations
that affect earlier work in this literature concerning the definition of types - such as partial
observability, the ad hoc selection of circumstances, the curse of dimensionality and
unobserved type-specific heterogeneity - that may lead to either upwardly or downwardly
biased estimates of IOp. We apply the latent class approach to quantify IOp in allostatic
load, a composite measure of our biomarker data. Using data from Understanding Society,
we find that a latent class model with three unobserved types best fits the data and that
these types differ in terms of their observed circumstances. Decomposition analysis shows
that about two-thirds of the total inequality in allostatic load can be attributed to the direct
and indirect contribution of circumstances.

34
2.4 Social Statistics: The better understanding of society - methodological innovation
on Understanding Society
Tuesday 3 September 10.10am

Obtaining consent for the linkage of social media data with large-scale population
surveys

Tarek Al Baghal
University of Essex

In the light of issues affecting survey quality, like the increase in unit non-response, it has
been argued that social media data from sources like Twitter can be used as a viable
alternative. However, there are also a number of shortcomings with Twitter data such as the
potential for it being unrepresentative of the wider population, and the inability to validate
exactly whose data it is you are collecting. A useful way forward is to combine survey and
Twitter data to supplement and improve both, but this requires consent. This study explores
these consent decisions. Our findings suggest that consent rates for data linkage are
relatively low, and this is in part mediated by survey mode, where face-to-face surveys have
higher consent rates than web versions

35
2.4 Social Statistics: The better understanding of society - methodological innovation
on Understanding Society
Tuesday 3 September 10.10am

Do Income Summary Screens Improve Income Data Quality?

Paul Fisher1, Jonathan Burton1, Tom Crossley1, Annette Jackle1, Alessandra Gaia2
1
University of Essex, 2 University of Milan-Bicocca

Income data collected as part of household surveys are critical for the study of material living
standards. Survey respondents are known to misreport their income, but the types of error
are not well-documented. We experiment with the use of an Editable Summary Screen
(ESS) during data collection to improve income data quality by a) testing two version of the
ESS in a large scale panel survey and b) shed light on the types of reporting errors made by
survey participants by classifying the revisions

36
2.5 Methods and Theory: Modern Fisherian perspectives on inference
Tuesday 3 September 10.10am

Modified maximum likelihood estimation through adjusted scores

Nicola Sartori1, Michele Lambardi di San Miniato1, Nancy Reid2

1
University of Padova, 2 University of Toronto

For regular parametric problems, modified maximum likelihood estimates can be obtained by
simple adjustments of the score equation. The choice of the adjustment determines the
properties of the resulting estimator. The most notable example is given by Firth's bias
reduced estimator. Following the same idea, Kenne Pagui et al. (Biometrika, 2017) proposed
a different adjustment that leads to a componentwise median unbiased estimator. Here, we
propose a new adjusment that provides an estimator such that each component, or block of
components, is asymptotically equivalent to the corresponding estimator obtained from a
modified profile likelihood, thus correcting for the presence of the remaining components,
treated as nuisance parameters. This estimator partly retains the equivariance property of
the maximum likelihood estimator and, as mean and median bias reduced estimators, has
the same asymptotic distribution.

37
2.5 Methods and Theory: Modern Fisherian perspectives on inference
Tuesday 3 September 10.10am

Conditioning and Randomisation in Selective Inference

Alastair Young
Imperial College London

Selective inference is concerned with performing valid statistical inference when the
questions being addressed are suggested by examination of data, rather than being
specified before data collection. Adjustment for double use of the data through `conditioning
on selection’ has been proposed as a route to valid inference in a frequentist framework. The
Bayesian standpoint on selective inference is less clear, but it may be argued that
conditioning on selection is required unless the selection takes place on the parameter
space as well as on the sample space. In this talk, we will argue that classical ideas of data
randomisation and appropriate conditioning are necessary to yield a theoretically justified
and practically appealing Bayesian inference in the selective context.

38
2.8 Communicating & Teaching Statistics: Recent innovations in school statistics:
remediations
Tuesday 3 September 10.10am

Recommendations for assessment of statistics in A level Mathematics

James Nicholson
Durham University

Statistics emerged as a discipline to address pressing practical problems. In the UK, this has
not been reflected in school statistics curricula, where students often work with small-scale
invented data to develop mastery of statistical technique. Recent curriculum reforms set out
to improve this situation; students are expected to work in class with a large authentic data
set. However, in the associated high-stakes assessment, there has still been very little
emphasis on statistical skills such as interpreting data and drawing conclusions, and a great
deal of emphasis on procedural skills, and on factual recall to demonstrate that they have
worked with the pre-release data. Apart from questions related to the pre-release large data
set for each qualification, there were no questions in any of the specimen papers or the live
papers in 2018 and 2019 that use real data.Contexts are often rather banal. One of the
reasons for this may be that mathematics does not have any embedded contexts in the way
that statistics assessed within Biology, Psychology, Politics, Social Sciences etc. does.
Some of the assessment is not fit for purpose and much of it does not allow students to see
why statistics is an important and relevant discipline.I will talk briefly about some examples
of poor assessment practice and some curriculum issues which may exacerbate the
problems with assessment, but the main focus will be the recommendations for assessment
of statistics which have been developed by a sub-group convened by the A level Contact
Group of the Royal Society’s Advisory Committee on Mathematics Education.

39
2.8 Communicating & Teaching Statistics: Recent innovations in school statistics:
remediations
Tuesday 3 September 10.10am

Using real data across the curriculum at secondary level

Darren Macey
Cambridge Maths

Despite the increase in the amount of ‘real data’ which is accessible now through the
internet, many teachers find it difficult to locate datasets which are suitable for use with the
content of the mathematics curriculum. Although curriculum documents talk aspirationally
about the use of technology in statistics, there are few resources provided which stimulate
the classroom practice. I am working on developing two types of resources which I hope can
help make it easier to use real data in schools, both in mathematics and across other
subjects.One is to identify, and make available, real datasets which have some numerical
variables and some categorical variables, where the full dataset can be disaggregated into
groups by a categorical variable and then summary statistics, or graphical displays can be
constructed for the different groups to allow meaningful comparisons. Currently, when
summary statistics or graphs of a dataset are constructed, usually nothing more is done – it
is an end in itself. These new curated datasets could be used in mathematics to provide the
practice required in learning statistical techniques, with an opportunity to develop skills in
describing and interpretation that are an increasingly important aspect of the statistical
expertise we wish students to develop. They could also be used in other subject areas to
which the dataset is relevant.The second is to produce data visualisations which allow
students to engage with datasets involving more variables than they usually meet in the
mathematics curriculum, but which are commonplace (at least implicitly) in social science
and humanities disciplines. I will show an example, based on a visualiser developed by the
SMART Centre at Durham, along with curriculum resources which could be used in
mathematics classrooms as well as in disciplines where inequalities in educational
attainment are explicitly studied.

40
3.1 Medical: Papers from the Journal of the Royal Statistical Society
Tuesday 3 September 2.10pm

Landmark linear transformation model for dynamic prediction with application to a

longitudinal cohort study of chronic disease

Yayuan Zhu1, Liang Li2, Xuelin Huang2

1
University of Western Ontario, 2 University of Texas MD Anderson Cancer Center

Dynamic prediction of the risk of a clinical event by using longitudinally measured

biomarkers or other prognostic information is important in clinical practice. We propose a
new class of landmark survival models. The model takes the form of a linear transformation
model but allows all the model parameters to vary with the landmark time. This model
includes many published landmark prediction models as special cases. We propose a
unified local linear estimation framework to estimate time varying model parameters.
Simulation studies are conducted to evaluate the finite sample performance of the method
proposed. We apply the methodology to a data set from the African American Study of
Kidney Disease and Hypertension and predict individual patients’ risk of an adverse clinical
event.

41
3.1 Medical: Papers from the Journal of the Royal Statistical Society
Tuesday 3 September 2.10pm

Semiparametric Model for Bivariate Survival Data Subject to Biased Sampling

Jin Piao1, Jing Ning2, Yu Shen2

1
University of Southern California, 2 The University of Texas MD Anderson Cancer Center,
Houston, USA

To better understand the relationship between patient characteristics and their residual
survival after an intermediate event such as the local cancer recurrence, it is of interest to
identify patients with the intermediate event and then analyze their residual survival data.
One challenge in analyzing such data is that the observed residual survival times tend to be
longer than those in the target population, since patients who die before experiencing the
intermediate event are excluded from the identified cohort. We propose to jointly model the
ordered bivariate survival data using a copula model and appropriately adjusting for the
sampling bias. We develop an estimating procedure to simultaneously estimate the
parameters for the marginal survival functions and the association parameter in the copula
model, and use a two-stage expectation-maximization algorithm. Using empirical process
theory, we prove that the estimators have strong consistency and asymptotic normality. We
conduct simulations studies to evaluate the finite sample performance of the proposed
method. We apply the proposed method to two cohort studies to evaluate the association
between patient characteristics and residual survival.

42
3.1 Medical: Papers from the Journal of the Royal Statistical Society
Tuesday 3 September 2.10pm

Adaptive design in surveys and clinical trials: similarities, differences and

opportunities for cross-fertilization

Michael Rosenblum1, Peter Miller2, Benjamin Reist3, Elizabeth Stuart4, Michael Thieme2,
Thomas Louis4
1
Johns Hopkins University, 2 US Census Bureau, Washington DC, USA, 3 US Department of
Agriculture National Agricultural Statistics Service, Washington DC, USA, 4 Johns Hopkins
Bloomberg School of Public Health, Baltimore, USA

Adaptive designs involve preplanned rules for modifying an on-going study based
onaccruing data. We compare the goals and methods of adaptation for trials and surveys,
identifysimilarities and differences, and make recommendations for what types of adaptive
approachesfrom one domain have high potential to be useful in the other. For example,
clinical trials couldbenefit from recently developed survey methods for monitoring which
groups have low responserates and intervening to fix this. Clinical trials may also benefit
from more formal identification ofthe target population, and from using paradata (contextual
information collected before or duringthe collection of actual outcomes) to predict participant
compliance and retention and then tointervene to improve these. Surveys could benefit from
stopping rules based on informationmonitoring, applying techniques from sequential
multiple-assignment randomized trial designsto improve response rates, prespecifying a
formal adaptation protocol and including a datamonitoring committee.We conclude with a
discussion of the additional information, infrastructureand statistical analysis methods that
are needed when conducting adaptive designs, as well asbenefits and risks of adaptation.

43
3.3 Environmental / Spatial Statistics: Using electronic health records to model spatial
variation in disease risk
Tuesday 3 September 2.10pm

Disaggregation of areal unit count data

Craig Anderson, Kamol Sanittham, Duncan Lee

University of Glasgow

Disease mapping is the field of epidemiology focusing on estimating the spatial pattern of
disease risk across a geographical region. Typically the region of interest is subdivided into a
set of administrative districts, with the disease data consisting of aggregated disease counts
at this district level. This means that inference is also generally restricted to estimating
disease risk at the district level.Such inference can, however, be susceptible to the
modifiable areal unit problem (MAUP), whereby the estimated risk surface can be affected
by the arbitrary choice of district boundaries. In reality, the district-level count is an
aggregation of point level disease cases, and it is possible that the results could be
completely different if a different spatial partition of the region was selected.In this research,
we aim to address this problem by producing “disaggregated” disease risk estimates based
on a regular 1km x 1km grid. We consider two main approaches – the first is based on
multiple imputation and the second on data augmentation. The method is illustrated using an
application to respiratory hospital admissions in Glasgow, Scotland.

44
3.3 Environmental / Spatial Statistics: Using electronic health records to model spatial
variation in disease risk
Tuesday 3 September 2.10pm

Spatio-temporal Modelling of Trends in Benign Prostate Hyperplasia drugs

prescriptions in Scotland

Federico Andreis1, Ashleigh Ward1, Emanuele Giorgi2

1
University of Stirling, 2 University of Lancaster

Over the last decade, increased access to official prescription records has provided a wealth
of opportunities for the application of advanced statistical methodology in order to better
inform public health policies. In this talk, our interest is on modelling the spatial and temporal
variation in the prescriptions for the treatment of Benign Prostate Hyperplasia (BPH)
symptoms in the Scottish population. We first describe how to extract general practice-level
data from the NHS Scotland Open Data portal and then link these to socio-economic
indicators (e.g. measures of deprivation and rurality of the patients base). In the second part
of the talk, we present a spatio-temporal analysis of the monthly number of daily doses for
drugs belonging to two groups (α-1 blockers and 5-α reductase inhibitors) prescribed by GPs
in Scotland over a period of 4 years. Our results show an increasing trend in the number of
BPH-related prescriptions, which is consistent with both a higher prevalence of prostate-
related conditions and an improved detection thereof, likely ascribable, respectively, to the
ageing of the population and the increased publicity about prostate cancer screening. While
trends are fairly similar across health boards, baseline rates of prescriptions for the two drug
groups tend to differ, possibly reflecting local effects due to the influence of the Area Drug
and Therapeutics Committees. Lastly, the inspection of the spatial residuals of the model
seems to point at the existence of geographic areas (of non-strictly administrative nature)
that exhibit similar prescribing behaviours.

45
3.3 Environmental / Spatial Statistics: Using electronic health records to model spatial
variation in disease risk
Tuesday 3 September 2.10pm

A spatially discrete approximation to log-Gaussian Cox processes for modelling

spatially aggregated disease counts data

Olatunji Johnson, Peter Diggle, Emanuele Giorgi

Lancaster University

The increased availability of electronic health records (EHRs) opens up new opportunities to
study variation in disease risk. However, the use of EHRs for geospatial analysis is
hampered by the lack of information on the residence or full postcode of the patients which is
due to confidentiality reasons. Instead, the Lower Super Output Area (LSOA) of residence of
the patients is usually provided in the UK. This leads to the aggregation of health outcome
over such partitions. As variation in disease risk occurs in a spatial continuum irrespective of
the format in which the data are available, in this talk, we discuss first that log-Gaussian Cox
process (LGCP) can be considered as a natural statistical paradigm for modelling spatially
aggregated disease count data, and second discuss the proposed computationally efficient
discrete approximation to LGCP model. Furthermore, we discuss an application to COPD
emergency admission covering admissions from 2012 to 2018 in South Cumbria and North
Lancashire, England, UK. We carry out a spatial prediction of the incidence and unexplained
risk of COPD emergency admission. We identify areas with high unexplained risk. And
advise that urgent attention should be given to those areas by deploying specialist
community COPD nurses as this will help save valuable hospital resources. The proposed
methodology is implemented in the open-source R package SDALGCP.

46
3.5 Methods and Theory: Pseudo likelihood theory and methods
Tuesday 3 September 2.10pm

Robustness of inference for GLMMs under model misspecification

Helen Ogden
University of Southampton

Generalised linear mixed models (GLMMs) are a widely-used class of models, but they
make some strong assumptions which are mostly made by convention, and are not easy to
verify. It is therefore important to understand the impact on the resulting inference if the true
data generating process does not obey these assumptions. I will consider the impact of
various types of model misspecification on pseudo-likelihoods and other methods for
conducting inference in GLMMs. Often a full joint model for all the data is not needed to
specify a pseudo-likelihood, which might rely instead on some low-dimensional marginal
distribution or a small number of moments. It may be easier to write down a sensible model
for these small components than it is to specify a full joint model, and pseudo-likelihoods are
often described as being relatively robust to misspecification of the full joint model,
compared with the full likelihood. I will evaluate the extent to which this is true, using GLMMs
as an example model class.

47
3.5 Methods and Theory: Pseudo likelihood theory and methods
Tuesday 3 September 2.10pm

Towards fully-efficient estimation

Ioannis Kosmidis
University of Warwick

The inverse of the Fisher information matrix in a likelihood problem is i) the variance-
covariance matrix of the asymptotic distribution of the maximum likelihood (ML) estimator; ii)
the dominant term in the expansion of the finite-sample variance of the ML estimator; and iii)
the "lowest" achievable variance-covariance that an unbiased estimator can achieve.
"Lowest" here is used to indicate that the difference of the inverse Fisher information from
the variance of any unbiased estimator is a positive definite matrix. These three
characterizations and the asymptotic unbiasedness of the ML estimator are key justifications
for the wide-spread use of the latter in statistical practice. For example, standard regression
software typically reports the ML estimates alongside with estimated standard errors coming
from the inversion of the Fisher information matrix at the estimates. Nevertheless, the use of
that pair of estimates and estimated standard errors for inference implicitly assumes,
amongst other things, that the information about the parameters in the sample is large
enough for the estimator to be almost unbiased and its variance to be well-approximated by
the inverse of the Fisher information matrix. In this talk, we present results from work-in-
progress on a novel estimation framework that aims to bridge the finite-sample gap between
estimates and the estimated variance-covariance matrix. We also show results from
inferential settings that are well-used in statistical practice.

48
3.8 Business, Industry & Finance: Contemporary Challenges in Industrial Statistics
Tuesday 3 September 2.10pm

Next generation methods for industrially-focussed earth-observation imagery

Louise Lloyd
Rezatec

Rezatec combines data science with satellite imagery and geo-spatial data to deliver cloud-
based analytics to global customers owning and operating land-based assets.The new and
growing Data Science team at Rezatec investigates and applies a large range of models to
satellite and geo-spatial data. Current work includes identifying different crops and tree
species, predicting water leaks and forecasting global commodity harvests a year in
advance.This presentation will focus on the world of forestry, and in particular, the massive
forests in the USA and Canada. These are too big to observe regularly from the ground, but
we can observe them from space. Using freely available high-resolution satellite imagery
and some ground data for training, we can identify the number, species and volume of trees
across large areas of forest with an accuracy greater than 80%. Typically, these results are
achievable with some straightforward machine learning algorithms, but accuracies above
this are hard to achieve and there are some situations where the algorithms fail completely.
Our ‘next-generation techniques’ combine these machine learning algorithms with statistical
and ecological models. They encompass knowledge about plant communities and tree
habitats in the hope of achieving interpretable models and accuracies greater than our
current levels.

49
3.8 Business, Industry & Finance: Contemporary Challenges in Industrial Statistics
Tuesday 3 September 2.10pm

Novel methods for senor-based streaming data

Idris Eckley
Lancaster University

The advent of low-cost sensors in the business and industrial landscape has resulted in easy
access to high-quality data streams. These streams, if appropriately harnessed, can provide
valuable insights into previously unseen operational aspects, and the opportunity to unlock,
automate and transform the management of key processes. In parallel, such developments
can also result in the need to develop new statistical methodology, for example in the areas
of time series, changepoint analysis and anomaly detection. This talk will provide an
introduction to the topic, introducing various case studies that highlight recently developed
work in this area.

50
3.8 Business, Industry & Finance: Contemporary Challenges in Industrial Statistics
Tuesday 3 September 2.10pm

Personalised marketing: challenges and recent advances

Arnoldo Frigessi
University of Oslo

Marketing is personalised when every user/customer is approached in a unique way, in an

effort to present products/items/services which are of special interest for exactly her/him.
Representing and understanding the personal preferences of each unique user, on the basis
of highly incomplete data, is the first important challenge. Inference exploits patterns of
similarity between users. The second challenge is to perform prediction of the user’s actions,
in order to optimise personalised marketing. We will describe two projects, one in
recommender systems (where accuracy of prediction and diversity of offers are important)
and social-network based marketing (where user preferences spread like infections over the
network).

51
PD2 Professional Development: Data FAIRification using R/Rstudio workflows
Tuesday 3 September 2.10pm

Data FAIRification using R/RStudio workflows

Brendan Palmer, Darren Dahly

HRB Clinical Research Facility Cork

Research data are often lost or discarded at the end of a study, despite their continued
value. Funders are cognisant of this loss, so researchers are increasingly being asked to
prepare their data according to the FAIR Guiding Principles
(https://doi.org/10.1038/sdata.2016.18). For many researchers it is unclear how such
initiatives can be incorporated into their current day-to-day workflows. The overriding goal of
this workshop is to provide attendees with a reproducible workflow based on R/RStudio that
adheres to the core FAIR principles of findability, accessibility, interoperability and reuse.

The workshop will be organised and delivered by Drs Dahly and Palmer of the HRB CRF-C,
School of Public Health, University College Cork. Drawing on their experience providing
statistical support for clinical investigators, and recognising the increased requirements for
FAIR data, they have developed an R/RStudio workflow, that runs from project initiation
though to automated report generation. This workflow also includes steps to ensure data are
‘FAIRified’ as the workflow culminates in downstream application of globally resolvable
unique, persistent identifiers (e.g. DOI) and resultant long term preservation.All materials will
be made freely available online in advance of the session. Participants will be invited to
follow the instructors on their own laptops as we take a ‘raw’ sample data set (in .csv format)
and subsequently clean, ‘FAIRify’ and publish the data object.

The workshop will be broken into four main sections:Introduction to open science and FAIR
data stewardship.Building research projects using a defined folder structure coupled with R-
projects.'FAIRification' of sample data as part of the data cleaning processPersistent DOI
assignment.

The main learning outcome from this session will be a reproducible workflow that attendees
can take back to their current workplace and use as a first step towards data ‘FAIRification’
of their own project outputs.

52
Rapid Fire Talks 1
Tuesday 3 September 3.40pm

A Bayesian Model Averaging Approach to g-Parameter Priors Elicitation

Saheed Afolabi
Ibadan, Nigeria

A special technique that measures the uncertainties embedded in model selection processes
is Bayesian Model Averaging (BMA) which depend on the appropriate choices of model and
parameter priors. As important as parameter priors' specification in BMA, the existing
parameter priors based on fast increasing sample sizes compared to the number of
regressors in a model give low Posterior Model Probability (PMP). Therefore, this research
aimed at eliciting a modified g-parameter priors to improve the performance of the PMP and
predictive ability of the model. From the functional form of the g-priors used; the tools of
BMA like Bayes Theorem, Bayes Factor (BF), Posterior Model Probability (PMP), Prior
Inclusion Probability (PIP) and Shrinkage Factor (SF) through the modified g-parameter
priors gj = established the superiority of the consistency's conditions and asymptotic
properties of the prior(s) using the Fernandez, Ley and Steel (FLS) models (1 & 2); and
respectively with as sample sizes. The result from the analysis revealed that the
performance of PMP was reliable with the least standard deviations (0.1994 SD 0.0411) and
(0.1086SD0.000) for model 1 and model 2 respectively; and it was convergent with the
highest means (0.5378Mean0.9577) and (0.8342Mean1.000) for model 1 and model 2
respectively. For the three modified g-parameter priors, the best reliability occurred when n =
100; 000 for Model 1 and Model 2 with (0.0631, 0.0521 and 0.0411) and (0.00, 0.00 and
0.00) respectively; also, the best convergence occurred with (0.9343, 0.9460 and 0.9577)
and (1.00, 1.00 and 1.00) for Model 1 and Model 2 respectively when n = 100; 000. The
predictive performance affirmed the goodness of the elicited g-parameter priors when n = 50
for Point prediction with (2.302, 2.357, 2.357); and when n = 100; 000 for Overall prediction
with (2.332, 2.334, 2.335) which were all closed to the LPS threshold 2.335 according to
BMA specification.

53
Rapid Fire Talks 1
Tuesday 3 September 3.40pm

How GSK is helping to increase statistical capabilities in sub-Saharan Africa

Lindsay Kendall, Annie Stylianou, Nandita Biswas, Agbor Ako

GSK

The demand for qualified experienced statisticians in sub-Saharan Africa is very high,
growing, and outnumbers the supply. In response GlaxoSmithKline (GSK) has committed to
help increase statistical capabilities on the continent as part of the Africa Non-Communicable
Disease (NCD) Open Lab.The Africa NCD Open Lab was established 5-years ago as part of
a series of GSK’s strategic investments to provide sustainable support for scientific research
in the field of NCDs in sub-Saharan Africa. We are currently working in partnership with
funders, researchers and academic groups to share expertise and resources to increase the
scientific understanding behind the unique attributes of NCDs in the African population. How
are GSK statisticians involved?We currently provide statistical support to 19 projects
receiving GSK funding and co-funding with Medical Research Council (MRC) UK and MRC
South Africa. Support ranges from consulting with the project teams on design issues,
approaches to minimise bias, sample size calculations, protocol development and analysis
considerations.We are currently providing both funding and support for 6 PhD students, 5
MSc students (another 5 planned for 2019), 3 internships (another 3 planned for 2019) and 1
fellowship (another 2 planned for 2019 and 2020) in collaboration with the sub-Saharan
Africa Consortium for Advanced Biostatistics Training (S2ACABT), the London School of
Hygiene and Tropical Medicine (LSHTM) and the African Institute for Mathematical Sciences
(AIMS).We hosted a week-long workshop in 2017 for 18 African statisticians at GSK’s
Research and Development facilities in the UK. We rolled out focused statistical, research
and NCD-specific training sessions - and provided an opportunity for networking.Other
initiatives are also in development!So, what’s next?We plan to continue the current statistical
support GSK provides in sub-Saharan Africa. Moving forward GSK welcomes suggestions,
partnership and collaborations!

54
Rapid Fire Talks 1
Tuesday 3 September 3.40pm

Valuation of preference-based measures: could borrowing strength from existing

countries’ valuations produce better estimates

Samer Kharroubi
American University of Beirut

Background: Valuations of preference-based measure such as EQ-5D and/or SF6D have

been conducted in different countries. There is a scope of borrowing strength from existing
countries’ valuations to generate better representative utility estimates. This is explored
using two case studies modelling UK data alongside Japan samples to generate Japan
estimates.

Methods: Data from two SF-6D valuation studies were analyzed where, using similar
standard gamble protocols, values for 241 and 249 states were devised from representative
samples of Japan and UK general adult populations, respectively. Two nonparametric
Bayesian models were applied to estimate a Japan value set, where the UK results were
used as informative priors in the first model and subsets of the Japan dataset for 25 and 50
health states were modelled alongside the full UK dataset in the second. Generated
estimates were compared to a Japan value set estimated using Japan values alone using
different prediction criterion.

Results: The results allowed the UK data to provide significant prior information to the Japan
analysis by generating better estimates than using Japan data alone. Also, using Japan data
elicited for 50 health states alongside the existing UK data produces roughly similar
predicted valuations as the Japan data by itself.

Conclusion: The promising results suggest that the existing preference data could be
combined with data from a valuation study in a new country to generate preference weights,
thus making own country value sets more achievable for low-middle income countries.
Further research and application to other countries and preference-based measures are
encouraged.

55
Rapid Fire Talks 1
Tuesday 3 September 3.40pm

Statistical reproducibility for (multiple) pairwise tests in pharmaceutical product

development

Andrea Simkus
Durham University / AstraZeneca

In pharmaceutical discovery and development, current decisions of statistical tests play an

important role. Statistical reproducibility is another property, like power, that provides
valuable information and extends our decision-making capacity. We concentrate on a real
world test scenario, which employs two pairwise tests: the t-test and its nonparametric
counterpart, the Wilcoxon Mann-Whitney test. The scenario involves 6 test groups, which
are given an increasing dosage of a drug and a chosen observation is subsequently
recorded. The aim of the test is to decide what dosage of the drug is the most effective one.
The main motivation is the study of this scenario and its reproducibility. Our aim is to answer
the question: Would a repeat of the experiment lead to the same test decision? We employ
nonparametric predictive inference (NPI) for reproducibility for the two pairwise tests. The
study of reproducibility of the t-test is a novelty. We also consider reproducibility for the
overall decision. We adopt NPI bootstrap for calculating reproducibility of both tests as
opposed to calculating exact lower and upper probabilities. NPI bootstrap is an alternative to
standard bootstrap and it is based on a repeated application of Hill’s assumption A(n). We
present two approaches. Firstly, we look at separate pairwise comparisons and their
reproducibility by creating NPI bootstraps for both test groups, applying the pairwise
comparison on those and counting how many times we get the same test decision.
Secondly, we study the reproducibility of the decision for the actual test scenario when we
carry out multiple comparisons, again by using NPI bootstrap. Among others, we conclude
that in this scenario both pairwise tests lead to similar reproducibility. When we move from
doing separate to multiple pairwise comparisons, reproducibility probability goes significantly
down. We also examine different ways of presenting statistical reproducibility and we
introduce them.

56
Rapid Fire Talks 1
Tuesday 3 September 3.40pm

Assessing local chlamydia screening performance by combining survey and

administrative data to account for differences in local population characteristics

Nathan Green1, Ellie Sherrard-Smith1, Clare Tanton2, Pam Sonnenberg3, Catherine Mercer3,
Peter White1
1
Imperial College London, 2 LSHTM, 3 University College London, Imperial College London

Reducing health inequalities requires improved understanding of the causes of variation.

Local-level variation reflects differences in local population characteristics and health system
performance. Identifying low- and high-performing localities allows investigation into these
differences. We used Multilevel Regression with Post-stratification (MRP) to synthesise data
from multiple sources, using chlamydia testing as our example. We used national probability
survey data to identify individual-level characteristics associated with chlamydia testing and
combined this with local-level census data to calculate expected levels of testing in each
local authority (LA) in England, allowing us to identify LAs where observed chlamydia testing
rates were lower or higher than expected, given population characteristics. Taking account
of multiple covariates, including age, sex, ethnicity, student and cohabiting status, 5.4% and
3.5% of LAs had testing rates higher than expected for 95% and 99% posterior credible
intervals, respectively; 60.9% and 50.8% had rates lower than expected. Residual
differences between observed and MRP expected values were smallest for LAs with large
proportions of non-white ethnic populations. London boroughs that were markedly different
from expected MRP values ( 90% posterior exceedance probability) had actively targeted
risk groups. This type of synthesis allows more refined inferences to be made at small-area
levels than previously feasible.

57
Rapid Fire Talks 1
Tuesday 3 September 3.40pm

Counterfactual Analysis Using Censored Duration Data

Andres Garcia-Suaza1, Miguel Delgado2

1
Universidad EIA, 2 Universidad Carlos III

We propose standardization techniques for the duration distribution in a population with

respect to another taken as standard using right censored data, which forms a basis for
counterfactual comparisons between distributional features of interest. Alternative
standardizations are based on either a proportional hazard semiparametric specification or a
nonparametric specification of the underlying conditional distribution. Applications to the
restricted mean survival time and the hazard rate are discussed in detail. The proposal is
applied to the counterfactual analysis of spells of unemployment duration gender gaps in
Spain between 2004-2007. The behavior in small samples is investigated using Monte Carlo
experiments.

58
Rapid Fire Talks 1
Tuesday 3 September 3.40pm

Drug utilisation reporting using administrative claims data in Ireland

Lea Trela-Larsen1, Cara Usher2, Laura McCullagh2, Michael Barry2, Cathal Walsh1
1
University of Limerick, 2 National Centre for Pharmacoeconomics, Ireland

Introduction: In Ireland the National Centre for Pharmacoeconomics (NCPE) assess

evidence about the health-benefit, cost-effectiveness (value for money) and budget impact of
new drugs compared to currently available treatments. These assessments inform funding
decisions for new medicines. Currently there is a lack of assurance post-reimbursement that
the predictions made in these assessments were realistic. Analysis of national administrative
pharmacy claims data could provide an important source of evidence on how these drugs
are used in practice. This evidence would be useful to inform NCPE assessments and the
review of funding decisions for these drugs. We aimed to engage with stakeholders and elicit
from them the information they deem necessary to fulfil their roles in the drug reimbursement
process. This engagement informed the development of drug utilisation reports.

Methods: Stakeholders completed an initial questionnaire to establish their role and

information requirements. Their responses helped us to develop our analyses and reports.
Draft drug utilisation reports, based on the analysis of national pharmacy claims data, were
developed using R. Stakeholders were then asked to provide feedback on these draft
reports.

Results: Stakeholders rated information on expenditure (9.2), health outcomes (9) and
patient numbers (8) as the highest for importance on a scale from 0 ‘not at all important’ to
10 ‘very important’. Draft reports included information on patient prevalence and incidence
over time, patient demographics, mean dose and treatment duration. Stakeholders found the
initial drafts useful and clear as high-level reports, but asked for more detail in certain areas.

Conclusions: Stakeholders saw drug utilisation evidence as an important additional source of

data to inform decision making. Pharmacy claims data can provide information on
expenditure and patient numbers, but not health outcomes. Data linkage to patient registries
could provide information on health outcomes. We will seek further feedback from
stakeholders to assess how they use our reports.

59
Rapid Fire Talks 2
Tuesday 3 September 3.40pm

Calculation of Relative Threshold Levels for the Capacity of Benefits from the
Arthroplasty Surgery using the Quantile Polynomial Regressions

Sujin Kang1, Jonathan Cook2, Andrew Price3

1
Imperial College London, 2 University of Oxford, 3 Nuffield Department of Orthopaedics,
Rheumatology and Musculoskeletal Sciences, University of Oxford, , , , , , ,

Objective: To calculate threshold levels of the pre-operative patient-reported scores that

patients could have benefit by estimating centiles of the study population.

Method: The polynomial based quantile regression were used to estimate an individual’s
pre-operative patient-reported score based on their probability of gaining a meaningful
improvement after hip or knee replacement. The fractional polynomial logistic regression
were fitted additionally to examine the benefit of the baseline covariates. Participants:
209,761 patients who underwent hip replacement, and 222,933 patients who underwent
knee replacement. Primary outcome measures were the change (post – pre-operative) of the
Oxford Knee Score (OKS) and the Oxford Hip Score (OHS), respectively.

Outcomes: The 3rd degree polynomial based quantile regressions to the probability of
achieving a meaningful improvement were estimated and validated. The upper thresholds
were estimated as being a pre-operative score of 40 for hip and 41 for knee (48 is the
maximum score). The predicted (i.e., relative) threshold levels from the 20th quantile
regressions (i.e., 80% population coverage) were 25 for the OKS and 32 for the OHS; they
showed high sensitivity and poor specificity compared to the threshold levels from the 10th,
30th and 50th quantile regressions. The predicted threshold levels from the 50th quantile
regressions (i.e., 50% population coverage) were 36 for the OKS and 38 for the OHS. As the
threshold is reduced, the probability of meaningful improvement increased towards
maximum probability values.

Conclusion: The polynomial based advanced modeling approaches have been developed
using the large routine patient-reported data; it allowed the pre-operative Oxford scores to be
used as a guide for referral. The models showed nicely differing proportions of individuals
and their expected outcomes with good performance, but also implied the uncertainty about
where an individual patient will end up.

60
Rapid Fire Talks 2
Tuesday 3 September 3.40pm

Mapping the uncertain future of longevity: an ensemble approach for forecasting

mortality

Mark Hancock, Guangquan Li, Pete Philipson

Northumbria University

Life expectancy has been on the rise in most countries due to the continuous development in
healthcare over the past century. This is positive. However, with the rising average age of
human, current plans for pensions and healthcare may need to be revised to remain
affordable for a country. To inform appropriate changes to these plans, mortality forecasts
need to be reliable with various sources of uncertainty incorporated. The goal of this project
is to develop a model ensemble approach to forecast human mortality. Based on Bayesian
model averaging (Hoeting et al. 1999), our approach forecasts mortality by probabilistically
combining the forecasts from a suite of models so that model uncertainty is accounted for. In
a conventional single-model forecasting approach, this source of uncertainty is ignored.
Here, we propose a mixture modelling approach to estimate the posterior model
probabilities, which are then used as the model weights. Each forecast model contributes
towards the final forecasts whilst the estimated model weights quantify the relative
contributions of these model.A feature of the proposed mixture approach is that the fitting of
the individual models and the estimation of their model weights are carried out within a
single fitting. This contrasts the two-stage approach proposed by Kontis et al. (2017) where
the model weights are estimated through a cross validation setting. In a situation where
there are only a limited number of time points with data available, the single fitting feature
from the mixture approach becomes beneficial.To compare the forecast performance of the
mixture and the two-stage approaches, a simulation study is performed, assessing the
performance of both point and interval forecasts. Our findings show promising results for the
mixture modelling approach, whilst both BMA techniques are shown to perform more reliably
than using a single-model approach.

61
Rapid Fire Talks 2
Tuesday 3 September 3.40pm

Multivariate Correction for Attenuation of a Congeneric Measurement Model without

Correlated Errors: A Test of the Bock and Petersen Approach

Scott Colwell
University of Guelph

Researchers often employ multi-item measures to test complex theoretical relationships

among constructs of interest. These items are often estimated as a congeneric
measurement model where items load on to their respective latent variables with differing
degrees of precision and error. However, when testing the relationships among the
constructs, latent variables are often replaced with composite variables to ease computation.
Psychometric research shows that measurement error—the amount of observed variance
that cannot be statistically attributed to the construct or known source of measurement
error—causes correlations to be attenuated.While attenuation in correlations due to
measurement error might be less problematic if all correlations in a network were attenuated
to the same degree, observed correlations are usually based on different underlying
relationships with different levels of measurement error variance. The result is that
correlations are usually attenuated to varying degrees, thereby distorting the observed
pattern of relationships relative to the underlying pattern of relationships. This becomes a
formidable problem when drawing conclusions about patterns of relationships between
constructs that are not influenced by distorting effects of psychometric artifacts such as
measurement error variance. Spearman (1904) established a bivariate correction for
attenuation between two variables with uncorrelated errors. However, Bock and Petersen
(1975) proposed that this approach is limited to only the bivariate condition in that in the
multivariate context, it may produce an indefinite correlation matrix that is not admissible for
modeling the between construct relationships. While Bock and Petersen (1975) derived a
multivariate correction for attenuation, it has yet to be tested in any extensive ways. As a
result, this research employs a Monte Carlo simulation to test the Bock and Petersen
approach with a congeneric measurement model without correlated errors. Results show the
Bock and Petersen approach limits the amount of relative bias in the correlation estimates
and furthermore ensures that the resulting correlation matrix is at least semi-positive definite.

62
Rapid Fire Talks 2
Tuesday 3 September 3.40pm

On Parameter Estimation of the Hidden Gaussian Process in perturbed SDE.

Li Zhou1, Yury Kutoyants2

1
Shandong University, Weihai, 2 Le Mans University, Le Mans, France

We present results on parameter estimation and non-parameter estimation of the linear

partially observed Gaussian system of stochastic differential equations. We propose new
one-step estimators which have the same asymptotic properties as the MLE, but much more
simple to calculate, the estimators are so-called "estimator-processes". The construction of
the estimators is based on the equations of Kalman-Bucy filtration and the asymptotic
corresponds to the small noises in the observations and state (hidden process) equations.
We propose conditions which provide the consistency and asymptotic normality and
asymptotic efficiency of the estimators.

63
Rapid Fire Talks 2
Tuesday 3 September 3.40pm

Model Averaging in a Multiplicative Heteroscedastic Model

Alan Wan
City University of Hong Kong

This paper develops a model averaging method for combining maximum likelihood
estimators of unknown parameters in a multiplicative heteroscedastic model. A weight
choice criterion based on minimising a plug-in estimator of the squared prediction risk of the
model average estimator is developed, and the asymptotic optimality of the resultant
estimator is established. Our estimator takes explicit account of the model of
heteroscedasticity for the error process as well as the uncertainty about the regressor choice
in both the mean and variance functions of the model, whereas all existing studies of model
averaging are based on estimators that either ignore heteroscedasticity, or assume there is
no uncertainty governing the regressor choice in the variance function. Results of our
simulation and real data analysis show that more efficient estimates than those produced by
heteroscedasticity-robust FMA estimators can be obtained using this new estimator, which
also performs better than the traditional AIC and BIC model selection estimators.In recent
years, statisticians have gravitated to the use of heteroscedasticity-robust standard errors for
the OLS estimator as an outgrowth of the popularity of this technqiue. However, as
emphasised by Leamer (2010), this does not necessarily imply that one should no longer
model heteroscedasticity because by ignoring heteroscedasticity one can potentially forego
substantial gains in estimator's efficiency. As discussed in the Introduction, possible
efficiency loss if heteroscedasticity is modeled incorrectly is one major reason for
statisticians to shy away from modeling heteroscedasticity, and instead favour the simpler
solution offered by the robust standard errors approach.Model averaging offers a way
around the problem of model mis-specification by hedging estimates against the very bad
models. For this reason, model averaging habitually delivers more accurate estimates of the
unknown parameters in finite samples and should be viewed as an alternative to the present
status quo of using the less efficient OLS estimator combined with heteroscedasticity-robust
standard errors .

64
Rapid Fire Talks 2
Tuesday 3 September 3.40pm

Robustness of Space-Filling Orthogonal Array Based Composite Design to Missing

Observation

Abimibola Oladugba, Ezechukwu Okeke

Department of Statistics, University of Nigeria, Nsukka

Robustness of Space-Filling Orthogonal Array Based Composite Design to Missing

ObservationAbimibola Victoria Oladugba1 & Ezechukwu Fidelis Okeke21,2 Department of
Statistics, University of Nigeria, NsukkaMissing observation occurs whenever a valid
observation is not available for anyone of the experimental unit. In this work, a new space-
filling orthogonal-array based composite minimax loss designs which are robust to one
missing observation and also enable parameters to be estimated without loss of efficiency
were proposed. The designs were constructed using a two-level factorial design and a three-
level orthogonal array with space-filling properties such as uniformity and maximin distance.
The proposed designs and other composite designs such as space-filling orthogonal-array
based composite designs, orthogonal array composite designs, small composite designs
and central composite designs were compared based on D-efficiency and precision of
regression coefficient estimates by calculating the generalized scaled standard deviations for
full model, linear, quadratic, and bilinear terms respectively. It was observed that space-
filling orthogonal-array based composite minimax loss designs performed better than other
composite designs for all cases.

65
Rapid Fire Talks 2
Tuesday 3 September 3.40pm

Easy and intuitive multivariate normal quadrant probabilities

Julia Crook
Mayo Clinic

There are well-established formulae for the calculation of multivariate normal quadrant
probabilities in the cases of 2 and 3 dimensions. We re-arrange these formulae to be in a
simpler and more intuitive form that is easier to remember and implement. We do this using
a transformed version of the correlation: ρ* = ρ*(ρ) = arcsin(ρ) /(π/2). This transformation is
monotonic and such that ρ* takes slightly attenuated but similar values to ρ: values of ρ* ,
like ρ, are in the range -1 to 1 and are equal to ρ when it is 0, 1 or -1. Two other special
cases that are easy to remember are: ρ* (1/2) = 1/3 and ρ* (-1/2) = -1/3. The simplified
formulae are, with usual notation: P(Z1>0, Z2>0)= 1/4 (1 + ρ12*), and P(Z1>0, Z2>0, Z3>0)
= 1/8 (1+ ρ12*+ ρ13*+ ρ23*). For 4 dimensions, we can extend this to create an exact
formula that applies when one of the pairwise correlations is zero. For higher dimensions we
have similar exact formula that also require some of the correlations to be zero. These
formulae are intuitive and appealing, and subsequently easy to remember and apply.In
addition to the situations where quadrant probabilities can be obtained exactly, they can also
be useful for rule-of-thumb or approximate calculations. For example, from the above it is
easy to see that if Z1 and Z2 are bivariate normal with mean zero and correlation 0.5 then
P(Z1>0, Z2>0) = 1/3 and hence that P(Z1>0, Z2<0) = 1/6 . Similarly for a trivariate normal
where all three pairwise correlations are 1/3, we have P(Z1>0, Z 2>0, Z3>0) = 1/4 along with
P(Z1>0, Z2>0, Z3<0) = 1/12. The simple formulae, as well as aspects of their derivation,
may be particularly useful in a teaching setting.

66
Rapid Fire Talks 3
Tuesday 3 September 3.40pm

Program evaluation and causal inference for distributional and functional data:
estimation of the effects of retirement on health outcomes

Andrej Srakar
Institute for Economic Research (IER), Ljubljana and Faculty of Economics, University of
Ljubljana

Statistical analysis of complex, i.e. non-standard data is gaining ground. Analysis of

compositions, intervals, histograms, distributions and functions has become more and more
common in contemporary statistics and econometrics. Despite several types of regressions
existing for symbolic data (e.g. Billard and Diday, 2002; 2006; Dias and Brito, 2015; Irpino
and Verde, 2015), causal inference has not been studied so far adequately here.
Furthemore, only slowly is it gaining ground using functional data (Zhao et al., 2018). We
develop a statistical theory for using instrumental variables with symbolic distributional data,
related to the prevailing usage of regressions in such situations, the so-called "two
components" Irpino and Verde model (Irpino and Verde, 2015). We apply the findings to a
pressing problem in the analysis of the aging process: the effects of retirement on health
outcomes. Charles (2004), Neuman (2008), Latif (2013), Insler (2014) and Eibich (2015)
conclude that retirement may lead to significant health improvements, but other studies find
negative retirement effects (e.g. Dave et al., 2008; Behncke, 2012). We use a panel dataset
of Survey of Health, Ageing and Retirement in Europe (SHARE) in Waves 1-6. To address
reverse causality in the relationship of retirement and health behaviours we use as an
instrument changes in retirement age (Komp, 2017). A novelty in the approach is that we
treat countries as units and the variables are aggregated over countries. In this manner, we
estimate the effect of the distribution of retirement across countries on distribution of health
outcomes over countries (the instrumental variable is distributional as well). We extend the
analysis for functional data and aggregate the variables to nonparametric functions, later
used in functional linear models. As program evaluation and causal inference has so far not
been studied with distributional data (and very seldom with functional data) the article is a
significant step ahead in the analysis of complex data.

67
Rapid Fire Talks 3
Tuesday 3 September 3.40pm

Estimating the correlation between bivariate survival endpoints with semi-competing

risks

Yinghui Wei, Lexy Sorrell, Małgorzata Wojtyś

University of Plymouth

Our objectives are to estimate the correlation between bivariate semi-competing risk time-to-
event endpoints. Traditional methods to calculate the correlation cannot be used due to the
common censoring of time-to-event endpoints. We develop a copula-based approach to
estimate the association between semi-competing risk endpoints, where a terminal endpoint
can censor the non-terminal endpoint but not vice-versa. Copula based likelihood functions
are derived to describe the survival outcomes and the association between them. We
estimate the association parameter of a copula and the hazard functions of the marginal
distributions, subsequently transform the association parameter to a correlation coefficient.
We use several copulas to estimate the association parameter between endpoints, and use
information criteria to choose copula models. The proposed methods are applied to a clinical
data set and evaluated by an extensive simulation study.

68
Rapid Fire Talks 3
Tuesday 3 September 3.40pm

Joint modelling of multiple primary outcomes in clinical trials with missing data.

Victoria Vickerstaff, Gareth Ambler, Rumana Z. Omar

University College London

Background: In clinical trials, sometimes multiple primary outcomes are specified. The
outcomes maybe of different types, say a mix of survival and continuous outcomes. For
example, in a trial investigating the effect of a health intervention on cannabis users, the
primary outcomes may be the time to psychiatric relapse and the level of cannabis in the
urine. These outcomes are typically associated. Joint models can be used to link survival‐
type outcomes with continuous outcomes and could provide better insights into the
intervention effect. The survival and continuous outcomes may be analysed using a survival
submodel and a longitudinal submodel, respectively. These submodels can be linked. One
approach to link the submodels is to share parameters between them. Another approach is
to use joint random effects.

Objectives: This study evaluates the performance of joint models in terms of bias and
efficiency of the estimated treatment effects. The results are compared to the estimates
obtained when analysing the outcomes separately.

Methods: Several simulation scenarios were investigated by varying the strength of the
association between the outcomes and the level of missing data. Joint models which share
parameters or have joint random effects were implemented using the R packages: JoineR,
jointModel and FrailtyPack.

Results: The results show that when the outcomes are analysed separately, parameter
estimation for the survival outcome is typically biased. The bias is reduced when using the
joint models. The joint models had increased standard errors for the estimated treatment
effect on the survival outcome compared to analysing the outcomes separately. When there
is strong association between the outcomes, the joint random effects models and the model
that utilised shared parameters between the longitudinal submodel and the survival
submodel performed best in terms of the mean square error of the estimated intervention
effects on the survival outcome.

69
Rapid Fire Talks 3
Tuesday 3 September 3.40pm

Comparing methods of defining hyposmia in a large incident cohort of patients with

Parkinson’s disease

Sofia Kanavou1, Michael Lawton1, Vanessa Pitz2, Yoav Ben-Shlomo1, Donald Grosset2
1
University of Bristol, 2 University of Glasgow

OBJECTIVE: Impaired olfaction (hyposmia) is a common feature in patients with Parkinson’s

disease (PD), yet there is no standard method to define it. A comparison of four published
methods of defining hyposmia using the Tracking Parkinson’s cohort, could explain why this
is difficult to be established.

METHODS: The 40-item University of Pennsylvania Smell Identification Test (UPSIT) or

Sniffin’ Sticks (Sniffin) were used to measure olfaction but harmonised, as cut-points for
Sniffin were converted from UPSIT values using the Item Response Theory method. The
proportion of hyposmic patients was calculated using 4 previously reported definitions:1)
Scores ≤ 23 for those aged ≥60 years and ≤ 28 for those <60; 2) Severity grading absolute
values, using gender-corrected cut-points; 3) Scores ≤15th centile for normals, age and
gender-corrected; and 4) scores ≤15th centile for normals, uncorrected. Control group with
age and gender matched to our PD cohort was simulated from published normative data.
Between method agreement was measured by Cohen’s kappa and Gwet’s AC1.

RESULTS: The proportion classified as hyposmic varied between 75.3-98.3% for 1,674
recently diagnosed PD patients and 14.5-52.8% in the simulated healthy group, depending
on method. The level of agreement varied significantly: Cohen’s kappa ranged from 0.10 to
0.65. Gwet’s AC1 ranged from 0.7 to 0.85 and was often substantially higher than the kappa:
1) 0.74>0.11 (Method 1 vs Method 2); 2) 0.85>0.19 (Method 2 vs Method 4). Observed
concordance matches better the agreement levels of Gwet’s AC1. It estimates agreement
probability more accurately, whereas kappa is affected by the prevalence of hyposmia and
produces more biased results. The method of defining hyposmia should be carefully chosen
when setting research questions and comparing results across different studies of PD.

70
Rapid Fire Talks 3
Tuesday 3 September 3.40pm

Assessment of alternate data-sources to meet the challenges of timelier and accurate

registration of fact-of-death, as needed by record-linkage studies.

Paula Curnow1, Charlotte Carr1, Oliver Smith1, Sheila Bird2

1
NHS Digital, 2 MRC Biostatistics Unit, Cambridge

High quality statistics, analysis and advice help Britain make better decisions. The challenge
for producers of statistics is to get the right data and information at the right time into the
hands of those who use them to make a difference and improve care and health outcomes
During the H1N1 influenza pandemic in 2009, the UK’s Scientific Advisory Group in
Emergencies could not access timely information on the deaths of infants (0-4 years) and
young children (5-14 years) in England and Wales because of the late registration of inquest
deaths. Accordingly, in 2010, evidence by the Royal Statistical Society to the Inquiry into
Scientific Advice in Emergencies by the House of Commons Science and Technology
Committee highlighted the need for England and Wales to achieve the timely registration of
all deaths. In 2017 Professor Sheila Bird, working with NHS Digital and Office for National
Statistics, undertook analyses to investigate if information on informal data of death on NHS
Digital’s Personal Demographics Service (PDS-DOD) might be used as a key source for
more timely data than afforded by the date of death as ultimately registered by Office for
National Statistics. Initial analysis explored the comparative timeliness and accuracy of these
respective data sources and concluded that, for deaths in 2011-14, PDS-DOD was
insufficiently available and insufficiently accurate. However recommendations of additional
analyses were made to understand why. This talk describes the results of these most recent
analyses, for deaths in 2011-16. By linking two rich national person level data sources, we
consider the influence that the primary use and business processes associated with each
has in relation to timeliness and accuracy of these death-dates reported within. Does data
entry error or demographic characteristics of the departed influence timeliness? Finally, we
reflect on how the results inform onward use for monitoring and research both now and in
the future.

71
Rapid Fire Talks 3
Tuesday 3 September 3.40pm

Optimisation using emulation in disease modelling. How to use potential

improvements to identify new design points.

Daria Semochkina
University of Southampton

The primary interest of this project is in optimising computationally expensive (and possibly
competing) noisy objective functions in disease modelling and decision-making. A decision
maker potentially could be interested in minimising deaths and minimising the cost of an
intervention. Those would most likely be competing goals.The simplest emulation-based
optimisation recipe goes like this: we build a sampling plan, we then calculate the responses
in these points and fit an emulator to this data, then we can locate the input parameters that
are as close to the true minimum of the function as we can make it. However, the
optimisation task is not complete until we validate this function value against the true,
expensive function itself. We then can, and indeed should retain more of our computational
budget so as to allow this search and update process to be repeated many times, adding
multiple so-called infill points. When performing a search for a new candidate, we wish to
position the next infill point at the value which will lead to an improvement on our best-
observed value so far. One possibility is to calculate a so-called expected improvement at a
finite set of points and select the best one. This, however, implies that our current minimum
observation is the correct representation of the function’s value at that point. If there is noise
present, using only the noisy observations might be risky, since the noise may introduce
errors in the ranking of the observations. In this talk, I will discuss different approaches to
emulation-based optimisation that are designed to overcome this drawback in applications to
optimal decision making.

72
Rapid Fire Talks 3
Tuesday 3 September 3.40pm

Analysing time-to-event data with recurrent events

Christiana Kartsonaki
University of Oxford

We consider individuals which are followed up over time and some event of interest is
observed. The event may occur more than once during follow-up and observation ends
either at censoring or death. Such data frequently arise in medical studies, but often only
time to the first event is considered as the outcome of interest. For example, in a study of
cancer patients, patients may have more than one recurrences, or in a study of risk factors
for pancreatitis individuals may have several hospitalisations for pancreatitis. We consider
various models for assessing the associations of explanatory variables with recurrent events,
including exponential based, Weibull based and semiparametric models. The resulting
estimates and their precision is compared and the different models are illustrated using a
dataset on colorectal cancer recurrences.

73
Rapid Fire Talks 4
Tuesday 3 September 3.40pm

Research and Application of Data Governance Architecture Based on Activity Theory

Huaihai Hui1, Des McLernon2, David Allen2

1
Chinese Academy of Sciences /University of Leeds, 2 University of Leeds

With the rapid increase in the amount of data generated by enterprises and the
diversification of data sources, the complexity of the utilization of such massive and multi-
source heterogeneous data is greatly enhanced. Therefore, the issue of how to carry out
data governance has become a very important topic. In this research, Activity Theory（AT）
is proposed as a methodological and analytical framework to study Data Governance
Architecture in the Big Data Platform. First, the study defines the data, and then designs the
data classification and data model construction methods. Data can be called the Subject in
AT. Second, this study proposes how to develop data management policies, how to design
data architectures, and how to set data standards and manage data quality, which is the
Rule in AT. Third, the study designed a data management organization consisting of
decision makers and people directly related to data, which is called a Community in AT.
Fourth, this paper allows the data owner, data supplier, data manager, and data user to be in
the Division of Labor of AT. Fifth, this study developed the Tools mentioned in AT, including
data acquisition tools, data cleaning tools, data integration tools, data search tools, data
analysis tools, and data visualization tools. Sixth, this study presents the results of a data
governance architecture that includes data assets, data products, data services, and data
distribution, which are defined as Objects in AT. Finally, this study runs this data governance
architecture as a complete system on the big data platform of an eighth-largest insurance
company in the world.

74
Rapid Fire Talks 4
Tuesday 3 September 3.40pm

Analysis of clickstream data

Ryan Jessop
Clicksco

Online user browsing generates vast quantities of typically unexploited data. Investigating
this data and uncovering the valuable information it contains can be of substantial value to
online businesses, and statistics plays a key role in this process. The data takes the form of
an anonymous digital footprint associated with each unique visitor, resulting in 10^6 unique
profiles across 10^7 individual page visits on a daily basis. Exploring, cleaning and
transforming data of this scale and high dimensionality (2TB+ of memory) is particularly
challenging, and requires cluster computing. We consider the problem of predicting
customer purchases (known as conversions), from the customer’s journey or clickstream,
which is the sequence of pages seen during a single visit to a website. We consider each
page as a discrete state with probabilities of transitions between the pages, providing the
basis for a simple Markov model. Further, Hidden Markov models (HMMs) are applied to
relate the observed clickstream to a sequence of hidden states, uncovering meta-states of
user activity. We can also apply conventional logistic regression to model conversions in
terms of summaries of the profile’s browsing behaviour and incorporate both into a set of
tools to solve a wide range of conversion types where we can directly compare the predictive
capability of each model. In real-time, predicting profiles that are likely to follow similar
behaviour patterns to known conversions, will have a critical impact on targeted advertising.
We illustrate these analyses with results from real data collected by an Audience
Management Platform (AMP) - Carbon.

75
Rapid Fire Talks 4
Tuesday 3 September 3.40pm

Remote fault detection: identify faulty refrigeration units using signal processing and
machine learning on smart metering data

Phuong Pham, Laura Shemilt, Kaelon Lloyd, Terry Phipps, Sean Stephenson
Centrica

Background: There is a wide variety of household appliances which are operated using
electricity and these electrical appliances are prone to failure, malfunction or degradation in
operation efficiency to different degrees. Appliance faults or malfunction can lead to risks of
power loss or fire. Therefore, early faults detection is important to avoid any consequences.
Fault detection requires a good understanding of appliance consumption's profile. While
using individual monitoring device for each appliance can be expensive, using Non-Intrusive
Load Monitoring techniques on the large volume of smart metering data can help to
disaggregate energy consumption for appliance from total household consumption at lower
cost.

Objective: We have developed a prototype method to detect faulty fridge freezers using a
residential smart meter dataset.

Methods: The input data, the household power signal, is sampled at low frequency (10s
interval) and no extra sensors were installed. In the first stage, fridge's power signal is
isolated from total household level. Due to the nature of cyclical patterns in fridge's power
profile, a mixture of signal processing technique and machine learning, such as spectral
analysis and grid search, can be used. In the next stage, the extracted fridge cycle is
examined against historic behaviour to detect any outliers from the normal state. Outlying
points were detected using several methods including forecasting (arima), global outlier
detection (modified z-score) and local outlier detection (local outlier factor, density-based
spatial clustering of applications with noise and relative density outlier score). Outliers, which
relate to faults in the fridge, can be used to detect breakdowns remotely. The algorithm was
developed in R using OOP paradigm to allow fully production.

Conclusion: Fridge’s cycle was derived from total household consumption and it was used in
a model combines of different outlier detection techniques. The output from the model
enable engineers to identify faulty appliances remotely.

76
Rapid Fire Talks 4
Tuesday 3 September 3.40pm

Data Science Education, Skills and Industry in Europe

Berthold Lausen1, Alexander Partner1, Stephen Lee2, Henrik Nordmark3, Mahdi Salhi3,
Christopher Saker12
1
Department of Mathematical Sciences, University of Essex, 2 Mathematics in Education
and Industry (MEI), 3 Profusion – a data science and marketing services company

Classification societies focussed on methodologies as for example classification/supervised

learning, clustering/unsupervised learning and multidimensional scaling defining the
mathematical foundations of the emerging discipline data science, since the Classification
Society was founded in London 1964. A special issue of the Journal of Data Science and
Analytics (Flach et al., 2018) has papers devoted to current issues in Data Science viewed
from a European perspective which were in the European Data Science Conference
(EDSC), an invitation only event organised by the late Professor Sabine Krolak-Schwerdt
and her team in November 2016 in Luxembourg as the inaugural conference of the
European Association for Data Science (EuADS). In this context we discuss challenges and
needs for data science education and skills.The landscape for school mathematics in the
United Kingdom is very different now compared to even just five years ago. During the
period of recent changes in the UK school system we have seen the emergence of
undergraduate (BSc) and postgraduate taught (MSc) qualifications in Data Science. For
example in 2014 the Department of Mathematical Sciences and the School of Computer
Science and Electronic Engineering at the University of Essex have introduced a BSc in
Data Science and Analytics and an MSc in Data Science. The curriculum of these courses
covers compulsory modules from computer science and mathematical sciences, introducing
students to a range of mathematics and statistical topics as well as computing skills such as
artificial intelligence, deep learning, information retrieval, programming, and text analytics. In
this talk we will also review curricula of data science related university degrees, consider
how their content matches industry expectations and discuss plans for further
developments.Flach, P, Spiliopoulou, M, Allegrezza, S, Böhmer, M, Hess, B, Lausen, B
(2018), Introduction to the special issue on Data Science in Europe, International Journal of
Data Science and Analytics 6:163–165.

77
Rapid Fire Talks 4
Tuesday 3 September 3.40pm

Adjusting reviewer scores for a fairer assessment via multi-faceted Rasch modelling

Caterina Constantinescu
The Data Lab, University of Edinburgh

Selecting submissions for a conference can be viewed as a measurement problem: in

principle, organisers aim to accept the ‘best' submissions, but the precise manner in which
this is achieved can vary considerably. It is also common to involve multiple reviewers in the
process, and it may not always be the case that all reviewers manage to rate all
submissions. Hence, there is a chance that some particularly harsh reviewers may rate the
same submission (and put it at a disadvantage), or some more lenient reviewers may
happen to rate the same submission and propel it higher in the ranking. A solution to this
issue is offered by multi-faceted Rasch models, which view submission scores as a function
of not just the quality of the submission in itself, but also reviewer severity. This allows to
adjust submission scores accordingly, providing a fairer measurement process.
Conveniently, the R package `TAM` allows to estimate this type of model. In this talk, I will
walk you through an example of how `TAM` was used on data collected as part of the review
process for a data science conference in Scotland.

78
Rapid Fire Talks 4
Tuesday 3 September 3.40pm

The Use and Interpretation of Statistics in Medical Research: An Evaluation of Medical

Students' Attitudes

Alaa Althubaiti
King Saud bin Abdulaziz University for Health Sciences

Objective: To evaluate the medical students’ attitudes toward statistics use and interpretation
in medical research and to assess sex, age and year of course differences in attitudes
toward statistics.

Methods: The Attitudes Toward Statistics in Medical Research (ATSMR) survey for health
students was administered. The survey assessed students' attitudes by responding to 32
items using a 7-point scale.

Results: 327 medical students participated in the survey with a 54.5% response rate.
Although students appeared to appreciate the value of statistics in their professional career,
they had negative to neutral attitudes about their feelings towards statistics, their own
intellectual knowledge and skills in statistics, and the difficulty of the subject. Students aged
23 years and above perceived statistics to be more difficult than younger students, and they
showed less positive feelings towards statistics. Male students showed more positive
feelings toward statistics than females.

Conclusion: It is important for medical educators to allocate the adequate mechanisms and
design targeted education in statistics courses that stimulate medical students' interest in the
field. Innovative ways to deliver the course contents should be explored to improve student’
learning experince.

79
Rapid Fire Talks 4
Tuesday 3 September 3.40pm

Is there a curse of Aaron Ramsey?

Anthony Masters
Nationwide Building Society

Various newspapers and websites have claimed that there is a 'curse of Aaron Ramsey'.The
suggestion is that goals by the Arsenal & Welsh footballer are usually followed by the deaths
of famous people.The intention is to test this theory.I collate two lists of notable deaths in the
UK, from the Daily Mirror and the BBC. This is compared against game data on Aaron
Ramsey, from Soccer Base. Date ranges from a single day (same day) to four days (on the
same day, to within three days) are considered.Based on one football season (August 2017
to May 2018), there does not appear to be such an effect. Ramsey scored eight goals in this
season, only 2 were followed by notable deaths (from the Daily Mirror list) within one day.
This is a lower rate than all two-day periods over that season (45%). Looking at three days
after a Ramsey goal, six goals were followed by a notable death, only slightly higher the
overall rate in that season (69%).Further seasons will be analysed and included in the
presentation.

80
Rapid Fire Talks 5
Tuesday 3 September 3.40pm

STATISTICAL ANALYSIS OF GENETIC RELATIONSHIP OF NIGERIAN AND KENYAN

SHEEP POPULATIONS

Asugha KESTER UGOCHUKWU

FEDERAL UNIVERSITY OF AGRICULTURE, ABEOKUTA, NIGERIA

This study evaluates, the genetic relationship of Nigerian and Kenyan-sheep populations
using Ovine-50k SNP-chip. Ten samples each were collected from four indigenous Nigerian
(Balami and Uda) and Kenyan (Red Maasai and Dorper) sheep breeds and genotyped using
the GeneSeek Genomic Profiler Indice HD Bead chip Quality control was carried out using
the Gen Abel package implemented in R software based on the following criteria; MAF 1%,
SNP Call Rate 90%, Per Individual Call Rate 90%, HWE, IBD ≥ 90%. The descriptive marker
and per id summary function of Gen Abel was used for observed Heterozygosity and
expected heterozygosity. The Analysis of Molecular Variance (AMOVA) was also carried out
on the different markers. Results from this study indicated that, a total of 47 SNPs on the X
chromosome were removed for violating quality control. The remaining 818 SNPs on the X
chromosome all passed the 90% threshold of per individual calling rate. Also, a total of 835
SNPs on the autosomal chromosome were removed for violating quality control. The
remaining 33,176 SNPs on Autosomal chromosome all passed the 90% threshold of per
individual calling rate. The sheep population in two major clades and two subclades
corresponding to the sampled population indicating each population holds their distinct
genetic background with no relationship within and among population, and infers that the
sheep breeds from both countries exhibit different ancestral origin. A 70.13% variation within
country, 42.56% variation observed among individuals within populations compare to
60.24% variation within population and 42.23% within individuals.The genetic differentiation
for autosomes showed the highest differentiation of 0.113 was observed between Red
Maasai and WAD while the lowest differentiation was observed between Red Maasai and
Dorper with 0.062 compared to that of sex chromosome with a record of 0.144 observed
between Red Maasai and WAD while the lowest differentiation was observed between Red
Maasai and Dorper with 0.066.

81
Rapid Fire Talks 5
Tuesday 3 September 3.40pm

THE IMPACT OF AGRICULTURAL PRODUCTIVITY ON ECONOMIC GROWTH IN

NIGERIA

SAHEED ABIDEMI Agboluaje

The Polytechnic, Ibadan, Nigeria

The role of agricultural products in the overall development of a nation cannot be quantified.
It is not only seen as a key to poverty reduction and vehicle for promoting equity, fairness
and social justice but also helps to supply the essential economic ingredients which are
necessary condition for sustainable economic growth. The main objective of this work is to
carry out an empirical investigation on the impact of agricultural productivity on economic
growth in Nigeria, using annual time series data from 2003 to 2017. The Ordinary Least
Square regression method was used to analyze the data. The results revealed that there
exists a high, positive cause and relationship between agricultural productivity and economic
growth. All the variables both dependent and independent includes, the GDP contribution of
the agricultural sector, gross expenditure on agriculture and gross access to bank loans had
a positive impact in the Nigeria economy and were also in relation to economic growth using
the Karl’s Pearson correlation coefficient. The findings show a strong impact on agricultural
policy in Nigeria. Therefore, in order to improve the activities of the agricultural sector,
government should assist the farmers with special incentives such as provision of adequate
funding and infrastructural facilities such as good roads, pipe borne water and electricity to
mention but a few.Keywords: Agricultural sector, Economic growth, Gross Domestic
Product (GDP), Regression.

82
Rapid Fire Talks 5
Tuesday 3 September 3.40pm

A performance comparison between empirical variograms in achieving the best valid

variogram

Esam Mahdi
Qatar University

Modelling the statistical autocorrelations in spatial data is often achieved through the
estimation of the variograms, where the selection of the appropriate valid variogram model
(especially for small samples) is crucial to achieving precise spatial prediction results from
kriging interpolations. To estimate such a variogram, traditionally, we first compute the
empirical variogram (the traditional Matheron or the robust Cressie-Hawkins or the kernel-
based nonparametric approaches). In this article, we conduct numerical studies comparing
the performance of these empirical variograms. In most situations, the nonparametric
empirical variable nearest-neighbor (VNN) showed better performance than its competitors
(Matheron, Cressie-Hawkins, and Nadaraya-Watson). The analysis of the spatial
groundwater dataset used in this article suggests that the wave (hole-effect) variogram
model fitted to the empirical variable nearest-neighbor, VNN, variogram is the most
appropriate choice. This selected variogram is used with the ordinary kriging model to
produce the predicted pollution map of the nitrate concentrations in groundwater dataset.

83
Rapid Fire Talks 5
Tuesday 3 September 3.40pm

Holistic approach to defining climate in ecological studies

Michel d. S. Mesquita1, Jan Ove Bustnes2, Igor Eulaers3, Carla Vivacqua4, André Pinho4
1
Bjerknes Centre for Climate Research, 2 Norwegian Institute for Nature Research, Norway,
3
Aarhus University, Denmark, 4 Federal University of Rio Grande do Norte, Brazil

The study of climate ecology has often focused on indices, such as the North Atlantic
Oscillation (NAO), that capture a dynamic feature of the climate system. However, such
indices may not represent the multivariate nature of the problem, nor local processes that
could be at play. Their focus has also limited understanding of the cause and effect in
ecological studies. We have created an alternative approach that: emulates the local climate,
reduces complexity, and links variables to mechanisms. Our study focuses on the effect of
climate variability in the ecotoxicology of tawny owls. These owls are exposed to
environmental pollutants, which affect their overall health and stress levels. These pollutants
can be transported through the atmosphere, or through the food web (mice, lemming,
insects). This study is unique in that it uses a holistic approach to quantifying ‘climate’ for
ecological studies, i.e.: as a composite atmosphere-land process, influenced by large scale
dynamics, that affects the local environment.The data come from the ERA-Interim reanalysis
project for the period 1979 to 2017. The monthly means of 11 variables were used, which
include atmospheric, snow, and soil processes. We have reduced the dimension through the
use of Principal Component Analysis (PC). Three PCs were retained based on screeplot and
Kaiser criterion analysis. We have interpreted each PC loading as follows: PC1 represents
thermal conditions; PC2 indicates (fair) weather conditions; and PC3 (unfavorable) weather
conditions. Seasonal variability is captured through the interaction between PC1 with PC2
and PC3.Each PC was then related back to the large-scale dynamics. In conclusion, our PC-
based multivariate index not only captures the NAO index, but it also captures the
thermodynamics conditions in the region. It takes into account the full environmental
conditions by using atmospheric, soil, and snow variables. Furthermore, it reveals a robust
and alternative approach to defining ‘climate’ in ecology studies.

84
Rapid Fire Talks 5
Tuesday 3 September 3.40pm

Issues in earthquake modelling.

Zak Varty1, Jonathan Tawn2, Peter Atkinson2, Stijn Bierman3

1
STOR-i Centre for Doctoral Training, 2 Lancaster University, 3 Shell Global Solutions

Earthquakes can be caused both by the motion of tectonic plates and by human activity.
Both types are inherently difficult to model and predict because the physical processes that
cause them are complex, often disputed and difficult to measure accurately. A statistical
approach to seismicity modelling provides a natural and popular way to accommodate these
uncertainties. The most common approach in the literature is to use the Epidemic-Type
Aftershock Sequence (ETAS) model. This model describes earthquake locations and
magnitudes using a marked, self-exciting point process. A parametric or semi-parametric
model for the conditional intensity function of this point process can be inferred from an
observed earthquake catalogue using either frequentist or Bayesian methods. The choice of
parametric forms within this function are typically chosen in line with empirical relationships
known as the modified Omori and Guttenberg-Richter laws.We present alternative, more
flexible, forms for these laws that allow for improved inference, application to smaller
catalogues, and better representation of uncertainty in the final fitted model. Proper inclusion
of this uncertainty is of paramount importance when deciding how to act in order to mitigate
seismic hazard in the future.

85
Rapid Fire Talks 5
Tuesday 3 September 3.40pm

Forecasting agricultural product and energy prices: A simulation-based model

selection approach

Robert Kunst1, Adusei Jumah2

1
Institute for Advanced Studies, 2 Central University, Accra, Ghana

The aim of our contribution is twofold. First, we study whether and to what degree the
dynamic interaction between commodity prices and energy prices can be exploited for
forecasting. Second, we present informative examples for the simulation-based forecast-
model selection procedure. Apart from prediction by competing specifications tobe selected
from a small choice set, we also explore forecast combinations constructed from a
continuum in the same framework. The simulation-based method explicitly permits letting the
forecast model choice depend on the intended time horizon of the forecast. With regard to
classical Granger causality, the evidence supports a causal direction from food prices to fuel
prices, without feedback and somewhat in contrast to our expectations. This causal link,
however, only benefits forecasting accuracy at relatively large sample sizes. Similarly, clear
evidence on considerable seasonal patterns cannot be fused to a seasonal time-series
model that outperforms non-seasonal rivals. Finally, the simulation experiments generally
favor the handling of all price series in first differences.

86
Rapid Fire Talks 5
Tuesday 3 September 3.40pm

Meta-analysis using simple methods successfully derives the big picture for
exemptions of fisheries landing obligation

Mickael Teixeira Alves, David Maxwell, Thomas Catchpole

Cefas

The recent reform of the EU Common Fisheries Policy (CFP) includes a landing obligation of
all catches of commercial fish. This policy is however subject to exemptions for “species for
which scientific evidence demonstrates high survival rates, taking into account the
characteristics of the gear, of the fishing practices and of the ecosystem”. The need for
timely evidence led us to synthetize results from past projects that collected data on fish
vigour at the point of discarding, to inform measures to improve the survival chances of
discarded fish. We analysed fishery and environmental data associated with plaice vigour
from ten projects conducted over two years in seas around England and Wales. The
analysis included basic plots, correlations, multivariate clustering methods and classification
trees. Mixed effects cumulative logit models were fitted to subsets of the data to estimate the
direction and strength of association between vigour and covariates by project. The
combined results of the approaches helped to draw overall conclusions on the factors
influencing the vigour of discarded plaice, highlighting the effects of the fish size and
temperature. Despite strong heterogenous results between projects, this work illustrates how
a combination of statistical tools and careful interpretation can provide scientific evidence for
policy purposes.

87
Rapid Fire Talks 5
Tuesday 3 September 3.40pm

Generalized Regression Control Chart for Monitoring Crop Production in Nigeria

using Asymmetric Distribution

Olatunji Arowolo, Mathew Ekum, Samuel Sogunro

Lagos State Polytechnic, Ikorodu

Recently, Nigeria focused on Agriculture as a way to diversify her economy. Crop

production, which is a proxy to measure agricultural output is considered important. So,
monitoring and controlling crop production (output) among states in Nigeria including FCT is
key. This study aimed at using the generalized regression control chart for this purpose. The
specific objectives include, identifying the effect of each contributing variable (predictor
variable) to crop production for a specified period; setup generalized regression control chart
for crop production in the specified period; examine the points out of control and monitor the
cross-sections (states) that could be responsible for the deviation from expected crop
production limits.The generalized regression control chart was used rather than the usual
conventional control chart. The conventional control chart does not put into consideration
factor(s) that affect crop production (response variable). The generalized regression control
chart however, considers the factor(s) that affect crop production over a specified period and
collapsed the assumption of normality, allowing the usage of other flexibility distributions
other than normal distribution. Therefore, Weibull distribution was used as the underlying
distribution rather than the normal distribution, because it fitted the crop production data
better than normal distribution and it is an asymmetric distribution, which is non negative.
The data were collected from National Bureau of Statistics (NBS). The result of data fitness
showed that Weibull distribution fit the data better than normal distribution. All the predictor
variables, farm-gate price, area planted and fertilizer usage, all have significant effect on
crop production. The result of the generalized regression control chart showed that crop
production is not in control. This implied that some states production is very low compared
to others. It is recommended that the states with low crop production should be encouraged
to embrace crop production and government should do a proper monitoring of crop
production across the state.

88
Rapid Fire Talks 6
Tuesday 3 September 3.40pm

Convergence and heterogeneity in global diets

Thai Le
Bournemouth University

Worldwide obesity has almost tripled since 1975. This trend is the consequence of
demographic, epidemiological and nutrition changes that have taken place as countries
develop and become more globalised. This research examines the global trends in food
consumption for main food categories since 1961 using data from the Food and Agriculture
Organisation of the United Nations (FAO). Preliminary data review discloses a steady rise in
food availability worldwide for the past 50 years, coupled with significant alterations in the
structure of the global diets in terms of both macronutrients and individual food aggregates.
Evidence shows that national diets are evolving over time and across countries in ways that
are both similar yet distinct giving rise to patterns that can be investigated statistically by the
means of cluster analysis. This research is the first attempt in food economics literature to
present the application of an innovative space-time clustering technique inspired by fuzzy
logic and copula functions. Agglomerations of countries characterised by similar dietary
trends are identified and the findings are further analysed to uncover both economic and
cultural factors that most matter in explaining the pace of dietary change as well as the
convergence that is observed globally. The findings will inform the public policy debate on
the relationship between diet and obesity and provide evidence to those interested in
formulating national policies to promote healthy diets.

89
Rapid Fire Talks 6
Tuesday 3 September 3.40pm

Population health in a digital age: the use of social media and wellbeing in Wales

Jiao Song, Alisha Davies

Public Health Wales

Technology and health survey is the first nationally representative population survey for
Wales to examine usage of different social media platforms with state of health and mental
wellbeing in Wales, across population groups. A random probability sampling approach was
used to identify a nationally representative household sample of 1,240 individuals aged 16
years and above living in Wales. We used Chi-square and Fisher-Freeman-Halton tests
followed by binary logistic regression to investigate the differences on usage across social
media platforms and functions across demographic and health groups. We will use
estimated marginal means to report the adjusted proportions of individuals who reported
each usage. 14.6% of over 16 Welsh population has no internet access at home. 63.5% are
super users who are on at least one social media platform on a daily basis or a few times a
day. Facebook, YouTube and WhatsApp are the most favourite platforms overall. People like
spending time on social networking (Facebook and LinkedIn) and sending messages on
WhatsApp. Over 90% of younger generations (16-29 and 30-39) are super users while the
proportions for 60-69 and 70+ are 46% and 16.3% respectively. The differences of social
media usage across Welsh Index of Multiple Deprivation quintiles are statistically significant
at p=.05 but not for gender. Inequalities and differences in social media usage with health
status and mental wellbeing are currently being investigated, using self-reported state of
health (0-100), Short Warwick-Edinburgh Mental Well-being Scale and long-term health
conditions as indexes.

90
Rapid Fire Talks 6
Tuesday 3 September 3.40pm

Social Class and Language Mastery effect in Proficiency Tests

Kaizo Beltrao1, Elza Nascimento2, Monica Mandarino2, Monica Guerra2, Renato Cardoso2
1
EBAPE FGV, 2 Cesgranrio Foundation

Bourdieu and Passeron defended the thesis that the school was the main locus to legitimate
and perpetuate class differences. This is reinforced by the multiple proficiency tests used in
the literature to monitor public policies, which privilege the use of the formal language as part
of the instruments. We propose to use a hierarchical model with ENADE’s results on
standard Portuguese grades, using as covariates, indicators of students’ socioeconomic
status and economic independence. These two indicators were extracted from the
questionnaire filled by the students previously to the exams using Optimal Scaling and
Principal Components.According to the INEP, a department of the Ministry of Education who
conducts the exam, the National Assessment of Student Achievement (ENADE) assess
undergraduate programs through an “exam administered to students who are finishing these
courses in higher education institutions throughout Brazil. The programs are grouped in
three representative areas and each year one group is assessed, meaning that each
program is assessed every three years.” Graduating students of all courses take the exam
every third year. All knowledge areas with at least 100 courses and two thousand students,
mandatorily participate in the exam. These exams, besides a section on specific knowledge
related to the professional area being assessed, poses some questions of general interest.
The section of general interest is composed of two short answer questions plus eight
multiple choices questions. The short answer questions are graded for content and for the
use of standard Portuguese. The Portuguese grade takes into consideration three
components: orthography; lexical and syntactical. Though questions contents in different
exams are not strictly comparable, we hypothesized that the use of formal Portuguese in the
answers is. Confirming Bourdieu and Passeron hypothesis, we find out that socioeconomic
status do have an impact on language proficiency, but at the University level it is not so
strong as measured at lower educational levels.

91
Rapid Fire Talks 6
Tuesday 3 September 3.40pm

“Severity to those who confess?”: Evidence from China’s 6,876 Cases of Intentional
Injuries

Mengjie Xu, Fang Wang, Liang Guo

Shandong University, Weihai

“Leniency to those who confess, severity to those who resist” is the oldest and most
important tenet of contemporary Chinese criminal justice, which was one of the outstanding
features of the Chinese communist revolutionary tradition. The goal is to encourage the actor
to repent and confess his/her crimes, and in return, receive reduced punishment to
encourage him to turn over a new leaf. However, there is a wide perception that confession
usually means the suspect will “rot in jail”. This study intends to verify whether it is true that
“severity to those who confess” and if so, provide an explanation behind this counter-intuitive
phenomenon. We randomly selected trial verdicts of 6,876 intentional injury cases from the
databases of China’s supreme court over the course 2014-2017. Each verdict was read and
manually coded by three law students. We estimated a series of linear regression models
with the length of prison sentence as the dependent variable, confession as the key
independent variable. After controlling 25 variables that can have an effect on the length of
prison sentence, we found that in general, confession is positively associated with the
dependent variable (0.16, p<0.01). however, if a suspect has turned him/her-self in, then the
length of prison sentence can be reduced (-0.20, p<0.05). The interactive effect between
confession and turning-oneself-in is not statistically significant. That is to say, confession
tends to receive an increase in punishment. Many confessions may be police-induced. The
finding reflects the poor investigatory capacity of judicial organs and the poor protection of
suspects’ right to remain silent. Our finding indicates that there is a pressing need for the
authority to keep the mandatory electronic recording of interrogations and to protect the
rights of criminal suspects and defendants against the possibility of judicial abuse of power.

92
Rapid Fire Talks 6
Tuesday 3 September 3.40pm

Does Job Insecurity Increase the Likelihood of Getting Married? Evidence from 2,123
Chinese Adults

Chaoying Fu, Xiaoli Xing, Mengjie Xu, Liang Guo

Shandong University, Weihai

Prior studies in the US and the UK find that the decline of stable jobs and the rise in casual
employment mean that men and women are now less likely to tie the knot, stay married and
have children within wedlock. Our study investigates the link between job insecurity and
people’s marital status (single, married, and divorced) using a sample of 2,123 adults in
China. We estimated a series of multinomial logistic regression models with marital status as
the dependent variable and job insecurity as the key independent variable. Job insecurity is
measured by a dummy variable (1 for a person has signed a fixed-term employment
contract, and 0 otherwise). We also control for the effects of workfare and job satisfaction as
well as demographical variables. The results show that job insecurity leads to a higher
probability of getting married. That is, marriage seems to be a mean for the Chinese people
to cope with the destabilizing effects of insecure jobs. The reason why our finding is different
from that in the US and the UK may lie in the cultural difference between East and West. In a
collectivist country like China, people tend to trust their potential partners and believe that
the emotional and psychological commitment of marriage will serve as the solid foundation
for the couple to get their share of weal and woe in life. In addition, we estimated the
moderating effect of people’s education level. The effect of job insecurity on getting married
tends to be smaller among university-educated individuals than those who have completed
only secondary studies. Likewise, workfare also tends to reduce the impact of job insecurity
on getting married. These two moderators confirm that when living in an unstable job
situation, marriage can help individuals who have less “resources” to deal with challenges
together.

93
Rapid Fire Talks 6
Tuesday 3 September 3.40pm

The Dark Side of Community-level Social Capital: Does Civic Participation Hurt Mental
Health in China?

Xiaoming Lin, Liang Guo, Chunyi Chen

Shandong University at Weihai

Social capital is a crucial determinant of mental health issues at the individual level, the
aggregate level, and their interactions. This study focused on civic participation, a type of
structural and horizontal social capital, to examine whether it has an impact on self-rated
mental health in rural and urban China. Followed the nested modelling strategy, we
estimated a series of two-level random-coefficient linear regressions to explain the variations
in mental health status, both at the individual level and at the community level by taking the
hierarchical structure into account. Using data from our nationally representative sample
includes 10,968 Chinese respondents from 130 county-level communities, we found that the
civic participation in 13 activities was positively associated with the mental health at the
individual level in urban areas, whereas the impact was not significant in rural areas. That is
probably that higher civic participation in urban areas could be manifested through
progressive coaching mechanism or knowledge dissemination process to fellow individuals,
thereby leading to enhanced self-esteem among individuals. On the contrary, at the
community level, enhanced civic participation was found to give rise to reduced mental
health in urban areas. An explanation for this could be based on the network resources
approach, suggesting that mediated by the subjective social status, a higher level of
community civic participation could create extra responsibility or overwhelming burden on an
individual. Especially in China, fear of losing “face” has been considered as a key deterrent
in an individual’s access to social capital and as a source of psychological stress. The
results provide an extended lens to analyze the predictors of mental health in China and
emphasize the importance to set relevant policy mechanisms to strengthen social networks.

94
Rapid Fire Talks 6
Tuesday 3 September 3.40pm

Rethinking Reliability of Psycho-political Indicators in Women Elites for Multi-country

Surveys

Rachel Gregory
University College Cork

Methodological issues of surveying elites are well documented, including low response rates
and response bias. Compounding these problems is the reliability of responses for women in
measures, such as political knowledge and political interest. Samples from voters suggest
that under certain conditions respondents construct a self-presentation fitting of sex-role
stereotypes. Not only has this not been formally tested in elite populations, but the
implications of sex-specific response bias in elites could alter policy decisions when
attempting to form policy to overcome discrepancies in democratic representation. Using
data from the 2014 PartiRep MP Survey, this research seeks to analyse the reliability of
women’s responses for direct measures of political ambition in comparison to signalling
behaviours of ambition by creating an indexed variable of psycho-political behaviours. A
comparison between men and women both within and across fourteen European countries
finds that while a direct measure of ambition is valid for male respondents, a more reliable
measure for female respondents develops from indexed behaviours, suggesting higher
response bias for women elites in comparison to male elites. At the same time, it raises the
question of possible overestimation of psycho-political indicators from male respondents.
Although ambition can be understood as only one psycho-political indicator impacting
women’s representation in democratic institutions, this finding questions the reliability of
other measures requiring explicit responses from women elites. Variation in this finding
between countries cautions the use of statistical information in making sweeping policy
change on a larger scale.

95
Rapid Fire Talks 7
Tuesday 3 September 3.40pm

Election prediction using financial models

John Fry1, Matt Burke2

1
Manchester Metropolitan University, 2 University of East Anglia, , , , , , , ,

We discuss approaches to use financial techniques to try and predict the results of recent
US and UK elections. A combination of different options pricing and econophysics methods
are used. A simple adjustment to the method of Taleb is introduced. However, this mainly
reproduces results obtained using raw polling numbers. An adjustment based on the JLS
model accounts for the possibility that polls may systematically misprice the probability of
certain outcomes.

96
Rapid Fire Talks 7
Tuesday 3 September 3.40pm

FORECAST ERROR: HOW TO PREDICT AN ELECTION: PART 1: POLLS

Timothy Martyn Hill

Barclays Corporate

The "Forecast Error" series of articles in Significance started examining election predictors in
2015. Each article considered many predictors, but each article covered just one election. In
2018 we started a new chapter in the "Forecast Error" series where we examine an
individual class of predictor more closely across many elections. This presentation will cover
possibly the most prominent: opinion polls.Opinion polls are not intended as predictors but
they are frequently used as such. We note the global accuracy work of others, then we focus
on the UK by detailing the evolution of British polls since 1937. We examine the final poll
accuracy, and the accuracy per month before the election, for British polls since 1945. We
then list some investigative tools, consider future developments, and draw conclusions.

SEE ALSO* https://www.significancemagazine.com/politics/595-forecast-error-how-to-

predict-an-election* https://www.significancemagazine.com/files/Forecast-error-polls.pdf

97
Rapid Fire Talks 7
Tuesday 3 September 3.40pm

Admin data and public perception/views of how we use it.

Neil Henderson
NISRA

This rapid fire talk will briefly explore public perception of Government-held administrative
data and its use. This talk is provided in the context of administrative data that the Census
Office of Northern Ireland holds and explores the Census Office engagement with the wider
public. The talk will discuss key concerns raised by the public as identified through previous
public engagement through the use of studies and focus groups. Issues like consent, data
protection, individual privacy, ethics and using data appropriately were mentioned. Public
support for the use of admin data is intrinsically linked to trust; trust in organisations, trust in
data protection systems and the perceived benefit to society. Clearly our role must be to
satisfy public demands, to improve levels of trust and to clearly and simply explain the
benefits to society of Government use of administrative data. Minimising communication
about the use of admin data would be counter-productive and would lead people to suspect
the government of trying to hide something. This needs to be balanced with the concern that
the subject matter is so complicated that it would be difficult to achieve a level of
understanding amongst the general public that would prevent people from worrying unduly
about their privacy and the security of their personal information.

98
Rapid Fire Talks 7
Tuesday 3 September 3.40pm

Connected Open Government Statistics

Bill Roberts
Swirrl IT Limited

The strategy of the ONS is "Better statistics, better decisions": that objective of using
statistical data more effectively is common across a range of public sector organisations.
However, applying statistical data to important questions of public policy or allocation of
resources usually requires data from a range of sources. Users report that it is often hard to
find the right selection of available data to feed into analysis. Different data providers are
inconsistent in their approach, making combination or comparison of data difficult.Swirrl is
working with the ONS and the Government Statistical Service in the 'COGS' project
('Connected Open Government Statistics') to develop practical solutions to these
problems.At the heart of our approach is delivering data so that it works well in software,
combined with the use of standards to enable interoperability. The challenge is choosing or
designing standards and then getting everyone to follow them. The project has reviewed and
prototyped standard approaches for structuring and exchanging data, building on the
standards of the web.We now have a set of processes and software tools to support them,
that government data publishers can follow to make their statistical data accessible in a
richer way. Interchange processes are based around CSV files following the 'Tidy Data'
structure. Re-usable codelists and variable definitions are prepared and made easily
findable. Behind the scenes, the standards of Linked Data are used to store and organise
the data in an interoperable way. The project is liaising with Reproducible Analytical
Pipelines teams at various departments, to integrate this new dissemination approach with
modern techniques for preparation of statistical releases.This talk will explain the details of
the approach and report on early experiences of implementation.

99
Rapid Fire Talks 7
Tuesday 3 September 3.40pm

Bad evidence encourages bad policy

Paul Marchant
Leeds Beckett University

Policy that is based on evidence is surely the rational choice. However if the evidence put
forward is flawed, more harm than good may result from a policy. It may be hard to get an
unjustified policy changed, because of the illusion of an evidential basis. In addition,
personal investment by those having encouraged the policy may hinder its change. High
standards of research are essential to ensure only trustworthy evidence is used when
deciding policy. Policy makers must be attuned to the need for separating wheat from
chaff.In this presentation, the flaws in some studies used to encourage expenditure will be
identified.Some particular errors include:Definitively ascribing cause from correlation in an
observational study with a weak design. Performing cosmetic analyses on grouped and
averaged data, rather than individual cases, achieves an implausibly high coefficient of
determination (R2). The study is cited elsewhere in a bid to acquire huge government
funding.In another report, that of an RCT, the control arm is not used, because it is ‘not
statistically significant’, when generating the size of the intervention effect. However, clues in
the report enable an estimate the effect of the omission. Other statistical misdemeanours are
also committed. The work again encourages public money to be spent on a false
promise.Existing best practice would go some way to reduce risks of believing bad evidence
through having: 1) a transparent public protocol written before embarking on the research
and 2) open research data to check work and to enable better analyses.A way of protecting
against continuing with implementing flawed policies would be to have, a pre-defined,
scientifically rigorous plan to check the effects of the roll out of any policy.

100
Rapid Fire Talks 7
Tuesday 3 September 3.40pm

Measuring the Economy - An ONS online book

Georgia Tasker-Davies, Jonathan Athow, Ed Palmer

ONS

Measuring the Economy The Office for National Statistics is creating an online book
‘Measuring the Economy’ for use by universities to support undergraduate and postgraduate
economic statistics teaching. The book aims to convey the importance of best-practice, real-
world economic measurement, and as such to support academics and students in
understanding the issues surrounding the measurement of the modern economy. Chapters
are being written by leading experts in their subject area, with the aim being to bring their
knowledge and expertise together in a single output. Our aim is for the chapters to be
standalone, so that they can be used a la carte while, also maintaining a common and
coherent approach. We are looking to develop teaching resources for each chapter- case
studies, exercises and test questions for example, once the chapter has been published in
Beta. As soon as these additional resources have been produced we will make them
available, alongside the chapter, for use and comment. Our aim is to publish the chapters, in
Beta, between May and December 2019. During this period, and up to March 2020, we are
inviting feedback and review from Stakeholders. We wish to collaborate with these
stakeholders to shape the project and ensure the final product is beneficial to the University
community. We will engage with this community regularly throughout the project, gaining
continual feedback and review of each chapter as it is published. Publication of the online
book, in alpha, is anticipated to be September 2020. The ‘Measuring the Economy’ intention
is also to publish a regularly update hard copy thereafter. Chapters:IntroductionMeasuring
inflationNational Income and ExpenditureSectoral AccountsSupply-use FrameworkThe
Labour MarketProductivity and GrowthInequalitySub-National StatisticsTrade and Balance of
PaymentsThe Financial SectorThe Environment – including Natural CapitalHard to Measure
SectorInnovation StatisticsPublic Finance

101
Rapid Fire Talks 7
Tuesday 3 September 3.40pm

Faster Indicators of UK Economic Activity by using over a hundred million VAT

returns

Luke Shaw, Louisa Nolan, Jeremy Rowe, Andrew Sutton, Daniel Ollerenshaw, Stephen
Campbell, Ioannis Kaloskampis
Office for National Statistics (ONS)

There is a current great appetite for faster information on UK economic activity. Indeed, the
Independent Review of Economic Statistics (Bean, 2016) stated that “the longer a decision-
maker has to wait for the statistics, the less useful are they likely to be”. The faster indicators
of UK economic activity project is one of the Office for National Statistics’ innovative
responses to this demand. The project identifies close-to-real-time large administrative and
alternative data-sets that are related to important economic concepts. From these data-
sources, we develop a set of timely indicators that allow early identification of potential large
economic changes. Data-sets included thus far are: HM Revenue and Customs Value
Added Tax (VAT) returns, ship tracking data from automated identification systems for UK
waters, and road traffic sensor data for England. Here we focus on indicators built from VAT
returns data. We have constructed monthly and quarterly diffusion indices built using
turnover and expenditure data from VAT returns, and several indicators based on VAT
reporting behaviour. We discuss the methodology behind the indices and present results-to-
date. We caution that care should be used in interpreting these indicators, and they are
supplementary to, not a proxy for gross domestic product. However, the suite of indicators
shows promise in identifying large changes to economic activity. We find that, at the time,
the VAT quarter-on-quarter turnover diffusion index would have identified the first quarter of
the recession which begun in 2008 five months before it was seen in official estimates. Since
April 2019 we have been publishing these indicators monthly, within a month of the end of
the period of interest. This is one month in advance of official GDP estimates.

102
4.1 Contributed - Medical: Survival Analysis
Wednesday 4 September 9am

Landmark analyses of survival benefit associated with statin prescription

Ilyas Bakbergenuly, Elena Kulinskaya, Lisanne Gitsels

University of East Anglia

Objective: Statins are prescribed for primary and secondary prevention of cardiovascular
disease, however the threshold of cardiac risk at which to prescribe statins is still
controversial, especially at older ages where everyone would be eligible. The research
objective was to dynamically predict the survival benefits associated with statin therapy over
the course of 25 years in patients residential in England or Wales.

Methods: Primary care records from The Health Improvement Network (THIN) were used.
The cohort included 110,243 patients who turned 60 between 1990 and 2000, were neither
diagnosed with cardiovascular disease nor prescribed statins. The cohort was followed up
until January 2017, where the medical history was updated every half a year. Landmark
analyses were carried out by fitting adjusted Cox’s proportional hazards regressions of the
hazard of all-cause mortality associated with current statin prescription at each landmark
from age 60 to 85 (51 time points).

Results: The initial results show that statin therapy is associated with increasing survival
benefits by older ages. By age 65, statin therapy became significantly associated with
survival (age 60 HR=0.88 (0.67-1.18) and age 65 HR=0.86 (0.78-0.94)). This survival
benefit remained approximately the same for the next ten years (age 70 HR=0.85 (0.79-
0.91) and age 75 HR=0.81 (0.74-0.89)), after which greater benefit was gained (age 80
HR=0.72 (0.62-0.84) and age 85 HR=0.57 (0.35-0.91)). The survival benefit of statin therapy
significantly differed by the birth cohort but not by sex or cardiac risk. The 1936-40 cohort
showed larger survival benefits of statin, probably due to development of more efficient
drugs. Furthermore, current prescription was a stronger predictor than cumulative proportion
of statins over time and age at first prescription.

Conclusions: After adjustment for cardiac risk and related medical history, it appears that
statin therapy is especially beneficial at older ages. Clinicians may want to use this new
information when managing cardiac risk of older patients.

103
4.1 Contributed - Medical: Survival Analysis
Wednesday 4 September 9am

Identifying biomarkers to predict pancreatic cancer

Christiana Kartsonaki
University of Oxford

Pancreatic cancer has the worst overall prognosis of all cancers, with a 5-year survival less
than 5%. Several metabolic and lifestyle factors are associated with pancreatic cancer risk,
but there is need to identify biomarkers that may help with risk prediction and early diagnosis
of pancreatic cancer. We aimed to identify protein biomarkers in blood which are associated
with being diagnosed with pancreatic cancer up to several years later.We designed a case-
subcohort study within the China Kadoorie Biobank, a prospective cohort study of over
500,000 Chinese adults with blood samples collected at baseline, to examine the
associations between circulating proteins and the risk of developing pancreatic cancer. We
used the OLINK immuno-oncology proteomics assay to quantify 92 biomarkers in 700
pancreatic cancer cases that accumulated over about 8 years of follow-up and a randomly
sampled subcohort of 700 individuals. We used Cox proportional hazards models with the
Prentice pseudo-partial likelihood to assess the associations between proteins and
pancreatic cancer risk. Time in study was used as the timescale. We examined the shape of
the associations using splines and by splitting protein values into groups. We used forward
selection and the method by Cox and Battey (2017) to identify sets of proteins that predict
pancreatic cancer risk.We identified several proteins that are associated with pancreatic
cancer risk. Sets of proteins identified from different variable selection approaches largely
overlapped. Sensitivity analysis with different weighting methods, using age as the timescale
or adjusting for other risk factors did not substantially change the conclusions. Including the
identified markers into models with other risk factors yielded a modest improvement in the
discriminatory ability of the model.We identified biomarkers which may help predict risk of
pancreatic cancer up to several years after measurement and which may contribute to
understanding its aetiology.

104
4.1 Contributed - Medical: Survival Analysis
Wednesday 4 September 9am

Is the Restricted Mean Survival Time Approach an Alternative to the Time-Dependent

Cox Model When Hazards are Non-Proportional?

Bee-Choo Tai1, Zhaojin Chen1, Joseph Wee2

1
National University of Singapore, 2 National Cancer Centre

Objective: We evaluate the performance of the restricted mean survival time (RMST)
approach as an alternative to the Cox time-varying covariate (TVC) model for quantifying the
treatment effect when there is non-PH.

Methods: To simulate non-PH, survival times were generated using Gompertz distribution
with shape parameter α = 0.1, scale parameter γ = 0.25, for n = 200 and 500 with 1000
replications. The censoring distribution was generated from U(0, 6), and treatment B(1, 0.5).
The life-expectancy ratio (LER) was used to quantify the effect of treatment. Assuming LER
from Gompertz distribution as the gold standard, the RMST estimates from the flexible
parametric survival model with 3 knots and 1 df were compared to those of Cox TVC in
terms of bias and mean squared error (MSE) for t = 1, 3 and 5 years. The two methods were
also compared using data from a randomised clinical trial of patients with nasopharyngeal
cancer (SQNP01).

Results: The RMST and Cox TVC estimates were close to the true LER for all t, although the
former had slightly larger bias and MSE. When applied to the SQNP01 data, LER of both
models increased with time. At 1-year, the estimates based on RMST and Cox TVC were
0.99 (95% CI 0.95 to 1.02) and 1.00 (95% CI 0.96 to 1.04) respectively. Their 3- and 5-year
estimates of 1.04 (95% CI 0.96 to 1.13) and 1.04 (95% CI 0.97 to 1.14), and 1.16 (95% CI
1.04 to 1.30) and 1.18 (95% CI 1.05 to 1.32) respectively were also progressively larger.

Conclusion: The RMST approach yielded slightly larger bias and MSE as compared to Cox
TVC when the hazards are non-proportional. However, estimate based on RMST may be
more appealing to clinicians because of its ease in interpreting the magnitude of survival
benefit.

105
4.2 Contributed - Official and Public Policy: Alternative data sources
Wednesday 4 September 9am

Estimating the impact of automation

Andrea Lacey, Anna Ardanaz-Badia

Office for National Statistics

Automation has become more relevant in recent years due to enhanced technology,
including data science and AI, which allows various processes to be automated. This is a
matter of both public and policy interest, as the introduction of new technologies will have
implications for the UK labour market. We have utilised an OECD study that assesses the
probability of automation of the tasks that people undertake within certain occupations, and
applied the UK outputs to the Annual Population Survey. Using this, we’ve produced
estimates of the risk of automation for occupations in 2011 and 2017 by various
demographics. In addition, we’ve looked at the skills within an occupation that are at high
and low risk of automation, and what skills may be ‘transferrable’. This presentation will
discuss the methodology used to produce estimates of automation in the UK, and the results
of our analysis, including what jobs are at risk of automation, and who is most likely to be
employed in these roles. We will also explore regionally where automation may affect.

106
4.2 Contributed - Official and Public Policy: Alternative data sources
Wednesday 4 September 9am

Young People’s Earnings Progression and Geographic Mobility

Bonang Lewis, Thomas Odell

Office for National Statistics

This presentation will be based on the published article ‘Young people’s earnings
progression and geographic mobility’. Innovative new Census linked to administrative data
was used to connect longitudinal income data to personal characteristics, allowing more
granular and intersectional analysis on how multiple forms of earnings progression differed
for young people between 2011/12 and 2015/16. Regression analysis identified
characteristics of people more likely to escape low pay. We also investigated patterns of
movement between local authorities for young people, and how earnings growth varied
depending on starting area and different city-region destinations. This work informed policy
thinking around the Department for Work & Pensions Labour Market strategy and provided
evidence for the Race Disparity Unit, to target funding based on ethnic disparities affecting
education to labour market transitions. The work was a successful feasibility test for this
unique admin linked dataset as it provided helpful new insights.

107
4.2 Contributed - Official and Public Policy: Alternative data sources
Wednesday 4 September 9am

Advancing the methods for administrative and transactional data in official statistics

Hannah Finselbach, Lucy Tinkler

Office for National Statistics

The Office for National Statistics (ONS) is transforming to put administrative and alternative
data sources at the core of our statistics. Combining new sources with surveys will allow us
to meet the ever-increasing user demand for improved and more detailed statistics. Our
history is mainly of indirect usage - in small area estimation, calibration weighting, census
quality assurance, and multi-source migration patterns. More recently we've become more
ambitious - and successful - in directly replacing (and augmenting) our high-quality survey
statistics with admin data. Like many national statistics organisations, we want to exploit the
rich data sources available from transactions by businesses and households with
government agencies and digital platforms. However, using this data involves addressing a
range of statistical challenges, including those described by Hand (2018)*. ONS has
established a methodological research programme, to develop a theoretical framework to
effectively use and integrate new data sources. We aim to work collaboratively across
academia, government and private sector on a variety of projects, to address some of these
statistical challenges, including:

• Measuring and communicating uncertainty of integrated or administrative data

• Linking multiple sources for multiple uses and analysis
• Measures the quality of data linkages
• Managing discontinuities within the data
• Creating synthetic data to enable researcher access to confidential data
• Dealing with definitional differences between admin sources and survey data
• Identifying and addressing under or over coverage
• Ensuring confidentiality and privacy of linked dataIdentification and correction of
errors in big/streamed data

This talk will outline our achievements to date, our plans, and the challenges that remain to
be solved.

*Hand, D., “Statistical challenges of administrative and transaction data.” Journal of the
Royal Statistical Society A, 2018

108
4.3 Contributed - Applications of Statistics: Applications 1
Wednesday 4 September 9am

Supervised Classification of Linear Synchronous Motor Vehicle State in a Smart

Factory.

Jill Daly
CIT

The research presented in this thesis examines how to use Vibration and Magnetometer
sensor data to identify where on a manufacturing production line a Linear Synchronous
Motor vehicle is situated and the type of movement of the vehicle at that position. This paper
seeks a solution to this problem by identifying the state of the vehicle on the track from the
data observations generated by a data capture experiment. The results from this research
will be part of a predictive maintenance pipeline in a smart industry environment.In the
absence of a labelled dataset, work was undertaken to label approximately 1.5 million
sensor data observations. This labelled dataset was used for iterative model training,
validation and feature engineering. Due to the large amount of data available, it was possible
to keep a section of data separate, so that it could be used as unseen validation data. This
validation data was additional to the 10 fold cross validation used. Model performance was
measured using accuracy and specificity because type 2 (false positive) errors were
identified as the most costly errors. Multiclass One-v-Other ROC was used as a further
measure of model performance.The algorithm adopted was the Supervised Learning
Random Forest classifier. The final Random Forest model was trained using data for
Vibration and Magnetometer combined, with all speed settings. The features used were
rolling aggregation mean and standard deviation, derived from the Z_AXIS, and with FFT
applied to those aggregate values. This model provided a training accuracy of 85.6% and a
validation accuracy (i.e.: generalisation) of 95.7%.

109
4.3 Contributed - Applications of Statistics: Applications 1
Wednesday 4 September 9am

Modelling Road Accidents in Edinburgh Using Hidden Markov Models

Valentin Popov1, Glenna Nightingale2, Andrew Williams3, Paul Kelly4, Ruth Jepson2
1
University of St Andrews, 2 School of Health in Social Sciences, Medical School,
Edinburgh, UK, 3 European Centre for Environment and Human Health, University of Exeter
Medical School, 4 Physical Activity for Health Research Centre, University of Edinburgh

Empirical study of road traffic collision rates is challenging at small geographies due to the
relative rarity of collisions and the need to account for secular and seasonal trends. In this
paper we demonstrate the application of Hidden Markov Models (HMMs) to successfully
describe road traffic collision time series using data from the city of Edinburgh (STATS19) as
a case study. We apply various other time series models to these data and demonstrate the
superiority of the HMMs for this particular case. Our model discrimination approach, based
on the Akaike Information Criterion indicates that the best performing model is a 2-state
HMM with Negative Binomial state-dependent distributions. The model accounts for the
diminishing trend observed in the data and, contrary to other competing models, indicates
teh presence of seasonality effects. In addition it shows good forecast ability. To date,
Hidden Markov models have not been used to model road traffic data from the UK. The
application of HMMs to such routinely collected data may be beneficial to natural
experiments or evaluations of interventions and policies that seek to impact traffic collision
rates.

110
4.3 Contributed - Applications of Statistics: Applications 1
Wednesday 4 September 9am

Dynamic Spatial Sampling in Semiconductor Manufacturing

Seán McLoone1, Gian Antonio Susto2

1
Queen's University Belfast, 2 University of Padova

Semiconductor manufacturing typically involves silicon wafers undergoing hundreds of

different steps over several weeks to build up the desired complex nanoscale structures.
Metrology is essential for monitoring the performance of these processes to ensure that both
spatial variation (intra-wafer) and temporal variations (inter-wafer) do not exceed the
demanding tolerances required for modern semiconductor devices. However, metrology is a
high cost, time consuming, and non-value added operation, hence, due to commercial
pressures, standard practice is to keep metrology to a minimum by adopting both spatial and
temporal sampling protocols. Here we consider the problem of optimising spatial dynamic
sampling plans such that the number of locations measured on each wafer is minimised
while retaining the capacity to: (1) generate an accurate reconstruction of the wafer profile,
and; (2) to detect previously unseen process behaviour in a finite time horizon.The solutions
we have developed are data driven, and involve the analysis of historical metrology data
using Forward Selection Component Analysis (FSCA), a greedy search based unsupervised
variable selection technique. Using this algorithm, a number of strategies for generating
dynamic sampling strategies are explored, and the trade-off that exists between wafer profile
reconstruction accuracy and the ability to detect previously unseen anomalies
assessed.Results for both simulated and industrial metrology case studies confirm the
efficacy of the proposed FSCA based dynamic sampling methodology, with substantial
reductions in the number of measurement sites that need to be measured possible for
processes that exhibit significant spatial correlation in inter-wafer variability. In particular,
analysis of a metrology dataset from a chemical vapour deposition process used in the
manufacture of hard disk drives shows that measurements per wafer could be reduced from
50 to 7 while maintaining wafer profile reconstruction accuracy levels of greater than 99%.

111
4.4 Contributed - Social Statistics: Populations
Wednesday 4 September 9am

Measurement Error Model to Correct the Inconsistencies in Migration Flow Data for
South America

Andrea Aparicio-Castro1, Arkadiusz Wiśniowski1, Francisco Rowe2, Mark Brown1

1
University of Manchester, 2 University of Liverpool

The intra-regional migrant stocks in South America doubled from 2 in 1990 to 4 million
people in 2015. This evidences that a great number of migrants of the region changed their
migration strategy towards South America. To gain an understanding of the spatial patterns
and trends underpinning the increasing appealing of immigrants to South America, reliable
data on migration flows are crucial to measure the evolving migration links between
countries of origin and destination.While migration flow data are available from individual
countries, they often incomplete and/or incomparable between countries and over time. Data
sources differ not only in how they define migrants and their population coverage, but also in
their systematic bias, accuracy and measurement methods, such as censuses, population
registers and surveys.To address these issues, this paper aims to develop a measurement
error model to correct inconsistencies in migration flow data for South America. Multiple
migration data tables from a variety of sources are combined in a measurement model to
correct and measure true (unobserved) migration flows. The model captures inconsistencies
in duration of stay, systematic bias in measurement, population coverage, accuracy and
variability of reported flow data from each data source. The resulting outcome is a set of
synthetic estimates of migration flows with measures of uncertainty.

112
4.4 Contributed - Social Statistics: Populations
Wednesday 4 September 9am

Combining health information systems data and probability survey data to monitor
health coverage indicators in low-resource settings

Caroline Jeffery1, Marcello Pagano2, Baburam Devkota1, Joseph J Valadez1

1
Liverpool School of Tropical Medicine, 2 Harvard T.H. Chan School of Public Health

Objectives: Delivering health services in low-resource settings partly depends on accurate

Health Information Systems(HIS) data. The quality of these records matters because poor
quality leads to poor judgements and outcomes. Unlike probability surveys, which are
representative of the population and carry accuracy estimates, HIS do not. They may not
even detect gaps in service coverage and leave communities exposed to unnecessary
health risks. Using data from Benin, Madagascar and India, we improve informatics by
introducing a novel method, the Annealing Technique (AT), for combining HIS and survey
data that improves the accuracy of health coverage indicators at the district and sub-district
levels, where management improvements are made, and provides an all-important measure
of the accuracy of the indicators.

Methods: The AT estimator is a weighted average between the administrative and the
survey estimators. The two weights are inversely proportional to their respective variances.
We apply AT to 3 datasets, where the survey data were collected using Lot Quality
Assurance Sampling: polio vaccination and vitamin A distribution in 19 communes of Benin
and 3 districts of Madagascar (2015); and 10 indicators measuring pre-natal, delivery and
post-natal care in 25 blocks in Bihar, India (2016).

Results and Conclusion: For Benin and Madagascar, the annealed and survey estimates of
the communes or districts differ by no more than 3%, while decreasing the standard error
(SE) by 1% to 6%; the HIS and annealed estimates are quite different, with 3 to 29 times
larger SE. For Bihar, the annealed and survey estimates of the blocks differ by no more than
10%, while decreasing the SE by 0.05% to 32.4%. The HIS and annealed estimates can
differ by up to 84.2%, with 1.4 to 32.3 times larger HIS SE. The AT methodology creates
more accurate estimators, together with measured HIS errors, marking a new role for
probability sampling and HIS for tracking the public’s health.

113
4.4 Contributed - Social Statistics: Populations
Wednesday 4 September 9am

Modelling and forecasting UK fertility using Bayesian Generalised Additive Models

Joanne Ellison, Ann Berrington, Erengul Dodd, Jonathan J Forster

University of Southampton

Aggregate UK fertility data available at the population level can only give limited information
about the patterns of variability of age-specific fertility rates. Survey data provides a rich
source of information through fertility histories of individuals and their corresponding
characteristics, which can help to gain a better understanding of the underlying variability of
fertility rates. By modelling the fertility histories of women surveyed in Wave 1 of the UK
Household Longitudinal Study (Understanding Society), we investigate the dependence of
birth events on selected variables as well as information derived from the fertility histories
themselves. Generalised Additive Models (GAMs) allow the incorporation of covariates and
interactions between them as smooth terms, where the precise form of the smoothness is
purely data-driven. Fitting Bayesian parity-specific logistic GAMs to the survey data, we learn
about the variability of fertility as a function of age, cohort, time since last birth, highest
educational qualification and country of birth. Following model selection, we find that
educational attainment is important for determining the likelihood of transitions to lower order
births, whereas country of birth significantly influences the odds of transitioning to higher
order births. Also, the chosen models decrease in complexity as parity increases. There is
the potential to combine inferences from this detailed individual-level data with the coarser
population-level data for the purposes of forecasting.

114
4.5 Contributed - Methods & Theory: Methods Showcase
Wednesday 4 September 9am

Large numbers of explanatory variables

Heather Battey1, David Cox2

1
Imperial College London, 2 Nuffield College, University of Oxford, , , , , , , ,

In the context of regression with a large number of potential explanatory variables and
relatively few study individuals, a key point is emphasised: if there are several models that fit
the data essentially equally well, one should aim to specify as many as is feasible. This view
is in contraposition to that implicit in the use of variable selection methods, which produce a
single model effective for prediction but potentially misleading for scientific understanding. I
will discuss a different approach whose aim is essentially a confidence set of models. The
talk is based on joint work with David R Cox:Cox, D.R. and Battey, H.S. (2017). Large
numbers of explanatory variables, a semi-descriptive analysis. Proc. Nat. Acad. Sci., 114
(32), 8592-8595.Battey, H.S. and Cox, D.R. (2018). Large numbers of explanatory variables:
a probabilistic assessment. Proc. R. Soc. Lond. A., 474The work is motivated by genomics.
However, the ideas are mathematical in basis and apply more widely.

115
4.5 Contributed - Methods & Theory: Methods Showcase
Wednesday 4 September 9am

Spaed rankings, and how to assess them --- with application to predicting the Premier
League

David Firth
University of Warwick

This is methodology inspired by an attractive --- and now quite commonly seen --- method
for summarising a predictive distribution on the possible rankings of N items. An example,
predicting end-of-season standings in Premier League football, can be seen at
https://twitter.com/DectechSports/status/1099241216221286401. The prediction is
presented as a nonnegative NxN matrix with row and column sums all equal to 1 (or 100%) -
-- a 'doubly stochastic' matrix. This provides a readily understood summary of the predictive
distribution. In the Premier League example, each row is a probabilistic forecast of the
finishing position for a given team, and each column predicts which teams could ultimately
occupy a specific league position. When a doubly stochastic matrix is used in this way we
will call it a 'spaed ranking'.(Spaed is the past tense of an old Scottish verb 'spae', meaning
predict or foretell. Its use here is motivated by the need to distinguish a spaed ranking from a
full predictive distribution over the set of all possible rankings. A spaed ranking can be
viewed as the expectation of such a predictive distribution.)We develop the use of a proper
scoring rule (such as the familiar Brier score), to compare spaed rankings and also to assess
the performance of a spaed ranking in and of itself. The latter assessment is possible
because a spaed ranking gives probabilistic forecasts for several outcomes simultaneously.
We show how to calibrate any chosen scoring rule through its sampling distribution under
the 'canonical' predictive distribution consistent with a spaed ranking. The sampling
distribution is analytically intractable, but is amenable to accurate approximation through an
intuitive MCMC scheme.The methods are general --- not limited to sports applications! --- but
will be illustrated here through application to published spaed rankings in Premier League
football.

116
4.5 Contributed - Methods & Theory: Methods Showcase
Wednesday 4 September 9am

Inference under unequal probability sampling with the Bayesian exponentially tilted
empirical likelihood

Andrew Yiu, Robert Goudie, Brian Tom

MRC Biostatistics Unit

In this presentation, we consider estimating a population outcome mean in the presence of

unequal probability sampling. We work under two settings. The first assumes a selection
mechanism determined by the data collector, but the analyst is provided only with partial
design information in the form of sampling probabilities for the selected individuals. The
second is an observational setting where the selection mechanism is unknown but is
assumed to be independent of the outcome conditional on a set of fully observed covariates.
Certain problems in missing data and causal inference, such as estimating the average
causal effect of a binary treatment with no unmeasured confounders, can be framed in this
perspective. In both cases, semiparametric estimators incorporating inverse probability
weighting have attractive large sample properties such as local efficiency and double
robustness. However, the reliability of these estimators is less justified theoretically in small
sample inference, where the use of an informative prior in a Bayesian approach may be
advantageous. We provide an inferential framework that offers the practical benefits of
Bayesian statistics, along with the desirable asymptotic guarantees of semiparametric
estimators. Central to this method is the exponentially tilted empirical likelihood, which
enables Bayesian analysis of moment condition models. We use the fact that many
semiparametric estimators can be defined as the solutions to unbiased estimating equations,
and formulate a set of moment constraints to calibrate posterior behaviour to the estimator of
choice. We prove Bernstein-von Mises theorems for a broad class of moment constraints,
establishing large sample equivalence with the frequentist approach. In particular, the
frequentist coverage of posterior credible intervals converges to the nominal level, and the
posterior concentrates around the limiting value of the estimator. Simulations verify these
theoretical results and suggest that our approach outperforms existing Bayesian proposals.

117
4.6 Contributed - Medical: Identification and Prediction
Wednesday 4 September 9am

Sample size requirements for validating risk prediction models for binary outcomes

Chen Qu, Gareth Ambler, Rumana Omar, Menelaos Pavlou

UCL

Objective: Risk prediction models for health outcomes are used in the clinical management
of patients and assessment of institutional performance. Given their importance in health
care, it is essential that the predictive performance of a risk model is assessed adequately in
an external validation study. However, there has been limited research into the sample size
requirements for validation studies and hence we aim to provide guidance to researchers
through this work. We propose precision-based methods to calculate the minimum sample
size required for an external validation study for binary outcomes.

Methods: The performance of a risk model is often assessed using the calibration slope and
c-statistic, hence it is important to estimate these measures with acceptable precision. We
derived approximate analytical expressions for the standard error of both estimates in terms
of the sample size (N), c-statistic (C) and outcome prevalence (P). The accuracy of the
analytical expressions was evaluated in a range of simulation studies varying the values of
C, P and N.

Results: The simulation results show that the proposed expressions provide good
approximations to the standard errors. That is, the expressions may be used to estimate the
precision of the c-statistic and calibration slope for a given sample size, and hence can be
used to determine a suitable sample size for a prospective external validation study. For
example, our analytical expressions suggest that a sample size of 1800 is required to
estimate the c-statistic with a standard error of 0.02, when the outcome prevalence is 10%
and the c-statistic is 0.72. With that sample size it is possible to estimate a calibration slope
of 1 with a standard error 0.11.

118
4.6 Contributed - Medical: Identification and Prediction
Wednesday 4 September 9am

The use of period analysis techniques as an alternative approach to estimate post-

transplant kidney survival outcomes

Chloe Brown
University of the West of England / NHS Blood and Transplant

Objectives: Organ transplantation is an area that is constantly evolving, with improved

survival observed over the last ten years. Nationally, kidney transplant outcomes are
monitored and survival estimates, obtained from traditional survival analysis methods, are
informative for patients and clinicians. It is therefore important that survival estimates are up-
to-date, reflective of current practice and applicable for a recently transplanted cohort. Period
analysis is a potential alternative approach which is thought to obtain more up-to-date
survival estimates. Here, different approaches for obtaining unadjusted and risk-adjusted
five-year kidney graft survival estimates are compared to determine which approach is the
most accurate and appropriate.

Methods: Data on adult first deceased donor kidney transplants performed between January
2006 and December 2018 were obtained from the UK Transplant Registry. Data were
analysed as shown in Table 1. Five-year graft survival estimates from each approach were
compared to the actual observed five-year survival. Cox proportional hazards modelling and
a period approach to Cox modelling will be compared in obtaining risk-adjusted survival
estimates.

Table 1:

Time period Method Analysed as at Name

2006-2008 Kaplan-Meier December 2013 Approach 1
2006-2013 Kaplan-Meier December 2013 Approach 2
2006-2013,
window period Period December 2013 Approach 3
2011-2013
2011-2013 Kaplan-Meier December 2018 Actual observed survival

Results:The different approaches produced survival estimates within 1% of the actual

observed survival. The unadjusted five-year survival estimates for the three approaches
were 85.7% (1), 86.4% (2) and 86.8% (3) respectively, which compared to an actual
observed survival of 86.3%.

Conclusion: Although there appears little difference between survival estimates obtained by
each approach, period analysis does possess other advantages. By only considering the
survival experience in a recent window period, this enables survival estimates of recently
transplanted patients to be obtained and earlier detection of changes in survival following
changes in practice. Comparison of a standard and period approach to Cox modelling is
currently in progress.

119
4.6 Contributed - Medical: Identification and Prediction
Wednesday 4 September 9am

Analysis of paediatric visual acuity using Bayesian copula models with sinh-arcsinh
marginal densities

Mario Cortina Borja1, Julian Stander2, Luciana Dalla Valle2, Charlotte Taglioni3, Brunero
Liseo4, Angie Wade1
1
Great Ormond Street Institute of Child Health; University College London, 2 School of
Computing, Electronics and Mathematics; University of Plymouth, 3 Dipartimento di Science
Statistiche; Università degli Studi di Padova, 4 Dipartimento Metodi e Modelli per l'Economia,
il Territorio e la Finanza; Sapienza Università di Roma

It is important that children who are visually impaired are identified early enough to start
corrective treatment. Visual acuity, which quantifies each eye’s spatial resolution capacity,
was measured in the left and right eyes of over 2700 children, together with each child’s age
(ranging from 3 to 8 years). The distribution of visual acuity is asymmetric and kurtotic thus
needing four parameters to be adequately modelled so the sinh-arcsinh distribution specified
the marginals of the bivariate model for measurements from both eyes. A flexible bivariate
statistical model was developed using copulas to understand how visual acuity changes with
age in each eye separately, and how the dependence between left and right eye visual
acuity is related to age. It was found that as age increases, visual acuity improves and the
dependence between visual acuity in the left and right eyes becomes stronger, meaning that
children’s eyes become better and more similar with age. The bivariate statistical model
also allowed the identification of children with unusual sight, distinguishing those who are
atypical in both eyes when they are considered together from those who are outliers in one
or both eyes when they are considered separately, as in current practice. This yields an
innovative tool that enables clinicians to recognise children with unusual sight who may
otherwise be missed.

120
4.7 Contributed - Data Science: Misc
Wednesday 4 September 9am

Big data and machine learning at the ONS: developing innovative temporal
disaggregation techniques

Geoffrey Megardon, Andrew Sutton, Ciaren Taylor, Daniel Ollerenshaw, Simona

Astadurova, Tingting Peng, Philip Wales
Office for National Statistics

The ONS is undergoing an important transformation driven by the leitmotif: better data for
better decision. An important milestone of the transformation is to understand how combined
administrative data sources (as opposed to our survey data alone) and data science
methods can improve our output statistics. With this goal in mind, the Economic Microdata
Research team is creating a longitudinal database gathering different administrative data
sources. One recurring problem that we face is producing monthly outputs from mixed, often
lower frequency data. We are working on VAT returns data from HMRC, which consist of the
turnover and expenditure of all the VAT registered businesses in the UK. We use a temporal
disaggregation method on quarterly returners that is constrained by using a relevant group of
monthly returners as indicators.Currently, the Standard Industry Classification (SIC) is
assumed to be a good grouping to use when constructing indicators. Our work explores two
ways to improve upon using SIC: 1) using hierarchical clustering so that the structure of the
data itself is used to infer the best indicators, 2) using the high-dimensional Trade data from
HMRC in combination to the VAT data to get a better measure of the similarity between two
businesses. We will be presenting our methods, and the different challenges we faced while
working with a high-dimensional, mixed and large dataset. We will show comparisons of our
results against those obtained using SIC.

121
4.7 Contributed - Data Science: Misc
Wednesday 4 September 9am

On Simulating Ultra High-Dimensional Multivariate Discrete Data

Alfred Schissler
University of Nevada, Reno

In this era of Big Data, it is critical to realistically simulate data to conduct informative Monte
Carlo studies. This is often problematic when data are inherently multivariate while at the
same time are (ultra-) high dimensional. This situation appears frequently in observational
data found on online and in high-throughput biomedical experiments (e.g., RNA-
sequencing). Due to the difficulty in simulating realistic correlated data points, researchers
often resort to simulation designs that posit independence --- greatly diminishing the insight
into the empirical operating characteristics of any proposed methodology. A major challenge
lies in the computational complexity involved in simulating these massive multivariate
constructions. In this paper, we first review high-dimensional multivariate approaches and
discuss relative merits of the approaches. Then we propose a fairly general procedure to
simulate high-dimensional multivariate distributions with pre-specified marginal
characteristics and a covariance matrix. Finally, we apply our method to simulate RNA-
sequencing data sets (dimension > 20,000) with heterogeneous negative binomial
marginals.

122
4.7 Contributed - Data Science: Misc
Wednesday 4 September 9am

Performance of calibrated non-positive definite covariance matrix under a

classification problem

Ronald Wesonga
Department of Statistics, College of Science, Sultan Qaboos University

Calibration methods to deal with covariance matrix via matrix nearness problem are
explored. The classification problems is studied to assess correct classification rates (CCR)
in the presence of the data missing at random (MAR) mechanism. We explore some
calibration methods. Given an arbitrary square matrix X of order n, the Frobenius norm is
used. The nearest symmetric semi-definite matrix P0 (X) = Qdiag{max(λ1, 0), ..., max(λn,
0)}QT is assessed against another calibrated form Pc(X) = Qdiag{max(λ1,c),...,max(λn,c)}QT
to evaluate information recovery from the former. A novel algorithm will be presented to deal
with the problem of CCR in the presence of MAR problem using the calibrated covariance
matrix. Results based on simulations and actual data will be presented and discussed.

123
4.9 Contributed - Environmental / Spatial Statistics: Climate
Wednesday 4 September 9am

A bivariate spatiotemporal model to estimate the occurrence of ozone and PM10

emergency alerts in Mexico City

Eliane R. Rodrigues1, Philip A. White2, Alan E. Gelfand2, Guadalupe Tzintzun3

1
Universidad Nacional Autonoma de Mexico (UNAM), 2 Duke University, 3 Instituto Nacional
de Ecologia y Cambio Climatico

Mexico City is among the many cities suffering from high levels of pollution. In order to
reduce population exposure and, therefore, the health hazard that is a consequence of this
exposure, Mexico City has implemented a series of preventive measures. Among them are
emergency alerts which are declared whenever high levels of ozone and/or PM10 occur. In
the present talk a bivariate spatiotemporal model is considered to predict local pollution
emergencies and to assess compliance to Mexican ambient air quality standards. Hourly
ozone and PM10 measurements from 24 stations across Mexico City from 2017 are
analyzed. Using this model, future pollutant levels using current weather conditions and
recent pollutant concentrations may be predicted as well as the regional maxima needed to
estimate the probability of future pollution emergencies. We discuss how predicted
compliance to legislated pollution limits varied across regions within Mexico City in 2017. We
find that predicted probability of pollution emergencies is limited to a few time periods. In
contrast, we show that predicted exceedance of Mexican ambient air quality standards is a
common, nearly daily occurrence.
This is a joint work with Philip A. White, Alan E. Gelfand, and Guadalupe Tzintzun.

124
4.9 Contributed - Environmental / Spatial Statistics: Climate
Wednesday 4 September 9am

A new statistical approach to forecasting non-stationary climate indices

Philip Sansom, Daniel Williamson, David Stephenson

University of Exeter

We present a new statistical model for forecasting non-stationary climate indices based only
on observations and compare it to state-of-the-art dynamical seasonal forecast models and
existing methods for estimating potential predictability. The new model separates a time
series of a climate index in to long-term trends (including multi-decadal signals), seasonal
behaviour, day-to-day variability, and observation errors. Simple non-stationary statistical
models are proposed for each component so that the state of each can be simultaneously
estimated at any time. The rate of change in each component is also estimated directly from
the time series. Periods of enhanced predictability due to transient changes in either the
mean or day-to-day autocorrelation of the climate index can also be identified and exploited
for prediction.We apply the proposed statistical model to the problem of seasonal forecasting
for the North Atlantic Oscillation (NAO). The statistical model identifies a period of enhanced
predictability of the NAO between December and March, extending into November and April,
due to transient changes in the mean of the NAO which we attribute to external forcing.
Forecasts of the winter (Dec-Jan-Feb) mean NAO achieve a correlation with the
observations greater than 0.5, making the statistical model comparable to the state-of-the-art
GloSea5 seasonal forecast model. The forecast skill is actually greater in spring (Mar-Apr-
May) than in winter. Limited forecast skill is also found in summer and autumn due to the
presence of long-term trends.The statistical model can also be used to study potential
predictability in climate indices. We estimate that around 60% of the inter-annual variability in
the winter mean NAO is attributable to an externally forced signal, 30% to accumulated day-
to-day variability, and 10% to long-term trends. These estimates are surprisingly close to
those from earlier potential predictability studies which confounded the predictable signal
with changes in the long-term mean.

125
4.9 Contributed - Environmental / Spatial Statistics: Climate
Wednesday 4 September 9am

Quantifying uncertainty in climate projections based on emergent constraints

Philip Sansom
University of Exeter

We present a new Bayesian statistical model and web-based tool for quantifying uncertainty
in climate projections based on emergent constraints. Emergent constraints are a popular
but controversial method of constraining projections of future climate based on simulations
from ensembles of climate models and historical observations.We explain the statistical
assumptions that lead to the standard model for emergent constraints. We highlight the
different sources of uncertainty present when exploiting emergent constraints. We argue that
the standard model for emergent constraints underestimates the uncertainty in projections
by ignoring uncertainty about how constraints estimated from ensembles of climate models
will manifest in the real world. We propose a new Bayesian statistical model for emergent
constraints that accounts for this missing uncertainty. The new statistical model is a simple
generalisation of the standard model for emergent constraints that allows for more realistic
judgements about how emergent constraints will manifest in the real world. The new
statistical model has been implemented as a web-based tool that enables quick and simple
visualisation of the effects of accounting for the missing uncertainty. We demonstrate both
the statistical model and the accompanying web-based tool by application to emergent
constraints on equilibrium climate sensitivity.

126
Keynote 4 - Barnett Lecture
Wednesday 4 September 10.10am

Data lakes from the data deluge- a digital environment vision

Marian Scott
Glasgow University

The landscape of environmental data comprises outputs from process models, data
generated from earth observation, and from in situ (static and mobile) sensor networks, and
citizen science contributions including from social media feeds- so everything is anchored in
time and space (where we may need to account for sampling effort, bias, detectability). In
many instances, we measure proxies of the environmental variable of interest, and often, we
still persist in thinking of the environmental media in silos. Learning from these many data
sources is challenging, our objectives may be prediction so a black box might be sufficient,
but in terms of explainability and transparency, will not not adequate. Variation, variation,
and more variation clouds our perceptions of the state of the environment (driven by data
quality, by measurement uncertainty and by natural variation, we know each particle of soil is
different at a fundamental level, each parcel of air or water similarly different). Our
understanding of fundamental environmental process may be high-level and profound, but
may include detailed biology, chemistry or geology, so that expert knowledge is contributing
a further form of data. Are we foolish to think we can learn the rules that govern our world
from our imperfect observations? In this presentation, I will reflect through a number of
examples on what we might mean by digital environment and what that vision is.

127
5.1 Medical: Precision Medicine
Wednesday 4 September 11.50am

Dynamic modelling of single-case (n-of-1) data: challenges and novel applications

Rute Vieira
University of Aberdeen

Introduction: Single-case studies are increasingly recognised as a valid and efficient

mechanism for making individualized evidence-based treatment decisions. Statistical
analyses of N-of-1 data require accurate modelling of the outcome variable while accounting
for its distribution, time-related trend and error structures (e.g. autocorrelation) as well as
reporting readily usable effect sizes for clinical decision making. A substantial number of
statistical approaches have been documented but no consensus exist on which method is
most appropriate for which kind of design and data.

Methods: In this paper we discuss, from a statistical perspective, N-of-1 studies and describe
a regression method for the analysis of N-of-1 data, borrowing ideas from longitudinal and
event history methodologies which explicitly incorporate the role of time and the dependence
of future on past. The aims include identifying predictors of response, describing adaptive
changes over time, or predicting future behaviour given prior history.

Results: Different applications of the method are presented, specifically to identify predictors
of physical activity during retirement transition and predictors of photoprotection behaviour in
patients with Xeroderma Pigmentosum, a rare inherited condition characterized by an
extreme sensitivity to ultraviolet rays from sunlight. Our approach was shown to be flexible,
adaptable to different types of outcomes and capable with dealing with the different
challenges inherent to N-of-1 modelling.

Conclusions: N-of-1 studies are feasible and helpful in identifying individualised predictors of
behaviour and facilitating the development of individualised intervention plans. Dynamic
modelling has the potential to expand access of N-of-1 researchers to robust and user-
friendly statistical methods.

128
5.1 Medical: Precision Medicine
Wednesday 4 September 11.50am

Predicting disease progression in neurodegenerative diseases with high phenotypic

variability

Frank Dondelinger
Lancaster University

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 a high-dimensional
mixture model approach for jointly solving the problem of data-driven estimation of clinical
phenotypes and prediction of disease progression. Longitudinal dynamics are captured via a
mixed model approach, and we take into account both the distribution of the response and
the distribution of the covariates for estimating the disease phenotypes. We demonstrate the
performance of our method 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 in both cases joint inference of the subtypes and
predictors improves the prediction performance, and hence the clinical usefulness of our
results.

129
5.1 Medical: Precision Medicine
Wednesday 4 September 11.50am

Methodological challenges for precision public health

David Wright1, David Taylor-Robinson2, Frank Kee1

1
Queen's University Belfast, 2 Institute of Population Health Sciences, University of
Liverpool

Objectives: Considerable debate surrounds the usefulness of applying precision medicine

techniques to the sphere of public health. We highlight some of the methodological
challenges which must be addressed to ensure that the benefits of “precision public health”
outweigh the harms.

Challenges: Building an evidence base for use of genomic or detailed digitally collected
phenotypic data in public health faces three main challenges. First, to identify a target for
intervention, risk of a given outcome must be defined for each individual. Widely used
epidemiological methods focus on estimating population level parameters which do not
translate readily at the individual level. Current analyses often fail to account for complex,
interdependent influences on individual health or for reporting biases, residual confounding,
collinearity and measurement error when assessing risk factors. Second, there is a need to
adopt new methods to design and test the effectiveness of individual level interventions.
Conventional parallel group randomised controlled trials are insufficient; n-of-1 studies and
micro-randomised trials offer potential solutions. Intervention design should be theory-based
and clearly define what represents a clinically important difference for each individual. Third,
a framework is needed to assess unintended consequences and potential harms of precision
public health evaluations. As more people use “wearable” technologies we are likely to face
more “incidental” findings that have limited “actionability” leading to unnecessary
interventions. Allowing apps and AI to substitute for insightful professional intervention in
support of behaviour change may open the door to commercial interests and risks of harm.
Most psychological studies of the impact of biological or behavioural risk profiling have
focussed on single behaviours - risks of information overload or inducing adverse
compensatory behaviours have generally not been studied.

Conclusions: There is a need for statisticians to engage with these challenges to allow
robust evaluation of precision public health interventions before they are implemented more
widely.

130
5.1 Medical: Precision Medicine
Wednesday 4 September 11.50am

Robust joint modelling: a new approach to handle time-varying outlier impacts

Laura Boyle1, Özgür Asar2, Lisa McFetridge3, Jonas Wallin4

1
The University of Adelaide, 2 Acıbadem Mehmet Ali Aydınlar University, 3 Queen's
University Belfast, 4 Lund University

Joint modelling approaches, which simultaneously analyse a longitudinal and survival

process, can be used to understand the relationship between repeated measurement data
and patient survival in medical research. Medical data commonly contain outlying values,
which can occur for a number of reasons, such as patients who are experiencing a period of
stabilising in response to a new treatment. To handle such data, robust joint modelling
approaches have been developed, utilising t-distributional assumptions to down-weigh the
detrimental impact of longitudinal outliers on parameter estimation. However, these
approaches do not currently account for scenarios where the pattern of outliers varies over
time.This research advances joint modelling methodology by allowing the degrees of
freedom parameter for the longitudinal residuals to vary through time using a natural cubic
spline. Results from a simulation study will be presented, in addition to application of the
method to two motivating medical datasets –a cohort of patients with liver disease, and a
cohort of renal patients from Northern Ireland. Through comparison of a range of robust joint
models, this research not only stresses the need to properly account for longitudinal outliers,
a practice not commonly done in the literature, but highlights the bias from not properly
estimating the degrees of freedom parameter. Each technique presented can be fitted using
the R software package “robjm”.

131
5.2 Official & Public Policy: Feeding the beast - Satisfying user demand for precision,
timelines and speed!
Wednesday 4 September 11.50am

How does a Government Statistical Agency respond to the challenge of developing

and monitoring a societal wellbeing framework?

Niall O'Neill
NISRA

Part 1: A Wellbeing Framework for Northern Ireland - How does a Government Statistical
Agency respond to the challenge of developing and monitoring a societal wellbeing
framework?The Draft Programme for Government (2016-21) and Northern Ireland Civil
Service (NICS) Outcomes Delivery Plan set the strategic context for central government in
Northern Ireland, while Community Development Plans set the context in our local
government districts. A wellbeing framework of 12 Outcomes and 49 population indicators
overarches all of this work.The framework is heavily dependent on the application of official
statistics produced by NISRA and, with data at its core, NISRA statisticians in all government
departments are responding to the increased user demand and focus that the framework
has brought.This presentation will cover NISRA’s role in the development of the framework,
including work to date in addressing gaps where data for proposed indicators did not exist.
We will also consider how NISRA has met the challenges of reporting change over time
through the establishment of a Technical Assessment Panel.

132
5.2 Official & Public Policy: Feeding the beast - Satisfying user demand for precision,
timelines and speed!
Wednesday 4 September 11.50am

How NISRA has been informing EU Exit discussions

Chris Ganley
NISRA

The purpose of the presentation is to provide a brief overview of some of the work NISRA
have been undertaking over the last 2 years to inform the BREXIT debate. A number of
projects have been undertaken in order to answer queries from policy colleagues around
issues such as:What is the value of trade with the EU?Who are our biggest trading
partners?What is the profile of NI businesses involved in trade with the EU?What products
do we sell to / buy from X, Y, Z?What impact would a change in the tariff regime have on NI
trade?Who’s most likely be impacted by changes to trading arrangements? How many
people cross the NI-ROI border?How many vehicles cross the border?How many EU
Nationals work in NI?How much freight is crossing the Irish Sea?Our work has fed into HM
Government Position papers, internal briefings for policy colleagues in the Department for
the Economy, Senior Civil Service and Department for Exiting the EU. We have also
published slide packs and papers to make the data accessible to the public to ensure that
there are informed debates around BREXIT.

133
5.2 Official & Public Policy: Feeding the beast - Satisfying user demand for precision,
timelines and speed!
Wednesday 4 September 11.50am

Forecasting Emergency Care Occupancy

Eugene Mooney, Kieran Taggart, Michael O'Donnell

NISRA

It is well known that our Emergency Care Departments are experiencing significant
challenges and pressures, as a result of increased demands and tighter budgets. Despite
the wealth of information routinely published informing health care professionals and the
wider public of these pressures, it has been of little use in assisting them to better plan their
services.This presentation will outline how NISRA statisticians have utilised this existing
information to forecast occupancy levels in emergency departments during each hour up to 6
weeks in advance, and enable Health & Social Care Trusts to better plan their unscheduled
care services.It will also demonstrate the use of technological developments that allow this
information to be almost completely automated and presented in a user friendly format,
whereby the user can adjust the information for their own specific needs.

134
5.3 Application of Statistics: Data in the NHS - facilitating quality improvement in
healthcare
Wednesday 4 September 11.50am

Monitoring perinatal mortality in the UK

Bradley Manktelow
University of Leicester

Audit data plays an important role in patient safety and quality improvement in healthcare.
Care providers (e.g. hospitals, surgeons) are able to compare their clinical performance to
other providers, specified benchmarks or their own previous performance. It is important that
such information is presented in a manner that allows clear interpreted and can facilitate the
improvement of the quality of care.Over the past 15 years, funnel plots have become a
standard graphical method with which to present summary statistics for individual healthcare
providers. Derived from SPC control charts, funnel plots are constructed to show the value of
the summary statistic together with control limits indicating the ‘expected’ range of these
values. Usually the control limits are selected to allow a precise probabilistic interpretation.
However, the use of funnel plots raises several issues, including the plausibility of the
assumptions and the use of statistical criteria to identify clinical outliers.One existing
alternative approach is to use multilevel models to estimate empirical Bayes estimates of the
values of the summary statistics. This approach can potentially allow the inclusion of clinical
criteria into the identification of outliers.These two methods will be described and illustrated
using mortality data from MBRRACE-UK, the national surveillance programme for stillbirths
and neonatal deaths in the UK. The potential advantages and challenges of the methods will
be discussed, together with the interpretation of the summary statistics.

135
5.3 Application of Statistics: Data in the NHS - facilitating quality improvement in
healthcare
Wednesday 4 September 11.50am

Linked cluster-randomised trials of audit and feedback: A “split-block” design

Rebecca Walwyn
University of Leeds

Feedback interventions are frequently tailored across a variety of healthcare settings. The
AFFINITIE programme evaluated feedback interventions used in audit and feedback cycles
embedded within a UK national comparative audit programme of blood transfusion practice.
Two feedback interventions (“content” and “follow-on support”) were evaluated in the context
of both a surgical audit and a haematology audit in two linked cluster-randomised, 2x2
factorial trials, one for each audit topic. The national audit data was used as baseline and
outcome data. As eligible, consenting NHS trusts across the entire country were involved in
both trials, we considered whether trusts should maintain their allocations across trials or be
re-randomised to interventions for the second trial. Patients and health professionals were
largely separate across the trials. Researchers were interested in the size of the intervention
effects in each setting, but also in whether these effects generalise across settings. Design
of experiments literature has developed largely independently of clinical trials. Multi-stratum
designs include cluster-randomised, split-plot and split-block designs as special cases.
Intervention packages, or selected components, are randomised to clusters, possibly at
more than one level, and outcomes are measured on patients. We will comment on the
issues around using national audit data in clinical trials. We will then illustrate a range of
possible multistratum designs, using our motivating example, and discuss the implications
for equipoise, contamination, balance, convergence and power of using a split-block design
to understand the generalisability of interventions across settings as part of a single research
programme. Finally, we will outline the limitations of existing multistratum designs for use in
a clinical trials context.

136
5.3 Application of Statistics: Data in the NHS - facilitating quality improvement in
healthcare
Wednesday 4 September 11.50am

Quasi-Experimental designs for quality improvement research in Acute Ischemic

Stroke Care

Abdel Douiri
King's College London

The decision-making process in healthcare is still predominantly driven by descriptive

analytics which only superficially explore the information contained within the data. This talk
will focus on methods for stroke quality of care improvement and will endeavour to provide
strategies and tools to implement the use of statistical methods and designs (e.g: quasi-
experimental design) to identify performance gaps, suggest possible solutions and inform
and evaluate quality improvement initiatives. A study example using propensity scores and
national real-world data to evaluate Intravenous thrombolysis with alteplase, one of the few
approved treatments for acute ischemic stroke care, will be presented and challenges will be
discussed.

137
5.5 Methods and Theory: Theoretical advances in experimental design
Wednesday 4 September 11.50am

Design and analysis of experiments testing for biodiversity

R. A. Bailey
University of St Andrews

It is now widely believed that biological diversity is good for the environment. One way that
ecologists test this is to place random collections of species in mini-environments and then
measure some outcome. Others use a carefully chosencollection of species.Is the outcome
affected by the number of different species present, or is it just the number of members of
each species that matters? And is there an interaction between these? Are the answers
affected by the temperature, or by the complexityof the environment?I have been working
with a group of fresh-water ecologists on the design and analysis of such experiments. Our
subsets of species are carefully chosen, not random. We design the combinations of these
subsets with levels of other factors.We also fit a nested family of plausible models to the
data. Our results suggest that the underlying model is not diversity at all. One of my crucial
inputs has been the use of Hasse diagrams as a way of understanding a complicated family
of plausible models for the expectation of the response.

138
5.5 Methods and Theory: Theoretical advances in experimental design
Wednesday 4 September 11.50am

Construction of Blocked Factorial and Fractional Factorial Designs

Janet Godolphin
University of Surrey

Generating matrix methods are used to construct 2n and 2n-p factorial designs with single
and double confounding, that is, with one or two forms of blocking. For designs with one
form of blocking, the construction of single replicate designs in blocks of size 2q, enabling
estimation of all main effects and maximising the number of estimable two factor interactions
is demonstrated. For the situation in which a specific subset of the two factor interactions is
of interest, methods on proper vertex colouring from Graph Theory are exploited to inform
the generator matrix and yield a suitable design.Designs with two forms of blocking are
represented as one or more rectangular arrays with rows and columns corresponding to
blocking factors. Templates are given for generating matrices which yield single replicate
constructions to estimate all main effects and the maximum number of two factor
interactions. The construction approaches are extended to accommodate fractional designs.

139
5.6 Communicating & Teaching Statistics: Show me the stats
Wednesday 4 September 11.50am

Show me the question

Andrew Wright
Novartis

Graphics are often designed “bottom up”. That is, we start from the data and decide how
best to display it. However, since a graph is most impactful if it is designed to answer a
specific question, it is more meaningful to start from the question itself. To this end, a “top
down” approach is suggested. We call this approach “Question-Based Visualizations” (QBV).
The QBV approach to developing visualizations encourages teams to first align on the key
questions they want to address with their data. Then, and only then, it is decided how to best
create a graph (or set of graphs) to answer this question. This subtle yet important shift in
the process of graph creation greatly improves the impact of the visualizations. In cases
where a question is best answered by a combination of information from disparate dataset
domains (e.g., vital signs and biomarkers), this is also more easily done “top down”; the
“bottom up” approach tends toward separate displays for separate domains. Visualizations
that are developed using this question-based approach will be illustrated in two different
settings: a static graphic for inclusion in a report and interactive graphics embedded within a
Shiny app.Putting the questions front and center when developing visualizations of our data
can help us to better influence our teams.

140
5.6 Communicating & Teaching Statistics: Show me the stats
Wednesday 4 September 11.50am

Look at graphs

Allan Reese
Independent consultant

Why, despite many sources of good advice including Tufte's books and a tutorial by
Vandemeulebroecke et al (to appear in CPT:PSP), do bad graphs continue to appear in print
and during presentations? A corollary of the assumption "graphs communicate" is that
people glance at them and move on. Secondly perhaps, "familiarity breeds contempt". When
I devised a university module in Graphical Interpretation of Data (GID), colleagues in the
Statistics department were happy to agree to assist at practical sessions: "I've used graphs
all my career so I'm familiar with them." Once they attended, the general response was, "I've
never thought about graphs that way before!" Students, from a variety of disciplines, often
signed up in the expectation of learning some snappy tricks with Excel. They had been
beguiled by advertising along the lines of, "Buy our software and with a few clicks you'll be
creating presentation masterpieces." How do we improve? How long do you/should you
spend looking at an individual graph? Should you start with the overall pattern or the detail –
the wood or the trees? The central premise of GID is that there are no fixed rules, no
absolutes. As a communication medium, better graphs – and better understanding of what
you are doing with them – come from practice and reflection. The lack of revision (redrafting)
is, I suggest, a major reason for poor graphs. Graphs exist in the context of their data and
analyses. Many published graphs appear to have come straight from software with all the
default settings. The surrounding text has been reviewed and redrafted many times – why
not the graphs?In this short session, I will critique a few example graphs to emphasize the
GID approach. I look forward to comments from the panel and audience.

141
5.6 Communicating & Teaching Statistics: Show me the stats
Wednesday 4 September 11.50am

Say one thing: numbers in the news

Robert Cuffe
BBC

When the aim is to communicate with millions at a glance, graphics need to be clear. Robert
will share examples from BBC online and TV graphics and some insight from audience
research about what works.

142
5.7 Data Science: Professional Ethics in Data Science
Wednesday 4 September 11.50am

How the ASA Ethical Guidelines for Statistical Practice comprise data science and
“data ethics”

Rochelle Tractenberg
Georgetown University and the Collaborative for Research on Outcomes and -Metrics

The American Statistical Association (ASA) first established its Ethical Guidelines for
Statistical Practice in 1999. In 2014, the ASA Committee on Professional Ethics set out to
revise these Guidelines. These were revised in 2016 and 2018. Lessons learned included
how to balance “completeness” with shortness of attention spans; the need for applicability
of the Guidelines to any person who works directly or indirectly with data; and the fact that
Guideline principles may conflict in any given case, so rather than “learn the Guidelines”,
ethical practitioners need to learn how to use the Guidelines. Recent worldwide interest in
data science and data ethics creates challenges for teaching statistics and statistical ethics,
but also creates new “data analysts” who are not, and who would not identify themselves as,
“professional statisticians”. If applicability of the Guidelines is subtly shifted from "ethical
statistical practice" to “ethical quantitative practice”, it can comprise data science and data
ethics.

143
5.7 Data Science: Professional Ethics in Data Science
Wednesday 4 September 11.50am

Data Science: Professional Ethics in Data Science

Leone Wardman
ONS

Data science methods are increasingly common within analytical professions and come with
unique ethical challenges. The Royal Statistical Society (RSS) Data Science Section and the
Institute and Faculty of Actuaries (IFoA) collaborated to explore the practical and ethical
implications of using data science in accordance with their respective strong professional
values, and overarching commitment to the public interest. In 2018, joint regional workshops
were held with a broad range of data science professionals to discuss key questions for data
science, including the responsibility of data scientists to society and practices for a good
data science ‘workflow’. A joint focus group considered the findings from the workshops,
along with a range of existing data ethics frameworks relevant to data science and Artificial
Intelligence, to develop globally applicable ethical and practical guidance for our respective
members when using data science. This session provides a background to the RSS and
IFoA collaboration and an exciting insight into the proposed joint ethical guidance. In
particular, it explores five overarching ethical themes and corresponding working practices to
assist members in working ethically with data science.

144
5.8 Business, Industry & Finance: Embedding Data Analytics in the Business Process
Wednesday 4 September 11.50am

Using data to become a clearer, quicker and tougher regulator

Johanna Hutchinson
The Pensions Regulator

Increasing poverty in old age, a growing retired population and a fluctuating economy have
led to considerable pension reforms in government. High profile cases of deficits in pension
scheme from BHS, Carillion, Tata Steel and Toys R Us, alongside the launch of obligatory
work-based pension schemes have kept the Pensions Regulator in the UK’s media. The
proactive regulation of pension schemes has required a new approach, an investment in
technology and a dependency on data. In this presentation I will discuss how we built data-
driven regulatory processes, provided operational efficiencies and new insights using data
science, data management and data governance techniques.

145
5.8 Business, Industry & Finance: Embedding Data Analytics in the Business Process
Wednesday 4 September 11.50am

Uncertainty in real-world applications of AI

Dongho Kim
PROWLER.io Limited

Real world applications of AI often present challenges such as uncertainty, and multiple
agents with different goals operating in dynamic environments. In order to mitigate this, we
at PROWLER.io pursue a principled machine learning approach combining probabilistic
modelling and decision theory.In this talk, we will describe some examples of applied AI
where there's not only uncertainty, but also unknown values that need to be worked with. We
motivate the need for an integrated approach that combines several sub-disciplines to solve
problems such as scheduling transport when stock levels are unknown and unobservable.

146
5.8 Business, Industry & Finance: Embedding Data Analytics in the Business Process
Wednesday 4 September 11.50am

IceCAM – the Iceberg Crop Adaptive Model – Minimising food waste by adapting
growing programmes to the weather

Iain Flint
G's Growers

Some fresh produce like lettuce and celery cannot be stored long-term and so must be sold
the week it is harvested. This means growers must anticipate how temperature, light and
other environmental factors will affect their crops to sow the right amount of crop months in
advance to meet their sales programme on any given week. To solve this, G’s Growers have
developed a crop growth model – IceCAM – that reliably predicts harvest date in response to
the weather in a given year for G's main crop Iceberg lettuce, as well as Gem and Romaine
lettuce, Spinach, and Radish. This allows the growers to target higher yields with the
confidence that they will meet their sales programme, meaning less land and resources are
required to produce the same amount of food. This talk will discuss the progress of the
model so far as well as recent developments including development of a biological crop
model, learning from remote sensing images, and Monte Carlo simulation of forecast
weather conditions to make optimal programme decisions under uncertainty, as well as
future ambitions for the project.

147
5.9 Environmental / Spatial Statistics: Climate Change
Wednesday 4 September 11.50am

Bayesian Additive Regression Trees for palaeoclimate reconstruction

Andrew Parnell
Maynooth University

Learning about past climate is of key importance in discovering how fast climate can change.
Most reconstructions of past climate use proxy data (pollen, coral, tree rings, etc) which are
treated as highly noisy weather stations. Statistical methods are required to calibrate the
proxy data to learn about their relationship to climate, and also to model the way that climate
might change over time.A large number of different statistical methods have been proposed
to perform reconstruction, across the full gamut of statistical techniques. In this talk I will
conduct a brief review of some of these and then propose a new one based on the method
of Bayesian Additive Regression Trees (BART). The key development includes that of
extending the BART model to multivariate responses with non-ignorable missingness. No
previous knowledge of BART is necessary, but attendees should be familiar with basic
Bayesian model building and fitting for e.g. linear regression models.

148
5.9 Environmental / Spatial Statistics: Climate Change
Wednesday 4 September 11.50am

Statistics for Heatwaves and Extreme Waves in a Changing Climate

Jonathan Tawn
Lancaster University

Understanding how environmental risk factors change over time is vital for forward planning
and infrastructure design. If interest is in the behaviour of heatwaves or extreme wave
events in 50 years’ time, standard extreme value methods using only observational data
cannot be applied with any confidence given the level of change in climate that is anticipated
over this time scale. Additional information from climate models is essential. However, using
global climate models induces complications linked to their low spatial resolution, the choice
of climate change scenario, and the sensitivity to climate model formulation. We will illustrate
the problems of dealing with these features for estimating risk by presenting new
downscaling methods for extreme wave events and a framework for heatwave modelling
which allows for different climate change effects on the marginal distribution and duration
properties of heatwaves. In our applications to North Sea waves and French heatwaves we
find that there could be a significant increase in risk, with extreme events increasing at a
faster rate than mean values.

149
6.1 Medical: Bridging public understanding of health and data sharing
Wednesday 4 September 2.20pm

Communicating about patient data - some experience from Understanding Patient

Data

Natalie Banner
Understanding Patient Data

Understanding Patient Data is an independent initiative hosted at Wellcome that seeks to

make the uses of patient data more visible, understandable and trustworthy. There are
enormous potential benefits from using data collected in routine care better to improve
health, but many people also understandably have concerns about their privacy, and also
want to ensure data is used for public and social good. The language around data is often
technical, complex and inconsistent, with unclear rules across organisations and a lack of
transparency over what happens to patient data and why. UPD aims to help demystify the
use of data, providing accessible tools and resources for clinicians, researchers, and anyone
who wants to create trustworthy systems and talk about data use to patients and the public.
In this talk, Natalie will give an overview of UPD’s work to date, resources produced and
collaborations with other organisations, and explore some of the challenges of engaging with
people about data use in accessible ways.

150
6.1 Medical: Bridging public understanding of health and data sharing
Wednesday 4 September 2.20pm

Socialising expertise: people-centred data governance of health information

Stephanie Mulrine, Madeleine Murtagh, Mwenza Blell, Joel Minion, Mavis Machirori
Newcastle University

Across the UK and internationally, steps are being taken to revolutionise access to patient
data collected by health services like the NHS. Work is underway to provide and assemble
the necessary legal and technical expertise. Yet without public and patient perspectives on
the use of their data, the potential for unsuccessful, dangerous or discriminatory
consequences is a real risk.Focus group participants in the North East of England and North
Cumbria were presented with information about a proposed technological solution: the Great
North Care Record. A clear ethical framework of values emerged: respect, reciprocity,
fairness, agency, privacy, transparency and trust. Whilst there is broad public support, these
findings demonstrate concerns need to be addressed and a set of key values which must be
adhered to. Further work (currently underway) to develop a co-production approach will
inform the design and governance of the evolving process of health information exchange
and access.

151
6.1 Medical: Bridging public understanding of health and data sharing
Wednesday 4 September 2.20pm

The patient imperative to ‘use MY data’

Debbie Keatley1, Finian Bannon2, Hannah McKenna2

1
use MY data, 2 Centre for Public Health, Queen's University Belfast

use MY data is a movement of patients, carers and relativesuse MY data supports and
promotes the protection of individual choice, freedom and privacy in the sharing of
healthcare data to improve patient treatments and outcomes. use MY data endeavours to
highlight the many benefits that appropriate usage of healthcare data can make, to save
lives and improve care for all.use MY data aims to educate and harness the patient voice to
understand aspirations and concerns around the use of data in healthcare delivery, in
service improvement and in research, aimed at improving patient decision making, treatment
and experience.Our vision is of every patient willingly giving their data to help others,
knowing that effective safeguards to maintain the confidentiality and anonymity of their data
are applied consistently, transparently and rigorously.Debbie will talk about the growth of the
movement and current topics in data access and usage, highlighting problems in access to
routinely collected patient data, and how there may be too much risk-aversion to giving
access to data.Debbie will then highlight what is possible when data is used. Using the
example of routes/pathways to diagnosis she will share findings from the recently completed
Northern Ireland Pathways to Cancer Diagnosis study and will demonstrate how Routes to
Diagnosis studies in England have been a driver for positive change; demonstrating the
benefits of collecting and using patient data.

152
6.5 Methods & Theory: Statistical sparsity
Wednesday 4 September 2.20pm

Overfitting correction in multivariate survival analysis

Anthony Coolen
King's College London

Generalized linear multivariate models are ubiquitous in applied statistics. They include
logistic regression for binary classification, proportional hazards (Cox) models for time-to-
event data or ordinal class data, frailty models, and so on. In most of these models,
parameters are estimated using Maximum Likelihood (ML) regression. However, when the
data dimension p is comparable to the sample size N, as is very often the case in post-
genome medicine, maximum likelihood inference breaks down due to overfitting. This
prompted the introduction of the so-called regularized models, which are equivalent to point-
estimate approximations of Bayesian parameter inference (MAP regression), and require
undesirable hyper-parameter tuning. Overfitting causes not only excess noise that is not
captured by conventional p-values, but it also induces a strong inference bias, leading to
consistent over-estimation of associations. In this presentation we show principles and
applications of a recent formalism for modelling quantitatively the effect of overfitting in
generalized linear models with ML or MAP regression. This formalism is based on the so-
called replica method from statistical physics. It leads to precise formulae for the correction
of multivariate regression parameters and base hazard rates in the overfitting regime, and
for analytical formulae for optimal regularization hyper-parameters. In this presentation we
will focus mainly on its applications in multivariate survival analysis.

153
6.6 Communicating & Teaching Statistics: Can mathematics anxiety obstruct learning
statistics in university students?
Wednesday 4 September 2.20pm

An interactional and inclusive approach to enhancing students’ engagement with

statistics.

Meena Mehta Kotecha

The London School of Economics and Political Science

The world is increasingly driven by data science with statistics becoming increasingly
important in all areas of society. Statistical skills are pivotal in the global employment market
and university degree programmes are becoming progressively quantitative, involving
statistical analysis for virtually all disciplines. It is imperative that we design and deliver
statistics courses for non-specialists bearing this in mind, to prepare our students to face
today’s data driven work environment. I found mathematics anxiety (MA) (Richardson, &
Suinn, 1972) rather than statistics anxiety (SA) (Cruise, Cash, and Bolton, 1985) to be an
obstructing factor to non-specialists’ engagement with statistics courses. Approximately 78%
of non-specialists report MA (Kotecha, 2016). Aversive prior experiences with mathematics,
poor achievement in mathematics and a low sense of mathematics self‐efficacy have been
found to be meaningful antecedent correlates of SA (Zeidner, 1991). MA is found to be an
antecedent of SA and both forms of anxiety are found to be highly correlated (Paechter et al.
2017). I developed a research informed intervention to reduce MA reported by university
non-specialist university students. A mixed methods approach was used in this research.
The findings reflect a highly significant reduction in MA reported by non-specialist university
students after the intervention was implemented. While this result is promising, it also raises
important research questions that require rigorous inquiry. As Merton (1948) wrote “More is
learned from a single success than from the multiple failures. A single success proves it can
be done. Thereafter, it is necessary only to learn what made it work.” My current research at
the University of Cambridge is aimed at gaining an in-depth insight into the evident reduction
in MA and developing a theory that integrates the use of social media -which is an important
feature of the intervention used since September 2012 to present. This talk should trigger
thought provoking questions and discussion.

154
6.6 Communicating & Teaching Statistics: Can mathematics anxiety obstruct learning
statistics in university students?
Wednesday 4 September 2.20pm

Are graduate students scared of statistics? Statistics Anxiety: barriers, enablers and
policies.

Carlos Fresneda-Portillo
Oxford Brookes University

Statistics anxiety, defined as `the feelings of anxiety encountered when taking a statistics
course or doing statistical analyses; that is gathering, processing and interpreting’ (Cruise,
Cash, & Bolton, 1985), has become one of the global sustainable challenges (UN
Sustainable Development Goals, 4.c). Given that the growing demand on data analysis
experts and the fact that career choices are influenced by mathematics and statistics
anxiety, it becomes of paramount importance to gain deeper understanding about what
factors influence Statistics Anxiety. Research points out that as many as 80% of graduate
students may be suffering from Statistics Anxiety (Onwuegbuzie, 2004). In this talk, we will
aim to provide an overview on where we are with Statistics and Mathematics anxiety, and
what educational policies are underway for amelioration.

155
6.6 Communicating & Teaching Statistics: Can mathematics anxiety obstruct learning
statistics in university students?
Wednesday 4 September 2.20pm

How do medical students feel about statistics and data skills during a 10-week
individual project and what support do they access?

Jamie Sergeant
University of Manchester

At the University of Manchester, undergraduate medical students undertake a 10-week

student-selected project in their third year. Projects can either be “research” or “non-
research” (e.g. service evaluation or clinical audit) and the aims for all projects include the
critical review of literature, understanding research methodology and understanding basic
statistical analysis. Support available to students to help them meet these aims includes
lectures, statistical advice appointments and signposting to resources. But how confident
and well-prepared do students feel to tackle the statistical aspects of a project and does this
influence their choice of project? Once they have completed their project, do students feel
that their statistics and data skills have been enhanced? What about their level of
confidence? What sources of statistical support were students aware of and which did they
access during the project? This talk will describe an attempt to answer these questions by
asking students.

156
7.1 Contributed - Medicine: Clinical Trials
Wednesday 4 September 4.10pm

Optimal treatment allocations in sequential multiple assignment randomized trial

(SMART) design

Mirjam Moerbeek, Andrea Morciano

Utrecht University, the Netherlands

In an adaptive treatment strategy (ATS) subjects are assigned to a sequence of treatments.

The treatments that are available in a given stage depend on the previous treatments and
whether the subject responded to these treatments. A sequential multiple assignment
randomized trial (SMART) combines multiple ATSs with the aim to select the best one.
Consider as an example a SMART for overweight patients: they are either randomized to a
dietary intervention or a physical activity intervention in the first stage. Those who lost a
sufficient amount of weight in the first stage keep their intervention in the second. Those who
did not are randomized to either the other intervention or a combination of both
interventions.Equal randomization is often used in SMARTS, but this is not necessarily the
most efficient choice. The aim of the current study is to find optimal treatment allocations
under either a fixed total sample size or a fixed budget. It uses multiple objective optimal
design methodology to take into account multiple pairwise comparisons among ATSs, and to
put a weight of importance on each comparison. Optimal treatment allocations depend on
the response rates to current treatments, and these response rates are often not known a
priori. For that reason we derive robust optimal design using maximin methodology. We
compare our optimal design with equal randomization and show how much efficiency is lost
by using the latter. Our optimal design methodology is implemented in a Shiny app.

157
7.1 Contributed - Medicine: Clinical Trials
Wednesday 4 September 4.10pm

Do mixed neuropathologies affect cognitive decline and possible Alzheimer’s Disease

Clinical Trial failure?

Sumali Bajaj1, David X. Thomas2, Kevin McRae-McKee1, Christoforos Hadjichrysanthou1,

Frank De Wolf1, John Collinge2, Roy M. Anderson1
1
Imperial College London, 2 University College London

Background: Phase 3 clinical trials of potential disease modifying therapies for Alzheimer’s
Disease (AD) have all been unsuccessful to date. AD is characterized by ß-amyloid (Aß)
plaques and neurofibrillary (Tau) tangles in the brain. While it is widely known that AD is a
multifactorial disease and that patients frequently have other pathologies in the brain, the
extent to which these co-occurring neuropathologies (TDP-43, CAA and Lewy Body) can
affect cognitive decline is poorly understood at present.

Methods: We developed a statistical model for describing the effect of the three co-occurring
pathologies in patients with AD and non-AD pathology on the progression of cognitive scores
(Mini-Mental State Examination, MMSE; Clinical Dementia Rating Scale Sum of Boxes,
CDR-SB) over time. Longitudinal scores were modelled within a Bayesian framework using a
hierarchical regression model with random intercepts and slopes. We also performed beta
regression as part of our sensitivity analysis. Beta distribution was appropriate for our
bounded continuous, skewed outcome and heteroskedastic data. Posterior predictive checks
were performed for model validation.

Results: In AD individuals we found that TDP-43 and CAA co-pathologies were significantly
associated with a faster rate of cognitive decline, after adjusting for confounders. Having
TDP-43 pathology, even without AD, can drive cognitive decline at a substantially fast
annual rate (ß = -1.19; 95% CI: -1.71, -0.69 and 0.70; 95% CI: 0.35, 1.06 for MMSE and
CDR-SB respectively). We demonstrate a high prevalence of moderate-severe co-
pathologies (63% in our data had TDP-43 or CAA along with AD pathology) and show their
effect on cognition. These results suggest that even if a drug completely removed all
parenchymal Aβ and Tau neuropathology within a well-defined AD population, a high
proportion of individuals may continue to cognitively decline, potentially being another
possible explanation for the failure of clinical trials solely target ß Amyloid.

158
7.1 Contributed - Medicine: Clinical Trials
Wednesday 4 September 4.10pm

Using Statistical Modeling to Identify the Useful Surrogate Outcomes in Critical Care
Studies

Rejina Verghis1, Bronagh Blackwood1, Cliona McDowell2, Daniel McAuley1, Mike Clarke12
1
Queens University Belfast, 2 Northern Ireland Clinical Trials Unit

Introduction: In critical care, many confirmatory studies fail, and this is sometimes because
the outcome measures used in the earlier studies were not valid. This contributes to
research waste and the aim of this research was to identify and validate surrogate outcome
measures in critical care trials.

Methods: A systematic review generated a full list of outcome measures used in critical care
trials, which was shortened based on the frequency of the reports and clinical judgement.
Critical care clinicians and researchers were then surveyed to identify the importance of the
identified outcomes to their clinical decision making and research. The association between
the surrogate outcomes and mortality was tested using Cox and joint modelling on relevant
secondary data.

Results: Forty eight articles were purposively sampled until data saturation, to generate a list
of 247 different outcomes. This was reduced to a smaller list of 22 outcomes, which were
broadly classified as resource use outcomes, disease severity scores, routinely collected
physiology outcomes and biomarkers. SOFA, a disease severity score in critical care, was
one of the most frequently reported outcomes and was considered an influential outcome by
clinicians and researchers. In the HARP2 study, the hazard ratio (95% CI), for SOFA score,
was1.54 (1.18 to 2.01) and 1.38 (1.29 to 1.48) for the cox model and joint model
respectively.

Conclusions: A variety of outcome measures have been used in critical care research, and
these are analysed and reported in many different ways. The variability makes comparisons
between studies in systematic reviews time consuming and difficult or impossible but the
SOFA score seems promising. In the dataset tested, the joint model estimates were more
precise than the cox model estimates and these analyses will be repeated on additional
datasets, to investigate the change in SOFA and associated difference in mortality rate.

159
7.2 Contributed - Official & Public Policy: Children's health and wellbeing
Wednesday 4 September 4.10pm

Understanding the histories of children in care and links to outcomes with

longitudinal analysis.

Cecilia Macintyre1, Gillian Raab2

1
Scottish Government, 2 University of Edinburgh

‘Looked after children' [1]are children in the care of their local authority. There are many
reasons children may become looked after: they face abuse or neglect at home; they have
disabilities that require special care; they are unaccompanied minors seeking asylum, or
who have been illegally trafficked into the UK; or they have been involved in the youth justice
system. Some children are cared for at home, with regular contact with social services. In
other cases the child or young person is cared for in placements away from their normal
place of residence by foster or kinship carers, prospective adopters, or in residential care
homes, schools or secure units. The experience of children and young people in care vary in
terms of the length of time and the range of different placements. This presentation
considers the link between and a child’s experience in care and educational outcomes
derived from administrative data. The data on looked after children is based on annual
returns which are made to Scottish Government by local authorities[2]. Individual level data
has been collected since 2008 with details of each placement. The data on absence,
attendance and exclusions [3]is collected every second year from publicly funded
schools.Data returns covering a ten year period were combined, resulting in over 50,000
children and around 120,00 placements. The analysis of school attendance of children,
included explanatory factors of length and consistency of placement, age and gender.
Challenges which were addressed included incomplete linkage due to missing identifiers ,
and selective effects of data due to incomplete coverage of care history for children born
prior to 2008 when data collection started.The presentation will discuss how the Scottish
Government has worked with partners – local authority, academic and other public bodies -
to improve the quality of the data used in this project

[1] Children (Scotland) Act 1995

[2] https://www2.gov.scot/Topics/Statistics/Browse/Children/PubChildrenSocialWork

[3]
https://www2.gov.scot/Topics/Statistics/ScotXed/SchoolEducation/AttendanceAbsenceExclu
sions

160
7.2 Contributed - Official & Public Policy: Children's health and wellbeing
Wednesday 4 September 4.10pm

Predictors of Mental Disorders in Children. Analysis of the Mental Health of Children

and Young People in England, 2017

Jodie Davis, Tim Vizard

Office for National Statistics

This session presents the methods and findings from a logistic regression analysis of the
2017 Mental Health of Children and Young People (MHCYP) survey in England. The
relationships between characteristics of a child (factors) and the development of mental
disorders are complex. More than one factor may be associated with a child having a
disorder, and these factors may also be associated with each other. Our research aimed to
identify which factors were associated with mental disorders in children (2 to 16 years old),
after controlling for other factors at the same time. By utilising a backwards logistic
regression, the research explored associations between fourteen factors and the presence
of mental disorders in children. Factors were grouped into demographic, family and
socioeconomic characteristics of children, with separate models run for preschool, primary
school and secondary school aged children.The results from this research were published by
NHS Digital in March 2019, in a report authored by the Office for National Statistics (ONS).
The findings showed that factors such as parental mental health, family functioning and
receipt of welfare benefits were associated with mental disorders in children of all ages and
disorder types. However, the research also found some factors were associated with
particular age groups and disorder types. This session, presented by ONS, will provide:an
overview of methods adopted in this analysis, some of the key findings from the research,
and future analysis plans for the survey.The 2017 Mental Health of Children and Young
People Survey was commissioned by NHS Digital, and conducted by ONS, the National
Centre for Social Research and YouthInMind.

For more information on the survey, please visit: https://digital.nhs.uk/data-and-

information/publications/statistical/mental-health-of-children-and-young-people-in-
england/2017/2017.

161
7.2 Contributed - Official & Public Policy: Children's health and wellbeing
Wednesday 4 September 4.10pm

Why our our children lonely? Findings from multivariate statistics and in-depth
interviews with 10-15 year olds

Ellie Osborn, Charlotte Hassell, Eleanor Rees

Office for National Statistics

Loneliness is normal and transient, but physical, mental and social problems arise when it
becomes chronic. It has a profoundly negative effect on health at all ages and Hawkley and
Cacioppo (2010) called it “social equivalent of physical pain”. Loneliness is associated with
poor health outcomes for younger ages as much as for adults. As part of the Prime
Minister’s strategy to overcome loneliness, the Social Analysis branch in ONS undertook
research exploring factors associated with loneliness among children. We developed interim
National Measures of Loneliness and noted that there is currently much more extensive and
robust data available on loneliness in older people, but comparatively little for children (aged
10-15) and young people (aged 16-24). We have also focused on analysing new data
collected to address this gap to map out what loneliness in these ages looks like in the UK.
Following the published exploratory analysis last year, regression analysis has now been run
to further understand the relationship between the factors highlighted. Focusing on children,
we have been able to further assess the relationship between loneliness, poverty, the impact
of a child’s perception, and their parent’s life satisfaction. We have also revisited the
interviews to use the children’s own words to complement this.

We will focus on:

• Which factors are associated with children’s experiences of loneliness

• What children say in their own words about their experiences of loneliness
• How we can ensure children’s loneliness is addressed to avoid illness in later life
• How we work with charities and policy-makers to ensure our analysis improves
loneliness in children as well as older ages

The analysis is key for the strategy for loneliness, and ONS works closely with DCMS and
charities to ensure that the research has the greatest impact.

162
7.3 Contributed - Medical: Data Science I

Wednesday 4 September 4.10pm

Applying Statistical Learning to Nuclear Magnetic Resonance Metabolic Profiling to

Predict Spontaneous Preterm Birth

Juhi Gupta1, Angharad Care1, Bertram Müller-Myhsok12, Laura Goodfellow1, Ana Alfirevic1,
Zarko Alfirevic1, Lu-Yun Lian1, Marie Phelan1
1
University of Liverpool, 2 Max Planck Institute of Psychiatry (Germany)

Preterm birth (PTB) is defined by the World Health Organisation as birth prior to 37 weeks of
gestation. Complications of spontaneous preterm birth (sPTB) are the leading cause of
death in children under age 5. Associated risk factors include genetics, lifestyle, or infection.
Understanding the mechanism of sPTB can aid development of novel methods to prevent
PTB. This study aims to investigate the metabolome of PTB high-risk patients to identify
potential biomarkers, using statistical methods. Serum samples were collected at 16 and 20
weeks of gestation from “high risk” women (with a history of sPTB or preterm premature
rupture of membranes (PPROM) under 34 weeks gestation), and “low risk” parous women
(with previous births delivered at term). Pregnancy outcomes were categorised into sPTB,
PPROM, high-risk term (HTERM) and low-risk term (LTERM). Serum samples (n = 647)
were acquired using solution-state 1D 1H Nuclear Magnetic Resonance (NMR) technology
(Bruker Avance III). Vendor software (Topspin 3.1) was used for processing and quality
control of spectra with Chenomx and in-house standards for annotating and identifying over
40 different metabolites. MetaboAnalyst web tool and R programming were used to analyse
the spectral bins.Probabilistic neural network analysis using a single timepoint between
cases and controls, resulted in a predictive AUC of 0.89 (LOOCV). One-way ANOVA
calculated across the 4 phenotypic sub-groups, resulted in a subset of metabolites
discriminating between HTERM:LTERM, sPTB:LTERM and PPROM:LTERM (p < 0.05 post-
hoc=Tukey HSD). Metabolites included phenylalanine and acetate, both reported in current
PTB biomarker literature.Variability due to exogenous substances and environmental
factors, such as drug intake or diet, require further investigation. Future work includes
integration of metabolomics with genomics, transcriptomics and proteomics. Investigating
multiple omics data for this patient cohort will allow for a better understanding of the
pathophysiology involved in initiation of early labour.

163
7.3 Contributed - Medical: Data Science I
Wednesday 4 September 4.10pm

Interpreting Clinical Narrative Diagnosis Models with Sentence Importance

Mark Ormerod, Jesús Martínez del Rincón, Neil Robertson, Bernadette McGuinness, Barry
Devereux
Queen's University Belfast

Despite advances in the application of deep neural networks to various kinds of medical
data, extracting information from unstructured textual sources remains a challenging task.
Using a dataset of de-identified clinical letters gathered at a memory clinic, we evaluate
recurrent neural networks (RNNs) on their ability to predict patients’ diagnoses of ‘Dementia’,
‘Mild Cognitive Impairment’ or ‘Non-impaired’. This classification framework can also have
applications in the automatic identification of patients as candidates for clinical trials.After
showing that models trained on state-of-the-art sentence-level embeddings outperform both
word-level models and a recent benchmark model that fine-tunes a pre-trained general-
domain language model, we probe sentence embedding models in order to reveal
interpretable insights into the types of sentence-level representations the RNNs build.
Specifically, we take a measure of sentence importance with respect to a given class and
identify clusters of sentences in the embedding space that correlate strongly with importance
scores for each class. Extracting the most frequent phrases within each group of sentence
representations shows that the model is sensitive to sentences that cluster around semantic
concepts such as a patient’s level of geriatric depression and how independent the patient is
in their daily activities.In addition to showing which sentences in a document are most
informative about the patient’s condition, our method can identify the types of sentences that
can lead the model to make incorrect diagnoses.

164
7.3 Contributed - Medical: Data Science I
Wednesday 4 September 4.10pm

A systematic review shows no performance benefit of machine learning over logistic

regression for clinical prediction models

Evangelia Christodoulou1, Jie Ma2, Gary S. Collins23, Ewout W. Steyerberg4, Jan Y.

Verbakel567, Ben Van Calster45
1
KU Leuven, Belgium, 2 Centre for Statistics in Medicine, Nuffield Department of
Orthopaedics, Rheumatology and Musculoskeletal Sciences, Botnar Research Centre,
University of Oxford, 3 Oxford University Hospitals NHS Foundation Trust, 4 Department of
Biomedical Data Sciences, Leiden University Medical Centre, Leiden, the Netherlands, 5 KU
Leuven, Department of Development & Regeneration, Leuven, Belgium; 6 KU Leuven,
Department of Public Health & Primary Care, Leuven, Belgium; 7 Nuffield Department of
Primary Care Health Sciences

Objective: To compare performance of logistic regression (LR) with machine learning (ML)
for clinical prediction modeling, and to describe methodology and reporting.

Methods: We conducted a Medline literature search (1/2016 to 8/2017), and extracted

comparisons between LR and ML models for binary outcomes. We classified comparisons
as at low or high risk of bias, where the latter was defined as: incorrect or unclear validation,
or different methodology for LR vs ML in terms of number of predictors, variable selection, or
validation.

Results: We included 71 out of 927 studies. The median sample size was 1250 (range 72-
3,994,872), with 19 predictors considered (range 5-563) and 8 events per predictor (range
0.3-6,697). The most common ML methods were classification trees (30 studies), random
forests (28), artificial neural networks (26), and support vector machines (24). Sixty-four
(90%) studies used the area under the receiver operating characteristic curve (AUC) to
assess discrimination. Calibration was not addressed in 56 (79%) studies. In 48 studies
(68%), we observed unclear reporting or potential biases in validation procedures for one or
more algorithms. We identified 282 comparisons between a LR and ML model (AUC range,
0.52-0.99). For 145 comparisons at low risk of bias, the difference in logit(AUC) between LR
and ML was 0.00 (95% confidence interval, -0.18 to 0.18). For 137 comparisons at high risk
of bias, logit(AUC) was 0.34 (0.20 to 0.47) higher for ML.

Conclusions: Methodology and reporting in studies that compare LR and ML needs

improvement. We found no evidence of superior performance of ML in clinical prediction
modeling. Future research should focus on delineating the type of predictive problems in
which different algorithms have maximal value.

165
7.4 RSS Prize Winners: Statistical Excellence Award for Early Career Writing 2019
Wednesday 4 September 4.10pm

The flying bomb and the actuary

Luke Shaw1, Liam Shaw2

1
Office for National Statistics (ONS), 2 Nuffield Department of Medicine, University of Oxford

In the early morning of 13th June 1944, one week after the D-Day landings, the Nazi regime
launched a new weapon at London. The first Vergeltungswaffe 1 (Vengeance Weapon 1, or
V-1), hit a railway bridge in Mile End, killing six people and leaving 200 people homeless.
Over the following 9 months, more than 2,300 ‘flying bombs’ fell on London, killing an
estimated 5,500 people. The V-1 was the first ever operational cruise missile, capable of
delivering an 850kg warhead from a range of 250km at any time, day or night. It was
certainly a formidable weapon of terror, but if it could be accurately targeted at munitions
factories it would also become a strategic threat to the war effort. As the campaign
progressed, it was clear that the V-1s sometimes fell in clusters: was this the result of
random chance, or of precision guidance? The answer to this question was of critical
importance. Humans are good at seeing patterns, but statistical analysis can help us decide
whether those patterns are due to chance. Were the V-1s really targeted in clusters, or were
they falling at random? Answering this question means going on a journey through bomb
maps, statistics textbooks, and even a novel, following in the footsteps of a British
actuary.The authors recreated R.D. Clarke's classic textbook example of the Poisson
Distribution, using V-1 impacts, 75 years after Clarke did and found similar results. This was
through one author (Liam) meticulously adding all the impacts to a Google Maps layer from
bomb damage maps republished in 2015. They also performed some extra assessment of
randomness using bootstrapping, and found a less significant p-value when using all data
available; with North West London less-heavily bombed. Finally, they assessed V-2 impacts,
and looked into the history of Thomas Pynchon's use of the Poisson Distribution within his
novel "Gravity's Rainbow".

166
7.4 RSS Prize Winners: Statistical Excellence Award for Early Career Writing 2019
Wednesday 4 September 4.10pm

Polarisation and the popular vote in Switzerland

Maximilian Aigner
Université de Lausanne

The recent use of referendums in European representative democracies (Hungary 2016, UK

2016, Catalonia 2017) has highlighted fears regarding populism, but also led to greater
interest in direct democracy. In general, popular votes are often uncontroversial, but those
that generate significant attention and polarise voters often show 1) high participation and 2)
a close outcome (small margin). To highlight such votes, we combine both characteristics
into a polarisation score which is related to the Gini inequality index.Using data from
Switzerland, a country with direct-democratic features, we apply this score to the last
hundred years’ worth of referendums, highlighting the most polarising votes of each decade.
Summarising values by their maximum leads to a well-developed mathematical theory in
many situations, and we use results from Extreme Value Theory to study the probability and
time to recurrence of a Brexit-sized event in Switzerland.The results indicate that events
equally or more polarising than the 2016 UK referendum are quite frequent. However,
‘populist’ uses of referendums remain a common feature even in a country with a long
tradition of direct democracy.

167
7.4 RSS Prize Winners: Statistical Excellence Award for Early Career Writing 2019
Wednesday 4 September 4.10pm

A story about a tiny bot

Marco Antonio Andrade Barrera

National Autonomous University of Mexico

This work’s aim is to show a simple example of a working artificial intelligence application, a
bot which can emulate a very specialized human thinking. It is created using Object Oriented
Programming and a popular machine learning approach. All design steps are described,
starting with basic elements like the bot’s name and birth, following the data gathering and
the machine learning algorithms, up to reach with a bot able to classify stock chart patterns
with high accuracy and speed. The work ends pointing that currently, AI applications can
provide us tools or abilities to do very complex processes or at least avoid time- consuming
processes that are out of the human capabilities. In this respect, AI is improving human
possibilities.

168
7.5 Contributed - Methods & Theory: Survival Analysis I
Wednesday 4 September 4.10pm

Goodness-of-fit tests for the cure rate in a mixture cure model

Ursula U. Mueller1, Ingrid Van Keilegom 2

1
Texas A&M University, 2 KU Leuven

We consider models for time-to-event data that allow that an event, e.g., a relapse of a
disease, never occurs for a certain percentage p of the population, called the cure rate. We
suppose that these data are subject to random right censoring and we model the data using
a mixture cure model, in which the survival function of the uncured subjects is left
unspecified. The aim is to test whether the cure rate p, as a function of the covariates,
satisfies a certain parametric model. To do so, we propose a test statistic that is inspired by
a goodness-of-fit test for a regression function by Härdle and Mammen. We show that the
statistic is asymptotically normally distributed under the null hypothesis that the model is
correctly specified and under local alternatives. A bootstrap procedure is proposed to
implement the test. The good performance of the approach is confirmed with simulations.
For illustration we apply the test to data on the times between first and second birth.

169
7.5 Contributed - Methods & Theory: Survival Analysis I
Wednesday 4 September 4.10pm

On generalizing Banks’ smoothed bootstrap method for right-censored data

Tahani Coolen-Maturi, Asamh Al Luhayb, Frank P.A. Coolen

Durham University

Bradley Efron invented the bootstrap method in 1979, which has become a popular tool ever
since. In 1981, Efron introduced the bootstrap method for right-censored data, a particular
type of data often encountered in reliability and survival analysis, where for a specific unit or
individual it is only known that the event has not yet taken place at a specific time. In 1988,
Banks introduced a histospline smoothed version of Efron bootstrap for real-valued data.
The proposed method combines Banks’ bootstrap and the right-censoring A(n) assumption
introduced by Coolen and Yan in 2004. The right-censoring A(n) assumption provides a
partially specified predictive probability distribution for a future observation based on n past
observations. Banks’ bootstrap is restricted to finite support; in this paper, we overcome this
by assuming a distribution tail for the end interval(s). A simulation study has been conducted
to compare the proposed method performance with Efron’s bootstrap for right-censored
data, where coverage probabilities of bootstrap confidence intervals for the three quartiles
are compared. Our study shows that the proposed method performs better compared to
Efron’s method, in particular for small sample sizes.

170
7.6 Contributed - Communicating Statistics:
Wednesday 4 September 4.10pm

Civic Statistics: Big Ideas, Needs and Challenges. Why we need a new subdiscipline

Joachim Engel
Ludwigsburg University of Education

Effective citizen engagement with social issues requires active participation and a broad
range of skills, including the understanding of data and statistics about society and our
natural and social environment. We provide an overview of a subfield we call Civic
Statisticsthat had been explored by a recent strategic partnership of six European
universities under the Erasmus+ program of the EU. Civic Statistics is statistics about
important societal trends and about topics that matter to the social and economic well-being
of citizens. It sits at the crossroads of multiple disciplines including social science, education
and statistics. Hence, a multidisciplinary educational perspective is needed, stepping outside
the comfort zone of traditional statistics instruction.Understanding Civic Statistics is needed
for participation in democratic societies, but involves data that often are open, official,
multivariate in nature, and/or dynamic, which are usually not at the core of regular statistics
instruction. Many statistics classes and educational curricula are not designed to teach
relevant skills and improve learners' statistical literacy, despite the importance of engaging
learners and future citizens with data about social issues and their connections to social
policy. This talk gives an overview to characteristics of Civic Statistics and provides guidance
to specific teaching materials developed by the ProCivicStat project.

171
7.6 Contributed - Communicating Statistics:
Wednesday 4 September 4.10pm

Insightful Analytics – surreptitiously adding thoughtful Statistics to the glamour of

Data Science

Neil Spencer
University of Hertfordshire

Statistics is boring and old fashioned while Data Science is clever and glamourous. Perhaps
not the majority view at an RSS conference but it is what much of the outside world sees.
Degree programmes with titles associated with Data Science have mushroomed. In other
degree programmes, modules with exciting Data Science-related titles are included to attract
students.However, Statistics has not become irrelevant. While Data Science can provide
amazing graphics based on enormous amounts of data, it still relies on Statistics to provide
valid interpretation. While Data Science can build models that give excellent predictions, it
still relies on Statistics to give insight into underlying mechanisms. While Data Science can
classify cases using cunning algorithms, it still relies on Statistics to decide whether these
have meaning.But if Statistics is so important, how do we get people to learn and use it
without realising they are doing so? The same way we get dogs to take tablets – we trick
them, perhaps by crushing them up and mixing with something we know they are going to
gobble up. [N.B. the comparison between students and dogs is an affectionate one – dogs
are fun-loving, friendly and capable of being trained to do amazing things.]In this talk we
examine how a Statistics module can be transformed into a Data Science module in such a
way that none of the important Statistical elements are side-lined. We surreptitiously address
each Data Science topic in such a way that in order to overcome its shortcomings, Statistics
needs to be understood and employed. We will also challenge audience members to
suggest additional Statistical topics that can be hidden within a Data Science cloak. In time,
we believe that the term “Statistics” will be recognised as that which gives real meaning to
Data Science. However, in the meantime we avoid the “S” word and call it “Insightful
Analytics”.

172
7.6 Contributed - Communicating Statistics:
Wednesday 4 September 4.10pm

Principles of statistical visualisation for public policy audiences: learning lessons

from the past

Thomas King, Murray Dick2

-, 2 Newcastle University

Data visualisations that present data literally may be useful within a workflow or to
communicate simple patterns, but they often have more aesthetic than illustrative value.
More complex analysis requires visualisation of statistics derived from data, and many
specialised practical tools exist for different substantive disciplines. However, statistical
complexity abounds within niche policy areas, and so standard tools often do not exist; or
where they do, the policy audience may be unfamiliar. We offer some principles that are
based on theories of visual communication, for statistical visualisations aimed at public
policy audiences, who know established regularities very well. These principles comprise
three factors: The context in which information is communicated (pragmatics) that reduce
cognitive load and contribute to the framing of the message.The content of these messages
(semantics), concerning the statistical inferences designers wish to convey.The mode of
convincing the audience (rhetoric); in this case, the application of visual metaphor towards
embedding policy information (or disrupting prior knowledge or predispositions) amongst the
audience.Features of effective communication are contrasted where others have failed
despite being statistically sound; typically because their approach to visual communication
lacked a sense of familiarity (pragmatics), policy-relevance (semantics), or a sense of
sympathy (rhetoric) with their intended audience. We exhibit our own alternative examples
based on historic data. Before formulating the message (the data visualisation) it is first
necessary to reason through the communicative needs of the audience; in this case, key
stakeholders in policy. This involves a process that contrasts with the broadcast approach of
much data visualisation design. Once a robust inference is established, our approach is to
identify a figurative representation of the data, in order to capture context, content and a
mode of convincing, in order to persuade even audiences with prior knowledge, by means of
visualisations that are sympathetic to their motivations.

173
7.7 Contributed - Data Ethics
Wednesday 4 September 4.10pm

The future of privacy and confidentiality methods

James Tucker, Keith Spicer

Office for National Statistics

A dramatic increase in the volume and sources of data presents an unprecedented

opportunity to innovate. On the other hand, against a backdrop of increasingly sophisticated
attacks by intruders and serious consequences of data breaches, the data revolution
presents new challenges to protecting privacy and confidentiality, and ensuring that the
associated methods are fit for purpose. It is vital that the statistical community understands
and addresses these evolving challenges to provide a solid foundation for innovation and
better decision making to take place. This is not a straightforward task, and there is a need
to draw upon the full range of expertise in this fast-developing field.The Office for National
Statistics joined forces with leading experts in privacy and data confidentiality from across
the world to explore the latest advances in methods, such as synthetic data and differential
privacy. Emerging themes are discussed, including the implications of increased use of data
linking, and the potential of intruder testing techniques. Also outlined are the challenges
faced by statisticians in keeping pace with the latest technological developments, and areas
of future research, such as the practical applications of machine learning and artificial
intelligence.

174
7.7 Contributed - Data Ethics
Wednesday 4 September 4.10pm

The Minority Report: Fairness and Explainability in Machine Learning

Stuart Millar
Centre for Secure Information Technologies, Queen's University Belfast

This talk aims to raise awareness of bias in predictive algorithms and real-world impact. As
opposed to a hardcore technical discussion, we’ll consider machine learning from an ethical
point of view and shift the conversation from our models doing things right to doing the right
thing. Aimed at anyone working in machine learning, deep learning or data science in any
capacity, plus those interested in wider ethics, it features short interesting case studies from
Microsoft, Google, the Met and US police forces. Handling bias and explainability in our
models will be the difference between further breakthroughs in using machine learning,
particularly deep learning, in the real-world, or not. With media coverage in recent times
intimating that predictive algorithms can lead to, for example, racial, gender or sexual
discrimination, there presently is a real risk models will make decisions we can’t explain, and
thereafter the damage will be already done. We'll discuss how we can start to take
responsibility for mitigating bias when engineering our models and algorithms, including what
tangible frameworks might look like to do so, and as researchers where we can publish our
work helping to solve problems in this area. Ultimately, if we are dealing with predictive
algorithms that make decisions about human beings, we must be responsible, be kind and
be good citizens all at the same time. In theory, our models should have the capacity to act
without the bias and discrimination that has plagued society for generations, but it is down to
us to ensure this it the case and prevent this prejudice from becoming embedded in our tech,
deliberately or otherwise.

175
7.7 Contributed - Data Ethics
Wednesday 4 September 4.10pm

Data ethics in a changing and challenging global context

Peter Elias
Warwick University

In 2016 the OECD published a report on ‘Research Ethics and New Forms of Data for Social
and Economic Research’. The report (https://doi.org/10.1787/5jln7vnpxs32-en) compiled
over a two year period by statisticians, researchers, policy makers and experts in research
ethics from 12 countries and the European Commission, put forward 13 recommendations
designed to encourage research and statistical cooperation across international boundaries.
The report was endorsed by the OECD Global Science Forum, a major international body for
promoting scientific collaboration. This presentation provides information about these
recommendations and, crucially, examines what has happened as a result.

176
7.8 Contributed - Business, Industry & Finance: Forecasting currency rates and stock
market indices
Wednesday 4 September 4.10pm

Forecasting Risk Measures Using Intraday Data in a Generalized Autoregressive

Score (GAS) Framework

Xiaohan Xue
ICMA Centre, Henley Business School, University of Reading

A new framework for the joint estimation and forecasting of dynamic Value-at-Risk (VaR)
and Expected Shortfall (ES) is proposed by incorporating intraday information into a
generalized autoregressive score (GAS) model. The GAS model was proposed by Patton,
Ziegel, and Chen (2017) to estimate risk measures in a quantile regression framework. The
intraday measures considered in this study include four realized measures: the realized
variance at 5-min and 10-min sampling frequencies, and the overnight return incorporated
into these two realized variances. In a forecasting study, the set of newly proposed
semiparametric models is applied to 4 international stock market indices: the S&P 500, the
Dow Jones Industrial Average, the NIKKEI 225 and the FTSE 100, and is compared with a
range of parametric, nonparametric and semiparametric models including historical
simulations, GARCH and the original GAS models. VaR and ES forecasts are backtested
individually, and the joint loss function is used for comparisons. The GAS-FZ-Realized
models, especially the GARCH-FZ and GAS-1F extended with the realized variance at the 5-
minute frequency, outperform the benchmark methods consistently across all indices and
various probability levels.

177
7.8 Contributed - Business, Industry & Finance: Forecasting currency rates and stock
market indices
Wednesday 4 September 4.10pm

Dynamic functional time series forecasting of foreign exchange implied volatility

surfaces

Han Lin Shang

Australian National University

Functional time series have become an integral part of both functional data and time series
analysis. Recently, important contributions to methodology, theory and application for the
prediction of future trajectories and the estimation of functional time series parameters have
been made. This paper continues this line of research by considering both static and
dynamic functional principal component analyses to decompose a time series of functions,
and by adapting univariate, multivariate and multilevel functional time series methods to
forecast implied volatility surfaces in foreign exchange markets. On one hand, the
multivariate functional time series method first standardizes all series, then combines them
into a long vector for each time period. Via a functional principal component analysis, the
extracted functional principal components and their associated scores can be obtained from
a joint covariance structure consisting of all series. On the other hand, the multilevel
functional time series method uses multilevel functional principal component analysis of
aggregate and maturity specific data to extract the common trend and maturity-specific
residual trend among different maturities. Both multivariate and multilevel functional time
series methods are able to capture correlation among maturities, and they are shown to be
more accurate than the univariate functional time series for producing short- to medium-term
forecasts. To demonstrate the effectiveness of the proposed methods, we examine the
statistical significance of the three methods in forecasting daily EUR-USD, EUR-GBP and
EUR-JPY implied volatility surfaces for a number of maturity periods.

178
7.8 Contributed - Business, Industry & Finance: Forecasting currency rates and stock
market indices
Wednesday 4 September 4.10pm

Forecasting Cryptocurrencies Volatility with NonGaussian Garch Models

Massimiliano Giacalone1, Raffaele Mattera1, Roy Cerqueti2

1
University of Naples "Federico II" - Department of Economics and Statistics, 2 University of
Macerata

Financial time series usually are characterized by leptokurtic distribution, heavy-tails,

skewness, volatility clustering and heteroscedasticity. Therefore, in this case thenormality
assumption for the innovation’s distribution is violated and the development ofmethods to
handle the stylized facts became an important issue to model financial timeseries. In this
context, an important problem is to define useful and efficient statisticalmethods for
estimating and forecasting volatility.Despite the literature about volatility models for
cryptocurrencies is not well developed yet,cryptocurrencies themselves are one of the most
evident and interesting examples of theparticular characteristic of financial time series. In
particular, the previous literature tried tofind the best GARCH model ignoring the
assumptions about the innovation distribution,even if the characteristic of non-normality and
skewness of the cryptocurrencies returndistribution are well known among researchers.We
provide a comprehensive study about volatility estimation and forecasting, in terms of market
capitalization, studying the dynamics of exchange rate with US Dollar.In order to evaluate
the specification of each model we considered the Akaike InformationCriterion (AIC), that is
one of the most common criteria used in the previous literature whilefrom the point of view of
forecasting accuracy we used the Mean Square Error (MSE) andMean Absolute Error
(MAE).We conclude that, while from the point of view of goodness-of-fitwith the
cryptocurrency data, the best fitting distribution are for all the considered exchangerates the
t-student and the Generalized Error Distribution while from the point of view ofvolatility
forecasts we can get the most accurate predictions implementing a SkewedGeneralized
Error Distribution in estimating GARCH model.

Main References1. Katsiampa, P., 2017, Volatility estimation for Bitcoin: A comparison of GARCH
models,Economics Letters, 158, 3-6.2. Mattera, R. & Giacalone, M. (2018). Alternative distribution
based GARCH models forBitcoin volatility estimation, The Empirical Economics Letters, 17(11), 1283-
1288.

179
7.9 Contributed - Environmental / Spatial Statistics: Health Factors
Wednesday 4 September 4.10pm

Area interaction point processes for bivariate point patterns in a Bayesian context

Glenna Nightingale1, Janine Illian2, Ruth King1

1
University of Edinburgh, 2 University of St Andrews

In this paper we consider bivariate point patterns which may contain both attractive and
inhibitive interactions. The two subpatterns may depend on each other with both intra- and
interspecific interactions possible. We use area interaction point processes for quantifying
both attractive and inhibitive interactions in contrast to pairwise interaction point processes,
typically model regular point patterns. The ability to permit both attraction and repulsion is a
valuable feature and allows for the modelling of different forms of interactions in a given
community. The differentiation between intra- and interspecific interactions in one model
accounts for the fact that the presence of a second species may ``mask" or ``magnify"
existing intraspecific interactions. A Bayesian approach has been applied for estimating
interaction parameters and for discriminating between eight competing research hypotheses.
For the particular application to modelling the interactions of species in a highly biodiverse
forest, this study reveals posterior support for an interspecific interaction of attraction
between the two species considered and may serve to inform forest rehabilitation schemes
relating to this forest. Overall, knowledge of the interactions of key species in any given
forest would be invaluable to reforestation efforts if this forest is later ravaged by wildfires.

180
7.9 Contributed - Environmental / Spatial Statistics: Health Factors
Wednesday 4 September 4.10pm

Simulating disease control strategies for Bovine Tuberculosis (bTB) in Northern

Ireland

Emma Brown, Adele Marshall, Hannah Mitchell, Andrew Byrne

QUB

Bovine Tuberculosis (bTB) is endemic to Northern Ireland’s (NI) cattle population with herd
incidence rates steadily rising from 5.12% in 2010 to 9.61% in 2017 (DAERA, 2018). The
corresponding cost, comprising of testing and compensation for culled cattle, was
approximately £44 million in 2017 alone (NIAO, 2018).There is a need for mathematical
modelling of bTB in NI to inform how it will behave when control strategies are applied. This
research aims to introduce potential control strategies, apply them to an epidemic model
(constructed as part of this project and is the first for bTB in NI), and assess the change in
bTB incidence levels.The epidemic model to be used assumes hosts in the disease system
start as susceptible (S), become latently infected through exposure (E), and finally become
infectious (I). This SEI model assumes homogeneous mixing between hosts with density
dependent transmission to account for increasing herd densities. The SEI model was
extended to allow bTB to spread through cattle movements and for interactions between
cattle and wildlife reservoirs. Control strategies will be investigated where the model’s values
for the transmission, demography, and bTB-detection parameters will be decreased to
represent biosecurity, reducing herd density, and improving the infection detection. Other
perturbations include restricting cattle movements and wildlife reservoir interactions.
Strategies will also be combined to assess if multiple measures will be more effective than
singular tactics.Simulating bTB control strategies reduces the need for resource-intensive
field trials and provides a clear comparison between multiple control measure combinations.
Knowledge from this analysis will inform and optimise resource allocation for a future testing
regime. In doing so, herd incidence rates for bTB will become more manageable with the
long-term goal of achieving the European Union’s Officially-Tuberculosis Free (OTF) status
when incidence levels stay below 0.1% for 3 consecutive years.

181
7.9 Contributed - Environmental / Spatial Statistics: Health Factors
Wednesday 4 September 4.10pm

Investigating the importance of environmental factors to understand renal disease

attributed to uncertain aetiology

Jennifer McKinley1, Ute Mueller2, Peter Atkinson3, Ulrich Ofterdinger1, Siobhan Cox1,
Damian Fogarty4, Chloe Jackson1
1
Queen's University Belfast, 2 Edith Cowan University, 3 Lancaster University, UK, 4 Belfast
Health & Social Care Trust, Belfast, Northern Ireland

There are several factors which are known to cause renal disease including age, ethnicity
and pre-existing medical conditions. However, the UK Renal Registry (UKRR) and other
national systems record a noteworthy number of incidences of renal replacement therapy
(RRT) attributed to uncertain or unknown aetiology. Some have been attributed to
environmental factors. The World Health Organisation divides naturally occurring elements
into three groups on their nutritional significance in humans: 1) essential elements; 2)
elements which are probably essential; and 3) potentially toxic elements (PTEs). For
Northern Ireland, the Standardised Incidence Rates (SIRs) for patients starting RRT
between 2006-2016, as provided by the UKRR, show that the SIRs for uncertain aetiology
are up to 8% higher than would be expected based on Northern Ireland’s age-specific
incidence rates. Our study uses regional geochemical soil and stream water data to
investigate the relationship between the environment and end-stage renal disease. Due to
the constraints of the closed (compositional) nature of geochemical data, log-ratio and log-
contrast approaches are explored. Poisson regression analysis is used to investigate any
potential relationship between observed cases of end-stage renal disease with unknown
aetiology and environmental covariates (soils, stream waters and stream sediments). A
Tweedie model is investigated to account for the zero-inflated nature of the dependent
variables which include Super Output areas with no incidences of uncertain aetiology. The
results indicate that elemental associations are essential to appreciating the potential role of
environmental factors in the development of renal disease.

182
8.1 Contributed - Medical: Treatments and Interventions
Thursday 5 September 9am

Optimum weighting schemes when performing Matching-Adjusted Indirect

Comparisons

Dan Jackson, Kirsty Rhodes, Mario Ouwens

AstraZeneca

Indirect comparisons and network meta-analyses are standard methods to compare

treatments from multiple trials. A common situation is where a company has individual level
data on its own trial but only has published aggregate level data on a competitor’s trial. In
this situation, Matching Adjusted Indirect Comparison (MAIC) is an established way to adjust
for between-trial population imbalances (different distributions of observed patient
covariates). This methodology uses a method of moments, that does not require individual
patient data in the competitor trial, to estimate the propensity score using logistic regression.
Patients in the trial for which individual patient data are available are then weighted so that a
population adjusted indirect comparison can be performed. The choice of logistic regression
is somewhat arbitrary in this modelling, and it is natural to consider alternatives. We
developed a novel approach where we used the method of moments in conjunction with the
propensity score regression model that results in the largest possible effective sample size.
Analytical results are obtained for the simplest possible case where there is one matching
covariate, but more generally numerical methods must be used to perform the necessary
optimisation. We investigated the properties of our proposed method by applying it to a well-
known (but artificial) example from NICE Technical Support Document 18, and we also
performed a simulation study based on this example. Worthwhile gains in effective sample
size can be obtained in situations where considerable adjustment is required. We conclude
that alternative weighting approaches are worthy of consideration when performing a MAIC,
and our “optimal” weighting scheme should be considered for this purpose.

183
8.1 Contributed - Medical: Treatments and Interventions
Thursday 5 September 9am

Efficient Computation for Evaluation and Comparison of Phase I Oncology Study

Designs

Jun Takeda
Astellas Pharma Inc.

One of the main purposes for a phase I oncology trial is to determine the maximum tolerated
dose (MTD), which is defined as the maximum dose with acceptable dose limiting toxicity
(DLT). Typically, such a study includes a sequence of cohorts with pre-specified rules to
select the next dose (or to stop the study) and determine MTD. There are a variety of
oncology phase I study designs. Famous among them are the three plus three (3+3) design,
Bayesian continuous reassessment method (CRM) and Bayesian optimal interval (BOIN)
design. Because the selection of study design and the tuning of parameters are crucial to
planning a study, there is a need for methodologies to evaluate and to compare among
different designs and different settings of parameters.A common practice to evaluate a study
plan can be summarized as follows: Firstly, several scenarios of the dose-response are
assumed. Then data are simulated based on the study plan and whether a subject has a
DLT or not is simulated with a random sample from the binomial distribution corresponding
to the assumed dose-response scenario. Finally, simulations are summarized, e.g., with
operational characteristics. One of the biggest issue in this procedure is its computational
burden, especially when Markov chain Monte Carlo (MCMC) or numerical integration is
performed in a Bayesian study design. Researchers also encounter this issue when
comparing their newly proposed designs with already-existing ones.To mitigate such burden
and to gain computational precision, we propose an improved procedure. In the proposed
procedure, the binomial random sampling part is replaced with an exact probability
calculation, possible paths of stochastic process are identified, and the number of performing
MCMC is reduced with the likelihood principle. With the improvement of computational time,
we also propose random generation of dose-response relationship in order to make an
exhaustive comparison among study designs.

184
8.1 Contributed - Medical: Treatments and Interventions
Thursday 5 September 9am

Graphical representation and comparison of attributable fractions across multiple

disease risk factors

John Ferguson
NUI Galway

Population Attributable fractions (PAF) estimate the proportion of disease cases that might
be avoided if a disease-risk factor could be removed at a population level; for instance, the
reduction in the prevalence of lung cancer if nobody smoked. They are useful
epidemiological tools to compare disease burden across differing risk factors and help
suggest possible targets for health-interventions. In this talk, I describe appropriate
definitions and estimation approaches for PAF using the Neyman-Rubin causal model.
Several novel visualizations are suggested, displaying various forms of PAF across multiple
different risk factors (both discrete and continuous). It is hoped that these visualizations will
help prevent misinterpretation of PAF, and may assist practitioners in their search of good
risk-factor targets for health interventions.

185
8.2 Contributed - Official & Public Policy:Capturing inequality
Thursday 5 September 9am

Statistics to Ensure Welfare for Ethnic Minorities

Kevin Johansen

In 2008, the ministry with responsible for Sami affairs decided to establish an Expert Group
to edit and publish statistics on Sami issues.Before that, there were hardly any statistics
available to developing policies on Sami affairs to advance Sami culture, language and
welfare programs initiated by the ministry or the Sami parliament.The aim of our project is to
contribute to in-depth knowledge on Sami society through numbers and commentary. This
will help policy and law makers gain a broader knowledge for making qualified decisions on
policy issues affecting especially the Sami people.Sami people live in Norway, Sweden,
Finland and Russia and are designated by the UN as indigenous people. Sami people
altogether number count 60 000, although the numbers aren´t definitive and vary. Because
of this, statisticians haven´t had a big interest or knowledge in developing statistics on Sami
issues.Through this project, the Expert Group has written and edited a vast number of
articles on Sami affairs focusing on several important topics such as education for Samis,
Sami health services, gender perspectives, Sami languages, Sami culture, reindeer herding,
fisheries and so on.Our user surveys indicate that the publications are a crucial tool for
policy makers, bureacrats, researchers and the media in getting access to Sami data
sources.Indigenous people have the same need for Data and Statistics as everyone else.

186
8.2 Contributed - Official & Public Policy:Capturing inequality
Thursday 5 September 9am

Improving Public Confidence in Gender Pay Gap Reporting - An Overview of the

RSS's Recommendations

Nigel Marriott
Marriott Statistical Consulting Ltd

Gender pay gap data is now part of the public discourse on the employment and
remuneration of women in the UK workforce. The data has the potential to highlight good
and bad practice among employers and could be transformational in reducing the gender
pay gap. However, the first two years of published data has uncovered a number of issues
in the way the data is collected, calculated and uploaded by the 10,000+ organisations who
are required to submit such data. To avoid such issues undermining public confidence in
this data, the RSS has published 10 recommendations to improve the situation. These
recommendations cover 4 themes:-1. Making the published figures more consistent and
easier to understand2. Increase the accuracy of the data and reduce chances of erroneous
data3. Protecting the integrity of the data and reduce opportunities for gaming4. Keeping the
reporting burden on employers as light as possibleThis talk will explain why these 10
recommendations have been and the outcomes that the RSS hope to see if these are
implemented. The RSS has drawn heavily on the work done by Nigel Marriott who first
posted his thoughts in this article https://marriott-stats.com/nigels-blog/gender-pay-gap-data-
and-12-ways-to-improve-it/ . It is hoped that statisticians will support these
recommendations and publicise them widely to HR professionals, journalists and policy
makers.PS - As of April 4th, the RSS recommendations have not been published. The
expected publication date is late April.

187
8.2 Contributed - Official & Public Policy:Capturing inequality
Thursday 5 September 9am

Using tax data to better capture top earners in household income inequality statistics

Martin Shine1, Peter Matejic2, Dominic Webber1, Richard Tonkin1

1
Office for National Statistics, 2 Department for Work and Pensions

The Office for National Statistics (ONS) annually releases “Effects of Taxes and Benefits”
(ETB) which examines household disposable income across the UK based on the Living
Costs and Food Survey. An issue with survey data is that top earners are both
simultaneously under represented, and they under-report their income (Burkhauser et al.
2018). Inaccurate recording of top earners will inevitably result in a misleading picture of
inequality, and household income across the UK. As a consequence, ONS, in collaboration
with the Department for Work and Pensions (DWP), have explored methods to address
under-reporting and under-coverage in ETB. In developing the adjustment, it was decided to
build upon the world-leading “SPI adjustment” developed by DWP for the “Households
Below Average Income” (HBAI) series. The SPI adjustment supplements survey data with
administrative tax information contained within the tax data set “Survey of Personal Incomes”
(SPI) from Her Majesty’s Revenue and Customs (HMRC). The current SPI adjustment used
by DWP adjusts the top earning pensioners and non-pensioners separately. The average
incomes of individuals above a threshold in SPI is calculated, and this income is assigned to
the relevant rich individuals, replacing their survey-data based income.ONS adjustments
explore the effects of adapting the SPI methodology to vary the threshold above which
incomes in ETB are replaced and explores the use of multiple income bands above the
threshold, assigning varying incomes from the SPI data to the “rich” individuals rather than
one income for all of them. ONS recently released "Using tax data to better capture top
earners in household income inequality statistics", exploring the effects of applying this new
method of adjustment. This will be followed by a joint paper later in the year between ONS
and DWP, going into more detail on the effect of the adjustment. The adjustment method will
be developed to be used in headline ETB statistics from 2018/19.

188
8.3 Contributed - Applications of Statistics: Applications on animals and plants
Thursday 5 September 9am

A Bayesian analysis of animal movement data

Colin Gillespie
Newcastle University

Many of us have been struck by the inherent beauty of animals moving collectively. Starlings
gathering at dusk in huge numbers to perform the most mesmerising of ballets, the entire
flock moving as if some fluid object. Fish forming tight milling structures in defence against
predation. Though we understand the evolutionary benefits offered to individuals by group
behaviour, little is understood about how these structures are formed.Much work has been
invested in developing theoretical agent-based models which seek to explain emergent
behaviour by interactions at an individual level. However, these models have largely only
been verified with a comparison to empirical observation at a qualitative level, and a
thorough quantitative comparison between field data and theory has been lacking.This study
is based on observations of sheep herds. Using a DJI Phantom 3 drone equipped with a
video camera, we filmed sheep movement. The locations and movement of the sheep were
then extracted using computer vision methods. This data was then analysed using a
mechanistic flocking model - a type of Vicsek model. The model was fitted using a
hierarchical structure with a Bayesian framework.

189
8.3 Contributed - Applications of Statistics: Applications on animals and plants
Thursday 5 September 9am

An application of covariate-based constrained randomisation in livestock research

Andrew Mead1, Amy Thomas1, Jess Evans1, Bruce Griffith1, Mark Eisler2
1
Rothamsted Research, 2 University of Bristol

The North Wyke Farm Platform (a BBSRC National Capability, BBS/E/C/000J0100) is a

unique research facility for the study of grassland livestock systems. It consists of 15
hydrologically-isolated catchments, grouped into three 21ha farmlets which are managed
under different grassland farming systems. The platform is highly instrumented, providing
high-quality data on characteristics of water, air, soil, plants and animals, much at a high
temporal resolution.Herds of beef cattle and sheep are a key resource, used both to assess
the relative performance of the different farming systems and explore aspects of animal
management. As much research considers individual animals as the experimental unit, with
replication farming system treatments only replicated between years, a balanced allocation
of available animals to different treatments is essential.Covariate-based constrained
randomisation is a common design approach in cluster or group randomised clinical trials,
where cluster- or group-based covariates are used to allocate units to treatments. The
standard approach considers all possible permutations of units to (usually 2) treatments,
assessing differences in the covariate values of units allocated to each treatment, and
selecting at random from those permutations meeting a specified level of balance between
treatments.We describe the application of covariate-based constrained randomisation to the
allocation of 90 beef cattle across the three farmlets, accounting for variation in 3 continuous
covariates (age, weight, growth rate) and balancing the allocation for 5 breed/gender
combinations. Further complications included the late identification of 7 of the animals,
allowing for differences between sires, and interest in splitting each herd of 30 animals into 2
comparable sub-herds. We assess how the standard approach can be successfully adapted
to cope with more than 2 treatments, very large numbers of permutations, and these
additional complications.

190
8.3 Contributed - Applications of Statistics: Applications on animals and plants
Thursday 5 September 9am

Predicting Temporal Mutant-Treated Arabidopsis Thaliana Gene Expressions

Susana Conde, Daphne Ezer

The University of Warwick

Plants respond to their environment in a time-of-day dependent manner, and genetic

perturbations may influence how plants react to their environments. We are interested in
developing a model that predicts how plants will respond to a combination of genetic and
environmental perturbations. Specifically, our aim is to predict mutant-treated gene
expression over time, given temporal profiles of gene expression of the same genes in other
experimental conditions (mutant-control, wild-type-treated, wild-type-control).Our data set
comes from a ribonucleic acid (RNA)-seq experiment done in Arabidopsis thaliana plants in
which researchers collected gene expression data at 8 time points during a two-hour period
surrounding the dawn. The plants were exposed to two different temperatures. Additionally,
researchers measured gene expression in mutants of some genes known to be related to
light sensing processes.We compare the effectiveness of a linear regression and a function-
on-function regression models. The latter, which allows for the dependence structure of the
longitudinal measures for either response and/or explanatory variables, is estimated with the
boosting algorithm in the R FDboost package (Brockhaus et al 2015). The response variable
is the functional gene expression in the mutant-treated condition, and the explanatory
variables are the functional gene expressions of wild-control, wild-treated and mutant-
control.We fit our models to a set of m = 672 genes found to be early-morning expressed
DNA-binding proteins, after imputing missing values. Our results show that the functional
regression model outperforms the linear regression one with respect to the observed cross-
validated root mean squared error. However, neither approach considers phase variabilities
(i.e. time warping), which will be the phocus of future research.

References: Brockhaus, S., Scheipl, F., Hothorn, T., and Greven, S. (2015). The functional linear
array model. Statistical Modelling, pages 279–300.

191
8.4 Contributed - Medical: Data Science II
Thursday 5 September 9am

UNCERTAINTY IN DIAGNOSTIC TESTING WITHOUT GOLD STANDARDS

Nicholas Gray, Marco De Angelis, Scott Ferson, Ryan Jackson, Uchenna Oparaji, , , , ,
University of Liverpool, Institute for Risk and Uncertainty

Problems with classification appear in various fields, from structural health in engineering,
supervised learning in computer science, and patient diagnosis in medicine. Diagnostic tests
are often imperfect, yielding false positives and false negatives. Bayesian methods are often
used to estimate the probability that a patient has a condition, given the results of a medical
test, the statistical characteristics of the test and the prevalence of a disease. However,
these methods cannot be used without assuming that there is a reliable way to determine
the true classification. When designing diagnostic tests or classification algorithms, analysts
often refer to the ‘gold standard’ of evidence to calculate the statistical characteristics of the
new test. However, these gold standards can themselves be imperfect and have some
errors associated with them. For instance, for patients suspected of having Giant Cell
Arteritis there are several tests that can be conducted in order to detect the disease.
However, there is uncertainty in all these methods, including the biopsy which is called the
gold standard by some. We present a series of logical rules that should hold when
considering methods that aim to account for the error in diagnoses based upon an imperfect
gold standard. These rules logically define limits where the uncertainty is at its maximum and
minimum level. Ensuring that the uncertainty is never less than the uncertainty of the gold
standard, and the uncertainty of the result is irreducible when a gold standard is missing. We
will discuss a method to account for the uncertainty using imprecise probabilities. This
method makes use of the gold standard test sensitivity and specificity to update these
statistics for the new test. We use Bayesian methods to calculate the uncertainty associated
with the probability that someone has a disease when given a positive test result.

192
8.4 Contributed - Medical: Data Science II
Thursday 5 September 9am

Spatial statistical modelling of retinal images from patients with diabetic macular
oedema

Wenyue Zhu1, Jae Yee Ku1, Yalin Zheng1, Paul Knox1, Simon Harding1, Ruwanthi
Kolamunnage-Dona1, Gabriela Czanner2
1
University of Liverpool, 2 Liverpool John Moores University

Images are an essential aspect in monitoring and predicting diseases, especially in clinical
management of retinal diseases. Spatial context in retinal images is highly relevant but often
under-utilised. Current approaches involving analysing each location separately or analysing
all locations together, which ignore the possible spatial correlations.To investigate the
appropriate statistical approaches to account for the suitability of spatial error specification.
We proposed a statistical inference framework for retinal images, which is based on a linear
mixed effect model with a spatial error structure. The spatial topography was explained via
fixed effect and spatial error structures, and correlation between eyes from the same patient
was explained through nested random effects. We compared our method with multivariate
analysis of variance (MANOVA) in analysis of spatial retinal thickness data from a
prospective observational study, the Early Detection of Diabetic Macular Oedema (EDDMO)
study involving 149 diabetic participants at their baseline visit. MANOVA analysis suggested
that the overall retinal thickness of eyes with maculopathy are not significantly different from
eyes with no maculopathy (p=0.11), while our spatial framework can detect the difference
between the two groups (p=0.02).Simulation was conducted to evaluate our spatial statistical
model framework, which illustrate how spatial correlations can affect the inferences about
fixed effects. The simulation results confirmed the increased power of the statistical
inference, which demonstrate the advantage of spatial modelling to provide more powerful
statistical inference with power increase from 88.1% to 95.3%, 88.9% to 100% for moderate
or high noise correlations.Our spatial approach addresses the need of correct adjustment for
spatial correlations in ophthalmic image, and it can be extended into prognostic or predictive
modelling in other diseases or imaging technologies.

193
8.4 Contributed - Medical: Data Science II
Thursday 5 September 9am

Predicting severe complications after cardiac surgery: methods and challenges

Linda Lapp1, Matt-Mouley Bouamrane1, Kimberley Kavanagh1, Stefan Schraag2

1
University of Strathclyde, 2 Golden Jubilee National Hospital

With the aging population patients undergoing cardiac surgery are more complicated to
manage than ever before. Up to 50% of heart surgery patients experience postoperative
complications, which can have a detrimental impact on patients’ quality of life and healthcare
resources.In a previous study we compared various machine learning methods (random
forest, AdaBoost, Gradient Boosting Model and stacking) at predicting severe postoperative
complications based on variables recorded before surgery. Our results showed that
AdaBoost has the best overall performance (AUC=0.731). To improve the previously
developed prediction model, we are looking into including cardiac intensive care data into
our model.The data for this study comes from Golden Jubilee National Hospital, Scotland,
and includes 6834 patient records. Two clinical audit datasets are used: one including data
recorded at the preoperative clinic, such as patient characteristics, comorbidities and
information about surgery; another including patients’ vital signs and laboratory results
recorded at cardiac intensive care unit. The time-series data about vital signs are recorded
every other minute, and laboratory results are received at least twice a day.Although the
number of patients with severe complications is considerably low in this dataset (5.50%), the
size of intensive care unit data is large: the median number of days stayed in the ICU for
these patients was 5.5 days, as opposed to 22 hours for patients with non-severe
complications. We will discuss various methods that are appropriate for our classification
problem and the challenges that arise from analysing such vast amount of real-world health
data.This work is part of a larger study developing a clinical decision support tool predicting
severe postoperative complications in cardiac patients. Such support system could help a
clinician to identify patients who are at risk of having complications in order to allocate
resources or avoid high-risk treatments.

194
8.5 Contributed - Methods & Theory: Survival Analysis II
Thursday 5 September 9am

Fast Bayesian hazard regression under general censoring via monotone p-splines

Matthias Kaeding
RWI - Leibniz Institute for Economic Research

The baseline hazard is the major building block of the Cox model, giving the instantaneous
rate of failure, conditional on survival up to time t and covariate values of zero. Most
Bayesian nonparametric approaches model the log-baseline hazard, causing the need for
numerical integration for likelihood evaluation. We propose to model the integrated baseline
hazard of the Cox model via monotone penalized B-splines instead; giving an analytically
available likelihood, speeding up inference and eliminating approximation error. Left, right
and interval censoring can be accounted for. The advantages of Bayesian P-splines carry
over to the monotone case: (1) Fully automatic smoothness parameter estimation, (2)
sparseness of involved design matrices. The closed form expression for the cumulative
baseline hazard is used to extend inference beyond marginal effects on the hazard rate;
allowing fast computation of the conditional mean and survival function and the simulation of
random deviates. Inference is carried out using MCMC and posterior mode estimation. The
proposed approach is tested using a Monte Carlo simulation and applied on a large data set
of times until change of gas price.

195
8.5 Contributed - Methods & Theory: Survival Analysis II
Thursday 5 September 9am

Empirical likelihood comparison of t-year absolute risks

Paul Blanche
Department of Biostatistics, University of Copenhagen

In the competing risks setting, which is common in medical research, a key quantity is the t-
year absolute risk. It is also often called the cumulative incidence function at time t. In
oncology and cardiology, is often estimated with the non-parametric Aalen-Johansen
estimator. This estimator handles right-censored data and has desirable large sample
properties. Inference for comparing two absolute risks, via a risk difference or a risk ratio,
can be done via usual asymptotic normal approximations and the delta-method. However,
small sample performance of this approach can be modest. Especially (i) coverage of
confidence intervals can be poor and (ii) inference using risk ratios and risk difference can
lead to contradictory conclusions, in terms of significant differences. Therefore, we suggest
to use empirical likelihood ratio inference as an alternative. This method can be seen as an
extension of that of Thomas and Grunkemeier (J Am Stat Assoc, 1975) to the competing
risks setting. Interestingly, this approach systematically leads to consistent conclusions when
comparing absolute risks via either risk ratios or risk differences, in terms of significance.
Simulation results also suggest that small sample inference using this approach can be more
accurate. We detail how to compute confidence intervals and p-values using empirical
likelihood ratios and provide a software implementation. Novel technical results include
formulas and algorithms to compute constrained non-parametric maximum likelihood
estimates, from which likelihood ratios and inference procedures are derived. An analysis of
clinical data is presented to illustrate the main ideas.

196
8.5 Contributed - Methods & Theory: Survival Analysis II
Thursday 5 September 9am

Regularised survival analysis of lung cancer patient outcome using radiotherapy dose
data

Elizabeth Buckingham-Jeffery
University of Manchester & Highways England

In 2015, over 46,000 new lung cancer cases were identified in the UK. This led to over
35,000 deaths from lung cancer in 2016, making it the UK's most common cause of cancer
death. Radiotherapy can be used to destroy cancerous tissues by delivering ionising
radiation to cancer cells. The majority of lung-cancer patients treated with curative intent are
given radiotherapy and advanced imaging and computing technologies are used to design
and deliver these treatment plans. This leads to high-granularity radiotherapy dosing
data.Radiation toxicity, for example as a result of radiotherapy damaging healthy tissue, is
estimated to be associated with a significant proportion of excess mortality at one year post
treatment. We have therefore used survival analysis methods and voxel-level radiotherapy
dose data to investigate how radiation delivered to normal lung tissue affects survival in lung
cancer patients treated with radiotherapy.We found that due to the high dimensionality of the
dosing information it is necessary to use statistical regularisation methods, in particular a
fused Bayesian elastic net. Fitting this model to the data allowed us to identify a smooth
dose-effect curve. The fitted model was then used to predict how survival could change with
(hypothetical) improvements in the delivery of radiotherapy.

This work was completed with Thomas House (University of Manchester), Alan McWilliam
(University of Manchester and The Christie NHS Foundation Trust), and Marcel van Herk
(University of Manchester and The Christie NHS Foundation Trust).

197
8.6 Contributed - Methods & Theory: Time Series
Thursday 5 September 9am

Nonparametric clustering for spatio-temporal datasets

Ashwini Venkatasubramaniam1, Konstantinos Ampountolas2, Ludger Evers2

1
The Alan Turing Institute, 2 University of Glasgow

This non-parametric Bayesian approach to clustering for spatio-temporal datasets seeks to

identify spatially contiguous clusters that represent distinct temporal patterns. This Bayesian
method uses a modified non-sequential distance dependent Chinese restaurant process
(ddCRP) to accommodate challenges posed by spatial and network connectivity structures
in the graph. The modified ddCRP is also re-defined to enable the number of clusters in the
network to be controlled. In addition, a spatio-temporal precision matrix is defined to fully
account for within cluster spatial dependencies and the associated temporal patterns. The
clustering method uses a Metropolis within Gibbs sampler to explore all possible clustering
configurations, as composed of cycles and paths in the network, and infer the relevant model
parameters. This novel approach to clustering is able to model multiple dependency
structures and determine the number of clusters in a data-driven and computationally
efficient manner. This unique method is illustrated by applications to spatio-temporal
datasets recorded over a grid-style graph and a map-based network.

198
8.6 Contributed - Methods & Theory: Time Series
Thursday 5 September 9am

Model averaging of integer-valued autoregressive model with covariates

Jiajing Sun, Yuying Sun, Xinyu Zhang

University of Chinese Academy of Sciences, Academy of Mathematics and Systems Science

This work addresses the model averaging of the integer-valued autoregressive (INAR)
process with exogenous covariates. We extended the work of Hansen and Racine (2012)
and Gao et al. (2016), and developed leave-subjects-out cross-validation (LsoCV) based
model averaging estimator for autocorrelated count time series. We prove that the resulting
LsoCV estimator is asymptotically optimal by a criterion equivalent to that used byHansen
and Racine (2012). Moreover, the LsoCV estimators are consistent, provided at least one
candidate model is not under-fitted. Monte Carlo simulations show that the LsoCV estimators
can achieve significant eciency gains over existing model averaging methods.

199
8.6 Contributed - Methods & Theory: Time Series
Thursday 5 September 9am

Estimation of the effect lag of predictors in modelling temporal datasets with

application the effect of lifestyle and air quality on health outcomes in Leeds

Jeanine Houwing-Duistermaat, Haiyan Liu, George Aivaliotis

University of Leeds

The amount of temporal datasets containing relevant information to answer important

questions is increasing rapidly. Modelling the relationship between temporal datasets is
challenging. Current available models are limited. For example often there is a lag in the
effect of predictor on the outcome, e.g. change of lifestyle might not immediately affect your
health. Also the effect of historical values of predictors may diminish over time. The available
model for this type of question can only deal with one temporal predictor. We propose a
functional linear model to predict a temporal response using multiple functional and
longitudinal predictors including effect lags of these predictors. Specifically the regression
functions in our model are written as the expansion of a basis system (e.g. functional
principal components, splines). For a given set of time lags, the coefficients of these fixed
basis functions are estimated via optimizing a penalization criterion. The time lags are
estimated by simultaneously searching on a prior grid mesh based on minimization of the
prediction error. We studied the mathematical properties of the estimated parameters and
the accuracy of predicting outcomes based on our model via an extensive simulation study.
Finally we applied the methods to data on health (e.g. BMI, COPD), lifestyle (physical
exercises) and air quality (e.g. NO2) available at postcode level for the City of Leeds.Our
conclusion is that our method performs well and provides insights on the relationships
between lifestyle, pollution and health in the City of Leeds. The simulation study showed that
the asymptotic properties hold, that the parameter functions are well estimated and that in
70% of the replicates our method identified the correct time lags.

200
8.7 Contributed - Medical: Estimation and Performance
Thursday 5 September 9am

Estimation of proportions by group retesting

Stephen Walter, Graham Hepworth

McMaster University

To estimate the proportion p of an attribute in a population, sometimes the sampled

individuals can be tested in groups. This can yield substantial cost savings over individual
testing, especially if p is small. If one desires to increase precision, but it is impractical to
sample additional individuals, it may nevertheless be possible to retest groups formed from
the individuals within the groups that have tested positive at the first stage. Hepworth and
Watson (2017) have described several situations where obtaining additional individuals is
indeed impractical or prohibitively expensive. If the initial testing is non-destructive, a re-
testing approach may then prove useful. Potential applications include: transmission of
diseases by insect vectors; plant or animal disease assessment; prevalence of viruses in
mosquito populations; and human blood testing, especially for HIV prevalence. Hepworth
and Watson developed an estimator of p for their recommended retesting method, which
involves a random regrouping of individuals in the positive groups from the first stage, but
because of its analytic complexity, required simulation to examine its variance properties.
We have now developed two closed-form analytical expressions for the variance of the
second stage estimator, and we compare their performance with the simulated results.
Estimates of p from the two stages can be combined for a satisfactory overall estimate. We
show that our solutions give acceptable approximations in a reasonable range of
circumstances.

References: Hepworth, G. and Watson, R.K. (2017) Revisiting retesting in the estimation of
proportions by group testing. Communications in Statistics – Simulation and Computation, 46, 261–
274. Walter, S.D., Hildreth, S.W. and Beaty, B.J. (1980) Estimation of infection rates in populations of
organisms using pools of variable size. American Journal of Epidemiology, 112, 124–128.

201
8.7 Contributed - Medical: Estimation and Performance
Thursday 5 September 9am

Survival analysis of cancer patient based on their Genome profile

Khaled Mubarek
Prince Sattam bin Abdulaziz University

Non-small-cell lung cancer (NSCLC) is one of the main sources of death in industrialized
nations with an expanding rate around the world. As a result, scientists are now looking for
some of the risk factors for lung cancer which can be caused by certain changes in the DNA
of lung cells. One way to detect these changes is the copy number alteration (CNA); which is
a type of structural variation in the genome . It usually refers to the duplication or deletion of
DNA segments larger than 1 kbp. Like other types of genetic variation, some CNAs have
been associated with susceptibility or resistance to disease. As a result, CNA can be used to
predict the survival of cancer patients. Next-generation sequencing (NGS) technologies
produce high-dimensional data that allow a nearly complete evaluation of genetic variation.
With the advent of high-dimensional datasets, the following problem has been faced: the
number of covariates (in our study 13968) greatly exceeds the number of observations (85).
The results of our analysis indicate that we can incorporate the copy number alteration
profile to predict the survival time. We investigate a Cox proportional hazards model within a
random effects model frame-work using penalized partial likelihood to model the survival
time based on lung cancer patients' clinical characteristics as fixed effects and CNA profiles
as random effects. We use AIC to estimate $\sigma$ which parameterizes the covariance
variance matrix of the random effect. For the fixed effects the model indicates that age,
stageT3, and stageN2 are statistically significant. Finally, comparing the Kaplan-Meier
survival curves with model-based average survival function indicates that the model
estimation works reasonably well. Also we covered methods for checking the adequacy of a
fitted Cox model.

202
8.7 Contributed - Medical: Estimation and Performance
Thursday 5 September 9am

Assessing performance of survival risk prediction models: A review of traditional and

modern methods

David McLernon1, Ben Van Calster2, Laure Wynants3, Maarten van Smeden2, Ewout
Steyerberg2
1
University of Aberdeen, 2 KU Leuven, Leiden University Medical Center, 3 KU Leuven,
Maastricht University

Objective: To review methods to evaluate the performance of survival risk prediction models.

Methods & Results: Traditional measures for assessing performance in risk prediction
models include the Brier score to indicate overall model performance, the c-statistic for
discriminative ability, and calibration models. These methods have been well defined and
can readily be applied for the binary outcome setting. Censoring complicates the extension
of these methods to survival models, where we may assess performance at specific time
points or for the full survival distribution. The Brier score can be extended to the survival
setting. The popular extension of the c-statistic for censored data by Harrell discounts pairs
that cannot be ordered. However, this approach ignores the study specific censoring
distribution on which the population parameters may depend. Uno’s C-statistic is a more
recent method that models the censoring distribution and uses it to weight the uncensored
observations. Calibration-in-the-large is the difference between the mean observed risk and
the mean predicted risk, which can be estimated with a model-based approach using the
Poisson model. This approach also enables estimation of the calibration slope, i.e. the
overall effect of predictors for survival. A relatively new development is to assess the
potential impact on making better decisions with decision theoretic measures such as net
benefit. This requires the specification of a context dependent decision threshold, or range of
plausible thresholds. For illustration of the above methods, we present a case study with
internal and external validation of a model predicting recurrence in breast cancer patients
following surgery with 2982 women for development and 686 for validation.

Conclusion: Various methods are available to assess overall performance, discrimination,

calibration and net benefit, which can be applied for both internal and external validation
before using a survival prediction model in clinical practice. Assessing a model’s value for
making better decisions is essential requiring methods beyond traditional assessments.

203
8.8 RSS Prize Winners: Best presentations from YSM 2019
Thursday 5 September 9am

Relaxing the constant hazard assumption in a multi-state model in hospital

epidemiology

Micki Hill1, Paul Lambert12, Michael Crowther1

1
University of Leicester, 2 Karolinska Institutet

BACKGROUND: Multi-state models are being increasingly used to capture more complex
disease pathways. Exponential models have been suggested as an accessible approach to
obtain a quick understanding of the data; however, assuming time constant hazards is not
always plausible. A flexible spline-based approach has been proposed, where transition
hazards can be time-dependent and predictions can still easily be obtained through
simulation.

METHODS: Predictions obtained from an exponential model were compared to those

obtained via simulation from a Royston-Parmar model with four degrees of freedom. Metrics
of interest included: transition probabilities, expected length of stay, attributable mortality
(AM) and population attributable fraction (PAF). They (and corresponding confidence
intervals) were easily obtained for the spline-based approach using the multistate package in
Stata [1]. This work was performed on previously analysed hospital acquired infection (HAI)
data [2].

RESULTS: Despite clear deviations from the constant hazards assumption, the empirical
estimates of the transition probabilities were approximated reasonably well by the
exponential model. However, functions of the transition probabilities, e.g. AM and PAF, were
not well approximated and the spline-based approach offered considerable improvements in
these cases.

CONCLUSIONS: We conclude that methods and software are readily available for obtaining
predictions from multi-state models that do not assume constant hazards. The multistate
package in Stata facilitates a range of predictions with confidence intervals via simulation,
which can provide a more comprehensive understanding the data.

REFERENCES:[1] Crowther MJ, Lambert PC. Stat Med. 2017;36(29):4719-4742[2] von Cube M,
Schumacher M, Wolkewitz M. BMC Medical Research Methodology. 2017;17:111

204
8.8 RSS Prize Winners: Best presentations from YSM 2019
Thursday 5 September 9am

Calculating Avoidable Deaths for Cancer Patient Survival

Jamie Stokes1, Mark Rutherford2

1
University of Oxford, 2 University of Leicester

Background: Differences in cancer survival between population groups are well-

documented. The impact of these differences can be determined by estimating the number
of deaths that could potentially be avoided or postponed if these differences were to be
eliminated.

Methods: Established methods of estimating and modelling relative survival were explored
and contrasted in the context of lung cancer patients from the SEER 18 dataset to determine
which approach best complements calculation of avoidable mortality. Life table estimates
were combined with flexible parametric survival modelling techniques to calculate the
number of deaths that could potentially be avoided if survival differences between males and
females were to be eliminated. This was explored further, investigating the extent to which
age and stage at diagnosis contributes to avoidable mortality.

Results: Flexible parametric models incorporating fractional polynomials for the modelling of
age as a continuous variable produced smooth, interpretable and robust estimates of
avoidable mortality. Based on an annual cohort of 15971 male lung cancer patients,
eliminating survival differences between males and females would lead to postponing
approximately 1336 deaths beyond 2 years post diagnosis, representing 11.1% of the
deaths predicted to occur in this time. For the same annual cohort, matching the gender
distribution in stage at diagnosis would postpone around 378 deaths beyond 2 year post
diagnosis, representing 28.3% of the total between-sex avoidable mortality for this period.

Conclusions: Flexible parametric avoidable mortality analysis provides an interpretable

method of measuring progress in the reduction of disease survival deficits to inform public
health policy, with clear applications in early cancer diagnosis.

205
8.8 RSS Prize Winners: Best presentations from YSM 2019
Thursday 5 September 9am

In need of some simulation: the Monte Carlo method for sample size calculations

Nick Beckley-Hoelscher, Fiona Reid

KCL

Background: The design stage is one of the most critical aspects of a clinical trial. A major
role for the statistician at this stage is estimating the number of participants required. Most
widely-used sample size calculations are based on simple mean-based minimal clinically
important differences (MCIDs), and on distributional assumptions for which tractable formula
for the sample size exist. This framework may break down when calculating sample sizes for
non-mean-based outcome measures, complex designs, or situations where there is no
tractable formula. Simple, user-written simulations can provide a flexible and easily
implementable way of performing sample size calculations in a wide array of situations.

Methods: This presentation will go through the steps required to undertake a simulation
sample size calculation, including checks to ensure type 1 error rates are preserved. This
will begin with a simple example for illustrative purposes, before moving on to more complex
examples.

Results: It is shown how simulations allow for a high degree of user-specification to describe
the alternative hypothesis; in particular, describing the MCID in terms of distributional
properties other than the mean difference (e.g. percentile-based changes). The examples
outlined also show how, provided an alternative hypothesis and method of analysis are
specified, the procedure is easily implementable and reproducible.

Conclusion: Simulation-based sample size calculations are a powerful method of calculating

the required number of patients for a trial, which can be used as an additional independent
method for simple designs, or as a flexible method for more complex designs.

206
8.9 Contributed - Business, Industry & Finance: Applying statistics to improve
industry
Thursday 5 September 9am

Predicting the number of resources required to ensure 97% of the time emergencies
are reached in 1 hour.

Laura Thornley
Northern Gas Networks

Northern Gas Networks is a highly regulated business, which means targets are set for the
benefits or Customers as well as to ensure high Health and Safety precautions. One of these
measures is to ensure that Northern Gas Networks responds to a Gas escape within 1 hour
of being informed 97% of the time.Gas escapes vary wildly year to year, as well as having
seasonality and links to temperature, meaning time series analysis is used to help support
prediction. On top of this, extreme events such as heavy snow or large bursts can cause
wide spread events, which require significant number of resources to attend escapes at the
same time. The purpose of this analysis was to understand the number of gas escapes that
can be expected at any given time (fluctuations daily, weekly and yearly) to allow Northern
Gas Networks to plan the number of resources required to attend these escapes.Northern
Gas Networks covers the UK area from Berwick upon Tweed down to Doncaster and from
Todmorden to Scarborough. Therefore it is important to distribute the resources across the
region as it is not possible to travel from on end to another within the 1 hour regulation time
frame. As such, spatial analysis is also considered to allow for resources to be distributed in
the correct areas and to account for variances in travel time between these regions. A small
tolerance is allowed of 3% of escapes to be reached outside of the 1 hour
window.Uncertainty analysis has been conducted previously to understand the level of
resourcing required to meet the 97% targets and as well as the different levels (98%, 99%
and 99.9%). Using these levels of uncertainty allows the business to make informed
decisions around resourcing requirements and demonstrate evidence to an auditor that we
are managing the network adequately.

207
8.9 Contributed - Business, Industry & Finance: Applying statistics to improve
industry
Thursday 5 September 9am

A new link function to analyze the launch of technological products

Gloria Gheno
Innovative data analysis

In the last decades the most important companies of the technological sector announcethe
new products in a spectacular way to intrigue and induce the media to try to predict thenews
coming out. In this work I analyze how current and potential customers, based ontheir
specific characteristics, are interested in the leaked information, in particularconsidering, for
example, the technologic products owned by them and their brand loyalty.Indeed, for a
company knowing the target customers interested in launch of newtechnologies is essential
to be able to develop products which reflect their needs so as tofully meet their expectations.
To study the degree of interest in the advertising campaign, amonth before the launch I ask
a sample of people with which probability it assumes that anew product with particular
characteristics is launched. Then I compare these results withthe product actually put on the
market. I divide the study into two parts, in the first part Ianalyze how the subjects form their
own opinions, in the second part I evaluate thegoodness of the forecast. To study the factors
which affect these probabilities, I use theBeta regression and a link function which I
specifically created. I demonstrate the realgoodness of this new technique by comparing it
with others already present in theliterature using specific statistical tests. This work offers
companies of the technologysector the opportunity to save time and money by eliminating
the costs of development ofunrequired characteristics.

208
8.9 Contributed - Business, Industry & Finance: Applying statistics to improve
industry
Thursday 5 September 9am

Enhancing our understanding of hazard perception in driving

Sritika Chowdhury, Neale Kinnear, Mark Bell

Transport Research Laboratory

Hazard perception is a term used to describe the ability to predict dangerous situations on
the road. It is a critical skill that is directly related to safety outcomes. As there is no perfect
objective measure of danger within a driving scene, hazard perception is based on
subjective appraisal by the driver. IMRA Europe commissioned TRL, the UK’s Transport
Research Laboratory, and NTU, Nottingham Trent University, to utilise their hazard
perception expertise to develop a method to assess danger locations and levels within video
footage. One objective of the study was to establish whether physiological measures (of
non-conscious arousal) can add value to the validation of danger identification and level.
Twenty two drivers with over ten years of driving experience were recruited to view video
footage of day-to-day driving from the driver’s perspective. Forty one-minute clips and 40
dynamic-stills covered four hazard types: no hazard, precursor not leading to a hazard,
precursor leading to a hazard and immediate hazard. Synchronised eye tracking data and
physiological measures such as galvanic skin responses (GSR) were recorded while drivers
viewed the footage. Drivers were asked to identify the most hazard area of specific scenes
and provide a cognitive rating of risk for each video. Finally, drivers were asked to complete
a self-reported driving style questionnaire.Analysis involved clustering techniques such as
hierarchical clustering and k-means using bootstrapping to cluster eye tracking data and
click location; factor analysis of survey data in order to identify specific driving styles in the
sample; and exploratory analysis of GSR. A reverse regression model was applied to
determine if physiological measures can act as a proxy for determining the level of danger
present in a scene and add value to the validation protocol.The data are currently being
analysed and the results will be used to inform the development of future research and
technology to improve road safety.

209
PD8 Professional Development: Pre-plenary Overview Meeting
Thursday 5 September 9am

A very short introduction to hypergraph data

Simon Lunagomez Coria

Lancaster University

Network data has become an important resource for both describing phenomena and
formulating models in the natural sciences and engineering. Networks have been used to
represent systems where units communicate, cooperate or influence one another.
Traditionally, relationships encoded in network data involve at most two units. Hypergraphs
extend this idea by encoding relationships that involve more than two units (coauthorship of
academic papers is a clear example of this). For this talk, we will review different ways to
represent a hypergraph as a data structure. We will also discuss some of the modelling
approaches that rely on the concept of a hypergraph, either as an observation or as a latent
variable.

210
Keynote 6: Bayesian Categorical Matrix Factorization via Double Feature Allocation
Thursday 5 September 10.10am

Bayesian Categorical Matrix Factorization via Double Feature Allocation

Peter Mueller1, Yang Ni2, Yuan Ji3

1
UT Austin, 2 TX A&M, 3 U Chicago

We propose a categorical matrix factorization method to infer latent diseases from electronic
health records data. A latent disease is defined as an unknown cause that induces a set of
common symptoms for a group of patients. The proposed approach is based on a novel
double feature allocation model which simultaneously allocates features to the rows and the
columns of a categorical matrix. Using a Bayesian approach, available prior information on
known diseases greatly improves identifiability of latent diseases. This includes known
diagnoses for patients and known association of diseases with symptoms.For application to
large data sets, as they naturally arise in electronic health records, we develop a divide-and-
conquer MonteCarlo algorithm, which allows inference for the proposed double feature
allocation model, and a wide range of related Bayesian nonparametric mixture models and
random subsets.We validate the proposed approach by simulation studies including mis-
specified models and comparison with sparse latent factor models. In an application to
Chinese electronic health records (EHR) data, we find results that agree with related clinical
and medical knowledge.

Bayesian Double Feature Allocation for Phenotyping with Electronic Health RecordsYang Ni,
Peter Mueller, Yuan Ji https://arxiv.org/abs/1809.08988

Consensus Monte Carlo for Random Subsets using Shared Anchors,Yang Ni, Yuan Ji, and
Peter Mueller https://arxiv.org/abs/[to be posted]

211
9.1 Medical: The Northern Ireland Clinical Trials Unit (NICTU) and the SANDWICH Trial
Thursday 5 September 11.30am

Stepped Wedge Trial Design and analysis methods for the SANDWICH trial

Clíona McDowell
NICTU

Objective: The objective of this talk is to describe the following:Stepped wedge trial design
using the SANDWICH trial as an example.Sample size for the SANDWICH trial.Proposed
methods for analysis.

Methods: The app https://clusterrcts.shinyapps.io/rshinyapp/ developed by Karla Hemming

was used to calculate the sample size. The aim of the study is to evaluate whether there is a
difference in the duration of hours on ventilation before and after exposure to the
intervention. There will be censoring i.e. children moving to other units, children not weaned
before the unit transitions to the training phase, those who are not weaned at the end of the
20-month trial period, children at the time they have a tracheostomy, those not weaned by
90 days, or children who die. We will use survival analysis and estimate a hazard ratio for
the intervention effect. Our survival analysis will estimate the hazard of being extubated and
removed from mechanical ventilation.

Conclusions: Cluster randomisation is essential, as the intervention is delivered at the level

of the cluster (site) as the individual level components would be susceptible to contamination
if individually randomised. The stepped wedge design has been chosen over the
conventional parallel cluster design for the following reasons: there are limited number of
clusters available to allow detection of the important clinical effect at 90% power; units are
more likely to participate in the trial if they are guaranteed their unit will at some point receive
the intervention; it would be infeasible and more costly to deliver the intervention
simultaneously to units randomised to the intervention in a parallel design; and if the
intervention is found to be effective, knowledge translation will be easier as PICUs
participating can potentially continue after the trial, maximising the benefits of any effects to
the NHS and patients.

212
9.4 Social Statistics: Old dogs new tricks: respondent centred approach to mixed
mode social survey redesign at ONS
Thursday 5 September 11.30am

A respondent centred approach - Questionnaire content redesign for a mixed-mode

Labour Market Survey

Vicky Cummings, Alex Nolan

Office for National Statistics

This paper shares insights into how the ONS Social Surveys Transformation Research and
Design (R&D) Team has approached the task of redesigning the Labour Force Survey
(LFS). This work is being undertaken as part of ONS’s Census and Data Collection
Transformation Programme which aims to rebalance ONS's data collection activity toward
wider, more integrated use of administrative and other non-survey data sources, thereby
reducing our reliance on large population and business surveys. The Programme also
focuses on developing the capability to move any residual survey data collection online but
as part of a mixed-mode design The R&D Team are focusing on redesigning LFS content
for online collection and are taking a respondent centered approach to this work. In taking
this approach, the R&D Team have aimed to design a survey which is user-centred and
optimized for the respondent’s chosen mode of completion whilst still meeting the business
need to provide high-quality data. This paper outlines the methods employed; including how
face to face and telephone interviewers have been integral throughout the development
journey. As survey users themselves, the interviewers know where respondents struggle and
consulting them has assisted in ensuring existing issues are addressed.This paper explores
how the R&D Team has blended traditional cognitive interviewing with user research
methods to understand the mental models of respondents around survey topics, survey
completion behaviour and usability of survey tools. Through these principles, the full end to
end process has been developed and tested with the understanding that giving users the
most realistic experience during user testing results in higher quality evidence. By providing
practical examples, and drawing on evidence from extensive user testing, this paper
demonstrates that, provided respondents understand the concepts, there is scope to design
the questions differently to suit the mode and to potentially improve data quality.

213
9.4 Social Statistics: Old dogs new tricks: respondent centred approach to mixed
mode social survey redesign at ONS
Thursday 5 September 11.30am

A respondent centred approach – Social survey materials and respondent

engagement strategies for mixed mode collection

Tara McNeill, Natalia Stutter

Office for National Statistics

To successfully achieve the goal of introducing online data collection to its portfolio of
household social surveys, ONS is pursuing an innovative research approach, exploring the
end-to-end journey of a potential survey respondent. This includes the research and
development of respondent materials and engagement strategies which are paramount in
getting people to take part independently online. This talk discusses the approach to
developing this work and will share key findings.Respondent communications play a pivotal
role in multimode data collection, particularly where the approach is online first. Over the last
two years ONS has invested heavily in qualitative and quantitative research to explore the
impact of respondent materials and mailing strategies on up-take and response rates. The
respondent centred approach has demonstrated the potential to achieve a 59% uptake rate
for a lightly incentivised push-to-web voluntary social survey.To develop an effective
respondent engagement strategy, a blank page approach to materials for household social
surveys was taken. This blended traditional qualitative social research methods alongside
innovative methods and tools to create a suite of communications based on user insights.
The research and design strategy has followed Agile development principles, enabling the
iterative development of materials, adapting to changing user needs and maturing
government design standards.This research approach has assisted in identifying what
works. Through the combined qualitative research and quantitative experiments undertaken
as part of the transformation of the Labour Force Survey, valuable insights have emerged
about the best mailing strategy for increasing the timeliness of response. Other experiments
have explored the extent to which using regionally branded envelopes can break down
barriers to opening mail, and also whether attrition on longitudinal surveys can be mitigated
through between wave engagement and what medium is most effective. This talk will share
the experiences and lessons learnt from developing these materials and engagement
strategies.

214
9.4 Social Statistics: Old dogs new tricks: respondent centred approach to mixed
mode social survey redesign at ONS
Thursday 5 September 11.30am

Findings and evidence from large-scale mixed mode testing of transformed social
surveys

Colin Beavan-Seymour, Vicky Parker, Joe Herson

Office for National Statistics

ONS is working towards the introduction of online data collection as the primary collection
mode for its suite of social surveys, supplemented by traditional face to face and telephone
collection. To achieve this, the design for the Labour Force Survey (LFS) has been
transformed end-to-end, focussing primarily upon core Labour Market outputs, to produce a
shorter and simpler survey instrument. This redesigned LFS is known as the Labour Market
Survey (LMS).The LMS is focussed primarily upon providing data for core Labour Market
outputs by completely redesigning the content of the survey to make it more relatable and
understandable for respondents. This is particularly important for the online self-completion
mode. The questionnaire content has a more logical flow and reduces respondent burden
through a more efficient routing system. However, the design still has to deliver the same
core Labour Market data as the LFS across multiple modes. This has been a challenging,
but positive, process.The high profile nature of the LFS and its importance for policy making
means that a complete redesign needs to be thoroughly tested, and the impact of changes
understood. Since mid-2017, ONS has undertaken a series of iterative quantitative tests;
each has had its own distinct research objectives but has also reactively tested changes in
the LMS design based upon the successes of each previous test.The testing programme
has provided research towards the use of a new sampling frame, incentivisation strategies,
materials, online data collection engagement and take up, self-completion questionnaire
design, paradata, and national / regional mixed-mode response rates. This talk will discuss
how the design of the LMS has evolved since its inception, the testing that has taken place
so far, the challenges and opportunities it has presented, and will discuss some preliminary
statistical findings which compare LMS and LFS data.

215
9.5 Methods & Theory: Multi-Parameter Regression Survival Modelling
Thursday 5 September 11.30am

Penalised Variable Selection in Multi-Parameter Regression Survival Modelling

Fatima Jaouimaa1, Il Do Ha2, Kevin Burke1

1
University of Limerick, 2 Pukyong National University

Multi-parameter regression (MPR) modelling refers to the approach whereby covariates are
allowed to enter the model through multiple distributional parameters simultaneously. This is
in contrast to the standard approaches where covariates enter through a single parameter
(e.g., a location parameter). Penalised variable selection has received a significant amount
of attention in recent years: methods such as the least absolute shrinkage and selection
operator (LASSO), smoothly clipped absolute deviation (SCAD), and adaptive LASSO are
used to simultaneously select variables and estimate their regression coefficients. Therefore,
in this work, we develop penalised multi-parameter regression methods and investigate their
associated performance through simulation studies and real data; as an example, we
consider the Weibull model.

216
9.5 Methods & Theory: Multi-Parameter Regression Survival Modelling
Thursday 5 September 11.30am

Semi-Parametric Multi-Parameter Regression Survival Modelling

Kevin Burke1, Frank Eriksson2, Christian Pipper3

1
University of Limerick, Ireland, 2 University of Copenhagen, Denmark, 3 LEO Pharma,
Denmark

We consider a log-linear model for survival data, where both the location and scale
parameters depend on covariates (i.e., a "Multi-Parameter Regression" [MPR] approach),
and the baseline hazard function is completely unspecified. This model provides the fexibility
needed to capture many interesting features of survival data at a relatively low cost in model
complexity. Estimation procedures are developed and asymptotic properties of the resulting
estimators are derived using empirical process theory. Finally, a resampling procedure is
developed to estimate the limiting variances of the estimators. A practical application to lung
cancer data is illustrated.

217
9.5 Methods & Theory: Multi-Parameter Regression Survival Modelling
Thursday 5 September 11.30am

A Multi-parameter regression model for interval censored survival data

Defen Peng1, Gilbert MacKenzie2, Kevin Burke3

1
ICVHealth/CHEOS, 2 Formerly of Centre of Biostatistics, Dept. of Mathematics & Statistics,
University of Limerick, Limerick Ireland, 3 University of Limerick,

We develop flexible multi-parameter regression (MPR) survival models for interval censored
survival data arising in longitudinal prospective studies and longitudinal randomised
controlled clinical trials. A multi-parameter Weibull regression survival model, which is wholly
parametric, and has non-proportional hazards, is the main focus of the paper. We describe
the basic model, develop the interval censored likelihood and extend it to include gamma
frailty, in order to account for unmeasured covariates, and a dispersion model for the frailty
variance, to allow for multiple frailty distributions. We evaluate the models by means of a
simulation study and a detailed re-analysis of data from a prospective observational study,
the Signal Tandmobiel study. The results demonstrate that The results demonstrate that the
multi-parameter regression model with frailty is computationally efficient and provides an
excellent fit to the data.

218
9.6 Communicating & Teaching Statistics: STEM Showcase
Thursday 5 September 11.30am

STEM Showcase

Simon White1, Rejina Verghis2, Jennifer Rogers3

1
University of Cambridge, 2 Northern Ireland Clinical Trials Unit, 3 University of Oxford

Statistics, and more importantly the skill of statistical literacy, is fundamental not only to the
next generation of STEM (Science, Technology, Engineering & Mathematics) personnel but
also to wider society – we must be able to meaningfully critique and discuss the wealth of
information used to make policies and run our lives, by governments, business and
everyone. As statisticians, we also have a duty to inspire the next generation and to engage
the public with mathematics and statistics; to encourage enjoyment of the subject, to
enhance and enrich study beyond the curriculum, and to encourage unusual ways of
communicating our science.

This session will present examples of engaging statistical activities that have been
developed, tested and successfully used in schools and at science festivals to inspire the
next generation of statisticians. It will be relevant to statisticians volunteering in schools or
with the general public.There will be three talks and a panel discussion.

This first speaker will be Jennifer Rogers, an RSS William Guy Lecturer. This is a prestigious
volunteer role which recognises Fellows with a successful track record in undertaking school
outreach activities. Jen will discuss her experiences as the post-holder and outline her
lecture topic - making life saving decisions in clinical trials: how much evidence do we need?

The second speaker is Rejina Verghis from the Northern Ireland Clinical Trials Unit. Rejina is
based in Belfast and has recently registered as a STEM Ambassador. She will outline her
motivation for becoming an Ambassador, and the types of activities she hopes to become
involved in.

The final speaker will be Simon White from the University of Cambridge, and the RSS
Education & Statistical Literacy Committee. Simon will outline several hands-on statistics
activities which embed learning important statistical concepts within an approachable (and
hopefully enjoyable) context.

Organised by the RSS Education & Statistical Literacy Committee and RSS Young
Statisticians' Section

219
9.9 Methods & Theory: Statistical analysis of relational data
Thursday 5 September 11.30am

Identifying potentially overlapping communities from two-mode networks

Veronica Vinciotti1, Saverio Ranciati2, Ernst Wit3

1
Brunel University London, 2 University of Bologna, 3 Universita' Svizzera Italiana, , , , , , ,

Actor-event data are common in sociological settings, whereby one registers the pattern of
attendance of a group of social actors to a number of events. These data are often
transformed to actor-actor data in order to be further analysed by network models, such as
stochastic block models. This transformation and such analyses lead to a natural loss of
information, particularly when one is interested in identifying, possibly overlapping,
subgroups or communities of actors on the basis of their attendances to events. In this talk,
we will present an actor-event model for overlapping communities and develop a Bayesian
procedure for inference. We present a real application on a terrorist network, whose
attendance to events was monitored during a period of time.

220
9.9 Methods & Theory: Statistical analysis of relational data
Thursday 5 September 11.30am

A network approach to votes exchange in the Eurovision Song Contest

Silvia D'Angelo1, Thomas Brendan Murphy1, Marco Alfò2

1
University College Dublin, 2 Sapienza, University of Rome

The Eurovision Song Contest is a popular TV singing competition held annually among
country members of the European Broadcasting Union. In this competition, each member
can be both contestant and jury, as it can participate with a song and/or vote for other
countries’ tunes. During the years, the voting system has repeatedly been accused of being
biased by tactical voting; votes would represent strategic interests rather than actual musical
preferences of the voting countries. In this work, we develop a latent space model to
investigate the presence of a latent structure underlying the exchange of votes. Focusing on
the period from 1998 to 2015, we represent the vote exchange as a multivariate network:
each edition is a network, where countries are the nodes and two countries are linked by an
edge if one voted for the other. The different networks are taken to be independent
replicates of a conditional Bernoulli distribution, with success probability specified as a
function of a common latent space capturing the overall relationships among the countries.
Proximity denotes similarity, and countries close in the latent space are more likely to
exchange votes. If the exchange of votes depends on the similarity between countries, the
quality of the competing songs might not be a relevant factor in the determination of the
voting preferences, and this would suggest the presence of some bias. A Bayesian
hierarchical modelling approach is employed to estimate the parameters, where the
probability of a connection between any two countries is a function of their distance in the
latent space, network-specific parameters and edge-specific covariates.

221
9.9 Methods & Theory: Statistical analysis of relational data
Thursday 5 September 11.30am

Latent Space Representations of Hypergraphs

Kathryn Turnbull1, Simon Lunagomez1, Christopher Nemeth1, Edoardo Airoldi2

1
Lancaster University, 2 Fox School of Business, Temple University

Relational data describing interactions among a population arise in a multitude of disciplines,

including systems biology, neuroscience and marketing. There exists a broad literature
concerned of analysis of such data when the interactions are assumed to be pairwise.
However, in many real-world applications the interactions may instead occur between
several members of a population and, in this case, the data are more appropriately
represented by a hypergraph. As an example, consider a coauthorship network where an
interaction indicates which academics have contributed to a paper. It is common for a group
of authors larger than two to write an article jointly, which corresponds to a hyperedge. The
literature on statistical analysis of hypergraphs is relatively underdeveloped, and in this talk
we introduce a model which extends the latent space approach of Hoff et al (2002) for
graphs to the hypergraph setting. In this framework, the nodes of the hypergraph are
assumed to lie in a low-dimensional space and the hyperedges are modeled as a function of
the latent coordinates. Using a toolkit from stochastic geometry, we develop a
computationally efficient model with a convenient likelihood. Furthermore, we explore and
analyse the properties of this model, and use the latent space to perform predictive
inference.

This is join work with Simon Lunagomez, Christopher Nemeth and Edoardo Airoldi.

222
9.9 Methods & Theory: Statistical analysis of relational data
Thursday 5 September 11.30am

Methods of analysing and comparing networks

Sofia Olhede1, Patrick Wolfe2

1
EPFL & UCL, 2 Purdue University

Networks have become a key data analysis tool. They are a simple method of characterising
dependence between nodes or actors. Understanding the difference between two networks
is also challenging unless they share nodes and are of the same size. We shall discuss how
we may compare networks and also consider the regime where more than one network is
observed.

223
10.1 Contributed - Medical: Meta-analysis
Thursday 5 September 2pm

On the comparison of alternative models in dose-response meta-analysis using

summarized data

Nicola Orsini
Department of Public Health Sciences, Karolinska Institutet

Objectives: A linear-mixed effects model for the synthesis of multiple tables of summarized
dose-response data has been recently proposed. The main advantage is to include studies
contrasting just two doses into spline analysis. Our aim is to evaluate the ability of the
Akaike’s information criterion (AIC) to suggest the underlying dose–response relationship.

Methods: Ten individual studies of 1,000 persons each were simulated under the
assumption of either a linear (Shape 1) or non-linear (Shape 2) relationship between a
quantitative dose and the mean outcome. The dose followed a positive right-skewed
distribution. Tables of summarized data (mean dose, mean difference about referent, sample
size, standard deviation) were generated upon categorization of the dose into two-quantiles.
Every simulated dose-response meta-analysis was analyzed with a linear-mixed effects
model using two commonly used strategies: linear and restricted cubic splines with 3 knots.
Accuracy of the AIC was examined by calculating the proportion of times in 1,000
simulations the Shape 1 and Shape 2 were correctly identified by choosing the lowest AIC
among the modelling strategies. We next explored how this accuracy could improve by
randomly categorizing the dose using two or three quantiles.

Results: Under Shape 1, the proportion of times the lowest AIC correctly indicated the linear
function was 99%. This proportion did not change using a mix of two/three quantiles to
create tables of summarized data. Under Shape 2, the proportion of times the lowest AIC
correctly indicated non-linearity was 2%. Using a mix of two/three quantiles raised the
accuracy to 98%.

Conclusions: The simulations of dose-response meta-analysis based on summarized data

using linear mixed-effects models showed a good and stable accuracy of the AIC to correctly
identify underlying linear relationships with just two quantiles of the dose. Regarding non-
linear relationships, good accuracy of the AIC can be achieved with a random mix of two and
three quantiles.

224
10.1 Contributed - Medical: Meta-analysis
Thursday 5 September 2pm

An extended mixed-effects framework for meta-analysis

Francesco Sera, Antonio Gasparrini

London School of Hygiene and Tropical Medicine

Objectives: Meta-analysis has become a standard method to summarize evidence in clinical

research. Standard applications commonly consider a single effect size estimated from
independent studies. However, extensions to deal with more complex meta-analytical
problems have been presented (e.g. multivariate, multilevel, and longitudinal meta-analysis).
All these extensions can be described as cases where non-independence among
observation within studies creates more complex correlation structures. The aim of this
contribution is to develop a unified framework for meta-analysis based on linear mixed-
effects (LME) models, where the correlation between effect sizes is modelled through a
flexible random-effects structure.

Methods: We derive a general analytic formulation that includes, as special cases, all the
specific models mentioned above. We have defined (restricted) maximum likelihood (ML and
REML) estimators, with efficient computational strategies based on profiled methods that
alternate (restricted) iterative generalized least squares and Newton-Raphson procedures.
Inference is based on asymptotic multivariate normal distribution of (RE)ML estimates. The
analytic framework and the inferential procedures are implemented in the new R package
mixmeta.

Results: The modelling framework will be illustrated using three case studies. In particular,
we will present an application of multivariate meta-analysis on 24 trials that compare four
alternative interventions to promote smoking cessation. In a second example we consider a
multilevel meta-analysis with 38 observations within 20 trials, each evaluating the
association between thrombolytic therapy and short-term mortality risks in multiple sub-
groups of treatment delay after a myocardial infarction. Finally, we illustrate an application of
longitudinal meta-analysis using data on 17 randomized controlled trials comparing
treatments of malignant gliomas at 6, 12, 18, and 24 months since the start of the treatment

Conclusions: The definition of a unified framework for meta-analysis, complemented with the
implementation in a freely-available and fully documented software, will provide researchers
with a flexible tool for addressing non-standard pooling problems.

225
10.1 Contributed - Medical: Meta-analysis
Thursday 5 September 2pm

Synthesis of individual and aggregate level data using multilevel network meta-
regression: extension to general likelihoods

David Phillippo1, Sofia Dias2, Tony Ades1, Nicky Welton1

1
University of Bristol, 2 University of York

Standard network meta-analysis (NMA) and indirect comparisons combine aggregate data
(AgD) from multiple studies on treatments of interest, assuming that any effect modifiers are
balanced across populations. Individual patient data (IPD) meta-regression can relax this
assumption, but in many cases IPD are only available in a subset of studies. Multilevel
Network Meta-Regression (ML-NMR) is a recently-proposed method which extends the IPD
meta-regression framework by integrating the individual-level model over AgD covariate
distributions to incorporate data from IPD and AgD studies and avoid aggregation bias.
However, ML-NMR requires the aggregate-level likelihood to have a known closed form.
Most notably, this precludes the application of ML-NMR to synthesis of time-to-event
outcomes, which make up the large majority of population adjustment analyses to date.We
extend ML-NMR to handle individual-level likelihoods of general form, illustrating with two
examples – a real network of plaque psoriasis treatments with ordered categorical
outcomes, and a simulated comparison of time-to-event outcomes.We show how the
individual-level likelihood function conditional on the covariates is integrated over the
covariate distributions in each AgD study to obtain the respective marginal likelihood
contributions. Quasi-Monte Carlo numerical integration is used, making application
general.Joint synthesis of ordered categorical outcomes lead to increased precision
compared to separate models. ML-NMR achieved better fit than a random effects NMA,
uncertainty was substantially reduced, and the model was more interpretable. For the
simulated survival data, ML-NMR agreed closely with the known truth, with little loss of
precision from a full IPD analysis.ML-NMR is a flexible and general method for synthesising
evidence from mixtures of individual and aggregate level data in networks of all sizes.
Extension to general likelihoods, including for survival outcomes, greatly increases the
applicability of the method. Decision making is aided by the production of effect estimates
relevant to the decision target population.

226
10.2 Contributed - Official & Public Policy: Business and migration
Thursday 5 September 2pm

Design strategy for data collection in business surveys to reduce response bias

Dominic Brown, Gary Brown

The Office for National Statistics

In the current era of falling response rates and reduced budgets, National Statistical
Institutes (NSIs) are compelled to innovate in data collection strategies. Instead of a single
static approach, designs that are adapted to different populations, and respond to different
conditions, are increasingly being considered.The UK Office for National Statistics (ONS)
has survey-specific targets for response rates for the 80+ establishment surveys it conducts
annually. During the data collection phase of the survey cycle cases are prioritised for data
collection, known in the UK as ‘response chasing’. The research project reported in this talk
examines the optimal response chasing strategy – in terms of cost and quality – for the ONS
as well as exploring some innovative possibilities and assess the implications for production
of official statistics.

227
10.2 Contributed - Official & Public Policy: Business and migration
Thursday 5 September 2pm

Using linked administrative data to better understand the activity and contributions of
international migrants within the UK

Megan Bowers, Nicola Rogers, Becca Briggs

ONS

International migrants enter and leave the UK for a variety of reasons, stay for different
lengths of time and interact with society and the economy in different ways. As part of the
cross-Government Statistical Service migration statistics transformation programme, ONS
are making use of a greater range of data to improve our understanding of these complex
issues and provide more detailed and relevant statistics to our users. We also need to
provide clear insights into how different groups impact on our workforce, communities and
public services such as the NHS and schools. We recently published a research
engagement report which set out our ambition to produce new insights on the impact of
migration and showed progress towards our new approach for producing international
migration flows and population statistics using administrative data. A series of in-depth case
studies were presented demonstrating how we can use administrative data to identify
“activity” for the migrant population. Activity can be defined as an individual interacting with a
system, for example attending a hospital or claiming a benefit. We discovered differences in
activity depending on migrants’ nationality, finding evidence that EU nationals register more
quickly for a National Insurance Number than non-EU. We also found that around 70% of
international students departed for 12 months or more at the end of their study in England
and Wales, while a quarter extended their UK stay.The case studies put a spotlight on what
linking together specific data sources – such as immigration, education, health and tax
records – can tell us about international migration. This presentation will cover our latest
findings, why they matter, our plans to produce further insights and an opportunity to gather
user feedback on the transformation programme. It will provide further context to the RSS
conference session on transforming population, migration and social statistics.

228
10.2 Contributed - Official & Public Policy: Business and migration
Thursday 5 September 2pm

How are services traded internationally? UK developments measuring the trade in

services by modes of supply.

Daniel Cheung, Dean Scott

Office for National Statistics

With the services industry representing approximately 80% of UK economic activity,

following the EU referendum there has been a growing need for trade in services data by
modes of supply; detailing the methods in which services are traded globally. Through
international collaboration with organisations such as the US Bureau of Economic Analysis
and with other government departments, the Office for National Statistics (ONS) has been
leading the way with this research.The data allows for a much more detailed understanding
of how the UK service sector operates globally, assisting trade negotiations by highlighting
strengths and weaknesses within the sector and allowing negotiations to be targeted for best
impact. Analysts will benefit by having a greater understanding of the globalisation of the
UK’s services sector.There are four modes of supply as defined by the World Trade
Organisations General Agreement on Trade in Services (GATS): trade conducted remotely,
trade provided by a commercial subsidiary present within the host country, trade conducted
by the supplier traveling abroad to provide the service, or the customer traveling to the
supplier’s home country to obtain the service. ONS piloted new mode of supply questions as
part of the International Trade in Services survey: the UK’s biggest source of trade in
services data. Following this pilot, the new questions were included as part of the full annual
survey. This presentation will detail these recent developments and present the latest
findings, with ONS being one of the first countries to publish results on modes of supply.

229
10.3 Contributed - Applications of Statistics: Applications 2
Thursday 5 September 2pm

Probability reasoning in judicial fact-finding

Ian Hunt1, The Honourable Mr Justice Mostyn2

1
Monash University, 2 Royal Courts of Justice

We argue that the laws of probability help promote coherent fact-finding ("on-the-balance-of-
probability") and can avoid logical contradictions. We assume that the laws of probability
hold, that Bayes' formula is valid and that probability can be interpreted as subjective
degrees of belief. Our argument is essentially that an "elementary probabilistic model of
degrees of belief often contains just the right balance of accuracy and simplicity to enable us
to command a clear view of the issues and see where we [are or could be] going wrong"
(Horwich, 1993).For example, assume that there are three mutually exclusive and
exhaustive explanations for something that happened and that the judge reckons that the
probability of each explanation being true is less than 0.5. Also assume that on pain of
incoherence the judge ensures that the sum of her subjective probabilities is one. In this
case, no legal facts can be found on-the-balance-of-probabilities; to find otherwise would
imply a contradiction in terms of the laws of probability (in particular that the three
probabilities must sum to one) or require a post-hoc fix to the originally reckoned
probabilities. Worse than a mere contradiction or a fix-up of the odds, a finding of fact for an
event with probability less than 0.5 risks serious injustice.We analyse real cases in which
judges apply similar reasoning and a related appeal decision which admonishes judges to
avoid using the laws of probability in findings of fact (it goes as far as suggesting that a
reference to the probability of a past event is pseudo-mathematics). We respond to the
criticisms launched in the appeal judgment and set out a positive case for using probability
reasoning in judicial fact-finding. And we argue that probabilistic reasoning enlightens, in a
therapeutic sense, legal principles related to inherent probability, the Binary Method and the
blue bus paradox (calibrating subjective probabilities to statistical frequencies).

230
10.3 Contributed - Applications of Statistics: Applications 2
Thursday 5 September 2pm

Bayesian networks and chain event graphs as decision making tools in forensic
science

Gail Robertson, Amy Wilson

University of Edinburgh

Bayes’ theorem and likelihood ratios are used in forensic statistics to compare evidence
supporting different propositions put forward during court proceedings. There is widespread
interest among forensic scientists in using Bayesian network models to evaluate the extent
to which scientific evidence supports hypotheses proposed by the prosecution and defence.
Bayesian networks are primarily used to compare support for source-level propositions, e.g.
those concerned with determining the source of samples found at crime scenes such as hair,
fibres, and DNA. While comparing source-level propositions is useful, propositions which
refer to criminal activities (i.e. those concerned with understanding how a sample came to be
at the crime scene) are of more interest to courts. Less work has been done on developing
probabilistic methods to assess activity-level propositions, hence finding a method of
evaluating evidence for these types of propositions would benefit practitioners. Chain event
graphs have been suggested as a decision making tool to assess the extent to which
evidence supports event timelines proposed by the prosecution and defence, and may be
more appropriate for assessing activity-level propositions. In this study we used Bayesian
networks and chain event graphs to combine different types of evidence supporting activity-
level propositions from a real-world drug trafficking case. We compared the use of Bayesian
networks and chain event graphs in evaluating evidence from activity-level propositions
associated with the case and developed a framework for these to be used by practitioners
with non-statistical backgrounds. We found that chain event graphs were better suited at
evaluating evidence from complex event timelines put forward by the prosecution and
defence.

231
10.3 Contributed - Applications of Statistics: Applications 2
Thursday 5 September 2pm

Judging a book by its cover - How much of REF 'research quality' is really 'journal
reputation'?

David Selby, David Firth

University of Warwick

The Research Excellence Framework (REF) is a periodic UK-wide assessment of the quality
of published research in universities. The most recent REF was in 2014, and the next will be
in 2021. The published results of REF2014 include a categorical 'quality profile' for each unit
of assessment (typically a university department), reporting what percentage of the unit's
REF-submitted research outputs were assessed as being at each of four quality levels
(labelled 4*, 3*, 2* and 1*). Also in the public domain are the original submissions made to
REF2014, which include—for each unit of assessment—publication details of the REF-
submitted research outputs.In this study we address the question: to what extent can a REF
quality profile for research outputs be attributed to the journals in which (most of) those
outputs were published?The data are the published submissions and results from REF2014.
The main statistical challenge comes from the fact that REF quality profiles are available
only at the aggregated level of whole units of assessment: the REF panel's assessment of
each individual research output is not made public. Our research question is thus an
'ecological inference' problem, which demands special care in model formulation and
methodology. The analysis is based on logit models in which journal-specific parameters are
regularized via prior 'pseudo-data'. We develop a lack-of-fit measure for the extent to which
REF scores appear to depend on publication venues rather than research quality or
institution-level differences. Results are presented for several research fields.

232
10.4 Contributed - Social Statistics: Response Data
Thursday 5 September 2pm

Administrative Data and it's use to combat survey Non-Response

Matthew Moore, Alasdair Rutherford

University of Stirling

The issues arising due to non-response are problematic for many surveys. With an
international trend of dropping survey response rates (Zhang et al, 2013). This has led to
more researchers asking the extent to which can non-response can cause issues with
analysis, and what can be done to combat this non-response. There are a few ways to
approach this problem which can be grouped into either fieldwork or post-fieldwork
techniques. Non-response weighting is a common method used, aiming to reduce potential
non-response bias. However, the construction of these weights often rely upon untested
assumptions and low quality and quantity of data (Collins et al 2001). Administrative data
can provide new opportunities in survey design including the construction of non-response
weights allowing the production of better survey-based estimates. In this paper we explore
the potential of using administrative data from the sampling frame of a large social survey to
model non-response. The data used for this paper is part of the Healthy Aging in Scotland
(HAGIS) Survey. The HAGIS pilot wave collected a sample of approximately 1000
individuals aged over 50 in the Scottish mainland. HAGIS utilised administrative data in the
sampling with screening being provided to ensure households with at least one eligible
respondent (over 50) was selected. Administrative data was also collected available for the
full sample, and not just respondents to the HAGIS pilot wave. The administrative data
consists of individual demographics, hospital admissions, and prescription data held by NHS
Scotland. Survey response is modelled as a function of explanatory variables from the
administrative data in order to examine whether there is systematic demographic and health-
based non-response in the survey. We further examine whether this non-response is
causing significant biases and then how health data might predict non-response. Finally, we
discuss how these models could be used in improving conventional weighting techniques to
combat non-response in survey design.

233
10.4 Contributed - Social Statistics: Response Data
Thursday 5 September 2pm

On reproducibility of hypothesis tests based on randomised response data

Fatimah Alghamdi
Durham University

Randomised response techniques (RRT) are frequently used when data on possiblysensitive
information is being collected by using a survey. There are many different RRT methods and
strategies which seek to know the truth from the respondents with more efficiency and more
privacy and without embarrassment beside its ability to decrease the bias which can happen
by wrong answers; the first techniques was presented by Warner (1965), then Greenberg
model (1979) who has modified some properties in Warner.A question of interest in
hypothesis test scenarios is the reproducibility of the results: if the test was repeated, would
it lead to the same conclusion with regard to rejection of the null hypothesis? We address
this question for Warner's and Greenberg method. We use nonparametric predictive
inference, a frequentist approach based on only few assumptions, to derive lower and upper
probabilities for test reproducibility.This poses the challenging question of finding another
measurement, which is called the measurement of the lower reproducibility probability
(MRP), to compare between the Warner and Greenberg test.Greenberg model is more
efficient than Warner model and the measurement MRP and the area of non-rejection of the
measurement (AUMRP) of Greenberg test are higher as well.The work will continue to
explore more useful results and real applications about (RRT) and it can be developed in
another direction.

Reference[1] Warner, S. L. (1965). Randomized response: A survey technique for eliminating evasive
answer bias. Journal of the American Statistical Association 60(309), 63–69.[2] Greenberg, B. G.,
Abul-Ela, A. L. A., Simmons, W. R., Horvitz, D. G. (1969). The unrelated question randomized
response model: Theoretical framework. Journal of the American Statistical Association, 64(326),
520-539.

234
10.4 Contributed - Social Statistics: Response Data
Thursday 5 September 2pm

Response rates, fieldwork and survey quality: Does reissuing reduce non-response
bias?

Linda Hutcheson
Ipsos MORI

Traditionally, response rates have been used as a proxy of survey quality. High response
rates are assumed to indicate good quality survey data with less potential for non-response
bias. However, over the last twenty years, response rates have been declining. Most major
face-to-face random probability surveys have expended more and more fieldwork effort and
cost by increasingly reissuing non-responding sample at first issue to other interviewers to
try to maintain response rates.But does this reduce non-response bias? This study examines
the impact that reissuing has on key survey estimates from two of Scotland’s most important
population surveys: the Scottish Household Survey and the Scottish Crime and Justice
Survey. In these surveys, reissuing increases the response rate by around 8-10%, from 55-
60% to around 63-68%. The impact on a range of measures is examined by comparing the
published estimates from the weighted full survey sample with estimates from first issue
interviews only, weighted as if they were the final achieved sample. As well as examining the
impact at the overall national level, impact among key sub-groups is assessed, and results
from two waves of each survey are compared to examine whether impact is consistent
across different waves.Overall, the results show that the impact of increasing the response
rate through reissuing on national estimates is modest. Among sub-group estimates, while
the impact in absolute terms is larger, this is because these estimates themselves are less
precise as they are based on smaller sample sizes. Reissuing some forms of unproductive
case may have more impact on estimates than others.

235
10.5 Contributed - Methods & Theory: High-Dimensional Data
Thursday 5 September 2pm

A novel method for analysis of high dimensional data

Reza Drikvandi
Manchester Metropolitan University

High dimensional data are rapidly growing in many domains, for example, in microarray
gene expression studies, fMRI data analysis, large-scale healthcare analytics, text/image
analysis, natural language processing and astronomy, to name but a few. In the last two
decades regularisation approaches have become the methods of choice for analysing high
dimensional data. However, obtaining accurate estimates and predictions as well as reliable
statistical inference remains a major challenge in high dimensional situations. In this talk, we
introduce a novel method to overcome this challenge in high dimensional linear regression
models. The proposed method enjoys from an effective combination of the regularisation
methods such as lasso and the dimensionality reduction techniques such as PCA and
sparse PCA. The sparse PCA is scalable and reduces the computations dramatically, which
is very attractive for big data analysis. The proposed method is evaluated theoretically and
practically using simulated and real data analysis, and its performance is compared with
existing regularaisation methods. We show, via theoretical and simulation investigations, that
the proposed method produces accurate parameter estimates and predictions, and also
enables us to conduct statistical inference on regression coefficients. Finally, we briefly
discuss valid post-selection inference in high dimensional data using the proposed
methodology.

236
10.5 Contributed - Methods & Theory: High-Dimensional Data
Thursday 5 September 2pm

Bias reduction methods for binary regression with misclassified responses

Euloge Clovis Kenne Pagui, Alessandra Salvan, Nicola Sartori

Department of Statistics - University of Padoua

Methods for binary regression models that ignore misclassification of the response may lead
to a substantial bias on the maximum likelihood estimators of the regression parameters.
With known misclassification probabilities, a flexible maximum likelihood framework which
allows for misclassification of the responses is possible and has been applied in different
contexts. Unfortunately, with moderate sample information or high misclassification
probabilities, the bias of the maximum likelihood estimator can still be large and may result in
misleading inference. As a solution to this problem, we propose the use of the adjusted
score functions for mean and median bias reduction proposed by Firth (1993) and Kenne
Pagui et al. (2017), respectively. Both approaches, consist of adding a suitable adjustment
term to the score function and then solving the resulting adjusted score equation. The former
approach produces a second-order mean bias reduced estimator while the latter gives a
third-order median bias reduced estimator. The two methods do not require finiteness of the
maximum likelihood estimate and are effective in preventing infinite estimates. For the
implementation of mean and median bias reduction, a simple modification of existing
software for generalized linear models can be used. Simulation studies show that mean and
median bias reduced estimators have better frequentist properties than standard maximum
likelihood.

237
10.5 Contributed - Methods & Theory: High-Dimensional Data
Thursday 5 September 2pm

Refining clustered standard errors with few clusters

Gianmaria Niccodemi, Tom Wansbeek, Rob Alessie, Viola Angelini, Jochen Mierau
University of Groningen

White cluster-robust standard errors, based on OLS residuals, are typically underestimated
in case of few clusters and positive correlation of the error term within cluster. This often
implies misleading inference, especially on constant within cluster regressors. Common
methods to estimate unbiased clustered standard errors with few clusters include the so-
called CR2VE, also known as BRL (bias reduced linearization), and CR3VE, both based on
transformed OLS residuals. An alternative method is based on estimating the variance
components to directly control for within-cluster correlation of the error term, if the
covariance matrix of the error term is known to have a certain within-cluster parametrization
(e.g., random effects). In this paper we present three refinements for these methods. First,
we introduce CR3VE-λ, an adaptation for CR3VE that takes the different cluster sizes into
account to make the computed standard errors closer to unbiasedness. Second, we
introduce formulas to compute equivalent CR2VE and CR3VE that exploit lower order
matrices and that are therefore computationally more efficient. Third, we show how the
standard Swamy-Arora random effects model, implemented in most statistical packages
(e.g., Stata), can perform poorly with few clusters as its estimate of the cluster effect can be
negative. This problem can be mitigated by employing a different random effects estimator.
We illustrate our refinements by several Monte Carlo simulations that show that our CR3VE-
λ performs better than CR3VE in case of highly unbalanced clusters, that our equivalent
CR2VE and CR3VE are computationally much faster as the cluster sizes increase, and that
the simple alternative to Swamy-Arora random effects model may be adequate for the
computation of clustered standard errors with few clusters, regardless the nature of the
covariance matrix of the error term.

238
10.6 Contributed - Data Science: Rage Against the Machine Learning
Thursday 5 September 2pm

Treating missing data with machine learning

Vinayak Anand Kumar, Vahe Nafilyan, Alex Milroy

Office for National Statistics

Missing data can be problematic as they reduce the reliability and accuracy of statistics.
Imputation generates values and/ or units to produce a dataset that is more representative of
the population and concept of interest. Ideally, imputation methods should be tailored to
specific problems, including the nature of missingness. . Unfortunately, in practice imputation
models are not always empirically tested due to the large volume of data or timeliness
constraints. The Methodology Division of the Office for National Statistics (ONS) in
collaboration with the Data Science Campus investigate the use of supervised Machine
Learning (ML) techniques to carry out imputation.. We test several state-of-the-art ML
algorithms such as XGBoost and Generative Adversarial Networks (GANs) to impute
missing values and compare this to the standard approach. We also use ML techniques to
simulate missingness and create a test dataset that we then use to validate the imputation
methods. The presentation will cover the key concepts behind XGBoost and GANs and the
findings from this program of work.

239
10.6 Contributed - Data Science: Rage Against the Machine Learning
Thursday 5 September 2pm

Predicting Human States with Multimodal data utilising Machine Learning

Methodologies

Daniel Moore1, Christine Spencer1, Lucy Rutherford2, Gawain Morison2, Gary McKeown1,
Lisa McFetridge1, Ben Bland2
1
Queen's University Belfast, 2 Sensum

The future of in-car sensing technology is facing a radical change over the coming decade,
largely driven by the need for increased driver safety. In this endeavour various approaches
have been explored, from self-driving vehicles to active collision-avoidance systems.
However, the logic by which these technologies should best act to remove the driver’s
influence from the car is not well-defined. Therefore, our objective is to identify the driver’s
latent stress to potentially facilitate the automatic handover of driver control systems, and
automate the change of in-cabin environment to reduce this latent stress. Such a ‘reactive-
cockpit’ could be activated through various systems, including the vehicle’s infotainment
devices, HVAC (climate) control, speed-limiter and assisted-driving technologies.This work
explores a multimodal approach to detecting driver stress. Here, human physiology, facial
expression and voice are analysed together to enhance the prediction accuracy in a method
that is more robust than using one data type alone. This work details the end-to-end steps of
developing a real-time machine-learning based technique for predicting stress, and the
problems and caveats associated with such a project. Here, data is collected using a driving
simulator and real-world driving tasks, using protocols designed to stimulate the ground-truth
state. Blood pressure and self-report data are used to confirm the protocols’ effectiveness.
The data collected is then annotated on a continuous scale and used as the training dataset.
Features are derived from the raw signals, such as Heart Rate Variability and Blink Rate, as
predictors on which the model is trained. The results indicate that this approach is able to
effectively detect driver stress as confirmed by the self-report data and blood pressure
values. Furthermore, such a model may be used to identify prolonged periods of driver
stress, which can act as a trigger-point for the car’s safety and environmental control
systems, which in turn may increase driver safety and comfort.

240
10.6 Contributed - Data Science: Rage Against the Machine Learning
Thursday 5 September 2pm

Rage Against The Machine - The limits of maching learning for automatic product
classification

Jens Mehrhoff
European Commission

The talk contains an assessment of applying supervised learning classification algorithms to

automatically map scanner data to the official classification of sub-indices for consumer price
statistics. The main test objectives are summarised as follows:compare the family of
probabilistic supervised machine learning classification methods;use more or less detailed
product descriptions and identifiers as feature-generating variables; andconsider different
percentages of the data set as training data - from smaller and larger data sets.A major
challenge in this respect is textual feature engineering. The product descriptions are not
natural text but use specific vocabularies and rely on different kinds of shorthand. To this
end, all possible three-character strings (trigrams) are exploited.The performance measures
derived are precision and recall as well as the execution time. Particular emphasis will be
given to the interpretation of the results in terms of usability in the production process, i.e.
guidance on what works, what does not, and why.

241
10.7 Contributed - Medical: Hospital Outcomes
Thursday 5 September 2pm

Using funnel plots and CUSUM techniques to monitor hospital-standardised mortality

Chris Mainey
University Hospitals Birmingham NHS Foundation Trust

Patient mortality and its association with healthcare is a contentious topic, with a strong
academic history, and various entrenched views. Despite differing opinions, hospitals and
regulatory bodies routinely measure and monitor mortality rates, regarding them as ‘smoke
alarms’ related to quality of care. Many of the techniques used for monitoring were
described in an influential RSS paper by Spiegelhalter et al. (2012), and this talk describes
them in practice in the NHS.The risk of death varies between patients according to many
factors and these factors must be taken into account when monitoring mortality. Two of the
most common indicators: the Summary Hospital-level Mortality Index (SHMI) and the
Hospital Standardised Mortality Ratio (HSMR) are calculated as indirectly-standardised
ratios of observed and ‘expected’ deaths. They vary in their predictors and definitions, but
both use logistic regression to predict the risk of death per patient, and the sum of the risk
scores form counts of ‘expected’ deaths at organisations. These methods are commonly
presented as ‘funnel plots,’ and use control limits to identify outlier organisations. These
limits are based on the Poisson distribution and suffer from overdispersion, and this talk will
discuss adjustment techniques. Funnel plots are only suited to cross-sectional analysis and
other techniques are need to monitor over time. This talk will discuss risk-adjusted CUSUM
techniques, as used by NHS regulator the Care Quality Commission (CQC) and Imperial
College’s mortality monitoring system. It explains how they are calculated, the differences
between techniques, and how they are used in practice.Two alternative approaches will be
discussed: CUSUMs based on aggregated data using z-scores, and patient-level log-
likelihood charts. These charts have different processes for setting trigger thresholds and
involving simulation or approximation techniques. Examples of techniques will be shown
using R, and shortcomings such as the effects of overdispersion and false positives, will be
discussed.

242
10.7 Contributed - Medical: Hospital Outcomes
Thursday 5 September 2pm

So, you’ve been added to the lung transplant list, now what? An analysis of patient
outcomes from listing for a lung transplant

Rachel Hogg1, Antonios Kourliouros2, Jenny Mehew1, Mohamed Al-Aloul3, Martin Carby4,
James Lordan5, Richard Thompson6, Steven Tsui7, Jasvir Parmar7,
1
NHS Blood and Transplant, 2 Royal Brompton Hospital, 3 Wythenshawe Hospital, 4
Harefield Hospital, 5 Freeman Hospital, 6 Queen Elizabeth Hospital Birmingham, 7 Royal
Papworth Hospital

Objective: When listed for a lung transplant, patients have several outcomes on the list;
transplant, death, or removal. Demand for lung transplantation vastly exceeds the
availability of donor organs, translating into long waiting times and high waiting list mortality.
The purpose of this work was to identify risk factors associated with the different patient
outcomes on the list, with results being used as a tool to help patients understand their
individual risks of each outcome (in collaboration with the Winton Centre).

Methods: Adults first registered for lung transplantation in the UK between 1 January 2004
and 31 March 2014 were analysed. Cox regression models were developed for each
outcome to identify risk factors and their associated impact. Due to the presence of
competing risks, Fine and Gray models were used to provide parameter estimates and risk-
adjusted cumulative incidence functions for each event.

Results: At 2 years post-registration, 58% of the 2213 patients were transplanted and 22%
had died on the list. A range of factors were found to impact each outcome from the Cox
analyses. Chance of transplant varied by disease group and centre (p<0.001). Taller
patients (p<0.001) had an increased chance of transplant. Age had a non-linear effect on
death on the list with recipients aged 42-56 having the highest chance of death.

Conclusion: Patients on the lung transplant list are complex with a range of factors
influencing their chance of a transplant or other outcome. This work identified these factors,
informing the clinical community and the future direction of allocation policy, allowing for
better patient care. The work with the Winton Centre will improve information available to
patients regarding their potential outcomes.

243
10.7 Contributed - Medical: Hospital Outcomes
Thursday 5 September 2pm

“Tell us what the data say”: Understanding and misunderstanding control charts for
monitoring hospital outcomes

Peter Martin, James Bedford, Ramani Moonesinghe

University College London

Control charts have found increasing use in the monitoring of outcomes from medical
treatment. Devices such as exponentially weighted moving average charts (EWMAs) allow
clinicians to track their own or their hospital’s performance in real time. Statistically defined
control limits provide thresholds that help to separate the signal of potential poor
performance from the noise of uninformative variation. There are good reasons to think that
statistical monitoring can benefit patients: for example, had outcome monitoring been in
place, the serial murders committed by English GP Harold Shipman might have been
detected earlier. On the other hand, large-scale statistical monitoring systems will inevitably
cause some false alarms. If handled insensitively, ‘alarms’ may induce undue panic and
premature attribution of blame that can damage staff morale and engender distrust of the
control charts themselves.The statistical literature mainly deals with the technical
specifications and distributional properties of control charts. Yet there is little published work
on how these charts are used in practice, or how statisticians and non-statisticians ought to
collaborate to ensure effective implementation and meaningful interpretation. This talk will
present a case study of monitoring patient deaths after emergency bowel surgery in over
150 hospitals in England, which together perform over 20,000 operations per year. We will
describe the design of risk-adjusted EWMA charts, discuss challenges of implementation,
and present guidelines for clinicians and service managers that propose a course of action
to take when a control limit is exceeded. We will also discuss some common
misinterpretations of control charts, and the temptation to use them for purposes other than
what they are good for. Faced with demands to explain “what the data tell us”, a crucial task
for the statistician is to communicate uncertainty and to be clear about both the benefits and
the limits of what statistics can do.

244
10.8 RSS Prize Winners: Best presentations from RSC 2019
Thursday 5 September 2pm

Uncertainty Quantification for Offshore Wind Energy

Jack Kennedy
Newcastle University

Energy systems models, are complex, computationally intensive and have a large number of
inputs. The model inputs are typically unknowns to be quantified in uncertainty and
sensitivity analyses.We represent the unknown quantities by a high dimensional Bayesian
prior distribution to be elicited from engineers from academia and industry. From this prior,
many samples can be drawn and run through the model to understand the uncertainty
induced in the model outputs from the uncertainty in the inputs.However, since such models
require vast amounts of CPU time we utilise Gaussian process emulators - statistical
approximations to mathematical models - to facilitate computation. Further, we emulate
stochastic computer codes with dependent, multivariate output, which adds in an extra layer
of complexity in the appropriate description of the dependence and variability in the model
outputs.We present the application of emulation techniques to a stochastic offshore
windfarm simulation to better the understanding of windfarm output over the early life of the
farm; a critical period to ensure the financial viability of offshore wind.

245
10.8 RSS Prize Winners: Best presentations from RSC 2019
Thursday 5 September 2pm

Improving the Emulation of Stochastic Computer Models

Evan Baker
University of Exeter

Computer models are increasingly being used to learn things about the real world instead of,
or as well as, real life experiments. These computer models can take a long time to run,
which can make using them for their intended purpose difficult. To help with this, we can
create a statistical replacement of the computer model that is much much quicker to run (an
emulator), based on a few runs of the computer model. Emulating stochastic computer
models can be difficult; requiring a large number of runs from the computer model before the
mean and variance can be accurately estimated. I will discuss how deterministic
approximations of stochastic computer models can be leveraged to improve the emulation of
the stochastic models. Specific implementation details involve the summation of Gaussian
processes, which is a popular idea in many areas of emulation.

246
10.8 RSS Prize Winners: Best presentations from RSC 2019
Thursday 5 September 2pm

Analysis of clickstream data

Ryan Jessop
Clicksco

Online user browsing generates vast quantities of typically unexploited data. Investigating
this data and uncovering the valuable information it contains can be of substantial value to
online businesses, and statistics plays a key role in this process. The data takes the form of
an anonymous digital footprint associated with each unique visitor, resulting in 10^6 unique
profiles across 10^7 individual page visits on a daily basis. Exploring, cleaning and
transforming data of this scale and high dimensionality (2TB+ of memory) is particularly
challenging, and requires cluster computing. We consider the problem of predicting
customer purchases (known as conversions), from the customer’s journey or clickstream,
which is the sequence of pages seen during a single visit to a website. We consider each
page as a discrete state with probabilities of transitions between the pages, providing the
basis for a simple Markov model. Further, Hidden Markov models (HMMs) are applied to
relate the observed clickstream to a sequence of hidden states, uncovering meta-states of
user activity. We can also apply conventional logistic regression to model conversions in
terms of summaries of the profile’s browsing behaviour and incorporate both into a set of
tools to solve a wide range of conversion types where we can directly compare the predictive
capability of each model. In real-time, predicting profiles that are likely to follow similar
behaviour patterns to known conversions, will have a critical impact on targeted advertising.
We illustrate these analyses with results from real data collected by an Audience
Management Platform (AMP) - Carbon.

247
11.1 Medical: The latest methodological developments in network meta-analysis
Thursday 5 September 3.30pm

Network meta-analysis of joint longitudinal and time-to-event data

Maria Sudell, Catrin Tudur Smith, Ruwanthi Kolamunnage-Dona

University of Liverpool

Network meta-analysis of joint longitudinal and time-to-event dataJoint models provide

methods to analyse potentially linked longitudinal and time-to-event data (termed joint data).
Joint data might be longitudinal data complicated by dropout, time-to-event analyses
involving time varying covariates, or instances where related longitudinal and time-to-event
outcomes are both of interest. An example of joint data is blood pressure repeatedly
measured over time, and time until death. Methodology for joint models has been expanded
in recent years to meta-analytic joint models, which pool joint data from multiple data
sources. The next step in this methodology development is network meta-analytic joint
models, which would allow direct and indirect evidence to be utilised when assessing
multiple treatment options simultaneously, for related longitudinal and time-to-event data.
This talk will describe and discuss aspects of joint models that complicate a joint data
network meta-analysis (including time varying components, and relationship between
networks of treatment effects for the longitudinal and time-to-event components which are
dependent on the joint model association structure), and will introduce modelling approaches
to perform network meta-analyses of such joint data.

248
11.1 Medical: The latest methodological developments in network meta-analysis
Thursday 5 September 3.30pm

Individual Patient Data in a Network Meta Analysis: Is it worth the effort?

Cathal Walsh2, Joy Leahy1

1
Trinity College Dublin, 2 University of Limerick

The use of Individual Patient Data (IPD) in Network Meta Analysis (NMA) is becoming
increasingly popular. However, as most studies do not report IPD, most NMAs are carried
out using aggregate data (AD) for at least some, if not all, of the studies. We investigate the
benefits of including varying proportions of IPD studies in an NMA. Several models have
previously been developed for including both AD and IPD in the same NMA. We carried out
a simulation study based on these models to check the effect of additional IPD studies on
the accuracy and precision of the estimates of both the treatment effect and the covariate
effect. We also compared the Deviance Information Criterion (DIC) between models to
assess model fit. An increased proportion of IPD resulted in more accurate and precise
estimates for most models and datasets. However, the coverage probability sometimes
decreased when the model was mis-specified. The use of IPD leads to greater differences in
DIC, which allows us to choose the correct model more often. We analysed a Hepatitis C
network consisting of three IPD observational studies. The ranking of treatments remained
the same for all models and datasets. We observed similar results to the simulation study:
the use of IPD leads to differences in DIC and more precise estimates for the covariate
effect. However, IPD sometimes increased the posterior SD of the treatment effect estimate,
which may indicate between study heterogeneity. We recommend that IPD should be used
where possible, especially for assessing model fit.

249
11.1 Medical: The latest methodological developments in network meta-analysis
Thursday 5 September 3.30pm

Reference prediction to connect evidence networks

Howard Thom1, Joy Leahy2, Jeroen Jansen3

1
University of Bristol, 2 Trinity College Dublin, 3 Precision Health Economics

In the absence of head-to-head randomised controlled trials (RCTs), network meta-analysis

(NMA) can compare treatment effects across a network of connected RCTs. However, RCT
evidence for interventions of interest can be disconnected or provide only highly uncertain
estimates for relative effects. Single-arm studies, not includable in standard NMA, may also
be the only data available on important comparators. In such scenarios, healthcare decision
makers, such as the National Institute for Health and Care Excellence in the UK or National
Centre for Pharmacoeconomics in Ireland, still need to be able to compare treatment
effects.Population adjusted indirect comparison methods have been proposed for comparing
disconnected treatments, but these require individual patient data (IPD) which are often held
by vested interests. If only aggregate data are available, aggregate level matching (ALM)
may be used to connect networks by linking similar RCTs or single-arm studies. A further
approach is to predict response of the reference treatment in disconnected RCTs or single-
arm studies using a random effect on the baseline treatment in NMA. Random effects on the
baseline interfere with randomisation in the connected RCTs and, mostly for this reason, are
not generally recommended.We propose a refinement of random effects on baseline, termed
reference prediction, to avoid these defects. We build a reference prediction model, using
available covariates to improve the predicted response, but keep it separate from the
connected RCTs. In the connected network of RCTs, we still assume independent baselines
and preserve randomisation. For disconnected RCTs, we re-parameterise treatment effects
to be relative to reference and make appropriate multi-arm adjustment for correlations. We
demonstrate our method using a real NMA of directly acting oral anticoagulants (DOACs) for
atrial fibrillation, which we break into disconnected networks and single-arm studies. We
compare reference prediction with both naïve random effects on baseline and ALM.

250
11.4 Social Statistics: Could I ask you about your life? Future-proofing household
surveys
Thursday 5 September 3.30pm

Scottish Government Population Surveys and Core Questions

Ben Cook
Scottish Government

Recently, the four face-to-face population surveys in Scotland were harmonized and pooled
together, creating the potential for a combined sample. Statistics from the harmonized
subset of questions in the surveys is published as the Scottish Surveys Core Questions. The
resulting increase in sample size allows statistics to be reported for minority equality groups
(such as specific ethnic and religious groups) and for smaller spatial geographies. Analysis
of outcomes for these smaller groups of the population is not possible through individual
survey results. The presentation will discuss;- the harmonization of questions,- the
coordination of sampling for the individual population surveys, - the development of pooled
weighting methodologies for single and multi-year samples,- estimate variability between
source surveys. We also present some ongoing results of this unique project, which span
multivariate analyses of multi-year datasets, official statistics publications and topic reports.
www.gov.scot/sscq

251
11.4 Social Statistics: Could I ask you about your life? Future-proofing household
surveys
Thursday 5 September 3.30pm

Welsh Government National Survey - an example of a merged survey

Steven Marshall
Welsh Government

The Welsh Government and its partner organisations have historically carried out a range of
separate surveys of individuals which provide valuable information for policy decisions. Such
surveys are expensive to carry out and with downward pressures on cost further
exacerbated by falling response rates this prompted a rethink. A review was commissioned
in 2014 to consider options that would deliver both robust survey data and achieve value for
money. The review was asked to consider all options. As a result five surveys were merged
from April 2016 into a single new face-to-face survey – two Welsh Government surveys and
three surveys from three separate sponsored bodies. Three of these surveys were
conducted face-to-face while one was self-completion and one carried out via telephone. A
new combined questionnaire was developed to provide a coherent single whole and
designed to try and minimise context effects and discontinuities due to mode changes. The
new survey provided significant cost savings of more than £6 million over a 5 year contract
period. There have been other benefits beyond cost reduction. Particularly for sponsored
bodies the new combined survey allows a more flexible approach to data collection – so that
small amounts of data can be collected regularly with more detail every few years rather
than a single one off survey every few years. There are improvements in the range of cross
topic analysis now possible, from example having sports participation and health topics in
the same survey. The basic design of the new survey has retained much of the approach in
the previous Welsh Government national survey including the rotation of topics and sub
sampling to maximize the full range of topics collected over time as well as being responsive
to new data demands.

252
11.4 Social Statistics: Could I ask you about your life? Future-proofing household
surveys
Thursday 5 September 3.30pm

Social Surveys in Northern Ireland - past, present and future

Kevin Sweeney, Andrew McCormick, Brendan Morgan

Northern Ireland Statistics & Research Agency

Household Surveys have come a long way in Northern Ireland (NI), from paper to CAPI,
laptop to tablet, floppy disk to 3G, from 1 survey with a 1200 household sample to sampling
almost 5% of the addresses in NI annually. Under the last NI Assembly, the development of
a ‘Programme for Government’ modelled closely on the Scottish outcome based approach
led to increased pressure for high quality survey data based on probability samples with
good response and placing increased demand on survey operations to deliver.This
presentation will consider:

- past developments and their implications for the future;

- the current strategy for the development of on-line surveys with a case study of an on-line
Census Test;

- some current developments to improve survey efficiency including Achieving Cooperation

Training, improving address location, payment of interviewers, use of GPS data;

- the appropriate direction of travel in terms of transformation, translation and offering on-line
and mixed-mode options

- is it possible to stem the decline in household survey response

253
11.5 Methods and Theory: Inferential Machine Learning - Accelerating Statistical
Methodology through ML
Thursday 5 September 3.30pm

RKHS-based tests for Survival Analysis

Tamara Fernandez
University College London

Hypothesis testing is a fundamental problem in Machine Learning and Statistics. Over the
past decade, statistical tests based on embeddings over reproducing kernel Hilbert spaces
(RKHS) have become popular due to their excellent results when dealing with different types
of complex data. With the incorporation of new sampling methods in clinical research, it is
important that Survival Analysis methodologies are able to deal with complex data, thus the
application of RKHS-based methods becomes timely and relevant. While the RKHS
methodology has been successfully applied to different problems in statistics and Machine
learning, up to the best of our knowledge, it has not been systematically explored in the
context of censored data. In this talk we explore different approaches to extend RKHS-based
tests to censored data, focusing particularly in the general problem of testing for
independence between right-censored survival times and high dimensional covariates. As
opposed to the classical/uncensored approach of embedding probability distributions onto an
RKHS, we propose to embed hazard functions which are estimated from the data using the
Nelson Aalen estimator. Following this approach, we derive an analytically tractable test-
statistic and a computationally efficient Wild Bootstrap procedure to estimate the null
distribution quantiles. We perform an extensive empirical evaluation of our test-statistic for
proportional and time-dependent hazard functions showing excellent results in both settings.
We finalise by discussing some theoretical properties and by establishing connections to
classical approaches such as log-rank tests and the Cox-Score test.

254
11.5 Methods and Theory: Inferential Machine Learning - Accelerating Statistical
Methodology through ML
Thursday 5 September 3.30pm

Using variational autoencoders to learn efficiently embedded representations of

functions and their properties

Swapnil Mishra, Seth Flaxman, Samir Bhatt

Imperial College London

Many machine learning methods specify a function class and an algorithm to learn the "best"
function in that class, for example by minimizing the empirical risk. Function classes can
range from simple linear models to reproducing kernel Hilbert spaces (RKHS). For a number
of popular algorithms, such as Gaussian process regression, learning from the function class
comes with computational challenges, including computational burden, correlated
parameters and algorithmic complexity. Here we use variational autoencoders (VAE) to learn
low dimensional embeddings of function classes. We show that our VAE framework can
accurately learn complex function classes such as Gaussian, log-Gaussian Cox, and
Hawkes processes. We also show that our VAE framework can be used efficiently in both
Bayesian and non-Bayesian inference schemes.

255
11.5 Methods and Theory: Inferential Machine Learning - Accelerating Statistical
Methodology through ML
Thursday 5 September 3.30pm

A unified machine learning approach to time series forecasting applied to emergency

department demand

Michaela Vollmer, Samir Bhatt, Seth Flaxman, Graham Cooke, Ben Glampson, Luca
Mercuri, Swapnil Mishra
Imperial College London,

There were 24.8 million attendances at Accident and Emergency (A&E) departments in
England in 2018-2019 corresponding to an increase of 11 million attendances in the past ten
years. This steadily rising demand at A&E departments leaves the NHS under constant
pressure to provide a good quality of care while being highly productive. Managing hospital
demand efficiently requires an adequate knowledge of the future rate of admission. Using
admissions data for the past eight years from St Mary's and Charing Cross Hospitals in
London, we developed a novel predictive framework to understand the temporal dynamics
of hospital demand and we applied an exhaustive statistical analysis to daily presentations at
the A&E departments at St Mary’s and Charing Cross Hospitals, evaluating a range of
standard time series and machine learning approaches. Ultimately, we applied a novel
ensemble methodology to combine the outcomes of the best performing time series and
machine learning approaches in order to make very accurate forecasts of demand, 1, 3 and
7 days in the future. We found that while St Mary's and Charing Cross Hospitals both face
an average daily demand of 208 and 106 respectively and despite considerable volatility
around this mean, we were able to predict attendances at these emergency departments up
to a mean absolute error rate of 7 patients corresponding to a mean absolute percentage
error of around 3% which may help to save up to £900,000 at the A&E department at St
Mary’s Hospital alone every year.

256
11.6 Communicating & Teaching Statistics: Do we need a Q-Step initiative for
statistics training in biology, medicine and health?
Thursday 5 September 3.30pm

Adapting the Q-step scheme for use in undergraduate Medicine: What would we need
to do?

Margaret MacDougall
University of Edinburgh

The success of the Q-step scheme in improving employability for social science graduates
intent on pursuing careers requiring quantitative research skills has been recognized. This
resource-intensive venture has been made possible through generous funding from the
Economic and Social Research Council. The cross-institutional nature of the work provides
scope for exploring a sustainable model where good practice and content from educational
programmes are shared across institutions. The relevance of quantitative learning is
reinforced through provision of Q-step work placement schemes, which invite students, as
employees, to experience authenticity when applying quantitative, including statistical, skills
to solve genuine real-life problems. This paper briefly explores the ‘why?’ and ‘how?’ of
developing a similar initiative in statistical learning, with a focus on undergraduate medical
education. This will provide an opportunity to discuss the key players in approving
requirements for entry to medical school and determining required learning outcomes for UK
medical schools to deliver in ensuring the clinical competence of their new graduates. In
supporting the case for a medical statistics variant of the Q-step scheme, I will suggest
recently reported funded research and recommend, and open the floor for further discussion
on, potential funding sources to explore. However, I will also bring to the table an honest
impression of some key challenges which need to be tackled collectively across prospective
participating institutions. Further, I will acknowledge the opportunities and challenges
presented by a crowded medical curriculum, particularly in relation to guaranteeing
placement opportunities with adequate statistical content for all. Additionally, I will highlight a
range of resource challenges in attempting to adopt an integrative approach to statistical
learning throughout the medical curriculum, not to mention the need to convince curriculum
managers that the efforts are worth investing in. Delegates will be welcome to share their
viewpoints in response to the above proposal.

257
11.6 Communicating & Teaching Statistics: Do we need a Q-Step initiative for
statistics training in biology, medicine and health?
Thursday 5 September 3.30pm

How could we design a Q-Step initiative for biology, medicine and health?

Jamie Sergeant
University of Manchester

If we accept the premise of the session’s title, that we do need a Q-Step initiative for
statistics training in biology, medicine and health, or at least that such an initiative would be
desirable, how might we go about designing one? This part of the session will explore ideas
and preferences from delegates on what the objectives of a Q-Step-type initiative in this field
might be and how the statistics community could work together to achieve them.

258
Poster viewing A
Wednesday 4 September 1.30pm

Bayesian Networks to assess the impact of 20mph speed limit policies on road traffic
collisions and casualties

Glenna Nightingale1, Ruth Hunter2, Frank Kee2

1
University of Edinburgh, 2 Queen's University Belfast

This study involves the use of Belief Networks for the evaluation of the impact of the 20mph
speed limit policy in the cities of Edinburgh and Belfast. Belief Networks are independence
probability graphs which consist of nodes (key variables) and arcs (hypothesized links
between key variables). We propose the use of multiple logic models in combination with
the results from a “contribution trial” to construct these Belief Networks. Essentially, a
contribution trial is a workshop which brings together key domain experts to seek their views
on the causal pathways inherent in the project’s logic models. The Belief Networks will be
updated using empirical “evidence” from our speed/casualty/collisions data streams to
generate posterior probabilities of key outcomes. We aim to perform model discrimination in
a Bayesian framework to determine which Belief Network or cluster of Belief Network
receives the highest probabilistic support. The Belief Network with the highest probabilistic
support, and highest practical significance will be used for inference. We will consider
gateway strategies (Dmitrienko et.al, 2011) where relevant as much as possible. The Belief
Network with the highest probabilistic support and practical significance will be incorporated
into a Web app (using R Shiny) to make the compiled network accessible to policy makers
and public health professionals. Empirical evidence such as change in average speed, or
increased speed limit enforcement can be entered into the app, and the change in the
conditional probabilities of key outcomes such as traffic collision rates and casualties are
outputted.Overall we aim to develop a robust statistical tool using domain expert information
(in a contribution trial) and Bayes probabilities to equip decision makers in assessing the
impact of the complex intervention of the 20mph speed limit policy. The method developed
will be encapsulated in a Web app making the method easily accessible. The initial
prototype of the proposed app is now available for discussion.

259
Poster viewing A
Wednesday 4 September 1.30pm

GEOTECHNICAL PROPERTIES OF SOIL IN SOLID WASTE DUMP SITE

Micheal Olobadola
Funaab

Soil samples at various depths from different sampling points on an open dumpsite were
assessed in order to investigate the impact of municipal solid waste dumping activities on
some geotechnical properties of soil. The geotechnical parameters, which include moisture
content, permeability, grain size analysis, compaction and Atterberg limits, were assessed
using ASTM standard. Results showed that geotechnical parameters for dumpsite soils are
higher than the control samples. The liquid limit (LL) and plasticity index (PI) ranged from 26-
38 and 11-20 respectively for dumpsite soil while the mean values of LL and PI for control
soil samples were 31 and 16 respectively. Soil moisture content was significantly higher in
dumpsite soil (25.8 – 33.6%) than control soil (10.7%). The percentage passing (N200) sieve
test ranged from 2.9 – 19.5% for dumpsite soil and 4.9 - 18.8% for control samples. The soil
under the control treatment had least permeability value of 0.0043 cm/sec while a maximum
value of 0.0058 cm/sec was obtained in the dumpsite. Maximum Dry Density (MDD) and
Optimum Moisture Content (OMC) for dumpsite ranged from 1.5 – 7.7g/cm3 and 1.4 –
29.11% respectively. However, the MDD and OMC values for control soil were 1.3g/cm3 and
6.6% respectively. The settlement (compressibility) analysis revealed that soil samples in the
study area are of low/medium degree of expansion and non critical/marginal degree of
severity status for stability purpose. The Compression Index (CC) values ranged from 0.12 –
0.25, suggesting low to moderate compressibility over an engineering time. Most of analyzed
geotechnical parameters values fall with the specification limits for construction purpose. The
result obtained also showed that dumping activities have great effect on soil’s geotechnical
properties. The obtained geotechnical parameters results for soil samples in the study area
showed

260
Poster viewing A
Wednesday 4 September 1.30pm

Data Fusion of Cardio-Metabolic Risk Prediction for Cognitive Impairment in the

Middle-Aged: the CARDIA Study

Sujin Kang
Imperial College London

Background: There is a lack of evidence of the role of cardio-metabolic risk factors in brain
health from large cohort studies which are comprised of metabolites; specifically, from a life-
course perspective with increasing CVD risk profiles.

Aim: The primary aim of this study is to investigate associations between cardio-metabolic
risk factors and cognitive impairment with brain MRI-derived parameters in the Coronary
Artery Risk Development in Young Adults (CARDIA) Study at Year 30, and to investigate if
targeted urinary metabolites are mediators of the associations. Significant epidemiological
exposures and targeted urinary metabolites will be examined, and the information will be
used to develop a risk assessment tool for brain health (i.e., cognition).

Design and population: This present study focuses on the CARDIA Study sub-cohort of 606
participants in middle-aged (aged 48-60 years who completed cognitive tests) with high-
resolution nuclear magnetic resonance (NMR) spectroscopy data, 340 of them with liquid-
chromatography mass spectrometry (LC-MS) data and 280 of them completed the brain MRI
data at Year 30, 2015-16.

Methods: Bayesian models and advanced network classification models will be employed to
investigate metabolic phenotypes. Different types of significant (1) clinical, socio-
demographic, lifestyle risk factors and (2) metabolite markers which will be integrated (i.e.,
fusion) to provide a cardio-metabolic risk-specific integrated map of brain health.
Longitudinal analysis of the associations will be used to explore the possibility of developing
a novel, comprehensive risk assessment tool of brain health.

Discussion: This study will employ a novel biostatistical approach of a fusion analytic method
for the different types or levels data on the large-scale epidemiological study. The approach
may be able to recognise complex multivariate epidemiologic, pathologic and phenotypic
relationships across the disease network that are yet unidentified.

261
Poster viewing A
Wednesday 4 September 1.30pm

Statistical Analysis on Industrial Sector of Nigeria Economy

Iyabode Favour Oyenuga

The Polytechnics, Ibadan, Nigeria

Industrial sector generates revenue, creates services, brings about incomes and create
employment. Over the years, the Nigerian economy has recorded a decline in industrial
growth with some industries recording closures as a result of difficult operating environment.
Few studies have examined the role of industry in economic growth of Nigeria. The study
evaluated the industrial sector of Nigeria gross domestic product with the use of Secondary
data over a period of 27 years from 1991 to 2017 extracted from Central bank of Nigeria
bulletin.Multiple regression method was adopted to analyze the data in order to test for
individual and joint parameter, heteroscedasticity, multicollinearity, autocorrelation and
predictive power. The results revealed that all the explanatory variables contribute
significantly to the model. There is a very strong positive correlation between the dependent
variable (Industry) and each of the explanatory variables (Crude Petroleum and Natural Gas,
Solid Minerals and Manufacturing). The model has high predictive power of 94.4%, that is,
94.4% of the total variation of the dependent variable is explained by the regression model.It
was recommended that government should encourage the growth of industries in Nigeria
through establishment of favourable policies for industrial development.

262
Poster viewing A
Wednesday 4 September 1.30pm

Modelling and monitoring the environmental impacts of aquaculture

Michael Currie1, Marian Scott1, Claire Miller1, Andrew Berkeley2

1
University of Glasgow, 2 Scottish Environment Protection Agency

Aquaculture accounts for nearly half of the global fish supply and is an expanding industry
around the world. Current farming techniques deposit a range of wastes on the seabed
which are dispersed, creating a zone of impact that is monitored to ensure that agreed
quality standards are met. In Scotland, the Scottish Environment Protection Agency (SEPA)
use a particle tracking model called NewDEPOMOD to predict the impact of aquaculture on
the seabed. When investigating the performance of NewDEPOMOD, two challenges are
presented: (i) limited observational data are available to validate the NewDEPOMOD
predictions and (ii) the computational cost of running NewDEPOMOD.There is some
uncertainty as to the suitability of some of the default parameter values within
NewDEPOMOD, and hence a sensitivity analysis has been developed to assess the
uncertainty of these parameters and the impact on the NewDEPOMOD predictions. this work
will then be used to inform the development of a statistical emulator to assess the
environmental impacts of aquaculture without the computational cost of NewDEPOMOD.The
sensitivity analysis was conducted at 2 sites with different physical properties. In
collaboration with SEPA, 11 parameters were identified for the sensitivity analysis, and
plausible ranges for their values were determined based on literature and expert knowledge.
Latin Hypercube Sampling was used to create 100 different sets of values for the
parameters, accounting for interactions that exist between some of the chosen parameters,
and 100 runs of NewDEPOMOD were performed to account for a random walk element
within NewDEPOMOD. A Random Forest approach was then used to investigate which
parameters described most of the variation in the NewDEPOMOD predictions.It is clear from
the sensitivity analysis that the physical properties of a site play a key role in the influence of
the parameters on the NewDEPOMOD predictions.

263
Poster viewing A
Wednesday 4 September 1.30pm

Response Surface Approach to Optimizing Baking Parameters for Improved Quality

Loaf with Local Content in Nigeria

Polycarp Chigbu
University of Nigeria Nsukka, Nigeria

Objective: A study was proposed to determine the optimum combination of imported wheat
and local cassava flours, baking time and temperature. The aim is to increase local content
in bread production in Nigeria while attaining acceptable quality of bread loaf. The three
variables under consideration are Baking time, (in minutes), Baking temperature, (in degree
Celsius) and Wheat/Cassava mix concentration, (in percentage). Three important qualities of
baked loaf that were to be monitored are moisture content, (in percentage), volume of loaf,
(in cm3) and plastic limit, (in cm).

Method: Response surface methodology was considered appropriate since the study
requires the optimization of baking factors and the central composite design was used for
that purpose. Different variations of the CCD involving partial replications were applied and
evaluated using the fraction of design space graphs to identify, among the numerous design
options, the design that gives the best spread of minimum variance of prediction of
responses throughout the design region. Also, five axial distances, , were considered: the
spherical cuboidal , practical , rotatable and orthogonal .

Results/Conclusion: The CCD with cuboidal axial distance and with the star portion
replicated twice is recommended for the experiment.

264
Poster viewing A
Wednesday 4 September 1.30pm

Detection of voids in additive manufacturing using x-rays in projection space

Sherman Lo1, Julia Brettschneider1, Thomas Nichols2, Jason Warnett3, Gregory Gibbons3,
Mark Williams3
1
University of Warwick, 2 Oxford Big Data Institute, 3 Warwick Manufacturing Group

X-ray computed tomography (CT) can be used for defect detection in additive
manufacturing. The object is scanned at multiple angles to reconstruct the object in 3D
space. The process can be time consuming. The aim is to investigate if it is possible to
conduct defect detection from a single scan to speed up the quality control procedure. An
experiment was conducted, a 3D printed sample was manufactured with voids to see if they
can be detected.Photons behave randomly. Hence to do defect detection pixel by pixel,
uncertainty must be taken into account. A compound Poisson model was used to model the
grey values in a pixel. It assumes that photon arrivals are a Poisson process with Gamma
distributed energy. This resulted in a linear relationship between the mean and variance of
the grey value, which can be used for variance prediction and to quantify the
uncertainty.Software was used to simulate the scan and it was compared with the x-ray
acquisition under the face of uncertainty. Each pixel was treated as a hypothesis test. The
software, and the information provided to it, was not perfect which led to model
misspecification and incorrect inference. The empirical null filter was proposed. The filter
locally normalise each test statistic using robust estimators of the mean and deviation. This
reduced the number of false positives. Sensible inference was achieved on the dataset
provided.

265
Poster viewing A
Wednesday 4 September 1.30pm

Geometric Characterizations and Symmetric Relations between Standard Normal

Distribution and Inverse Mills Ratio based on Pythagorean Theorem

Shingo Nakanishi1, Masamitsu Ohnishi2

1
Osaka Institute of Technology / Computing Center, 2 Osaka University / Graduate School of
Economics and Center for Mathematical Modeling and Data Science

We dealt with the geometric characterizations between standard normal distribution and
inverse Mills ratio by circle and square from the viewpoints considering the height of
densities such as the ancient Egyptian drawing pictures by using the Greek Pythagorean
Theorem. First, we can clarify that the general solution as the second order differential
equation of the integral form of cumulative distribution function of standard normal
distribution and that of Bernoulli differential equation of inverse Mills ratio are tied on the
rotational symmetric figure. Especially, we can also confirm the integrals of cumulative
standard normal distribution are related to inverse Mills ratio geometrically. Second, from
these tendencies, we can also get the geometric and symmetric modified intercept forms for
maximal profits of winners, these losses of losers, and their banker’s fee. We can
understand that these equations are changing such as Pythagorean Theorem correctly
based on the relations between the probability points and these probabilities. If the
probability is 0.612003 which was found by Karl Pearson, we can show you that it is the
special point of standard normal distribution such as Sir David Roxbee Cox’s proposal. The
right triangle with 3:4:5 is also the third quantile of standard normal distribution. Third, we
can also realize there are many similar tendencies close to the relations between circles and
squares such as Vitruvian man by Da Vinci and Mandalas although there might not be
related to normal distribution directly and historically. The ancient Egyptian drawing styles
enable us to illustrate symmetric relations and geometric characterizations between standard
normal distribution and inverse Mills ratio with circle and square based on Pythagorean
theorem. We think that our ideas shall be contributed in the statistical modelling and these
evaluation fields since our suggested figures might be more easily to be understood and
powerful than we thought.

266
Poster viewing A
Wednesday 4 September 1.30pm

When impairment is a handicap: higher education in Brazil

Kaizo Beltrao1, Hugo Simas2, Moema Teixeira2, Ricardo Goes2

1
EBAPE FGV, 2 Cesgranrio Foundation

In 2016, the Brazilian Government passed a Law (13409) defining quotas for ethnic groups
and individuals with disability. This Law was actually a revision of a previous legislation, Law
12721, enacted in 2012. Since 1995, Inep (a department of the Ministry of Education) has
been collecting data on University courses and students enrollment through the Brazilian
higher education census. From 2009 on, this census individually lists and identifies all
students enrolled in University and College courses with information on age, gender,
disability, enrollment status, race/ethnic group, majors, scholarship and selection method for
entry in the course (e.g. through a national government exam, private exam, quotas). With
the student identification number, it is possible to follow students across different years and
check if they are still enrolled, if they droped-out, if they got transferred to another course, if
they died, if they are licensed, if they graduated, among other possibilities.This text intends
to follow up students with different disabilities. The anual higher education census lists
thirteen different disabilities: blindness, partial visual impairment, Deafness, partial hearing
impairment, physical deficiency, deafness combined with blindness, multiple deficiency,
intellectual deficiency, Asperger syndrome, Autism spectrum disorder, Rett syndrome, and
so on. Drop-out rates and average time to completion for each disability group and the
overall population controlling for availability of in campus resources for those with special
needs. We find out that students with special needs do take, in average, longer to graduate
than other students, but the expected time is mediated by the existence of in-campus
resources and programs. Some regions in Brazil are better equipped with these resources.
Data is modelled using a survival model.

267
Poster viewing A
Wednesday 4 September 1.30pm

New Second-Order Asymptotic Methods for Nonlinear Models

Gubhinder Kundhi1, Paul Rilstone2

1
Memorial University of Newfoundland, 2 York University

New higher-order asymptotic methods for nonlinear models are developed. These include
generic methods for deriving stochastic expansions of arbitrary order, methods for evaluating
the moments of polynomials of sample averages, a method for deriving the approximate
moments of the stochastic expansions with simplified expressions for the first four moments
and a third-order approximate Saddlepoint expansion. These techniques are applied to
improve inferences with the Two-Stage Least Squares estimator and the weak instruments
problem. It is well established that Instrumental Variable (IV) estimators in the presence of
weak instruments can be poorly behaved, in particular, be quite biased in finite samples. In
our application, finite sample approximations to the distributions of the IV estimators are
obtained using Edgeworth and Saddlepoint expansions. Higher order analytical corrections
provided by these expansions are used to analyze departures from normality. In a Monte-
Carlo experiment, the performance of these expansions is compared to the first order
approximation and other techniques commonly used in finite samples such as the bootstrap.

268
Poster viewing A
Wednesday 4 September 1.30pm

ON THE USE OF VARIABLE TRANSFORMATION IN ESTIMATING POPULATION

PARAMETERS OF THE STUDY VARIABLE

Chinyeaka Izunobi, Aloysius Chijioke Onyeka

FUTO, Federal University of Technology, Owerri. Imo State, Nigeria

This work is based on the use of variable transformation in estimating population parameters
of the study variable in Stratified random sampling scheme. The properties of the estimators
were obtained. Numerical illustration was carried out to verify theoretical results in the
literature and to compare the efficiencies of some selected estimators that utilized variable
transformation. Results of the study revealed that the use of variable transformation often
leads to increased efficiency over the sample mean estimator. It was also observed that
when using variable transformation, increased efficiency could be achieved by increasing the
number of transformed variables to be used in a given survey.

269
Poster viewing A
Wednesday 4 September 1.30pm

Modelling Baseline Hazards in a Cox Model using Gaussian Processes

Bindu Vekaria, Glen Martin, Thomas House

University of Manchester

There is growing interest in developing prognostic models that produce absolute measures
of risk for an outcome of interest (e.g. time-to-death). These provide an effective way of
communicating risks with patients and aid decision-making. In the context of time-to-event
outcomes, such models are usually based on the Cox model, where parametric distributions
(e.g. Weibull) are used to model the baseline hazard. However, these lack the flexibility
needed to model complex time-varying hazards. Alternatively, flexible parametric methods
model the baseline hazard using cubic splines, but such models are difficult to validate,
rarely used in practice and difficult to interpret.Joensuu et al. (2012) proposed an alternative
method utilising a Bayesian framework by modelling the log baseline hazard with a Gaussian
Process. While this is attractive since it models the baseline flexibly, it has only been applied
to scenarios where one can assume a piecewise constant hazard. Generally, this is not
representative of clinical settings.Therefore, we propose a novel piecewise linear Gaussian
Process, which can be extended to higher order polynomials, where necessary. Here, I
present the results from a simulation study that compares the predictive accuracy of our
method with existing methods, across a range of representative scenarios.

270
Poster viewing A
Wednesday 4 September 1.30pm

R: A Hitchhikers Guide to Reproducible Research

Brendan Palmer, Darren Dahly, Joseph Eustace

HRB Clinical Research Facility Cork

As funding agencies increasingly require adoption of open science practises in the research
that they fund, there is a pressing onus on researchers to respond. For many, awareness of
the tools, skillsets and resources available to achieve this is lacking. University subscriptions
to statistical analysis software, such as (but not limited to) SPSS, restricts the end user to a
defined suite of operations and functions, that may not be accessible at a later stage in the
research project or the researchers career. Furthermore, such programmes do not allow for
reproducible research to be faithfully applied.Conversely, the R programming language is
free to download, install and use. Together with the graphic user interface, RStudio, it is now
the favoured data analysis tool across a variety of fields, including statistics. Advances in R
software development, namely the reorganisation of popular R packages under the umbrella
banner of the “tidyverse”, has made R accessible to a wider audience. This has numerous
benefits, chief amongst them being that R, as a programming language, has become more
human readable, intuitive and accessible. Data importation, cleaning, exploration,
visualisation, analysis and reporting can be enacted through R. Combined, R allows for
reproducible workflows to be enacted, packaged and published.We conclude that senior
academics and programme coordinators adopt R modules as a core element of academic
programmes aligned with research activities. We propose this be implemented through
structured PhD training modules or as part of early career researcher continuing professional
development.

271
Poster viewing A
Wednesday 4 September 1.30pm

Variance Estimation in Dual Channel Microarray Applied on Split-Plot Design

Abimibola Oladugba, Ifeoma Ogochukwu Ude

Department of Statistics, University of Nigeria, Nsukka

A well-designed experiment is an important and fundamental step needed in planning a dual

channel microarray experiment used in measuring gene expression. Dual channel
microarray applied on split-plot design is complicated, due to the fact that important effects
may be confounding in the array during the process of pairing the samples. Forming an
appropriate model and using sufficient replication for estimating effects confounded with
array overcomes this challenge. In this work, three methods of sample pairing which we
called vertical loop method (design A), cross loop method (design B) and horizontal loop
method (design C) were used in pairing samples in a dual channel microarray performed in a
split-plot design in order to ascertain which method gives the minimal variance for the effects
of interest. Also, the numbers of replication were varied in order to check its effect on the
estimated variance. The brute force and analysis of variance methods were used in
estimating the variance estimates and components in each of the designs. The results
showed that design A had the least variance for comparing the mean difference between the
sub-plot-treatment, color, first and second levels of treatment B at each level of treatment A
and first and second levels of treatment A at each level of treatment B while design C had
the least variance for comparing the whole-plot treatment and the variance estimated
decreases as the number of replication increases. We therefore conclude based on the
results that the vertical loop method of sample pairing gave the least minimal variance for
comparing the effects of treatment comparison.

272
Poster viewing A
Wednesday 4 September 1.30pm

Major Errors in Research on Lighting and Public Safety

Paul Marchant
Leeds Beckett University

The poster gives a couple of examples of where statistical errors give rise to claims of
detecting very large beneficial effects in an area of public policy. However, the serious errors
committed coupled with the absence of checkable data, make such positive claims suspect.
The poster points out where the major errors reside and indicates, through detective work,
the consequences for the claims.

273
Poster viewing A
Wednesday 4 September 1.30pm

Interpretable drivers of sensitivity analysis for non-ignorable missing covariate in

linear regression models

Peng Yin1, Rong Zhu2, Jianqing Shi2

1
Chinese Academy of Sciences, 2 University of Newcastle upon Tyne

Goodness-of-test is not valid for non-ignorable missing data problems because of non-
identiability. The model structure for missing data is very complex and in many cases, both
ignorable and non-ignorable missingness mechanisms are plausible. Sensitivity of missing
data mechanism demands serious circumspection. In this paper we will utilize a likelihood-
based analysis platform to investigate the sensitivity of missing data mechanism (MDM)
modelling in the missing covariate problems. Our analysis will adopt an approximate
measure of incomplete data bias for parameter estimations, which characterizes the
magnitude of uncertainty and its gradient from a specied model. We will discuss the
sensitivity measures in linear regression models with missing covariate specically. How to
identify key interpretable sensitivity parameters is very important and this will be addressed
in quite details. Our discussion will show a well approximated and easy interpretable
formular for measuring incomplete data bias. Numerical examples will be presented in
simulation studies and real data example.

274
Poster viewing A
Wednesday 4 September 1.30pm

A randomized controlled trial of a school-based smokeless tobacco intervention to

improve knowledge and prevent adolescent tobacco use

Shafquat Rozi1, Nida Zahid1, Talat Roome2

1
Department Of Community Health Sciences, Aga Khan University, Karachi, Pakistan, 2
Dow University of Health Sciences, , , , , , ,

Objectives: To assess the effectiveness of intervention in improving knowledge, attitude and

perception regarding smokeless tobacco (SLT) use and its harmful effects and intention to
quit SLT among school going adolescents.

Methods: A school-based cluster randomized control trial was carried out in 18 secondary
schools targeting male and female students from grades 6 to 10 in Karachi. Intervention
comprised of a 30 minutes power point presentation, two posters, and one pictorial booklet
and a video game on the hazards of use of various tobacco products. Primary outcome was
knowledge about hazards of smokeless tobacco (SLT) and secondary outcome measures
were 1) Attitude about hazards of SLT and 2) Perception about hazards of SLT.

Results: We enrolled 738 participants in intervention group and 589 in the control group.
Mean score of knowledge significantly improved in intervention as compared to control group
(p value <0.01). Intention to quit was found to be proportionately higher (33%) in the
intervention group as compared to control group. Generalized estimating equations (GEE)
were used to assess the association of various factors with knowledge regarding harmful
effects of SLT use. Significant predictors of increase in knowledge score were found in
children: who had seen any anti SLT messages on social media in the past 30 days, who
were getting information regarding harmful effects of SLT use in school or textbooks and
who had friends using SLT.

Conclusions: A school-based intervention was effective in increasing knowledge regarding

the harmful effects of SLT use and intention to quit smokeless tobacco use among school
going adolescents. Introduction of such educational programmes on a regular basis in
schools or as part of school curriculum can have an impact on reducing prevalence of SLT
use.

Key words: Smokeless tobacco, cluster randomized trial, adolescents, school based
intervention

275
Poster viewing A
Wednesday 4 September 1.30pm

Multilevel weighted ordinal regression analysis to determine the association of

bullying with depression symptoms among school adolescents of Pakistan

Shafquat Rozi, Maryam Ali, Apsara Ali, Wajeeha Zahid

Department Of Community Health Sciences, Aga Khan University, Karachi, Pakistan

Purpose: In poor resourced country, youth violence is the most neglected problem to be
focused because of its low reporting rate and less awareness of it effect on psychological
wellbeing of adolescents. The aim of this study is to determine the association of bullying
status and other factors with depression symptoms among school going adolescents of
Pakistan.

Methods: We used secondary data from The Global School-based Student Health Survey
(GSHS) conducted in 2009. This was a school-based cross-sectional survey conducted on
nationally representative sample of adolescents from 8th to 10th grades with age 11–16
years. A two-stage cluster sampling technique was employed for selection of school and
children. The outcome variable was presence of depressive symptoms whereas, the primary
exposure was bullying victimization experienced by adolescents in last 30 days. Multilevel
weighted ordinal regression analysis was performed.

Results: A significant proportion of severely (48%) and moderately (46%) depressed

adolescents reported being bullied. Whereas, three fourth of mildly/non-depressed
adolescents (70%) reported having no exposure to bullying in past 30 days. A significant
interaction was found between bully and parent-child relationship indicating that adolescents
exposed to long term bully and have poor parent-child relationship had 3 times higher odds
of depression (OR: 3.35, 95% CI: 1.95, 5.75) as compared to adolescents with no exposure
to bully with a good parent-child relationship.

Conclusion: This particular study has identified the potential risk factors that will help in
formulating preventive strategies to address the toll of depressive symptoms experienced by
adolescents.

276
Poster viewing A
Wednesday 4 September 1.30pm

Complex interaction analysis by liquid association and its extension

Kerchau Li
Institute of Statistical Science, Academia Sinica

Data science has been fostered by the rapid accumulation of a growing number of large
open databases, each containing tens of thousands of variables pertinent to the knowledge
advance of its specific scientific domain. For example, in cancer research, the Cancer
Genome Atlas(TCGA) consortium has completed its mission of molecular-profiling over
11,000 tumors from 33 of the most prevalent forms of cancer. Here a general statistical
framework to enhance the information-distilling from the complex patterns of interaction
between variables is presented. We have earlier introduced Liquid Association (LA) for
inferring higher order of association between variables in a system (Li 2002, Proceedings of
National Academy of Sciences, USA). LA was originally introduced to study patterns of
gene-gene interaction that involve three genes at a time. LA aims at finding how the
correlation pattern between a pair of functionally associated genes may be dynamically
altered due to the influence of ever-changing but often-hidden cellular state. It is
computationally expensive to compute LA scores for all possible triplets in very large
datasets, but Graphic Processing Units (GPUs) can be used to speed up the LA computing.
There are several ways of extending LA, including background adjustment and dimension
reduction. We shall demonstrate how they can be used to exploit complex interaction and
broaden the scope of statistical multivariate analysis.

277
Poster viewing A
Wednesday 4 September 1.30pm

Likelihood approximation and prediction for large spatial and spatio-temporal

datasets using hierarchical matrices

Anastasiia Gorshechnikova1, Carlo Gaetan2

1
University of Padova, 2 Ca' Foscari University of Venice

Large datasets with irregularly sited locations are difficult to handle for several applications
of Gaussian random fields such as maximum likelihood estimation (MLE) and kriging
prediction due to a high computational complexity. For relatively large spatial and (or)
temporal dimensions the exact computation becomes unfeasible and alternative methods
are necessary. Several approaches have been proposed to tackle this problem. Most of
them assume a specific form for the covariance function and use different methods to
approximate the resulting covariance matrix. The aim is to approximate the covariance
functions in a format that facilitates the computation of MLE and kriging prediction with large
spatial and spatio-temporal datasets. For a sufficiently general class of spatial covariance
functions, a methodology is developed using hierarchical matrices. This technique involves
the partitioning of a matrix into sub-blocks according to specific given conditions with the
further approximation of the majority of these sub-blocks by low-rank matrices. The resulting
covariance matrix allows for computation of the matrix-vector products and matrix
factorisations in a log-linear computational cost followed by an efficient MLE and kriging
prediction. Numerical experiments are performed on simulated spatial and spatio-temporal
data to recover the true values of the parameters of the covariance matrix. This method is
further applied on a real dataset consisting of measurements of atmospheric carbon dioxide
mole fractions. The prediction accuracy and computational time are compared with other
methods. As a result, among the methods considered in this study, the approach presented
is the most efficient in terms of the root mean-squared prediction error and computational
time.

278
Poster viewing A
Wednesday 4 September 1.30pm

Predictive Modeling for Nursing-care Levels and Costs

Yasutaka Hasegawa1, Hideyuki Ban1, Tetsuya Moriike2, Hayato Nishikawa2, Satoshi Horie2
1
Hitachi,Ltd. Research & Development Group, 2 Hitachi Ltd. Government & Public
Corporation Information Systems Division

Background: The growth of nursing-care cost, which is caused by aging population, is one of
the most critical issues in Japan. Each local government is required to plan effective nursing-
care prevention measures utilizing healthcare data, and it is important to plan prevention
measures that consider not only the current nursing-care situation but also the future.

Objectives: In this study, we created a model to predict future nursing-care levels and costs
using nursing-care, medical and checkup data.

Method: First, using machine learning, we created a predictive model for future nursing-care
levels from 487 explanatory variables included in the data. Next, using a linear regression,
we created a model that predicts future nursing-care costs from the prediction result of
nursing-care levels and current nursing-care costs. To create a model for future nursing-care
levels, the following seven machine learning methods were tried; (a) Binomial logistic
regression, (b) Multinomial logistic regression, (c) Ridge regression, (d) Lasso regression,
(e) Elastic Net, (f) Gradient boosting, and (g) Support vector machine.

Results: We used data of 405,687 people from 2014-2015 to create predictive models. As a
result of the evaluation using test data of 81,136 people, it was confirmed that model (a), (d),
(e), and (f) can predict nursing-care levels after one year with more than 95% accuracy and
that nursing-care costs after one year can be predicted with less than 1% error. In addition,
the model (a) can analyze explanatory variables contributing to the occurrence and progress
of nursing-care, and it found that checkup and nursing-care services reduce the occurrence
and progress of nursing-care.

Conclusions: The model (a) is a high accuracy model that can explain factors that contribute
to the occurrence and progress of nursing-care, and confirmed that this model is the most
appropriate model for planning nursing-care prevention measures in the local government.

279
Poster viewing A
Wednesday 4 September 1.30pm

HyCOSM: a Hybrid-COupled-Stochastic Model for subseasonal climate prediction

Michel d. S. Mesquita12, Michel d. S. Mesquita, George U. Pedra3, Saurabh Bhardwaj4,

Santosh K. Muriki4, Lincoln M. Alves3, Iracema F. A. Cavalcanti3
1
Bjerknes Centre for Climate Research, 2 Future Solutions, Norway, 3 National Institute for
Space Research, INPE, Brazil, 4 The Energy and Resources Institute, TERI, India

The Intraseasonal scale is particularly important in climate science because it acts as a

modulator of weather systems. It is also one of the most difficult time scales to predict due to
the strong influence of large-scale atmospheric systems. We have developed a subseasonal
forecast model called HyCOSM: a Hybrid-COupled-Stochastic Model for subseasonal
prediction. HyCOSM allows for the prediction of extreme precipitation of up to 28 days. It
consists of three distinct steps: the first (S1) is the characterization of the intraseasonal
climatic pattern and the development of a diagnostic index through Maximum Covariance
Analysis (MCA); we call this step the Multivariate Intraseasonal Rainfall Index (MIRI). This
index is made of two scalar components, which are then modelled in the second phase (S2)
using a Vector Autoregressive (VAR) model. For the third step (S3), we have used Artificial
Neural Networks (ANN) to obtain the intraseasonal rainfall forecasts. Results from S1 show
that the MCA output captures 71% of the intraseasonal rainfall variability in India and 65%
over South America; it also represents reasonably well the rainfall signal, as well as the
atmospheric circulation and convective patterns; Furthermore, it describes the evolution
mechanism of atmospheric systems in both space and time, making it possible to create a
projection in a space-phase diagram. In S2, the VAR model captures the variability of each
scalar component up to 28 days in advance. The correlation coefficient between the forecast
and the observed index is larger than 0.7 (p < 2.2e-16). In S3, the ANNs provide a 28-day
rainfall forecast of extreme precipitation. Overall, the information generated through
HyCOSM can significantly help decision makers and farmers regarding extreme events,
short-term precipitation trends, and the evaluation of other rain-related variables.

280
Poster viewing A
Wednesday 4 September 1.30pm

Multi-step Regression Model for Estimating Hospital Length of Stay

Hirofumi Kondo, Kaoru Kobayashi, Yasutaka Hasegawa, Hideyuki Ban

Hitachi,Ltd. Research & Development Group

[Background and Objectives] The growth of medical cost caused by aging and increasing of
chronic disease patients has become one of the critical issues in the developed nations. To
solve this isuue, measures focusing on health promotion and disease prevention are
important. This time, in order to visualize disease risk and utilize it for the health business,
we developed a model that predicts the number of hospitalization days for the 8 major
lifestyle-related diseases based on medical check-up results and past medical history.

[Methods] There is a problem that the average and variance of hospitalization days differ
greatly depending on the characteristics of the disease. Therefore, we used multi-step
regression combining a logistic regression model using binomial distribution that predicts the
occurrence of hospitalization for each disease and multiple regression that predicts the
number of days of general hospitalization based on hospitalization risk values for eight
diseases. Using a data of about 101,466 people(Age 35~69) that can be tracable for 8
years, we have built a model to predict the Total number of hospitalization days in the future
5 years. However, cases where the total hospitalization days exceeded 100 days were
excluded because they were considered to be largely affected by unobservable factors.
Diseases targeted in this study are malignant neoplasms, cerebrovascular disease,
cardiovascular disease, diabetes, pancreatic disease, liver disease, hypertension disease
and kidney disease.

[Results] As a result of evaluation by 5 cross validation methods, it was confirmed that

hospitalization days in the future 5 years can be predicted with 4.1% as an average absolute
error (s.d. 2.9). Moreover, in the first layer model that predicts the occurrence of
hospitalization, it was confirmed that the discrimination performance by AUC was more than
0.7 or more in each disease(e.g. 0.8 for cardiovascular disease).

281
Poster viewing A
Wednesday 4 September 1.30pm

Comparison of two stage and one stage meta-analysis of joint longitudinal and time-
to-event data

Maria Sudell, Catrin Tudur Smith, Ruwanthi Kolamunnage-Dona

University of Liverpool

Context: Meta-analysis of joint longitudinal and time-to-event data is a growing area of

research. Recent publications have suggested methods for one or two stage meta-analysis
of such data. However (for meta-analyses generally) under certain circumstances, results
from one and two stage analyses can differ (an example being that one stage methods
calculate exact likelihoods, whereas two stage methods calculate approximate likelihoods).
Given the complex nature of joint longitudinal and time-to-event data, it is important to
establish how results from one or two stage approaches could differ, the extent of any
potential discrepancy, and the reasons for any discrepancies, whilst establishing guidelines
for best practice.

Objective(s): This research will compare results of one and two stage meta-analyses of joint
longitudinal and time-to-event Individual Participant Data (IPD) under a range of conditions,
and give recommendations for future research.

Method(s): Multi-study joint longitudinal and time-to-event data will be simulated under a
range of conditions including differing levels of heterogeneity and association structure. One
and two stage meta-analyses of these simulated datasets will be conducted, and the results
compared. Code to fit the models assessed will augment currently available joint modelling
software packages in R. Discrepancies between results from one and two stage approaches
will be identified and discussed, with reference to potential reasons for the discrepancies. A
demonstration in a real multi-study hypertensive dataset will also be presented.

Results: Results and conclusions will be presented and discussed at conference. Guidelines
for future meta-analyses of joint longitudinal and time-to-event IPD will be presented.

282
Poster viewing A
Wednesday 4 September 1.30pm

Estimation of Hospitalization Occurrence Using a Neural Network Based on

Longitudinal Clinical Features

Kaoru Kobayashi, Hirofumi Kondo, Yasutaka Hasegawa, Shuntaro Yui, Hideyuki Ban
Hitachi, Ltd

Japanese Health Insurance Societies provide health services aimed at promoting health and
preventing lifestyle-related diseases for their members. For effective health services, health
condition of the members should be grasped for past, present and future based on
healthcare data accumulated the society. Although we have developed prediction technique
for future lifestyle-related disease onset and medical cost of that, prediction of hospital
admission is also important because hospitalization has a large financial burden on the
members. Therefore, in this study, we developed a predictive model for hospitalization by
diabetes whose complication tends to be severe using longitudinal data of medical checkup
and past medical history. It is difficult that applying well known time series analysis methods
to the longitudinal data because the data is yearly data in other words the number of the
data points observed is very small. Therefore, in our proposed method, firstly we generated
longitudinal features which represent transition of health conditions from the longitudinal
data. Secondly, in order to perform efficient and effective feature extraction from large
dimensional explanatory variable vector consists of the longitudinal data and the longitudinal
features, the vector is divided based on medical knowledge into several groups. And then,
the divide vectors are compressed and hospitalization occurrence is predicted by multi
modal neural network. The proposed method applied to the 60,000 people's longitudinal
data which consists three years data for explanatory variables and five years data for
objective variables. As a result, the AUC of the proposed method was 0.90 and the number
of false positive reduced 12% from the GLM modeled using the longitudinal data. From the
result of investigation, this is because that the proposed method is more accurate than the
GLM especially for members whose health condition is worse non-linearly. In conclusion, it
was shown that the transition of health condition contributed to the improving accuracy and
the interpretability.

283
Poster viewing A
Wednesday 4 September 1.30pm

HealthyR Notebooks: Democratising open and reproducible data analysis

Riinu Ots
The University of Edinburgh

Introduction: An urgent priority for medical science is the empowerment of clinicians to

robustly perform their own reproducible data analyses. Traditionally, data analysis has
required complex and expensive software run on costly computer equipment. Furthermore,
R is traditionally known for having a steep learning curve. Recent developments in the R
ecosystem have made statistical programming more accessible, allowing us to efficiently
teach R to non-statisticians.

Methods: “Data notebooks” combine analysis and reporting in one document to provide an
intuitive interface between the researcher and their data analysis; this interface is the future
of open and reusable data analysis. Our solution called “HealthyR Notebooks” harnesses the
RStudio, R Markdown and the tidyverse packages, giving users access to state-of-the-art
data science algorithms.

Results: We have modified our established and successful course, “HealthyR: R for
Healthcare Data Analysis”, into a globally scalable, open access resource. This format not
only provides the necessary training foundation but delivers participants a fully functioning
toolbox with which they can immediately commence their own data analyses. The transition
from traditional to literate programming has been well received by pilot attendees. Take-
home message: R Notebooks are an intuitive and efficient way to teach and use R, the
HealthyR Notebooks will be a freely available resource by the end of this year.

284
Poster viewing A
Wednesday 4 September 1.30pm

Creating an online mathematics and statistics learning community

Rachel Hilliam, Gaynor Arrowsmith, Alexander Siddons, Derek Goldrei, Cath Brown
The Open University

There is an increasing focus within Higher Education on the wider student experience.
Student engagement and a well-developed community are associated with greater levels of
retention student satisfaction and success. It is therefore crucial not only to support students
at all stages of their student journey, but also to create a space where they can benefit from
peer support and interact with the wider mathematics and statistics (M&S) community.The
School of Mathematics and Statistics at The Open University has a long tradition of
engaging students outside the ‘classroom’ environment and has needed to adapt how this
support has been provided as the number of face-to-face opportunities has diminished. One
such initiative was to provide a subject forum with the specific remit of offering advice on
study planning – the module advice forum.The forum is hosted on an M&S Study website
which provides dedicated resources to help inform module choice and study planning, which
are generally not easy to access from pan-university websites. This includes lists of
examination results and student satisfaction ratings on individual M&S modules. These have
been collated over several years, so that students can see trends in order to help inform
their decisions.Resources are also available for students to self-assess their readiness to
start their next module, along with targeted advice about topics which they might need to
revise or alternative modules to consider. Careers advice and guidance specific to M&S is
also provided along with links to accreditation bodies such as the IMA and RSS.Discussions
in the forum have flourished creating a true learning community. This space also provides
students with the opportunity to engage more fully in issues such as curriculum development
and delivery of student support leading to improvements in the structure of M&S
qualifications, influencing the content of new modules, more effective assessment strategies,
and better ways of supporting students.

285
Poster viewing A
Wednesday 4 September 1.30pm

Making the most of intensive longitudinal data

Tanja Krone
TNO

We want to know everything that happens all the time, especially when doing research. One
way to adhere to this wish, is Ecological Measurement Assessment (EMA, also known as
Experiences Sampling Measurements). Using EMA, people are asked to answer questions
multiple times a day or week based on timepoints (e.g., every two hours) or on events (e.g.,
each meal), are asked to wear sensors (e.g. black carbon) or to measure variables
themselves (e.g., glucose). This gives a rich dataset combining different kinds of
measurements over time. However, the data is always less then perfect: up to 80% missed
measurements occur, due to forgetfullness, drop-outs and faulty equipment. Furthermore,
the designs are often based on convenience, allowing for collection on workdays, or
whenever a participant ‘feels like it’. Last but not least, several kinds of data-streams may be
combined, such as heart rate and questionnaires, measured respectively every minute and
three times a day. As such, EMA data poses an enormous challenge, both in seeing its
possibilities, and working within its limits.We aim to create a protocol advising on how to
collect and analyze EMA data, taking into account the different study designs, the preferred
options of the applied scientist that needs to work with them, the used methods, combining
different data-streams, and the handling of missing data. To this end, we will simulate
datasets that are practically feasible to collect yet statistically optimal in the data-richness,
and analyze them with several different techniques, combining practical research design with
innovative statistical modelling. Our aim is to compare both the designs tested and the
methods used, to find the optimal data-collection design and statistical analyses for the
diverse range of EMA data. We will present the detailed plans on how we handled this, the
designs and methods considered, and (hopefully) our first results.

286
Poster viewing A
Wednesday 4 September 1.30pm

Parallel Computation for Seeded MDS Based on K-means Clustering

Xiangyu Meng1, Jian Huang1, Finbarr O' Sullivan1, Liam Ó Súilleabháin2, Zhaoyan Xiu1
1
University College Cork, 2 Division of Research at Kaiser Permanente

Parallel Computation for Seeded MDS Based on K-means ClusteringMultidimensional

Scaling (MDS) is a generic name for a family of dimension reduction algorithms used to
configure points in low dimensional space to maintain pairwise distance between them as
accurately as possible. Although MDS is powerful for identifying the underline pattern of high
dimensional data, the time and space complexities are at least quadratic in the number of
points, N, which makes it computationally difficult or impossible for data sets when N is very
large. To solve this problem, we developed a seeded MDS approach. We also examine the
possibility of a seeded approach to use the parallel computation to speed up the
computation of the proposed seeded MDS.Our approach consists of three steps:Step1. An
MDS analysis is carried out to map n = K + 1 cases to a K-dimensional space as seeds. A
seed case can be selected as cluster centres obtained by clustering analysis of the
data.Step2. Using the seed values as a reference, an optimization procedure is developed
to map each non-seed case so that its distances to the seed points in the K-dimensional
space closely match its corresponding distances from these cases. Step3. A segmentation
method is applied to the dataset. Then we use a multithread method with the segmented
data in R to accelerate the optimization process by parallel computations. The accuracy of
the proposed algorithm is evaluated based on the cMDS method by total distance difference
- accumulating the distance of corresponding points, estimated by our approach and cMDS
respectively in K-dimensional space. The algorithm is implemented in R and demonstrated
on simulated and real datasets. The memory usage and computation time and accuracy as a
function of the number of cases N will be reported. The relative performance of different
approaches to the selection of seed cases will also be reported.

287
Poster viewing A
Wednesday 4 September 1.30pm

Reporting of planned power and statistical design in published digital education

intervention studies for health professionals: a cross-sectional analysis

Ram Bajpai1, Josip Car2

1
Research Institute for Primary Care & Health Sciences, Keele University, 2 Centre for
Population Health Sciences (CePHaS), Lee Kong Chian School of Medicine, Nanyang
Technological University, Singapore

Objective: The aim of this analysis was to assess the reporting of planned power and
statistical design in published digital education intervention studies for health professionals.

Methods: We performed a systematic search of Medical Literature Analysis and Retrieval

System Online, Excerpta Medica database, Cochrane Central Register of Controlled Trials
(Cochrane Library), PsycINFO, Cumulative Index to Nursing and Allied Health Literature,
Education Resources Information Centre, and Web of Science published between 2007 and
2016. A total of 874 intervention studies were identified and categorised into online-offline,
mobile digital education, and simulation-based modalities for pre and post-registration health
professions’ education. Of these, 242 studies were randomly selected for methodological
review and thematic analysis.

Results: Nearly two-thirds (63.2%) of studies published after 2011 and primarily from north
America (47.9%). More than two-third of these studies were parallel design (88%), single
centre-based (95%), and published in subject-specific journals (70.3%). Median number of
study authors was 5 (interquartile range [IQR]: 4 to 7), and very few studies (6.6%) reported
statistician as co-author. The median sample size was 72 (IQR: 12 to 120) participants.
However, less than one-third (30.6%; n=74) of the studies were explained their formal power
analysis steps. Of these, 72 studies reported type-I error and 68 studies reported desired
power explicitly. The preferred choice of desired power in these studies was 80% (n=55).
Only 10.3% of studies were adherent to the CONSORT reporting guidelines. Reporting of
power analysis was significantly (p<0.001) associated with the adherent of CONSORT
guidelines. Only 9% studies described their strategy to handle missing data in analysis.
Nearly half (45.9%) of the studies reported p-value statement, and 59.5% of studies reported
the statistical software used in their data analysis.

Conclusions: This analysis indicates the association between the reporting of power analysis
and adherent to CONSORT guidelines.

288
Poster viewing A
Wednesday 4 September 1.30pm

Harnessing repeated measurements of predictor variables: a review of existing

methods for clinical risk prediction

Lucy M. Bull, Jamie C. Sergeant, Mark Lunt, Glen Martin, Kimme Hyrich
The University of Manchester

Background: Clinical prediction models (CPMs) can predict the risk of health outcomes for
individual patients. The majority of existing CPMs are limited to only harnessing cross-
sectional patient information and are unable to harness longitudinal medical data available in
electronic health records (EHRs). Incorporating repeated measurements into CPMs provides
an opportunity to enhance their performance in practice as they inform how patients change
over time and are less sensitive to measurement error (1, 2). Our aim was to systematically
review the literature to understand and summarise existing approaches to harnessing
repeated measurements of predictor variables in CPMs.
Methods: Medline, Embase, and Web of Science were searched for articles reporting the
development of a multivariable CPM for individual-level prediction, and modelling repeated
measurements of at least one predictor. Information was extracted on: the method, its
specific aim, reported advantages and limitations, and software available to apply the
method.
Results: The search revealed 217 relevant articles. Three methodological aims for
harnessing longitudinal information were identified: enabling updated predictions over time;
inferring a covariate value at a pre-specified time, and accounting for the effect of covariate
change. Seven distinct methodological frameworks were identified. These frameworks range
from time-dependent covariates in survival analysis, through to joint models, and machine
learning algorithms.
Conclusions: Identified frameworks and the methods within them can vary by the aims
addressed, assumptions made, and their use of longitudinal information. Their applicability
will depend on the clinical question, regularity of follow-up visits, sample size, and variability
amongst the longitudinal predictors.
References1. Steyerberg EW, Moons KGM, van der Windt DA, Hayden JA, Perel P, Schroter S, et al.
Prognosis Research Strategy (PROGRESS) 3: Prognostic Model Research. PLOS Medicine.
2013;10(2):e1001381.2. Sweeting MJ, Barrett JK, Thompson SG, Wood AM. The use of repeated
blood pressure measures for cardiovascular risk prediction: a comparison of statistical models in the
ARIC study. Statistics in medicine. 2017;36:4514-28.

Winner of Best Poster at YSM 2019

289
Poster viewing B
Thursday 5 September 1.10pm

Bayessian Survival Analysis of Cervical Cancer using Weibull Regression Model

Serifat Folorunso
Department of Statistics, University of Ibadan

The modeling and analysis of lifetime for terminal diseases such as cancer is a significant
aspect of statistical work. This study considered data from 440 cases of patients treated from
cervical cancer between 2006 and 2015.This research is to apply a Bayesian modeling
approaches for prediction of patient survival covering both statistical methodologies for
Bayesian survival models on cervical cancer. The study achieves the following objectives: (i)
to describe and implement a Bayesian Weibull model to predict patient survival; (ii) to
observe the impact of censoring on fitting a Weibull regression model with covariates.We
investigate Bayesian survival analysis of cervical cancer in addition to a simulation study by
approximating posterior summary using Laplace Approximation and Laplace Demon function
with Posterior mode, Posterior mean, posterior standard deviation and quartiles by fitting
Weibull. From the MCMC algorithm with 100,000 iterations, discarding 10,000 as burn-in
and from censoring levels of 10%, 20%, 50% and 80%, where we ordered all 440 cervical
cancer dataset from smallest to largest, found the relevant percentile for the censoring time
and set the top 10%, 20%, 50% and 80%, respectively, as censored. Then, we fit new
dataset to estimate the model and ran posterior predictive checks to explore the uncertainty
about the predictions. Our results show that censoring has an effect on the performance of
the Weibull models in that, as the proportion of censoring increases, poorer parameter
estimates were obtained in terms of both bias and precision, depending on the “closeness”
of the components. Our study indicated that when the amount of censoring is small, very
little bias is likely to result and confirmed that an acceptable model of survival data can still
be obtained with light censoring up to 20%.

290
Poster viewing B
Thursday 5 September 1.10pm

Data Ethics as an Ontological Problem

Thomas King

While the character of artificial ‘intelligence’ and the nature of ‘bias’ are regularly aired, ‘data
ethics’ is taken for granted. This is compounded by the longstanding ethical basis of data at
the datum level - do we not already have a ‘data ethics’? Digital data and statistical data are
totally different ideas yet ‘data ethics’ is often used without defining what is meant by ‘data’.
This term, ‘data ethics’, causes confusion but it is about data used in bulk, for which both
statistical and computational perspectives have a bearing. Ethics is about opening questions
not right answers, and the frame proposed will help us to specify our questions about use
and impact of data.This work sets out a (2 x 2) frame for understanding the distinction from
data aggregated from the individual datum. This includes the separation of interpretation
from stochastic issues of measurement error, model selection, missing data and sampling.
Both of these work at individual and bulk levels e.g. causes (may) correspond to correlations
at the bulk level. Measurement of an attribute of an individual has corresponding construct
validity and reliability. Consent by individuals has worked well for direct interventions but is
incoherent for aspects of linkage to future data.Case studies are established as effective for
stimulating ethical deliberation, so examples are presented to illustrate this frame: individual
lives might be saved by a surgical or pharmaceutical treatment, but who are the people
saved by public health interventions? Coverage of sensors for data collection to be used in
prediction, allocation and training will lead to conservative inferences simply in response to
sparser sampling. Service efficiency improvement is popularly synonymous with cuts, but
should such concerns be used by the public to restrict such data usage? Such examples
lead directly to consideration of what one ought to do and evaluation of the impact of data
use on society.

291
Poster viewing B
Thursday 5 September 1.10pm

Forecast Error: The History of Political Betting in the United Kingdom

Timothy Martyn Hill

Barclays Corporate

The "Forecast Error" series of articles in Significance started examining election predictors in
2015. Each article considered many predictors, but each article covered just one election. In
2018 we started a new chapter in the "Forecast Error" series where we examine an
individual class of predictor more closely across many elections. This presentation will cover
political betting.

Political betting is frequently used as a predictor of elections. The United Kingdom has a
history of political betting that predates the 20th century. We briefly review the pre-20th
century wagers, then examine the Majorities betting market that existed before World War 2,
then examine fixed-odds betting on elections that evolved as high-street bookmakers
became normalised in the 1960s, then look at the 21st century with markets such as
exchange betting, spread betting, and prediction markets.

292
Poster viewing B
Thursday 5 September 1.10pm

Finite-Sample Bias in Pseudo-Out-of-Sample Testing and Cross-Validation

Ian Hunt
Monash University

This paper analyses finite-sample bias in cross-validation estimates of expected prediction

error. A significant risk of positive bias is identified against flexible models in the context of
limited sample sizes --- this bias has practical implications for assessing curve-fitting models
in finance and economics, for example when comparing regime-change models, or modern
machine-learning techniques, with less flexible model-averaging methods. The bias is also
positively related to the number of data-points left out for cross-validation purposes --- this
has implications for traditional pseudo-out-of-sample Diebold-Mariano tests (using “hold-out”
windows) because these tests rely on a special case of cross-validation that leaves out large
chunks of data.First, I give analytic results for general leave-“n”-out cross-validation.
Secondly, I render pseudo-out-of-sample tests as a special case of cross-validation. I
conclude that the bias against flexible models is higher for pseudo-out-of-sample-tests than
with leave-one-out and K-fold cross-validation.The analytic results add evidence to Diebold's
conclusion that “[t]he hunch that pseudo-out-of-sample analysis is somehow the ‘only,’ or
‘best,’ or even necessarily a ‘good’ way to provide insurance against in-sample over-fitting in
time-series model comparisons proves largely false'' (Diebold, 2015). Plugging-in better
estimates of over-fitting into the traditional DM-test statistic may help; better estimates in this
context include those based on the expected optimism methodology (Efron, 2004) and
cross-validation methods that leave out fewer data-points than pseudo-out-of-sample hold-
out windows. This plug-in strategy could also ameliorate the problem of low power (Hansen
and Timmerman, 2015) in pseudo-out-of-sample tests.I analyse both cross-sectional and
time-series models. The time-series section adds finite-sample results to Hyndman-et-al
(2018), who prove that leave-one-out cross-validation estimates of expected prediction error
converge in probability to true expected prediction error for auto-correlated time-series if the
estimated model adequately accounts for the true time-series dependence (pseudo-out-of-
sample testing appears to rely on an analogous assumption).A simulation study and an
application to forecasting equity market returns complete the paper.

293
Poster viewing B
Thursday 5 September 1.10pm

Modelling Categorical Data in Frequency Domain via Mutual Information Approach

Oluwaseun Wale-Orojo, Ajibola Soyonka, Olaniyi Olayiwola

FUNAAB

Phenomenon whose occurrences are described with count data in a multi-dimensional

contingency table is common in medical and epidemiological studies. This study thus
obtained the density function(s) of nonlinear vectors via mutual information approach and its
measure of dependence from multivariate exponential power distribution which is a member
of multivariate elliptical contoured family. The obtained model which accommodates for light
and heavy tailed member distributions depending on the shape parameter were used to
establish result for nonlinear vectors from multivariate Laplace, normal, uniform, students’ t
and Fisher distributions. Multivariate dependency of live birth on maternal age as the number
of pregnancy increases irrespective of maternal height advantage as claimed from previous
study, was established from the theoretical model obtained.

294
Poster viewing B
Thursday 5 September 1.10pm

Generalised additive models applied to tourism data

Caterina Constantinescu
The Data Lab, University of Edinburgh

In this talk I will discuss some of the functionality offered by R package `mgcv`. I will cover
two generalised additive models (GAMs) built using this package, and applied to time series
data measuring the monthly number of visitors at two historic sites in Edinburgh, Scotland:
Craigmillar and Edinburgh Castles. These GAMs achieved a close description of the
observed data, and largely respected the patterns of seasonality and other particularities that
these time series exhibited.One of these models was devised to forecast future visitor
numbers, and used the data to predict itself. A separate explanatory model was also
constructed, and included various additional predictors (considered to be plausible drivers of
tourism activity), e.g., temperature, hotel occupancy data, economic factors etc. Relevant
Google search trends were also included in the prediction, so I will briefly discuss the related
package, `gtrendsR`. Supporting functionality from packages `itsadug` and `mgcViz` will also
be discussed, as these were helpful for generating visualisations for the relationship
between each predictor included and the outcome, as well as generating confidence
intervals around the predictions.Results included a linear relationship (as mgcv will not
‘force’ the fit of a wiggly line when this is not supported by the data) between hotel
occupancy indicators and castle visitors, as well as the discovery of interesting non-linear
patterns over time, allowed to vary by castle as well as visitor ticket types.

295
Poster viewing B
Thursday 5 September 1.10pm

kth Inflated Poisson Regression Modelling - EM Algorithmic Approach

Sameera Viswakula, Srinal Fernando

University of Colombo

This research focuses on modelling inflated count data using Poisson regression models.
Poisson distribution is widely used to model count data. However, in many real-life
applications, frequency of some of the counts is significantly higher than that is expected by
the Poisson distribution (eg. zero, one, etc.). Therefore, inflated Poisson regression models
are used for these data. Commonly used statistical software offers only zero or one inflated
Poisson regression modelling mainly due to the complexity of the parameter estimation. In
this research, EM algorithms are derived to model inflated count data using k-inflated
Poisson regression models (k=0,1,2,…). A simulation study is conducted to assess the
validity of estimates. Real-life data sets are used to illustrate the usage of the newly
developed EM algorithm based R package for inflated count data.

296
Poster viewing B
Thursday 5 September 1.10pm

The Lindley-Weibull Power Series Class of Distributions and its Application to Life
Time Data

Boikanyo Makubate1, Broderick O. Oluyede2, Thatayaone Moakofi1

1
Botswana International University Of Science and Technology, 2 Georgia Southern
University

A new generalized class of distributions called the Lindley- Weibull Power Series (LiWPS)
class of distributions proposed and studied. The new class will be flexible, also with
desirable propeties including hazard function that shows increasing, decreasing and bathtub
shapes. This class of distributions generalizes the Lindley power series and the Weibull
power series class of distributions respectively. A special model of the LiWPS class of
distributions, the new Lindley Weibull Poisson (LiWP) is considered and some of of its
mathematical properties are obtained. The LiWP distribution contains several new and well
sub models including (Lindley Weibull Logarithmic (LiWL), Lindley Weibull Poisson (LiWP),
Lindley Weibull Geometric (LiWG) and Lindley Weibull Binomial (LiWB)). Maximum
likelihood estimation technique is used to estimate the model parameters followed by a
Monte Carlo simulation study. Finally an application of the LiWP model to a real data set is
presented to illustrate the usefulness of the proposed class of distributions. Goodness of fit
tests are used to compare the fit of the LBXII distribution with its sub-models and other
models for a given data set.

297
Poster viewing B
Thursday 5 September 1.10pm

Multi-stage models of cancer and disease

Anthony Webster
NDPH, University of Oxford

Since Armitage and Doll's publication of "a multi-stage theory of carcinogenesis" in 1954 [1],
the multi-stage model has underpinned our conceptual understanding of cancer. In the last
few years, researchers have started to apply the model to other diseases, such as
Amyotrophic Lateral Sclerosis (Motor Neurone Disease), providing new insights into how the
disease can arise and progress [2]. Here the multi-stage model is simplified and extended
through a simple mathematical recipe for composing independent, sequential and
cumulative models of carcinogenesis or other multi-stage failure models [3]. Relationships
between different published cancer models are clarified, and derivations of new and existing
results are simplified. Potential limitations of the model are discussed in the context of recent
cancer research. The result is a simple framework for combining biologically-motivated
models of each step in a disease’s progression, providing a mathematical toolkit to study the
failure of complex systems, biological or otherwise.

[1] P. Armitage and R. Doll, “The age distribution of cancer and a multi-stage theory of
carcinogenesis,” British Journal of Cancer, vol. 8, no. 1, pp. 1–12, 1954.

[2] A. Chio, L. Mazzini, S. D’Alfonso, et al. “The multistep hypothesis of ALS revisited The role of
genetic mutations”, Neurology, vol. 91, no. 7, pp. E635–E642, 2018.

[3] A.J. Webster “Multi-stage models for the failure of complex systems, cascading disasters, and the
onset of disease”, https://doi.org/10.1101/476242, in press, PLOS ONE, (2019).

298
Poster viewing B
Thursday 5 September 1.10pm

Spatial Analysis on Hepatitis in Punjab, Pakistan during 2010-14 and Interpolation of

Disease Rates

Majida Jawad1, Jawad Kadir2

1
University of the Punjab, 2 Lancaster University

Increasing rate of Hepatitis cases is world wise concerns. Pakistan is one the highest rated
countries in the world where not only hepatitis rate is so high but also increasing very rapidly.
Therefore, a microscopic spatial analysis of the hepatitis infection that can provide scientific
information for further intervention and disease control is needed. Tehsil wise percentage of
infection cases recorded by the Punjab, Bureau of Statistics in Multiple Indicator Cluster
survey were included in this study. Changing pattern of infection rates were observed by the
used of spatial autocorrelation and interpolation tools. Moran’s I plots and Hotspot analysis
were conducted to identify the high risk areas. The finding of the analysis showed that the
areas that are far away from the capital of the districts are more vulnerable as compare to
the others. The analysis also pointed out that the areas belong to the southern part of the
Punjab like Rajanpur and Hafizabad are the high-risk clusters.

299
Poster viewing B
Thursday 5 September 1.10pm

Constrained Empirical Bayes Estimation in Multiplicative Area-Level Models with

Risk Analysis Under an Asymmetric Loss Function

Elaheh Torkashvand1, Mohammad Jafari Jozani1

1
University of Notre Dame, 2 University of Manitoba

Consider the problem of benchmarking small area estimates under multiplicative models
with positive parameters where estimates are constrained to aggregate to direct
estimates for the larger areas. Constrained (hierarchical) empirical Bayes estimators of
positive small area parameters under many conventional loss functions can not be
expressed in closed forms and/or might not even exist. Under the conventional squared
error or weighted squared error loss functions benchmarked Bayes and empirical Bayes
estimators can take negative values. In this paper, we propose a loss function that
guarantees positive constrained estimates of small area parameters under multiplicative
models. The proposed loss function penalizes underestimation of the unknown parameters
of interest more than the squared error and the Kullback-Leibler (KL) loss functions. The
hierarchical and hierarchical constrained empirical Bayes estimates of small area
parameters and their corresponding risk functions under the new loss function are
obtained. Also, the asymptotic approximation of risk and its second-order unbiased
estimator are provided. We implement the Jackknife method in order to reduce the bias of
estimators of risk. Finally, the performance of the proposed methods is investigated using
simulation studies as well as a real data analysis for estimating asthma rate in
subpopulations.

300
Poster viewing B
Thursday 5 September 1.10pm

Predictive regression with p-lags and order-q autoregressive predictors

Harshanie Jayetileke1, You-Gan Wang1, Min Zhu2

1
Queensland University of Technology, 2 University of Queensland

This study considers the predictive regression models with p lagged predictors while the
predictor xt itself is an AR(q) model. Akin to the standard predictive regression models, the
ordinary least squares produces biased slope estimates. The existing methodology is
applicable in the cases, p = q =1 via plug-in approach, or more generally, the augmentation
approach introduced by Amihud, Hurvich and Wang (2010) in the case of p=q. We show
that, the augmented regression method works in general, for any p and q orders. Extensive
simulation studies indicate that, (i) the existing bias reduction methods become
unsatisfactory when the order q is misspecified, in terms of bias and the mean squared
errors; (ii) the correct identification of both p and q values are essential to successfully
remove the bias; and (iii) the general augmentation method can reduce the bias satisfactorily
and produces the smallest mean squared errors. The empirical application to stock market
data from S&P 500 Index illustrate, a procedure to identify the qth order of the predictors
such as dividend yield, earnings-to-price ratio, etc... and to obtain bias reduced estimates.

301
Poster viewing B
Thursday 5 September 1.10pm

Bayesian inference for GARCH Model for Cryptocurrency

Wantanee Poonvoralak
Chulalongkorn University, Sasin School of Management

Objective: The cryptocurrency financial market is known to be one of the most popular and
volatile financial market in recent years. It is different from others financial markets due to
that fact it can verifies transactions through its peer to peer network instead of using a
trusted central party. Hence, numbers of GARCH models using statistical inferences have
been used to examine the volatility effects of this currency. We propose the Bayesian
inference for GARCH model for this currency. We aim to provide empirical results to help the
BASEL committees to make more informed decisions about modelling in cryptocurrency and
to whether it should be uses as part of the regulatory market risk management.

Methods: We used the most recent cryptocurrency with time period from 28th April 2013 to
21st March 2019 from the top 2 as listed on the 21st March 2019, they are the Bitcoin (BTC)
and Ethereum (ETH) with market capitalization of $71,048,680,219 and $14,548,360,669
respectively. We firstly applied GARCH(1,1) to these data using statistical inferences and
found that estimated kurtosis and persistence parameters suggested a requirement of heavy
tail model distribution like the Student t-GARCH(1,1). This was not the case for Bayesian
Markov Chain Monte Carlo (MCMC) approach.

Results: We found cryptocurrency data to be similar to other financial data such as the
Foreign Exchange (FX) series in term of their volatility effects using GARCH modeling.
Bayesian inference can help capture the persistence parameters and the MCMC sampling
ensured finite kurtosis. Our results supported the decision that the cryptocurrency have
implications on market risk management for investors and that the BASEL committees
should consider it as part of regulatory risk management framework.

302
Poster viewing B
Thursday 5 September 1.10pm

A New Alternative to Cohen's Kappa Coefficient of Agreement

Tarald O. Kvalseth
University of Minnesota, USA

When two observers (raters) are independently classifying items or observations into the
same set of mutually exclusive and exhaustive categories, some measure of agreement
between the observers is often of interest. While a number of such measures have been
proposed, the so-called Cohen’s kappa, or simply kappa, has become by far the most
frequently used measure in various fields of study. The popularity of kappa is based on its
simplicity and appealing interpretation and the important fact that it corrects for the potential
agreement expected by chance alone. When the various probabilities or proportions of
agreement and disagreement between the two observers are represented in terms of a
square contingency table, the chance agreement of kappa is based on statistical
independence involving the marginal probability distributions. However, in spite of its
popularity, Cohen’s kappa suffers from some well-known limitations related to its
dependence on the marginal distributions. In particular, kappa takes on unreasonably small
values when the two marginal distributions are approximately symmetrical and highly uneven
(non uniform). Alternative measures have been proposed and aimed at trying to overcome
the kappa limitations. However, as argued in this present paper, the problems with Cohen’s
kappa as well as its proposed alternatives originate in their definitions of chance agreement
(or disagreement) based on the marginal distributions. The present paper proposes a new
kappa alternative that does incorporate chance agreement, but the chance agreement is not
based on the overall marginal probability distributions. It is, however, based on independent
chance classifications by the observers. The new measure does indeed overcome the
limitations of Cohen’s kappa and appears to have only desirable properties. A weighted
form of this measure for the case when the classification categories are ordinal is briefly
discussed. Also, statistical inference procedure for the new measure is outlined.

303
Poster viewing B
Thursday 5 September 1.10pm

classNB: a classifier based on the negative binomial distribution

Xiaoqiang Wang1, Tianyuan Yao1, Chunmao Huang2, Heng Ge1

1
Shangdong University (Weihai), China, 2 Harbin Institute of Technology at Weihai

Overdispersion is a widespread phenomenon in most count data sets. The negative binomial
distribution is commonly adopted to fit the over-dispersed count data. In current work we
devote to construct a toolkit, a R package called classNB for the classification problems
based on the negative binomial distribution. In this package, we considered the mixture
models, hidden Markov models, zero-inflated negative binomial models depending on the
different data structures. In order to accelerate the process of parameter estimations, we
develop an efficient Expectation Maximization (EM) algorithm which can avoid the typical
numerical solution embedding in the EM algorithm. We further theoretically proved the
convergence of the novel EM algorithm for the mixture models as well as the hidden Markov
models. The simulation studies demonstrated that our methods effectively reduce the
runtime of parameter estimations. Moreover, the proposed approaches also gain the same
classification performance as the typical methods. The simulated data and the real data
illustrate that the proposed toolkit, classNB has a good ability to classify the count data such
as the next generation sequencing read count data, the earthquake data, air population
data, etc.

304
Poster viewing B
Thursday 5 September 1.10pm

Analyzing diagnostic test data when sensitivity is not constant

Geoffrey Jones1, Wes Johnson2, Cord Heuer1

1
Massey University, 2 UC Irvine

The analysis of data from diagnostic tests commonly makes the assumption that the
sensitivity and specificity of the test(s) are constants. This is particularly important when
there is no gold-standard test and latent class analysis has to be used to estimate the
disease prevalence and test characteristics. For example the Hui-Walter model (Hui &
Walter, 1980) uses data from two conditionally independent tests in two distinct populations
to get an identifiable model for the six parameters – the assumptions of constant sensitivity
and specificity in each population are crucial for achieving identifiability. In practice there are
often scientific reasons for assuming that sensitivity, at least, will vary across populations.
We present a Bayesian latent class analysis of data from a study of prevalences of
leptospirosis in New Zealand beef herds to show how varying test sensitivity might be
modelled and analyzed. The results suggest that relevant scientific information, incorporated
not just into the priors but also into the model structure, can lead to a more appropriate
analysis.

Hui, S.L., and Walter, S.D. (1980). Estimating the error rates of diagnostic tests. Biometrics,
36, 167-171.

305
Poster viewing B
Thursday 5 September 1.10pm

Application of Interrupted Time Series in evaluating PUMA - an intervention to

improve early detection of in-patient deterioration in children’s hospitals

Katie Taiyari1, Kerry Hood1, Robert Turbey1, Devina Allen2, Damian Ronald3, on behalf of the
PUMA Study Team
1
Centre for Trial Research, Cardiff University, 2 School of Healthcare Sciences, Cardiff
University, 3 University of Leicester

Background and objectives: PUMA is a quasi-experimental design study aimed at

developing, implementing and evaluating the ‘PUMA Programme’ – an intervention to help
children’s hospitals improve their ‘early warning system’ and prevent avoidable mortality and
morbidity among in-patients. Aggregate monthly outcomes were collected for four sites
before (pre-), during (phase-in) and after (post-) implementation of the intervention. The
common approach when analysing data from such design using ITS modelling is to exclude
the phase-in data from the time series analysis, which is the period the intervention was
implemented. However, such an approach is not appropriate in a scenario where there is a
possibility that the performance may have been altered during the phase-in period. In such a
scenario, excluding phase-in data from ITS analysis would not only reduce the power of the
study but also reduce the chance of detecting the true change.

Methods: We examined three different approaches of evaluating ITS trajectory: (1) including
the phase-in period in the analysis, (2) considering phase-in period as post-intervention, and
(3) excluding the phase-in period. We compare Akaike information criterion (AIC) and
Bayesian information criterion (BIC) from each model to choose the appropriate approach.

Results and Conclusions: Data from four UK-based paediatric hospitals in the PUMA study
were collected from May 2015-October 2018, covering pre-intervention, phase-in and post-
intervention periods. Data was analysed using ITS modelling. The data was the aggregate
number of children in a given month who experienced at least one of following critical
events: mortality, unplanned admission to PICU, unplanned admission to PHDU, cardiac
arrest, and respiratory arrest. Data from each hospital were analysed separately using
Interrupted Time Series model. All the results will be presented.

306
Poster viewing B
Thursday 5 September 1.10pm

Latent class models for use in a clinical setting using cohort data

Charlotte Watson1, Hannah Lennon2, Nophar Geifman1, Andrew Renehan1

1
University of Manchester, 2 IARC

Objectives: Life-course approaches to determine risk factor exposure are increasingly

favoured over traditional ‘once-only’ epidemiological determination. One approach is latent
class trajectory modelling (LCTM), which clusters individuals’ changes in exposure over
time, and offers a public health strategy to identify early divergent adverse trajectories as
potential intervention targets. There is increasing use of LCTMs in mainstream
epidemiology, but often with poor model description, and an over-reliance on the BIC metric
for model selection. Here, we aimed to explore methods in deriving and validating LCTMs
from multiple cohorts to ensure that they properly represent observed patterns across
different populations.

Methods: We interrogated three cohorts, AARP (N: 321,827), PLCO (N: 147,488) and WHI
(N: 151,363), with longitudinal BMI as the exposure and cancer incidence as the endpoint of
interest. We extended our previously published work
(https://bmjopen.bmj.com/content/8/7/e020683) to develop (i) a scoping framework for
random-effect structures; (ii) testing multiple start points to obtain the global maximums; (iii)
facilitating model choice with alternative metrics (APPA, Odds of Correct Classification,
Entropy, Mismatch and DoF) to BIC. We developed visual model assessment tools, like
convergence plots.

Results: We illustrate a number of examples where deviation from model assumptions yield
very different classifications. We identified computational challenges for cubic models
greater than 4 classes with up to 20 start points. After arriving as preferred models, we show
that LCTM improve the performance characteristics of BMI exposure, compared with once
only BMI measures, however, this improvement is clinically modest. LCTM might best
identify specific sub-populations that have particularly high risk for cancer incidence.

Conclusions: The study highlights that model selection needs to be undertaken with care
and not based solely determined by lowest BIC. We emphasise that multiple start points
should be tested when using these models.

307
Poster viewing B
Thursday 5 September 1.10pm

Elicitation of prior beliefs for data lying in the simplex

Anastasia Frantsuzova, John Paul Gosling

University of Leeds

Right-stochastic matrices are used in the modelling of Markov processes (transition

matrices) with a property that their elements are non-negative and each row sums to one.
Similarly, data representing proportions of a whole can be unified into the compositional
framework with similar structural properties. If we consider the problem of estimating these
probabilities from a Bayesian standpoint, a natural choice of prior for the above multinomial
likelihood is the conjugate Dirichlet family. There also stands a question of subjectivity of the
prior and sensible ways to elicit such beliefs. We will present an overview of current
elicitation methods for prior beliefs with data lying in the compositional simplex space in light
of the Sheffield Elicitation Framework (SHELF). Additionally, we propose an additional step
to the elicitation procedure, which aims to encapsulate dependencies between elicited
marginal distributions and to allow the experts to refine or assure their beliefs by sensibly
allocating weights to regions of the simplex.

308
Poster viewing B
Thursday 5 September 1.10pm

Reactor Embrittlement Curves – A Cracking Puzzle

Caroline Pyke
National Nuclear Laboratory

Light Water Reactors (LWRs), such as Sizewell B in the UK, contain a steel Reactor
Pressure Vessel (RPV). The continued safe operation of LWRs, and their life extension,
requires improved understanding of the behaviour of the toughness of these steels with
increasing exposure to radiation - embrittlement. The condition of the RPV is monitored with
surveillance samples, which are representative of the materials and irradiation conditions of
the actual vessel. These are periodically withdrawn and their toughness tested using the
Charpy impact test. For each steel a series of samples are tested at different temperatures
and the relevant data recorded. The resulting data returns a sigmoidal curve for each set of
surveillance samples. Empirical models of embrittlement focus on the shift of the curve at an
individual point with increasing irradiation. However, this approach is limited as it relies on
good data from unirradiated samplesThe purpose of this work was to determine whether the
whole embrittlement curve could be established empirically using material composition data
and irradiation parameters. The curves were estimated using non-linear least squares fits to
describe the steel condition using three parameters. Ordinary linear least squares regression
was then used to estimate each of these parameters. One of these parameters, the upper
horizontal asymptote, otherwise referred to as the Upper Shelf Energy (USE), was
particularly well predicted. The USE is expected to decrease with increased exposure to
radiation. However, there can be considerable variability and/or bias in the test data and this
phenomenon is not always observed. Indeed, often positive shifts in USE are observed from
the raw data. From our models of USE we can estimate shift in USE when exposed to
various levels of irradiation. Comparing these predictions with observed shift in USE returns
some clear biases in the results.

309
Poster viewing B
Thursday 5 September 1.10pm

Fecal microbiota, Vitamin D, serum biomarkers of inflammation and diet: a complex

interactive network influencing colorectal cancer risk and prognosis.

Pietro Belloni1, Davide Serrano2, Chiara Pozzi3, Bruno Fosso4, Silvia Guglietta5, Harriet
Johansson2, Patrizia Gnagnarella6, Debora Macis2, Valentina Aristarco2, Nicola Segata7,
Federica Corso8, Federica Bellerba8, Maria Rescigno9, Sara Gandini8
1
European Institute of Oncology, 2 Division of Cancer Prevention and Genetics, European
Institute Of Oncology IRCCS, Milan, Italy, 3 Mucosal Immunology and Microbiota Unit,
Humanitas Research Hospital, Milan, Italy, 4 Institute of Biomembranes and Bioenergetics,
Consiglio Nazionale delle Ricerche, Bari, Italy, 5 Department of Microbiology and
Immunology Medical University of South Carolina, USA, 6 Division of Epidemiology and
Biostatistics, European Institute of Oncology IRCCS, Milan, Italy, 7 Department CIBIO,
University of Trento, Trento, Italy, 8 Department of Experimental Oncology, European
Institute of Oncology IRCCS, Milan, Italy, 9 Unit of Immunobiology of Dendritic Cells and
Immunotherapy, European Institute of Oncology IRCCS, Milan, Italy

Background Microbiota and Vitamin D are involved in several pathologies including cancer.

Methods A prospective case-control study (34 CRC patients and 33 age match-controls) was
designed to investigate microbiota composition, vitamin D indicators, circulating biomarkers
of inflammation and lifestyle were analyzed.

Results Our 16S metagenomic sequencing confirmed a higher abundance of Fusobacterium

nucleatum, Bacteroides fragilis, Parvimonas micra, Solobacterium moorei and
Porphyromonas sp. in CRC patients. F. nucleatum showed a seasonal trend with greater
abundance in winter when 25-hydroxyvitamin D (25OHD) is significantly lower (P=0.0002).
In a multivariate model, F. nucleatum was found to be significantly associated with 25(OH)D
(P=0.036) independently of CRC status and season. The markers of inflammation IL-6 and
hs-reactive C protein negatively correlated with Vitamin D but positively with F. nucleatum
(all P<0.05). Vitamin D receptor polymorphisms were found to correlate with CRC (FokI ff
were more frequent among cases than controls, P-value=0.03) and with the abundance of
microbiome members. Genera such as Parvimonas, Peptostreptococcus and Fusobacterium
were significantly enriched in CRC patients consuming a diet poor in fatty fish (source of
vitamin D) and rich in carbohydrates. Lastly, F. nucleatum was found to be significantly
associated with time to relapse (P=0.03) and multivariate models including F. nucleatum,
Vitamin D and Adiponectin identify CRC with an area under the ROC curve that patients
ranged from 80-90% after cross-validation and reached similar values predicting recurrence.

Conclusions These parameters may set the bases for a diagnostic tool capable to predict the
presence and recurrence of CRC.

310
Poster viewing B
Thursday 5 September 1.10pm

How clean is your takeaway?

Aura Popa
The Statistical Company

Consumers that are using food takeaway delivery apps like Just East, Deliveroo, Uber Eat
do not see the Food Hygiene Rating of the restaurant, but the rating of other consumers. In
October 2018, a BBC investigation found numerous zero rated takeaways on Just Eat, the
UK's biggest online platform. In some major cities, more than half of the zero rated
takeaways are on the platform.With the aim of comparing the Food Hygiene Rating Scheme
versus the rating given by the customer to see if there is any relationship between the two,
this poster will analyse the options of restaurants from three different takeaway delivery
apps, looking into 2 wards of Kensington and Chelsea Borough, London. The two different
areas, one of the poorest and one of the richest in UK according to the Index of Multiple
Deprivation, are chosen with the purpose of testing the hypothesis if there are differences
between the food hygiene and consumers’ rating. The insights will be displayed in a visual
format using infographics containing geographical heat maps through Spatial
Analysis.Should the law of Food Hygiene Rating Scheme be changed and ask all the take
away outlets to publish their score not on the door, but on these apps?

311
Poster viewing B
Thursday 5 September 1.10pm

A Bayesian approach to incorporate ecological species detection uncertainty into a

rarity index

Jafet Belmont1, Marian Scott1, Claire Miller1, Craig Wilkie1, Tom August2, Philip Taylor2,
Steve Brooks3
1
University of Glasgow, 2 Centre for Ecology and Hydrology, 3 The Natural History Museum

Occupancy models at a community level have been explored widely in the literature. These
models allow occupancy patterns to be estimated when the detection probability for the
species within the community is less than one. Individual species responses may vary widely
within a community, making the task of identifying the ecological processes of interest that
drive the observed occupancy patterns difficult, especially for rare and elusive
species.Several studies have developed methods to quantify the rarity of a species at the
community level, but none has accounted for detectability bias. Leroy, B. et al. (2012)
proposed a multiscale index of relative rarity (IRR) that allows species rarity to be compared
regardless of the spatial scale, geographic area or taxonomic group. The index is computed
by assigning a rarity weight to each species based on the observed occupancy and a rarity
cut off point determined by the user.For this work, a 2-stage modelling approach was
developed to estimate Leroy’s rarity index in a Bayesian setting while accounting for
detection bias. The index is computed at the first step as a derived quantity from the
estimated occupancy by using a multispecies occupancy model. The second stage takes
into account the first stage estimation uncertainty and models the responses as a function of
site-level ecological covariates to evaluate which drivers affect the composition of rare
species in a community. This approach was applied to a presence-absence dataset for
Dragonfly occurrences in lakes across the UK. A selection of species-level covariates were
used to estimate occupancy and detection probabilities. The estimated species occupancy
was then used to compute the rarity index which was modelled in a second step to evaluate
the effect that landscape connectivity and environmental stressor covariates have at different
spatial scales.

312
Poster viewing B
Thursday 5 September 1.10pm

Modelling the effect of body mass on seabird survival: A Bayesian MCMC approach

Zhou Fang1, David Elston1, Francis Daunt2, Kate Searle2

1
Biomathematics and Statistics Scotland, 2 Centre for Ecology & Hydrology

Many species of seabirds are vulnerable to changes in their local environment. Such
changes may directly cause mortality, but often instead are more subtle and affect the
energy balance of birds (for example by making it more difficult to find food), thus causing
changes in body mass which may alter survival probabilities. In order to obtain a clear idea
of such indirect impacts, it is therefore important to determine the relationship between bird
mass and over-wintering survival rates.Past studies in this area have used quite small data
sets and simple methods. We analysed larger data sets consisting of mass measurements
and capture-recapture or re-sighting records for four species – kittiwakes, puffins, guillemots
and razorbills.These data sets are challenging to model for a variety of reasons. Birds have
varying (and often quite small) probabilities of being recaptured or resighted, with the
possibility of trap dependence effects and changing degrees of recapture effort. Mass
measurements may be rare in frequency, or show different patterns of variability between
birds, between years, and in terms of the day of the year each was measured. Birds may
have different distributions of masses according to their sex, which is often not known.
Survival is also impacted by other effects due to each year’s environmental conditions and
the bird’s age.We defined a data augmented Bayesian MCMC model that attempted to
capture these effects, combining a model for between- and within-year variation in bird mass
with one for the survival of birds. This was fitted to each of the species, allowing us to obtain
new estimates for the relationship of bird mass with bird survival.

313
Poster viewing B
Thursday 5 September 1.10pm

How varying CD4 criteria for treatment initiation was associated with survivability of
HIV-patients?: retrospective analysis of health records from India

Ram Bajpai1, Himanshu Chaturvedi2

1
Research Institute for Primary Care & Health Sciences, Keele University, 2 National
Institute of Medical Statistics, Indian Council of Medical Research, New Delhi, India

Background: HIV treatment and care services were scaled up in India at national level in
2007 with objective to increase HIV-care coverage. CD4-based criterion (or HIV clinical
stage III/IV irrespective of CD4) was used for antiretroviral treatment (ART) initiation with
increasing threshold in later years.

Objective: This paper aimed to evaluate the survival probabilities by varying CD4 criteria
(CD4 ≤200 for the period 2007-08, CD4 ≤205 for the period 2009-11, and CD4 ≤350 from
2012 onwards) for ART among of HIV-positive patients.

Method: This retrospective analysis included 127,949 HIV-positive patients aged ≥15 years.
All patients were registered for ART treatment between January 2007 to December 2011
and followed-up till December 2013 in Andhra Pradesh state, India. Mortality rate per 100
person-years was calculated. Kaplan-Meier and Cox-regression analysis were used to
explore the association with all-cause mortality.

Results: Median age at the treatment initiation was 34 years (interquartile range [IQR]: 29 to
40), and 45.9% were female. Median CD4 count was 172 (IQR: 102 to 240) at the time of
treatment initiation, and 19.3% of them had ≤100 CD4 count. Mortality rate for the period
2007-08 (CD4 ≤200) was 8.5/100 person-years compared to 6.4/100 person-years at risk for
the period 2012 onwards (CD4 ≤350). Earlier thresholds for treatment initiation showed
higher risk of mortality (CD4 ≤200 (2007-08), adjusted hazard ratio [aHR]: 1.86, 95%
confidence interval [95%CI]: 1.68-2.07; CD4 ≤250 (2009-11), aHR: 1.67, 95%CI: 1.51-1.85)
compared to the later threshold of CD4≤350 for treatment initiation from 2012 onwards.

Conclusions: Increasing CD4 threshold for treatment initiation over time was independently
associated with the higher risk of mortality. The findings of this study provided critical insight
into developing an effective and efficient HIV-treatment, care, and support program, and
create a sustainable way to respond to the HIV epidemic in Andhra Pradesh, India.

314
Poster viewing B
Thursday 5 September 1.10pm

Modelling of long-term vegetation change in Scotland's native forests

Jacqueline Potts1, Alison Hester2

1
Biomathematics and Statistics Scotland, 2 James Hutton Institute

Forests play a key role in climate change mitigation, adaptation and delivery of a range of
ecosystem services. This study examines the relationship between long-term changes in
forest vegetation in Scotland’s native forests and contemporaneous changes in climate,
pollution and grazing. Plots, originally surveyed between 1958 and 1983, were resurveyed
in 2007, with plant species composition being recorded. Four types of forest habitat (ash,
acid oak-birch, base-rich oak-birch, and pine) were considered, both separately and as an
overall sample of ‘forest’ plots. Changes in species richness were analysed using a
generalised linear mixed model, and changes in species diversity and the percentage cover
of major species groups and individual species were also analysed. A test of multivariate
homogeneity was used to assess whether dispersion (variation between plots within habitat
groupings) had changed. Statistical models were also used to explore the effect of drivers
associated with climate (temperature, precipitation), pollution deposition (nitrogen, sulphur)
and the density of herbivores on the vegetation data. All four habitat types showed
compositional change. There was no evidence of homogenisation, with the opposite trend
being seen in most habitat groups. The analyses identified statistically significant effects
associated with climate, pollution and grazing, with deposition of reduced nitrogen, in
particular, being associated with species compositional changes. Notable species changes
seen between surveys include increases in pteridophytes (bracken) and declines in forb
cover, and a doubling in frequency and cover of Fagus sylvatica (common beech).

315
Poster viewing B
Thursday 5 September 1.10pm

Introducing an adaptation to the MRSea package to allow for non-overlapping spatial

ranges in a spatial interaction term; the case of the African Vultures.

Lindesay Scott-Hayward1, Monique Mackenzie1, Claudia Faustino1, Ortwin Aschenborn2

1
University of St Andrews, 2 University of Namibia

This poster presents a method for overcoming the issue of including a spatial interaction
term for non-overlapping spatial regions. The data consist ofgeo-referenced locations from
satellite tagged vultures, recorded between March 2015 and November 2018. To
characterise habitat preference, Bernoulli generalised additive models were fitted to nine
environmental covariates along with a spatial interaction term; s(x,y, year-month) to predict
probability of presence of the vultures. The bivariate smooth of coordinates was based on
the CReSS method (Complex REgion Spatial Smoother) with SALSA2D to determine the
number and location of knots (the flexibility of the smooth). The standard implementation in
the MRSea package for the interaction term uses the same candidate knot locations across
each level of the interaction but the parameter estimates may vary. For this analysis, there
were some year-month combinations whose spatial range was completely separate to others
and this created an estimation issue. The new framework proposed for such situations
allocates the flexibility of the spatio-temporal term to each factor level, in this case each
year-month. Rather than setting out say 300 candidate knot locations, evenly spaced
through the full geographical range of each bird (as incorporated in the MRSea package),
candidate knots were placed throughout the geographical range of each year-month. Thus,
each knot location was only relevant to a particular year-month and the parameters
associated with these were only effective within that year-month. This new framework
allowed model convergence and greatly reduced both the number of parameters estimated
and the computational time.

316
Poster viewing B
Thursday 5 September 1.10pm

Electrophysiological Signatures of Cognitive Impairment in Parkinson’s disease – A

Machine Learning Approach

Aoife Sweeney1, Barry Devereux1, Charlie Ong2, John McKinley3, Seamus Kearney3, Julia
Foy2, Brian Murphy4, Bernadette McGuinness13, Peter Passmore13
1
Queen's University Belfast, 2 South Eastern Health and Social Care Trust, Northern
Ireland, 3 Belfast Health and Social Care Trust, Northern Ireland, 4 Queen's University
Belfast/Brainwavebank Ltd.

Dry-EEG offers an inexpensive, non-invasive and faster method to assess cognition in aging
clinical groups. Cognitive impairment is prevalent in Parkinson’s disease (PD), with 50% of
patients developing dementia within 10 years (Williams-Gray et al., 2013). The presence of
mild cognitive impairment (MCI), particularly visuospatial dysfunction, has been associated
with increased dementia risk. EEG has previously been shown to have predictive value for
patient outcomes in mild cognitive impairment associated with Alzheimer’s disease (AD-
MCI), yet there is little information on the electrophysiological correlates of PD-MCI. A 12-
month follow-up study is being conducted. PD patients (n=60) and matched controls
complete a comprehensive neuropsychological assessment and a battery of EEG tasks.
These computerized cognitive tasks assess resting state activity, attention, language,
memory and visuospatial cognitive domains. The PD patient cohort consists of both PD-NC
(normal cognition) and PD-MCI patients. The primary outcomes for this study are the ability
of dry-EEG to differentially discriminate between cognitively normal and cognitively impaired
PD patients, and to detect changes in PD cognitive status over 12 months. We present
power spectral analysis and event-related potential components which are associated with
cognitive dysfunction in PD. Differences between groups are assessed using randomisation
tests with FDR-corrected pairwise comparisons. Correlation analysis with validated clinical
measures such as the MoCA and Hoehn and Yahr scores are performed. Machine learning
algorithms such as the ‘Elastic Net’ (Zou & Hastie, 2005), random forests and k-nearest
neighbour models are used to identify EEG features which most accurately discriminate
between patient subgroups. Finally, the quality of the EEG data obtained (in terms of signal
to noise ratio and the percentage of epochs rejected), plus patient acceptability of this
technology are measured.

317
Poster viewing B
Thursday 5 September 1.10pm

Birth-death models for population genetics

Anastasia Ignatieva12, Jotun Hein1, Paul Jenkins2

1
University of Oxford, 2 University of Warwick

Given a sample of genetic data obtained at the present, a question of interest is how to
reconstruct a possible genealogy relating the individuals through common ancestors in the
past. The coalescent is a standard model for genealogies in population genetics, but
standard formulations assume a constant, or deterministically growing, population size. The
dynamics of a population exhibiting exponential growth, such as viral populations, can more
naturally be modelled as a birth-death process, which captures the stochastic variation in
population size over time.The history of the population is then described by a birth-death
tree; bifurcations in the tree correspond to births and leaves correspond to deaths of
individuals. The genealogy of a sample of individuals alive at the present time can be
obtained by deleting extinct and non-sampled lineages from the full tree. The stochastic
process corresponding to this genealogy is termed the reconstructed process.The
reconstructed process, viewed backwards in time, has an inhomogeneous pure-death
formulation (with a time-dependent death rate). We discuss the properties of this process,
and show how this formulation can be used to derive distributions characterising the
genealogy of individuals sampled from the population.

Winner of Best Poster at RSC 2019

318
 Cortina Borja, Mario, 120
A
Crook, Julia, 66
Afolabi, Saheed, 53 Cuffe, Robert, 142
Aigner, Maximilian, 167 Cummings, Vicky, 213
Al Baghal, Tarek, 35 Curnow, Paula, 71
Alghamdi, Fatimah, 234 Currie, Michael, 263
Althubaiti, Alaa, 79
Anand Kumar, Vinayak, 239 D
Anderson, Craig, 44
Andreis, Federico, 45 Daly, Jill, 109
Aparicio-Castro, Andrea, 112 D'Angelo, Silvia, 221
Arowolo, Olatunji, 88 Davillas, Apostolos, 34
Davis, Jodie, 161
Derby, Steven, 19
B
Dondelinger, Frank, 129
Bailey, R.A., 138 Douiri, Abdel, 137
Baipai, Ram, 288, 314 Drikvandi, Reza, 236
Bajaj, Sumali, 158
Bakbergenuly, Ilyas, 103 E
Baker, Evan, 246
Banner, Natalie, 150 Eckley, Idris, 50
Barrera, Marco Antonio Barrera, 168 Elias, Peter, 176
Battey, Heather, 115 Ellison, Joanne, 114
Beavan-Seymour, Colin, 215 Engel, Joachim, 171
Beckley-Hoelscher, Nick, 206 Eze, Jude, 17
Belloni, Pietro, 310
Belmont, Jafet, 312 F
Beltrao, Kaizo, 91, 267
Besbeas, Takis, 32 Fang, Zhou, 313
Blanche, Paul, 196 Ferguson, John, 185
Bowers, Megan, 228 Fernandez, Tamara, 254
Boyle, Laura, 131 Finselbach, Hannah, 108
Brown, Chloe, 119 Firth, David, 116
Brown, Dominic, 227 Fisher, Paul, 36
Brown, Emma, 181 Fitz-Simon, Nicola, 15
Buckingham-Jeffery, Elizabeth, 197 Fleming, Michael, 3
Bull, Lucy, 289 Flint, Iain, 147
Burke, Kevin, 217 Folorunso, Serifat, 290
Frantsuzova, Anastasia, 308
Fresneda-Portillo, 155
C Frigessi, Arnoldo, 51
Chen, Chunyi, 12 Fry, John, 96
Cheung, Daniel, 229 Fu, Chaoying, 93
Chigbu, Polycarp, 264
Chowdhury, Sritika, 209 G
Christodoulou, Evangelia, 165
Clarke, Paul, 33 Galwey, Nicholas, 21
Colwell, Scott, 62 Ganley, Chris, 133
Conde, Susana, 191 Garcia-Suaza, Andres, 58
Constantinescu, Caterina, 78, 295 Geue, Claudia, 22
Cook, Ben, 251 Gheno, Gloria, 208
Coolen, Anthony, 153 Giacalone, Massimiliano, 179
Coolen-Maturi, Tahani, 170 Gillespie, Colin, 189

319
Glennie, Richard, 31 Kim, Dongho, 146
Godolphin, Janet, 139 King, Thomas, 173, 291
Gorshechnikova, Anastasiia, 278 Kobayashi, Kaoru, 283
Gray, Nicholas, 192 Kondo, Hirofumi, 281
Green, Nathan, 57 Kosmidis, Ioannis, 48
Gregory, Rachel, 95 Kotecha, Meena Mehta, 154
Griffin, James, 16 Krone, Tanja, 286
Gupta, Juhi, 163 Kundhi, Gubhinder, 268
Kunst, Robert, 86
H Kvalseth, Tarald, 303

Hancock, Mark, 61
L
Hannah, Jack, 6
Harron, Katie, 25 Lacey, Andrea, 106
Hasegawa, Yasutaka, 279 Lapp, Linda, 194
Henderson, Neil, 98 Larkin, Jason, 9
Hill, Micki, 204 Lausen, Berthold, 77
Hill, Timothy Martyn, 97, 292 Le, Thai, 89
Hilliam, Rachel, 285 Lee, Clement, 20
Hogg, Rachel, 243 Lewis, Bonang, 107
Houwing-Duistermaat, Jeanine, 200 Lewis, Jonathan, 7
Hui, Huaihai, 74 Li, Kerchau, 277
Hunt, Ian, 230, 293 Li, Yan, 2
Hutcheson, Linda, 235 Lin, Xiaoming, 94
Hutchinson, Johanna, 145 Lloyd, Chris, 13
Lloyd, Louise, 49
I Lo, Sherman, 265
Lunagomez Coria, Simon, 210
Ignatieva, Anastasia, 318 Lut, Irina, 26
Izunobi, Chinyeaka, 269
M
J
MacDougall, Margaret, 257
Jackson, Dan, 183 Macey, Darren, 40
Jaouimaa, Fatima, 216 Macintyre, Cecilia, 160
Jawad, Majida, 299 Mahdi, Esam, 83
Jayetileke, Harshanie, 301 Mainey, Chris, 242
Jeffery, Caroline, 113 Makubate, Boikanyo, 297
Jessop, Ryan, 75, 247 Manktelow, Bradley, 135
Johansen, Kevin, 186 Marchant, Paul, 100, 273
Johnson, Olatunji, 46 Marriott, Nigel, 187
Jones, Geoffrey, 305 Marshall, Steven, 252
Martin, Maria Cristina, 4
K Martin, Peter, 244
Masters, Anthony, 80
Kaeding, Matthias, 195 McAneney, Helen, 24
Kanavou, Sofia, 70 McCrea, Rachel, 30
Kang, Sujin, 60, 261 McDowell, Cliona, 212
Kartsonaki, Christiana, 73, 104 McKinley, Jennifer, 182
Keatley, Debbie, 152 McLernon, David, 203
Kendall, Lindsay, 54 McLoone, Sean, 111
Kenne Pagui, Eugene Clovis, 237 McNeill, Tara, 214
Kennedy, Jack, 245 Mead, Andrew, 190
Kester Ugochukwu, Asugha, 81 Megardon, Geoffrey, 121
Kharroubi, Samer, 55 Mehrhoff, Jens, 241

320
Meng, Xiangyu, 287 R
Merritt, Laura, 10
Mesquita, Michel d. S., 84 Reese, Allan, 141
Mesquita, Michel D. S., 280 Roberts, Bill, 99
Mews, Sina, 29 Robertson, Gail, 231
Millar, Stuart, 175 Rodrigues, Eliane, 124
Mishra, Swapnil, 255 Rogers, Jennifer, 219
Moerbeek, Mirjam, 157 Rosenblum, Michael, 43
Mooney, Eugene, 134 Rozi, Shafquat, 275, 276
Moore, Daniel, 240
Moore, Matthew, 233 S
Mubarek, Khaled, 202
Mueller, Peter, 211 Saheed Abidemi, Agboluaje, 82
Mueller, Ursula, 169 Sansom, Philip, 125, 126
Mulrine, Stephanie, 151 Sartori, Nicola, 37
Munro, Robyn, 5 Schissler, Alfred, 122
Scott, Marian, 127
Scott-Hayward, Lindesay, 316
N
Selby, David, 232
Nakanishi, Shingo, 266 Semochkina, Daria, 72
Niccodemi, Gianmaria, 238 Sera, Francesco, 225
Nicholson, James, 39 Sergeant, Jamie, 156, 258
Nightingale, Glenna, 180, 259 Shang, Han Lin, 178
Shaw, Luke, 102, 166
O Shine, Martin, 188
Simkus, Andrea, 56
Ogden, Helen, 47 Song, Jiao, 90
Oladugba, Abimibola, 65, 272 Spencer, Neil, 172
Olhede, Sofia, 223 Srakar, Andrej, 14, 67
Olobadola, Micheal, 260 Stokes, Jamie, 205
O'Neill, Niall, 132 Sudell, Maria, 248, 282
Ormerod, Mark, 164 Sun, Jiajing, 199
Orsini, Nicola, 224 Sweeney, Aoife, 317
Osborn, Ellie, 162 Sweeney, Kevin, 253
Ots, Riinu, 284
Oyenuga, Iyabode Favour, 262 T
P Tai, Bee-Choo, 105
Taiyari, Katie, 306
Palmer, Brendan, 52, 271 Takeda, Jun, 184
Parnell, Andrew, 148 Tasker-Davies, Georgia, 101
Peng, Defen, 218 Tawn, Jonathan, 149
Pham, Phuong, 76 Teixeira Alves, Mickael, 87
Phillippo, David, 226 Thom, Howard, 250
Piao, Jin, 42 Thornley, Laura, 207
Poonvoralak, Wantanee, 302 Torkashvand, Elaheh, 300
Popa, Aura, 311 Tractenberg, Rochelle, 143
Popov, Valentin, 110 Trela-Larsen, Lea, 59
Potts, Jacqueline, 315 Tucker, James, 174
Pryce, Gwilym, 11 Turnbull, Kathryn, 222
Pyke, Caroline, 309
V
Q
Varty, Zak, 85
Qu, Chen, 118 Vekaria, Bindu, 270

321
Venkatasubramaniam, Ashwini, 198 White, Simon, 219
Verghis, Rejina, 159, 219 Wright, Andrew, 140
Vickerstaff, Victoria, 69 Wright, David, 130
Vieira, Rute, 128
Vinciotti, Veronica, 220 X
Viswakula, Sameera, 296
Vollmer, Michaela, 256 Xu, Mengjie, 92
Xue, Xiaohan, 177
W
Y
Wale-Orojo, Oluwaseun, 294
Walsh, Cathal, 249 Yin, Peng, 274
Walter, Stephen, 201 Yiu, Andrew, 117
Walwyn, Rebecca, 136 Young, Alastair, 38
Wan, Alan, 64
Wang, Xiaoqiang, 304 Z
Wardman, Leone, 144
Watson, Charlotte, 307 Zeitlin, Jennifer, 28
Watson, Joe, 8 Zhou, Li, 63
Webster, Anthony, 23, 298 Zhu, Wenyue, 193
Wei, Yinghui, 68 Zhu, Yayuan, 41
Wesonga, Ronald, 123 Zylbersztejn, Ania, 18, 27

322
 Royal Statistical Society
12 Errol Street, London EC1Y 8LX
020 7638 8998
[email protected]

rss.org.uk