BAYSM 2108: Bayesian Young Statisticians Meeting · Dependent processes in Bayesian nonparametric inference Igor Prunster, Bocconi University¨ ... models, for example by using graphical

BAYSM 2108:

Bayesian Young Statisticians Meeting2–3 July 2018, University of Warwick

Chair:Sara Wade (University of Warwick) [email protected]

Local and Scientific Organisers:Raffaele Argiento (University of Torino) [email protected] Barons (University of Warwick) [email protected] Durante (Bocconi University) [email protected] Spano (University of Warwick) [email protected]

Administrator:Paula Matthews (University of Warwick) [email protected] Desai (University of Warwick) [email protected]

Contents

1 Administrative Details 3

2 Timetable 5

3 Keynote talks, in order of appearance 7

4 Contributed talks, in order of appearance 8

5 Posters 16

6 Participant List 27

Map of Campus 30

We kindly thank our sponsors:

2

1 Administrative Details

Webpage

Any further information and announcements will be placed on the workshop webpage:warwick.ac.uk/baysm

Getting Here

• Information on getting to the University of Warwick can be found atwarwick.ac.uk/about/visiting

• Parking permits can be acquired at no further cost at (must be booked at least 5days in advance)https://carparking.warwick.ac.uk/events/bayesian-young-statisticians-meeting-2018

Registration and Locations

• Registration is open 8:30–9:00, Monday 2 July, in the main atrium of the Zeemanbuilding. Tea and coffee will be available.

• Talks will be held in room MS.01, Zeeman Building.

• Breakfast is provided for those with on-campus accommodation; specifically forConference Park accommodation, breakfast will be at the Rootes Restaurant (firstfloor of the Rootes Building).

• Lunch is provided on Monday and Tuesday in the atrium of the Zeeman building(and undergraduate workroom).

• Workshop Dinner will be at 19:30–22:00 on Monday 2 July at the Rootes LearningGrid in the Rootes Building.

• Informal Meeting will be held at 17:30-19:30 on Sunday 1 July at the Grad Deck.

• Poster session and wine reception is on Monday from 17:00 to 19:00, in the atriumof the Zeeman building.

Accommodation

Conference Park: ensuite rooms on campus in the Arthur Vick block. Keys can be col-lected from 15:00 to 22:45 at Conference Reception in the Student Union Atrium. If youplan to arrive after 22.45, please contact Conference Reception to arrange late key collec-tion [email protected] or 02476 528910. All rooms have linen and toiletries.All bedrooms must be vacated by 9:30am on the day of departure. Keys can be left at Con-ference Reception (in the Students Union building), Rootes Restaurant (in Rootes Build-ing) or one of the boxes situated in the entrance halls of each residence. A luggage roomis available at Conference Reception.

Radcliffe/Arden: ensuite rooms on campus. Check-in time will be from 15:00 at Rad-cliffe/Arden reception and check-out will be before 10:00.

3

Internet Access

• Campus: Wireless access is most easily available via eduroam, which is supportedacross most of the Warwick campus. See eduroam.org.

• Accommodation: Wireless access is available, ask for log-in details at checkin-in.

Facilities

• Supermarket, Food and Drink Outlets: details (along with locations and openingtimes) can be found at warwick.ac.uk/services/retail

• Arts Centre: warwickartscentre.co.uk

• Sports Centre: warwick.ac.uk/sport

• Health Centre: uwhc.org.uk

• Pharmacy: Students Union Atrium. Open 9am - 6pm.

Telephone Numbers

• Emergency: Internal - 22222; External - 024 7652 2222

• Security: Internal - 22083; External - 024 7652 2083

• Department of Statistics: Internal - 574812; External - 024 7657 4812

Taxis

• Swift Cabs 024 7777 7777

• Trinity Street Taxis 024 7699 9999

4

2 Timetable

Talks will take place in room MS.01 in the Zeeman Building (Mathematics & Statistics), Univer-sity of Warwick. Poster session and lunches will be held in the Atrium of the Zeeman Building.

Sunday, 1st July

17:30-19:30: Informal Meeting at the Grad Deck (light snacks and drinks provided)

Monday, 2nd July

8:30 – 9:00: Registration and coffeeKeynote (co-chaired by Agnieszka Borowska and Oliver Stevenson)9:00 – 10:00: Kerrie Mengersen “Challenging Data and Bayesian Solutions.”10:00 – 10:30: Coffee breakContributed talks:Applications Discussant/Chair: Amy Herring10:30 – 10:50: Tommaso Rigon “A functional Pitman-Yor mixture model for

flight routes segmentation.”10:50 – 11:10: Charles Au “Analysis of Australian voters’ political attitudes: a

Bayesian Grade of Membership approach.”11:10 – 11:30 Alex Diana “Polya Tree Bayesian nonparametrics.”Medical Statistics Discussant/Chair: Deborah Ashby11:30 – 11:50: Iliana Peneva “A Bayesian Nonparametric Model for Integrative

Clustering of Omics Data.”11:50 – 12:10: Kristian Brock “A Phase II Clinical Trial with Efficacy and Toxicity

Outcomes and Baseline Covariates.”12:10 – 12:30 Nicolo Margaritella “Parameters clustering in Bayesian functional

PCA of neuroscientific data.”12:30 – 13:30: LunchContributed talks:Methods Discussant/Chair: Michele Guindani13:30 – 13:50: Gemma Moran “Spike-and-Slab Lasso Biclustering .”13:50 – 14:10: Matthew Sutton “Multivariate Bayesian Sparse Group Selection.”14:10 – 14:30 Jack Jewson “Principled Bayesian Minimum Divergence Infer-

ence.”Subjective Discussant/Chair: Jim Smith14:30 – 14:50: Fiona Turner “Ice Cores and Emulation: Learning More About Past

Ice Sheet Shapes .”14:50 – 15:10: Bjorn Schrinski “Towards Bayesian hypothesis testing of macrore-

alistic modifications of quantum mechanics.”15:10 – 15:30 Duco Veen “Using the Data Agreement Criterion to Rank Experts’

Beliefs.”15:30 – 16:00: Coffee breakKeynote (co-chaired by Marta Crispino and Maxime Rischard)16:00 – 17:00: Igor Prunster “Dependent processes in Bayesian nonparametric in-

ference.”17:00 – 19:00: Poster session19:30 – 22:00: Conference dinner, Rootes Learning Grid

5

Tuesday, 3rd July

Keynote (chaired by Hoang Le Vu and Tommaso Rigon)9:00 – 10:00: Yee Whye Teh ”Bayesian Nonparametric Modelling with Neural

Networks.”10:00 – 10:30: Coffee breakContributed talks:Bayesian Learning Discussant/Chair: Sebastian Vollmer10:30 – 10:50: Cedric Spire “Novel Bayesian Unsupervised Learning of System

Parameters using Dynamical Information.”10:50 – 11:10: Daniel Tait “Multiplicative Latent Force Models.”11:10 – 11:30 Nishma Laitonjam “Non-parametric Overlapping Community

Detection.”Spatio-Temporal Discussant/Chair: Barbel Finkenstadt11:30 – 11:50: Chiara Forlani “A Bayesian Space-Time Model to Integrate Spa-

tially Misaligned Air Pollution Data.”11:50 – 12:10: Gabriel Riutort-Mayol “Spatio-temporal Gaussian processes with

derivative information.”12:10 – 12:30 Cecilia Balocchi “Bayesian Spatial Clustering with Particle Opti-

mization: Crime in Philadelphia.”12:30 – 13:30: LunchKeynote (co-chaired by Amani Alahmadi and Leah South)13:30 – 14:30: Judith Rousseau “Using asymptotics to understand ABC.”Contributed talks:Computations Discussant/Chair: Jim Griffin14:30 – 14:50: Xiaoyue Xi “Bayesian Quadrature for Multiple Related Integrals.”14:50 – 15:10: Changye Wu “Discrete Piecewise Deterministic Markov Processes

using Hamiltonian dynamics.”15:10 – 15:30 Agnieszka Borowska “Semi-Complete Data Augmentation for Ef-

ficient State-Space Model Fitting.”15:30 – 16:00: Coffee breakFinance and Economics Discussant/Chair: Mark Steel16:00 – 16:20: Adam Smith “Bayesian Analysis of Demand Models with Random

Partitions.”16:20 – 16:40: Alexander Kreuzer “Bayesian inference for a single factor copula

based stochastic volatility model using Hamiltonian Monte Carlo.”16:40 – 17:00 Darjus Hosszejni “Efficient Bayesian Estimation of the Stochastic

Volatility Model with Leverage.”Keynote (co-chaired by Gemma Moran and Kees Mulder)17:00 – 18:00: Stephen Senn “Will Bayesian analysis deal with the “replication

crisis”?”18:00 – 19:00: Closing and Awards for Best Talk/Poster.

6

3 Keynote talks, in order of appearance

Challenging Data and Bayesian SolutionsKerrie Mengersen, Queensland University of Technology

The data of the present and the anticipated data of the future offer unprecedentedchallenges for modelling and analysis. In this presentation we will explore some of thesedata sources and types and discuss how Bayesian methods might be used to addressthem. We will do this in the context of some real-world case studies.

Dependent processes in Bayesian nonparametric inferenceIgor Prunster, Bocconi University

We consider models based on dependent completely random measures for inferentialproblems with complex dependence structures. Some of their marginal and conditionaldistributional properties are presented with focus on hierarchical constructions, but alsoadditive and nested structures will be discussed. These distributional results provideinsight into the inferential implications of the considered models and allow to deriveeffective sampling schemes. Popular nonparametric models are obtained as special cases.Illustrations related to species sampling problems and survival analysis are provided.

Bayesian Nonparametric Modelling with Neural NetworksYee Whye Teh, University of Oxford

In Bayesian statistics, there has long been a desire to make our models more flexibleto capture the complexities of data, by increasing the capacity and expressivity of ourmodels, for example by using graphical models and Bayesian nonparametrics. Neuralnetworks can also be thought of as ways to make our models more flexible, albeit indifferent ways. In this talk I will introduce neural networks and deep generative modelsto a statistical audience, and give a number of examples of how they can be used in setupsthat are otherwise Bayesian nonparametric in philosophy.

Using asymptotics to understand ABCJudith Rousseau, University of Oxford & Universite Paris Dauphine

Approximate Bayesian computations are used typically when the model is so complexthat the likelihood is intractable but data can be generated from the model. With the initialfocus being primarily on the practical import of this algorithm, exploration of its formalstatistical properties has begun to attract more attention. In this work we consider theasymptotic behaviour of the posterior obtained by this method and the ensuing posteriormean. We give general results on: (i) the rate of concentration of the resulting posterioron sets containing the true parameter (vector); (ii) the limiting shape of the posterior;and (iii) the asymptotic distribution of the ensuing posterior mean. These results holdunder given rates for the tolerance used within the method, mild regularity conditionson the summary statistics, and a condition linked to identification of the true parameters.I will show in particular that we have very different behaviours if the model is well ormis-specified. I will highlight what are the practical implications of these results on theunderstanding of the behaviour of the algorithm.Joint work with David Frazier, Gael Martin and Christian Robert.

7

Will Bayesian analysis deal with the “replication crisis”?Stephen Senn, Luxembourg Institute of Health

It has been alleged that science is suffering from a replication crisis and that P-valuesand the cult of significance are largely to blame. Various cures have been proposed byvarious authorities including replacing P-values by Bayes factors and adopting muchmore stringent levels for significance in an attempt to get significance tests to approx-imate Bayesian standards. I shall argue that although these proposals are not withoutinterest, failure to adopt them cannot be the explanation for any supposed replicationcrisis. Amongst matters that have been overlooked are 1) that the common standard ofreplication is inappropriate 2) that the problem of false negative results cannot be ig-nored 3) that classical P-values are very similar to a type of Bayesian analysis with a longpedigree and 4) that when striving to understand the frequency with which replicationfailures occur, frequentist properties are, after all, important. I conclude that replication ispoorly understood and that if there is a problem to cure, the cure will be as much culturalas technical.

4 Contributed talks, in order of appearance

Contributed Session I: Applications, 10:30 on Monday

A functional Pitman-Yor mixture model for flight routes segmentationTommaso Rigon, Bocconi University

A private company selling flight tickets is interested in clustering functional observa-tions to implement data-driven marketing strategies. We address this problem propos-ing a functional Pitman-Yor process with penalized B-splines and discussing its practicaladvantages in this specific application. To overcome computational difficulties, we em-ployed a variational Bayes approximation for tractable posterior inference. The proposedmodel is finally applied to a recent dataset.

Analysis of Australian voters’ political attitudes: a Bayesian Grade ofMembership approach

Charles Au, The University of Sydney

The Australian Election Study (AES) survey asks voters, who are eligible to vote inAustralian elections, questions regarding their political attitudes and interest in politicsand election campaigns. Political beliefs can be multidimensional and heterogeneous fordifferent voters. As the vast majority of variables on the AES is categorical, it is usefulto classify voters into groups based on certain characteristics. One method is to use theGrade of Membership (GoM) model, where voters could have partial membership (ratherthan full membership) in different groups (extreme profiles) and assigned GoM scores foreach profile. The Bayesian Markov chain Monte Carlo (MCMC) approach is adopted forparameter estimation of the GoM model. To facilitate the use of the Gibbs sampler, theGoM model is expressed hierarchically into a latent class model via data augmentation.This paper will look at grouping voters in the AES survey data in 2016 based on theirlevel of confidence in major societal institutions in Australia.

Polya Tree Bayesian nonparametricsAlex Diana, University of Kent

8

Many ecological sampling schemes do not allow for individual marking but provideinstead only the total count of individuals detected at each sampling occasion, as opposedfor example to capture-recapture scheme. In this paper we propose a novel approach tomodel count data in an open population where individuals can arrive and depart fromthe site during the sampling period. A Bayesian nonparametric prior known as PolyaTree is used for modelling the bivariate density of arrival and departure times. Thanks tothis choice, we can easily incorporate prior information on the density while still allowingthe model to flexibly adjust the posterior inference according to the observed distributionof the counts. Moreover, the model provides great scalability as the complexity does notdepend on the population size but just on the number of sampling occasions, making itparticularly suitable for data-sets with high numbers of detections. We apply the newmodel to count data of newts collected by the Durrell Institute of Conservation and Ecol-ogy, University of Kent.

Contributed Session II: Medical Statistics, 11:30 on MondayA Bayesian Nonparametric Model for Integrative Clustering of Omics Data

Iliana Peneva, University of Warwick

Cancer is a complex disease, driven by a range of genetic and environmental fac-tors. We have developed a Bayesian nonparametric model for combined data integrationand clustering called BayesCluster, which aims to identify cancer subtypes and addressesmany of the issues faced by the existing integrative methods. The proposed method canintegrate and use the information from multiple different datasets, and can capture richerbiological structures by using nonlocal priors. We incorporate feature learning becauseof the large number of predictors, and use a Dirichlet process mixture model approachto learn the number of subtypes. We investigate different approaches to speed up theMCMC computations to ensure that the inference is tractable. We apply the model todatasets from the Cancer Genome Atlas project of glioblastoma multiforme, which con-tains clinical and biological data about cancer patients with extremely poor prognosis ofsurvival. By combining all available information we are able to be better identify clinicallymeaningful subtypes of glioblastoma, opening the way for new personalised treatments.

A Phase II Clinical Trial with Efficacy and Toxicity Outcomes and BaselineCovariates

Kristian Brock, University of Birmingham

PePS2 is a phase II trial of the efficacy and safety of pembrolizumab in performancestatus 2 non-small-cell lung cancer patients. Previous studies have shown that efficacy iscorrelated with the extent to which PD-L1 is expressed in the tumour, and pretreatedness.There are few clinical trial designs that test co-primary efficacy and toxicity outcomes inphase II, and fewer still that incorporate baseline covariates. Thall, Nguyen and Esteypresent one such design but it has been scarcely used in trials. Their Bayesian model in-corporates terms to conduct a dose-finding study. This aspect is not required in PePS2because a candidate dose has been widely tested. We introduce a novel simplificationfor phase II that focuses on testing efficacy and toxicity whilst adjusting for baseline co-variates. The method shares information across cohorts. Simulations show it is far moreefficient than analysing cohorts separately. Using the design in PePS2 with 60 patients totest the treatment in six cohorts, we can expect error rates typical of those used in phase IItrials. However, we demonstrate that care must be used when specifying the models forefficacy and toxicity because more complex models require greater sample sizes for biasto be controlled.

9

Parameters clustering in Bayesian functional PCA of neuroscientific dataNicolo Margaritella, University of Edinburgh

The modelling of brain activity has a long history but only in recent years has it startedbenefiting from the extraordinary technological advances in the fields of neurophysiol-ogy and neuroimaging. With a remarkable amount of spatiotempo- ral data recordablesimultaneously from several parts of the brain, researchers are challenged to find mod-els that can capture meaningful patterns and structures be- hind such complexity. Thusmotivated, we aim at modelling neuroscientific data employing functional data analysis(FDA) within a Bayesian perspective; in partic- ular, we exploit the flexibility of an infinitemixture model for clustering functional principal component scores to account for curvedependencies. We show that the clustering of parameters in functional Principal Compo-nent Analysis (PCA) can be expressed in terms of a hierarchical model and offers a moregeneral approach than time series clustering, thus allowing for a much finer curves classi-fication. More- over, we present results from a simulation study showing improvementsin curves reconstruction, especially when the signal-to-noise ratio is low.

Contributed Session III: Theory and Methods, 13:30 on Monday

Spike-and-Slab Lasso BiclusteringGemma Moran, University of Pennsylvania

Biclustering has become a popular tool, particularly in the analysis of gene expressiondatasets. Such biclustering methods find subsets of genes which act similarly in only asubset of the samples. This is unlike usual clustering methods which utilize the entireset of genes to group samples, potentially missing important information. Biclusters ofinterest often manifest as rank-1 submatrices of the data matrix. This submatrix detec-tion problem can be viewed as a factor analysis problem where both factors and load-ings are sparse. In this paper, we propose a new biclustering method which utilizes theSpike-and-Slab Lasso of Rockova and George (2016) to find such a sparse factorization ofthe data matrix. This is achieved using a fast, deterministic EM algorithm that rapidlyidentifies promising biclusters. This method, called Spike-and-Slab Lasso Biclustering,outperforms other biclustering methods in a variety of simulation settings.

Multivariate Bayesian Sparse Group SelectionMatthew William Sutton, Queensland University of Technology

We consider sparse Bayesian estimation for a multivariate response variable in a lin-ear regression where covariates have grouping structure. For recovery of the underlyingsignal, spike and slab priors are known to give optimal results. However, these priorsoften come at the expense of too narrow credible sets and can be computationally ex-pensive. Alternative approaches such as the Bayesian group lasso provide a smootherprior but the full posterior can spread too much, leading to sub-optimal recovery rates.In this work we propose methods which combine the shrinkage effects of the Bayesiangroup lasso with the spike and slab to trade off between these two priors. We deriveefficient Gibbs sampling algorithms for our models and provide the implementation ina comprehensive R package called MBSGS available on the CRAN. The performance ofthe proposed approaches is compared to state-of-the-art variable selection strategies onseveral real and simulated data sets.

Principled Bayesian Minimum Divergence Inference

10

Jack Jewson, University of Warwick

In the M-open world, the decision maker must concern themselves with the KL-divergence minimising parameter in order to maintain principled statistical practice (Walker,2013). However, it has long been known that the KL-divergence places a large weight oncorrectly capturing the tails of the data generating process. As a result traditional infer-ence can be very non-robust. We build upon general Bayesian updating (Bissiri, Holmesand Walker, 2016) to propose a statistically well principled Bayesian updating of beliefstargeting the minimisation of any statistical divergence. We improve both the motiva-tion and the statistical foundations of existing Bayesian minimum divergence estimation(Hooker and Vidyashankar, 2014, Ghosh and Basu, 2016), allowing the well principledBayesian to target predictions from the model that are close to the data generating pro-cess in terms of some alternative divergence measure to the KL-divergence. We arguethat defining this divergence measure forms an important, subjective part of any statisti-cal analysis, allowing model specification to be decoupled from issues of robustness. Wecompare the performance of several different divergence measures for conducting simpleinference tasks on both simulated and real data sets. We follow this by demonstrating theimpact altering the target divergence can have in a high dimensional on-line changepointdetection scenario, where inferences using the KL-divergence can be shown to over detectthe number of changespoints under misspecification. Work supervised by Jim Q. Smithand Chris Holmes, collaborating with Jeremias Knoblauch and Theo Damoulas.

Contributed Session IV: Subjective Bayes and Prior Elicitation, 14:30 on Mon-day

Ice Cores and Emulation: Learning More About Past Ice Sheet ShapesFiona Turner, University of Sheffield

As computer power has increased, extensive research has been done on reconstructingice sheets at both poles. This is done using mathematical computer models of climatecombined with data and various physical constraints, providing us with estimates of theshape and size of the ice sheets. The focus is often on critical time periods, when the icesheets were most in flux and there is greatest variance in estimates.

Towards Bayesian hypothesis testing of macrorealistic modifications ofquantum mechanics

Bjoern Schrinski, University of Duisburg-Essen

We discuss a Bayesian approach to test a class of macrorealistic modifications to quan-tum mechanics known as collapse models. They provide an ad-hoc explanation for thequantum-to-classical transition by adding hypothetical nonlinear and stochastic terms tothe Schrodinger equation. In most of the literature, experimental tests of these modelsare discussed by presenting measurement results that rule out a volume of the underly-ing parameter space. While a frequentistic analysis of the data and the comparison withthe theoretical prediction is well established, a Bayesian approach would allow one to in-clude different, (un-)correlated observables and thus overall improve the hypothesis test.Tests of such “classicalizing” macrorealistic modifications can also be used as a naturaldefinition for the degree of macroscopicity reached in a given superposition experiment- one is more macroscopic than another if it rules out more modification parameters. Forthis, a standardized and consistent Bayesian treatment for all conveivable, or not yet con-ceived, kinds of experiments is highly desirable.

11

Using the Data Agreement Criterion to Rank Experts’ BeliefsDuco Veen, Utrecht University

Experts’ beliefs embody a present state of knowledge. It is desirable to take thisknowledge into account when doing analyses or making decisions. Yet ranking expertsbased on the merit of their beliefs is a difficult task. We show how experts can be rankedbased on their knowledge and their level of (un)certainty. This can be done via a pro-cess called expert elicitation. Elicitation entails the extraction of expert knowledge andtranslating this knowledge into the probabilistic representation (Hagan et al., 2006). Byusing a probabilistic representation can assess how accurately they can predict new data,and how appropriate their level of (un)certainty is. The expert’s specified probability dis-tribution can be seen as a prior in a Bayesian statistical setting. By extending the DataAgreement Criterion (DAC) (Bousquet, 2008), an existing prior-data conflict measure toevaluate multiple priors, i.e. experts’ beliefs, we can compare experts with each other andthe data to evaluate their appropriateness. The DAC is based on Kullback-Leibler (KL)divergences which ensures that prior-data conflict decisions are based upon the full prob-ability distributions. Furthermore the DAC incorporates a clear classification of prior-data conflict. Using this method new research questions can be asked and answered, forinstance: Which expert predicts the new data best? Is there agreement between my ex-perts and the data? Which experts’ representation is more valid or useful? Can we reachconvergence between expert judgement and data? We provided an empirical exampleranking (regional) directors of a large financial institution based on their predictions ofturnover.

Contributed Session V: Advances in Bayesian Learning, 10:30 on Tuesday

Novel Bayesian Unsupervised Learning of System Parameters, usingDynamical Information .

Cedric Spire, Loughborough University

Often, real-world systems are characterised by system behavioural/structural param-eters that are themselves not directly measured, but another variable that affects the sys-tem behaviour, is. In such situations, scientists seek values of the system parameters, atwhich a measured value of this observable is realised. If, a set of pairs of values of systemparameters and observable is available–in the form of training data–supervised learningcan be undertaken to learn the functional relationship between the system vector andthe observables, subsequently enabling parameter prediction. However, multiple real-world problems exist, (spanning Neuroscience, Astronomy, Finance, Engineering, etc.),in which training-data remains unavailable, and still, predicted value of the system pa-rameters is sought, at test-data on the observable. We present a novel Bayesian methodto deal with such unsupervised learning of a structural parameter vector of a stationarydynamical system, by invoking information on the dynamical evolution of the state spacepdf . The sought system vector is embedded into the support of the state space density ,and inference on the unknowns is performed with Metropolis-within-Gibbs. We illustrateour methodology to address a challenging astronomical problem of identification of thefraction of dark matter in a real galaxy, using the only available, missing test-data on twodifferent kinds of galactic particles.

Multiplicative Latent Force ModelsDaniel Tait, The University of Edinburgh

12

Bayesian modelling of dynamic systems must achieve a compromise between provid-ing a complete mechanistic specification of the process while retaining the flexibility tohandle those situations in which data is sparse relative to the model complexity, or a fullspecification is hard to motivate. Latent force models achieve this dual aim by specify-ing a parsimonious linear evolution equation with an additive latent Gaussian processforcing term. The resulting model is equivalent to a GP regression with mean and kerneldepending on the the linear dynamics. In this work we extend the latent force frameworkto allow for multiplicative interactions between the latent GPs and the state variables.This enables the embedding of additional geometric structure in the model at the expenseof the a tractable posterior distribution. To proceed with inference we introduce a sam-pling method derived from a truncated Neumann series expansion of the trajectories. Wedemonstrate the performance of our method on a dynamic system possessing both a highdegree of geometric structure as well as a tractable solution and see excellent agreementbetween our approximate method and the true conditional distribution.

Non-parametric Overlapping Community DetectionNishma Laitonjam, University College Dublin

In this paper we present an overlapping community detection algorithm by extendingthe Affiliation Graph Model (AGM) to a non-parametric model from which the numberof communities as well as community assignments, can be inferred. In previous work,a non-parametric extension of the AGM, which we refer to as GP-AGM, that employs aGamma-process prior has been proposed. In contrast, we combine the AGM likelihoodwith an Indian Buffet Process (IBP) prior and develop some sophisticated Metropolis-Hastings moves to fit its parameters. We compare our model on benchmark datasets, withthe GP-AGM as well as some other similar models, and show that our model performswell when the data contains clear community, rather than more general block structure.

Contributed Session VI: Bayesian Spatio-Temporal Analysis, 11:30 on Tuesday

A Bayesian Space-Time Model to Integrate Spatially Misaligned Air PollutionData

Chiara Forlani, Imperial College London

In air pollution studies, concentration levels from monitoring stations are usuallycombined with dispersion models to obtain complete spatial coverage. However, thesedata sources are misaligned in space and time. If misalignment is not considered, it canbias the results from the statistical inference. In this study we consider NO2 concentra-tion, and aim at improving predictions at a regular grid level combining two differentdispersion models, the Air Quality Unified Model (AQUM) and the Pollution ClimateMapping (PCM) model, in Greater London, for the years 2007-2011. We propose a jointmodel that reconstructs the spatial fields of the misaligned covariates from the numericalmodel outputs together with the latent field of the response variable (concentration levelat monitoring stations), accounting for different sources of uncertainty. The method is im-plemented through the Integrated Nested Laplace Approximation (INLA) and StocasticPatial Differential Equation (SPDE) framework. Our spatio-temporal model allows us toreconstruct the latent fields of each model component, as well as to predict daily maps ofpollution concentrations.

Spatio-temporal Gaussian processes with derivative informationGabriel Riutort-Mayol, Universitat Politecnica de Valencia

13

In modeling problems there is often a priori knowledge available concerning the func-tion to be learned. This information can be sometimes expressed in terms of the deriva-tives of the functions. In this work, we propose a spatio-temporal model that takes theinformation of derivatives into account by jointly modelling a spatio-temporal processand its derivative process using Gaussian processes. Experimental results obtained onMicrofading Spectrometry (MFS) observations collected on the surface of rock art paint-ings show effectiveness of the proposed modeling fitting the functions dynamics of thedata and improving their interpretability in specific applications.

Bayesian Spatial Clustering with Particle Optimization: Crime inPhiladelphia

Cecilia Balocchi, University of Pennsylvania

Understanding the relationship between change in crime over time and the geographyof urban areas is an important problem for urban planning. Accurate estimation of chang-ing crime rates throughout a city would aid law enforcement as well as enable studies ofthe association between crime and the built environment. Bayesian hierarchical model-ing is a promising direction since areal data, such as crime counts in neighborhoods overtime, require principled sharing of information to address spatial autocorrelation betweenproximal locations. Fully parametric models such as the CAR model of Besag [1974] in-duce spatial smoothing across geographically proximate areal units. However, in com-plex urban environments, there may be sharp boundaries intrinsic to the city geographyor population distribution that results in distinct clusters of areal units that can exhibitmarkedly different trends. Incorporating knowledge of this latent underlying partitionof areal units within our hierarchical model can improve parameter estimation. Typically,this partition is unknown a priori and the usual stochastic search techniques for that par-tition are computationally prohibitive since these searches must explore a vast discretespace of possible partitions. In this work, rather than directly sampling from the pos-terior distribution of partitions, we introduce an ensemble optimization procedure thattargets the partitions with largest posterior probability. We run several greedy searchesover the posterior distribution of partitions that are made “mutually aware” through anentropy penalty and an Variation of Information penalty that repels search trajectoriesthat appear headed to the same point. We apply our developed methodology to estimatechanges in crime throughout Philadelphia over the 2006-17 period.

Contributed Session VII: Computational Statistics, 14:30 on Tuesday

Bayesian Quadrature for Multiple Related IntegralsXIAOYUE XI, University of Warwick

Bayesian probabilistic numerical methods are a set of tools providing posterior dis-tributions on the output of numerical methods. The use of these methods is usuallymotivated by the fact that they can represent our uncertainty due to incomplete/finiteinformation about the continuous mathematical problem being approximated. In this pa-per, we demonstrate that this paradigm can provide additional advantages, such as thepossibility of transferring information between several numerical methods. This allowsusers to represent uncertainty in a more faithful manner and, as a by-product, provideincreased numerical efficiency. We propose the first such numerical method by extendingthe well-known Bayesian quadrature algorithm to the case where we are interested incomputing the integral of several related functions. We then prove convergence rates forthe method in the well-specified and misspecified cases, and demonstrate its efficiency in

14

the context of multi-fidelity models for complex engineering systems and a problem ofglobal illumination in computer graphics.

Discrete Piecewise Deterministic Markov Processes using Hamiltoniandynamics

CHANGYE WU, University Paris-Dauphine

In Hamiltonian Monte Carlo (HMC) algorithms, one refreshes the momentum at eachiteration and discards all expect for the last one the proposals proposed during leapfrogintegrators. Such refreshment introduces much randomness in terms of exploring the tar-get space, while discarding the intermediate proposals wastes much computation effort.In this article, we propose a randomness-controlled HMC algorithm, which can overcomeabove two potential drawbacks and coincides with the recently prevailing piecewise de-terministic Markov process sampler.

Semi-Complete Data Augmentation for Efficient State-Space Model FittingAgnieszka Borowska, University of Glasgow

State space models are an intuitive and flexible class of models, frequently used inpractice. Their flexibility, however, comes at the price of significantly more complicatedfitting of such models to data as the associated likelihood is typically analytically in-tractable. For the general case a Bayesian data augmentation (DA) approach is oftenemployed, where the true unknown states are treated as auxiliary variables and imputedwithin the MCMC algorithm. However, standard ”vanilla” MCMC algorithms may per-form very poorly due to high correlation between the imputed states, leading to the needto specialist algorithms being developed. The proposed method circumvents this ineffi-ciency by combining DA with numerical integration in a Bayesian hybrid approach. Thedeveloped algorithm permits standard “vanilla” sampling schemes to be applied for up-dating the imputed states that preform considerably better than the traditional approach.The proposed Semi-Complete Data Augmentation algorithm is then applied to differenttypes of problems demonstrating efficiency gains in empirical studies.

Contributed Session VIII: Applications in Finance and Economics, 16:00 onTuesday

Bayesian Analysis of Demand Models with Random PartitionsAdam Smith, University College London

Many economic models of consumer demand require researchers to partition sets ofproducts or attributes prior to the analysis. These models are common in applied prob-lems when the product space is large or spans multiple categories. While the partitionis traditionally fixed a priori, we let the partition be a model parameter and propose aBayesian method for inference. We build on previous covariate-dependent random parti-tion models to construct a new partition distribution characterized by a location partitionand scale parameter. This location-scale partition distribution is useful in two ways: (1)as a proposal distribution within the context of a Markov chain Monte Carlo routine; and(2) as a prior or random-effects distribution. Our method is illustrated within the contextof a store-level category demand model, where we find that allowing for uncertainty inthe partition is important for preserving model flexibility, improving demand forecasts,and learning about the structure of demand.

15

Bayesian inference for a single factor copula based stochastic volatility modelusing Hamiltonian Monte Carlo

Alexander Kreuzer, Technische Universitat Munchen

Single factor models are commonly used in finance to model the joint behaviour ofstocks. The dependence is usually modelled with a multivariate normal distribution.[6] provide an extension of this, the factor copula. The one factor copula requires thespecification of bivariate linking copulas. Thus resulting joint models can accommodatesymmetric or asymmetric tail dependence. To allow for multivariate time series we use aone factor copula together with stochastic volatility margins. We develop joint Bayesianinference for such models using Hamiltonian Monte Carlo within Gibbs sampling. Thuswe avoid the two step approach for mar- gins and dependence in copula models as sug-gested by [8]. The approach is illustrated for a portfolio of financial assets with respect toone day ahead value at risk forecasts.

Efficient Bayesian Estimation of the Stochastic Volatility Model with LeverageDarjus Hosszejni, WU Vienna, Austria

The sampling efficiency of MCMC methods in Bayesian inference for stochastic volatil-ity (SV) models is known to highly depend on the actual parameter values, and the effec-tiveness of samplers based on different parameterizations differs significantly. We derivenovel samplers for the centered and the non-centered parameterizations of the practicallyhighly relevant SV model with leverage, where the return process and innovations of thevolatility process are allowed to correlate. Moreover, based on the idea of ancillarity-sufficiency interweaving, we combine the resulting samplers in the hope of achievingsuperior sampling efficiency, irrespectively of the baseline parameterization. The methodis implemented using R and C++, with the help of Rcpp for easy interfacing between thetwo languages. Finally, we carry out an extensive comparison to already existing sam-pling methods for this model.

5 Posters

Modelling falls in the elderly: Chain Event Graphs for modelling publichealth interventions

Aditi Shenvi, Centre for Complexity Science, University of Warwick

Interventions in public health typically focus on some aspect(s) of the training of ser-vice providers such as doctors, nurses and pharmacists, or assessment methods, refer-ral pathways, treatment plans or post-treatment care of patients. Asymmetries in mod-els of such interventions manifest as context-specific conditional independence relationsbetween the variables of the model. One way of analysing such data would be to useBayesian Network (BN) technologies. However, in this setting, the lack of symmetry inthe problem means that this approach is rather contrived. An alternative would be tomodel it using the Chain Event Graph (CEG) which are much more sympathetic to theanalysis of tree-like structures that directly express both sequencing of events and contex-tual independence relations in their topology. In this paper, we model the pre-trial processof an intervention to reduce the risk and rate of falls in people aged over 65 years withCEGs which can directly model and represent the underlying process. We then comparethe CEG model to the more conventional BN model when applied to this setting.

16

Sparse Bayesian regression for “large n and large p”: accelerating Gibbssampler for logistic models via conjugate gradient methods

Akihiko Nishimura, University of California - Los Angeles

In a modern observational study based on healthcare databases, the number of obser-vations is typically in the order of 105 106 and that of the predictors in the order of 104 105.Although the sample size is large, it is rarely large enough to estimate the regression co-efficients without some sparsity assumptions. Sparse regression plays an important rolein such situations, and the Bayesian approach based on shrinkage priors has been shownto possess many desirable theoretical properties. Despite the rich literature on theoriesof Bayesian shrinkage priors, the progress in the required computational methods haslargely been limited to the p >> n case. The posteriors are amenable to Gibbs sampling,but a major computational bottleneck arises from the need to repeatedly sample fromhigh-dimensional Gaussian distributions, whose precision matrix Φ has a closed-form ex-pression yet is expensive to compute and factorize numerically. In this article, we presenta novel algorithm to speed up this bottleneck based on the following observation: we cancheaply generate a random vector b such that the solution of a linear system Φβ = b hasthe desired Gaussian distribution. An accurate solution of the linear system can then befound by the conjugate gradient algorithm with only a small number of the matrix-vectormultiplications by Φ, without ever explicitly inverting Φ. We apply our algorithm to an-alyze a data set from the OHDSI project — in which the design matrix is of size 72,489 by22,175 — and demonstrate an order of magnitude speed-up in the posterior computation.

Batch effect correction using Bayesian latent factor regression via non-localpriors

Alejandra Avalos Pacheco, University of Warwick

Batch effects are a source of experimental variation that is present in many largedatasets and adds difficulty to the task of integrating and summarising these data forexploratory analyses. In order to keep up with the large influx of biological data, avail-able due to high-throughput technologies, new dimensionality reduction techniques areneeded for effective understanding of this information. There is a need for innovativemethods that can integrate data from different batches while preventing these technicalbiases from dominating the results. We provide a model based on factor analysis andlatent factor regression, which incorporates a novel adjustment for variance that is dueto batch effects and is often observed in bioinformatics data. This model is extended byusing different sparse priors (local and non-local). Finally, a toy example and a motiva-tion case study based on ovarian cancer datasets are presented and discussed, along withdirections for future research.

Exact and Approximate Bayesian Methods of Sampling from PosteriorDistributions in ODE Models

Amani Alahmadi, Monash University

The behaviour of many processes in science and engineering can be accurately de-scribed by dynamical systems models consisting of a set of ordinary differential equations(ODEs). Often these models are associated with parameters that are difficult to obtainby experiment, in which case Bayesian inference can be a useful tool. In principle, ex-act Bayesian inference using Markov chain Monte Carlo (MCMC) techniques is possible,however, in practice, such methods may suffer from slow convergence and poor mixing.

17

To address this problem several approaches based on Approximate Bayesian Computa-tion (ABC) have been introduced, including Markov chain Monte Carlo ABC (MCMCABC) and ABC Sequential Monte Carlo (ABC-SMC). While the system of ODEs describesthe underlying process that generates the data, the observed measurements invariablyinclude errors. In this paper, we argue that several popular ABC approaches fail to ade-quately model these errors because the acceptance probability depends on the choice ofthe discrepancy function and the tolerance with out any consideration of the error term.We observe that the so-called posterior distributions derived from such methods do notaccurately reflect the epistemic uncertainties in parameter values, however, we proposeda solution for SMC ABC that can be used to include the error part and solve this problem.We illustrate the results in two ODE epidemiological models with simulated data and onewith real data.

Hypothesis test statistics in approximate Bayesian computationAnthony Ebert, ARC Centre of Excellence for Mathematical & Statistical Frontiers

Approximate Bayesian computation (ABC) samplers are useful for parameter infer-ence whenever (conditional on parameter values) is it possible to simulate data but it isnot possible to evaluate the likelihood. The approach of ABC is to accept parameter val-ues which produce simulated data which, when compared to the observed data is “close”.Frequently the notion of “close” is defined with reference to lower dimensional summarystatistics but this can cause problems when appropriate summary statistics are difficult todefine. A new approach is to use metrics on probability measures such as the Wassersteindistance and maximum mean discrepancy to define a distance between the full datasets.We propose a new class of distances based on transformations to the normal distribution.We show how hypothesis tests for normality can be used as an efficient technique fordistance estimation between datasets within an ABC sampler.

Non Linear Mixed Effects Models: Efficient MCMC Independent ProposalBelhal Karimi, Ecole Polytechnique

Population models are widely used in domains like pharmacometrics where we needto model phenomena observed in each set of individuals. The population approach canbe formulated in statistical terms using mixed effect models. When the conditional ex-pectation of the complete log likelihood is hard to compute, the Maximum Likelihoodestimates are obtained using a stochastic version of the EM algorithm. Yet, this methodimplies being able to sample from the posterior distribution of the parameters given theobserved data. A Markov Chain Monte Carlo procedure can be used to perform this sim-ulation. Our contribution consists in accelerating this posterior sampling in order to im-prove the overall parameter estimation algorithm convergence properties. For both con-tinuous and non continuous data models, we build a proposal for the MCMC procedurethat takes into account the multidimensional and covariance structure of the individualparameters. This proposal stems from an approximation of the true posterior distributionusing a Taylor expansion of the structural model, for continuous data models, or the loglikelihood otherwise, around the computable mode of the true posterior distribution. Wegive experimental results on real and simulated pharmacokinetics datasets showing theeffectiveness of our technique.

Simulation study of HIV temporal patterns using Bayesian methodology

18

Diana Isabel Cardoso Rocha, Center for R&D in Mathematics and Applications,University of Aveiro, Portugal

Viral load values and CD4+T cells count are markers currently evaluated in the clini-cal follow-up of HIV/AIDS patients. In this context, it is relevant to develop methods thatprovide a more complete temporal description of these markers, e.g. in between clinicalappointments. To this end, we combine a mathematical model and a Bayesian methodol-ogy to estimate trajectories from a set of observed values. Also, we construct a variationband containing the most central trajectories for one patient, by exploring the range ofvalues in the a posteriori distributions. The methods are illustrated with simulated data.

Modeling global surface temperatures in terms of climate forcing and along-memory stochastic process

Eirik Myrvoll-Nilsen, UiT - The Arctic University of Norway

Global surface temperatures are highly influenced by climate forcing variables likevolcanic eruptions, solar patterns and anthropogenic effects. These variables are impor-tant, both to model and predict future temperatures. In addition, a realistic stochasticmodel of temperatures need to account for chaotic atmospheric dynamics. This can beachieved by using fractional Gaussian noise (fGn), which is the increment process of frac-tional Brownian motion. An essential property of fGn is that it exhibits long-memoryproperties. This implies that the resulting temperature model, incorporating climate forc-ing variables with fGn, becomes computationally challenging. To speed-up calculations,we provie an accurate Gaussian Markov random field approximation of fGn, using aweighted sum of just a few first-order autoregressive processes. This gives a latent Gaus-sian model which can be analysed efficiently using R-INLA, providing estimates of thetemperature responses, the parameters of the fGn process and the contributed effects offorcing. Further, the temperature model is used to estimate the transient climate responsefor historical data and to give predictions of future global surface temperature responses,given different scenarios for the climate forcing variables.

A Conditional Autoregressive Model for estimating Slow and Fast Diffusionfrom Magnetic Resonance Images

Ettore Lanzarone, CNR-IMATI

The Intra-Voxel Incoherent Motion (IVIM) model is largely adopted to estimate slowand fast diffusion coefficients as biomarkers for different diseases. However, the referenceapproach to obtain the spatial maps of these coefficients is based on a voxel-by-voxelestimation, which neglects the spatial structure of the biological tissue and results in noisymaps. Thus, to get less noisy maps, we propose a Bayesian estimation approach thatexploits a Conditional Autoregressive (CAR) specification of the prior density. We bothconsider a pure CAR model and a mixture CAR model, and we compare the outcomeswith two reference approaches. Results show better maps under the CAR approach.

Comparison between Suitable Priors for Additive Bayesian NetworksGilles Kratzer, Department of Mathematics of Zurich University

Additive Bayesian networks (ABNs) are types of graphical models that extend theusual Bayesian generalized linear model (GLM) to multiple dependent variables throughthe factorisation of the joint probability distribution of the underlying variables. Whenfitting an ABN model, the choice of the prior of the parameters is of crucial importance.

19

If an inadequate prior is used, data separation, data sparsity and too weakly informa-tive priors lead to issues in the model selection process. In this work a simulation studybetween two weakly informative priors is presented. The first is a Gaussian prior, witha mean of zero and a large variance, currently implemented in the R-package ABN andthe second belongs to the Student’s t-distribution, specifically designed for logistic re-gressions. We compare the impact of the prior on the accuracy of the learned additiveBayesian network in function of different parameters. We then conclude by highlightingthe good performance of the Student’s t informative prior. Finally, driven by the results,we discuss future developments with so-called Diaconis-Ylvisaker conjugate priors forGLMs which will help to improve and solve the aforementioned issues.

Combining Genetically Informative Data with Questionnaire Data UsingPsychometric Models

Inga Schwabe, Tilburg University

Statistical analysis of genetically informative data like twin data or human DNA aimsat determining the relative contribution of nature and nurture in explaining individualdifferences. In behavior genetics, a subfield of psychology, inference is focused on esti-mating the heritability of a complex trait or characteristic (e.g., personality, quality of lifeor intelligence). Within this field, often, measurement instruments based on self-reportand observation such as questionnaires are used to collect data. Where in the last quarterof the 20th century, the focus was on using structural equation modeling (SEM) and re-porting Maximum Likelihood point estimates, the first decade of the 21st century has seenincreased use of Bayesian methods. This has opened up new modeling possibilities suchas inference on complex models that were not easily tractable using standard frequentisttechniques. These include also the simultaneous estimation of the genetic model and apsychometric item response theory (IRT) model to analyze the questionnaire data on itemlevel. The integration of such a measurement model is important since an analysis basedon an aggregated measure (e.g., a sum-score based analysis) can lead to an underestima-tion of heritability (van den Berg et al., 2007) or the finding of spurious gene-environmentinteractions (Schwabe et al., 2014; Molenaar et al., 2014). Although MCMC modeling ofthese models is easily carried out using off-the-shelf software packages like JAGS (Plom-mer, 2003) and BUGS (Lunn, Thomas, Best and Spiegelhalter, 2000), this new technologywith its richness of possibilities has not yet been embraced by the behavior genetics com-munity. This might be partly due to a lack of standard for reporting results and becausemost researchers in the field of behavior genetics have an applied background and aretherefore less familiar in learning a new programming language. To make Bayesian anal-ysis more accessible for this research field, I introduce the R package BayesTwin. Thepackage includes a wide range of genetic models where an IRT model was integratedinto the analysis to facilitate analysis on item level as well functions that plot relevantinformation or help determining whether the analysis was performed well.

On choosing mixture components via non-local priorsJairo Alberto Fuquene Patino, University of Warwick

Choosing the number of mixture components remains a central but elusive challenge.Traditional model selection criteria can be either overly liberal or conservative when en-forcing parsimony. They may also result in poorly separated components of limited prac-tical use. Non-local priors (NLPs) are a family of distributions that encourage parsimonyby enforcing a separation between the models under consideration. We formalize NLPs

20

in the context of mixtures and show how they lead to well-separated components thathave non-negligible weight, hence interpretable as distinct subpopulations. Importantlywe suggest default prior settings based on detecting multi-modal Normal and T mix-tures, and minimal informativeness for categorical outcomes where multi-modality isnot a natural consideration. We also give a theoretical characterization of the sparsityinduced by NLPs, derive tractable expressions and propose simple algorithms to obtainthe integrated likelihood and parameter estimates. Although the framework is genericwe fully develop multivariate Normal, Binomial and product Binomial mixtures basedon a family of exchangeable moment priors. The theory and underlying principles holdmore generally, however. Our results show a serious lack of sensitivity of the Bayesianinformation criterion (BIC) and insufficient parsimony of the AIC and a local prior coun-terpart to our formulation. The singular BIC behaved like the BIC in some examplesand the AIC in others. We also offer comparisons to overfitted and repulsive overfittedmixtures. In most examples their performance was competitive but depended on settingthe prior parameters adequately to prevent the appearance of spurious components. Thenumber of components inferred under NLPs was closer to the true number (when thiswas known) and remained robust to prior parameter changes, provided these remain inthe range of recommended defaults.

Parameter-tuning free MCMC for Sparse Hierarchical Non-stationary ModelsKarla Monterrubio Gomez, University of Warwick

We construct non-stationary hierarchical models based on stochastic parameters andGaussian Markov random fields. We choose the prior field to be Matern, and we studytwo hyperprior field models for the prior length-scale field. One choice is an Ornstein-Uhlenbeck hyperprior, that we formulate through an autoregressive AR(1) process. Theother, corresponds to the squared exponential hyperprior, which is frequently utilised.In addition, we treat measurement noise variance and length-scale hyperprior parame-ters as unknowns in the posterior distribution. In this type of hierarchical construction,hyperparameters and hyperpriors tend to be strongly coupled a posteriori, which chal-lenges Markov Chain Monte Carlo algorithms. We present a comparative evaluation ofthree equivalent MCMC sampling schemes, which are all free of parameter tuning. Theapproach here presented is applied to interpolation problems in both, simulated and realdata examples. The experiments, consider posterior consistency of the estimates with re-spect to discretisation methods and explore the efficiency of the sampling methodologiesunder different settings.

Mixtures of Peaked Power Batschelet Distributions for Circular Data WithApplication to Saccade DirectionsKees Tim Mulder, Utrecht University

Circular data is encountered throughout a variety of scientific disciplines, such as ineye movement research as the direction of saccades. Motivated by such applications,mixtures of peaked circular distributions are developed that fit markedly better on dis-tributions of saccade directions. The peaked directions employed are a novel family ofBatschelet-type distributions, where the shape of the distribution is warped by meansof a transformation function t(). Because the inverse Batschelet distribution features animplicit inverse that is not computationally feasible for large datasets such as eye move-ment data or more complex models such as mixtures, we develop an alternative calledthe Power Batschelet distributions that is easy to compute and mimics te behaviour of the

21

Inverse Batschelet distribution. Inference is performed in both a frequentist way, throughExpectation-Maximization (EM) and the bootstrap, and a Bayesian way, through MCMC.All parameters can be fixed, improving stability. Model comparison can be performedthrough information criteria or through bridge sampling in the Bayesian framework,which allows performing a wealth of hypothesis tests through the Bayes factor. An Rpackage is available to perform these analyses.

Posterior Consistency for n in the Binomial (n,p) Problem with BothParameters Unknown

Laura Fee Schneider, Georg-August University of Gottingen

We consider estimating the parameter n of the binomial distribution from k indepen-dent observations when the success probability p is unknown. There is a long history ofdiscussion about this problem and various estimators have been suggested in the last 80years. We contribute to this discussion by suggesting a family of Bayesian estimators andprove an asymptotic statement about posterior consistency valid for a broad class of priordistributions.

Latent topic analysis for document matching between job descriptions andresumes

Le Vu Hoang, Trinity College Dublin

The matching of suitable resumes or CVs with a specific job description (JD) is a dif-ficult task at large scale for HR departments or recruitment agencies due to the rich in-formation and ambiguous structure of both documents. Current solutions are predomi-nately to do this manually, which is costly, time-consuming and limited in terms of scale.In this work we look at the use of efficient topic modelling techniques to explore the struc-ture of JDs and resumes in a large database, and discuss how further research in bipartitematching methods would allow more automatic job to resume matching.

Unbiased and Consistent Nested Sampling via Sequential Monte CarloLeah South, Queensland University of Technology

We propose a novel sequential Monte Carlo algorithm (NS-SMC) which is a variantof nested sampling (NS) that does not suffer from the same theoretical or practical issuesas NS. Like NS, NS-SMC is robust to phase transitions in the underlying model. NS-SMC also produces consistent and, under certain conditions, unbiased estimates of themarginal likelihood. We compare the performance of NS-SMC and likelihood annealingSMC on several challenging and realistic problems. In performing these comparisons,we propose a novel approach for implementing and tuning these method efficiently inpractice.

This is joint work with Robert Salomone, Christopher C. Drovandi and Dirk P. Kroese.

Conjugate prior in the Mallows ranking model with Spearman distanceMarta Crispino, INRIA Grenoble, France

In this work we analyze, within the Bayesian framework, the Mallows rank model, apopular distance-based model over rankings. Previous work (Vitelli et al. 2017) enabledBayesian inference for the Mallows model with most of the right-invariant metrics usedin the literature. However, the analysis was limited to the use of the uniform prior over

22

the space of permutations, for the central parameter of the density, known as consensusranking. We present some theoretical insights on the prior elicitation problem for theconsensus ranking, and show the existence of sufficient statistics, in the special case whenthe Spearman distance is used as the metric in the model. Additionally, by exploiting thenotion of permuthoedron, we find an explicit form for the conjugate prior on the centralparameter of the distribution, and study its properties.

Bayesian Dynamic Tensor RegressionsMatteo Iacopini, Ca’ Foscari University of Venice

Multidimensional arrays (i.e. tensors) of data are becoming increasingly available andcall for suitable econometric tools. We propose a new dynamic linear regression modelfor tensor-valued response variables and covariates that encompasses some well knownmulti- variate models such as SUR, VAR, VECM, panel VAR and matrix regression modelsas special cases. For dealing with the over-parametrization and over-fitting issues due tothe curse of dimensionality, we exploit a suitable parametrization based on the parallelfactor (PARAFAC) decomposition which enables to achieve both parameter parsimonyand to in- corporate sparsity effects. Our contribution is twofold: first, we provide anextension of multivariate econometric models to account for both tensor-variate responseand covariates; second, we show the effectiveness of proposed methodology in definingan autoregressive process for time-varying real economic networks. Inference is carriedout in the Bayesian framework combined with Monte Carlo Markov Chain (MCMC). Weshow the efficiency of the MCMC procedure on simulated datasets, with different sizeof the response and independent variables, proving computational efficiency even withhigh-dimensions of the parameter space. Finally, we apply the model for studying thetemporal evolution of real economic networks.

Do house prices jump across school district lines in NYC? A Bayesiannonparametric approach to Geographic Discontinuity Designs.

Maxime Rischard, Harvard Statistics

Regression discontinuity designs (RDDs) are natural experiments characterized by thetreatment assignment being fully determined by covariates. Most research has focusedon one-dimensional cases, where units with a forcing variable lying on one side of athreshold value receive a treatment that the rest do not receive. More recently, situationswith multiple forcing variables have garnered interest. When these variables are spatialcovariatesthat is, when a treatment is applied to a region but not its neighborthe resultingnatural experiment is termed a geographic regression discontinuity design (GeoRDD).We propose a Bayesian nonparametric approach to GeoRDDs, based on Gaussian processregression. We address multiple nuances of having a functional estimand defined ona border with potentially intricate topology, particularly when defining and estimatingcausal estimands of the average treatment effect, and when testing for non-zero treatmenteffects. Finally, we illustrate our methodology on a dataset of property sales in NewYork City, and show evidence of a price discontinuity between two neighbouring schooldistricts.

Particle Monte Carlo methods for the integration of multiple datasets withapplication to genomic data

Nathan Cunningham, University of Warwick

23

Modern high-throughput technologies generate a broad array of heterogeneous data,e.g. gene expression and DNA methylation, providing distinct, but often complementary,information on the genotype. A primary challenge in analysing these data lies in discov-ering the means of combining these disparate data sources into a single, cohesive analy-sis. Uncovering group structure in genomic data is key to understanding how the geneticvariation between individuals translates into the variation we can observe or measure—the phenotype. Previous work in this area[Kirk et al 2012] proposed MDI (multiple dataintegration), a context-dependent clustering algorithm capable of integrating data froma range of different datasets and data types, allowing for the sharing of information be-tween datasets. In MDI, allocations are modelled according to a Dirichlet multinomialallocation mixture model. In our work, we present particleMDI, a clustering algorithmwithin the framework of MDI wherein cluster allocations are updated using a particleGibbs sampler, enabling faster mixing of the MCMC chain in comparison to MDI. Weexplore modifications to the traditional particle Gibbs algorithm, exploiting nuances pe-culiar to cluster analysis to improve the computational efficiency of the algorithm. Wedemonstrate the efficacy of our method in applications to clustering synthetic and realdatasets.

Modelling career trajectories of cricket players using Gaussian processesOliver Stevenson, University of Auckland

Generally speaking, a sportsperson’s career follows a typical trajectory. They begin ata young age with some raw but undeveloped ability, which improves over time as the ath-lete gains experience and participates in specialised training and coaching programmes.Eventually an athlete reaches the peak of their career, after which ability tends to declineuntil the athlete opts to retire from their chosen discipline. Additionally, fluctuations inability are common due to factors such as form, injury and other external circumstances.This is particularly relevant in cricket, where many players continually experience peaksand troughs in terms of individual performance, due to the nature of the sport. We fita Bayesian model that employs Gaussian processes to model how the abilities of NewZealand cricketers change across a playing career. Given the high dimensionality of themodel, and for ease of model comparison, nested sampling is used to fit the processes.The model allows for more precise quantification of a player’s ability at any given pointof their career, than traditional methods. The results provide coaches with more informa-tion at their disposal when determining whether or not a player is improving, which haspractical implications when comparing players and in team selection policy.

Mixture Modeling with Particle OptimizationSameer Deshpande, Wharton, University of Pennsylvania

We consider mixture modeling and clustering, in which we assume that each obser-vation was drawn from one of several distributions. Taking a Bayesian approach requiresus to study the posterior distribution over the space of all partition of n observations,a space whose dimension increases exponentially in n. Unfortunately, stochastic searchtechniques like Markov Chain Monte Carlo generally do not scale well to problems ofsuch size. In this work, we extend the particle optimization framework of Rockova (2017)from variable selection to clustering, targeting several clusterings with high posteriorprobability simultaneously. At a high level, this procedure works by running several“mutually aware” mode-hunting trajectories that repel one another whenever they ap-proach the same model. We then propose extensions of this framework designed to bettersummarize the posterior variability in clustering.

24

Message-Passing Monte CarloSamuel Power, University of Cambridge

For Markov Chain-based sampling from high-dimensional posterior distributions, ge-ometric approaches have been invaluable, with Hamiltonian Monte Carlo (HMC) in par-ticular seeing wide applicability and empirical success. More heavily geometric variantslike Riemannian Manifold HMC (RMHMC) are seen to be even more powerful on tar-gets with especially intricate geometry, but have seen limited scalability due to the com-putational costs incurred. In 2014, Zhang and Sutton introduced Semi-Separable HMC,which allows RMHMC to scale for certain hierarchical models. In this work, I extend thismethod to ‘Message-Passing Monte Carlo’, a variant which applies to the more generalmodel class of factor graphs. This algorithm is modular and scalable, while retaining thefavourable geometric properties of the original algorithm.

Assessing the accuracy of individual matched record with varying block sizesand cut-off values using a Markov Chain based Monte Carlo simulation

approachShovanur Haque, Queensland University of Technology

Record linkage is the process of matching together the records from different datasources that belong to the same entity. Record linkage is increasingly being used by sta-tistical, health, government and business organisations to link administrative, survey,population census and other files to create a robust file for more complete and compre-hensive analysis. Despite this increase, there has been little work on developing tools toassess the accuracy of linked files. Ensuring that the matched records in the combinedfile actually correspond to the same individual or entity is crucial for the validity of anyanalyses and inferences based on the combined data. Haque et al. (submitted) proposeda Markov Chain based Monte Carlo simulation approach (MaCSim) for assessing linkageaccuracy and used ABS (Australian Bureau of Statistics) synthetic data to illustrate theutility of the approach. Different blocking strategies were considered to classify matchesfrom non-matches with different levels of accuracy. In order to assess average accuracyof linking, correctly linked proportions were investigated for each record. The analysesindicated strong performance of the proposed method of assessment of accuracy of thelinkages. In this paper, this method is employed to evaluate the accuracy of linkages, us-ing varying block sizes and different cut-off values while minimizing error. The aim is tofacilitate optimal choice of block size and cut-off value to achieve high accuracy in termsof minimizing the average False Discovery Rate (FDR) and False Negative Rate (FNR).

On the Bayesian nonparametric modelling of extremesStefano Rizzelli, Bocconi University

The analysis of extremes of multiple variables is relevant in many real applications.The Extreme-Value (EV) theory provides an approximate distribution for normalized vec-tor of componentwise maxima, using asymptotic arguments. The dependence structureof such a distribution is characterized by the so-called Pickands dependence function(Pickands, 1981) and angular probability measure. Recently, Marcon et al. (2016) pro-posed a Bayesian nonparametric model for inferring both the Pickands dependence func-tion and the angular distribution of bivariate random vectors of componentwise maxima,using a Bernstein polynomial prior. The asymptotic properties of the inferential methodproposed by Marcon et al. (2016) have not been established yet. In this work we fill thisgap. We establish the almost sure consistency of such a Bayesian procedure, endowing the

25

space of Pickands dependence functions with the L1-metric and obtain almost sure con-sistency on the space of angular distributions as a by-product. Furthermore, we disccussan extension of this study to the more realistic scenario where the density associated tolimiting distribution only provides a pseudo-likelihood, seeking conditions under whichthe corresponding pseudo-posterior distribution is consistent.

Evaluating FAVAR with Time-Varying Parameters and Stochastic VolatilityTaiki Yamamura, Queen Mary University of London

This paper investigates the performance of FAVAR (Factor Augmented Vector Au-toregression) with time-varying parameters and stochastic volatility (TVP/SV-FAVAR)in capturing time variation in monetary policy transmission, in comparison to that ofsmall-scale TVP/SV-VAR. The analysis is conducted through Monte Carlo (MC)-basedexperiments using open-economy Dynamic Stochastic General Equilibrium (DSGE) asthe data-generating process. The experiments show that although TVP/SV-VAR does notadequately detect the time variation, TVP/SV-FAVAR does. This result is subsequentlyinterpreted in terms of information amounts of these two empirical models. Using theconcept of “informational deficiency” and a technique to compute it, both of which wereproposed in recent literature, I reveal that small-scale VAR easily suffers from two kindsof informational problems in identifying the time variation, whereas FAVAR can avoidthem by containing a sufficient quantity of common latent factors.

Estimating Average Treatment Effects in Evaluation Studies: Using DirichletProcess Mixtures

Zizhong Yan, Warwick, Southampton

This paper focuses on the estimation of the average treatment effect on the treated(ATT) in evaluation studies under unconfoundedness. As an alternative to traditionalmatching and reweighting methods, I propose a constrained Dirichlet process mixture ofnormals (DPMN) model to consistently estimate the covariates distribution in the treat-ment group and match the control units to the treated so that the distributions of co-variates are stochastically equivalent. I use this approach to build a matching estimatorand a reweighting estimator with desirable properties. First, the DPMN matching esti-mator meets the balancing property by construction. Second, since DPMN yields consis-tent estimates of the propensity score, the reweighting estimator is semi-parametricallyefficient. Traditional matching and propensity-score based methods are two-step ap-proaches, which may result in incorrect standard errors. In this paper, the whole algo-rithm is integrated into a single efficient Markov Chain Monte Carlo scheme: the result-ing marginal standard errors can account for errors arising from the first step estimation.I illustrate this new method with Monte Carlo experiments and an empirical applicationof the LaLonde(1986) data. The DPMN reweighting estimator is found to have a perfor-mance comparable to conventional reweighting estimators. I also find that the DPMNmatching estimator is less biased and more efficient than traditional matching estimators,as a result of improved balance.

Bayesian variable selection via Hopfield Networks and interaction selectionMatteo Vestrucci, University of Texas at Austin

26

Variable Selection can be seen as a subproblem of Model Selection: given a model,a set of p covariates, and a dependent variable, we are interested in selecting the “best”subset of covariates. Already for a small number of variables p, the number of candi-date models becomes very large (2p), and searching the solution space with Bayesiantechniques usually involves using slow MCMC methods that only partially explore theposterior distribution. In this poster we propose to use Hopfield Networks to quicklyfind the mode of the posterior distribution across the whole solution space. To analyzetheir performance we construct a Bayesian linear model and analytically calculate thejoint posterior distribution of each variable’s inclusion in the model, and apply it to dif-ferent scenarios. Preliminary empirical results suggest a polynomial rate of convergenceclose to O(p). After having identified the mode, we show a strategy that could be used toobtain posterior observations using rejection sampling, and proceed to describe how thegeneral framework can be expanded to do interaction selection.

27

6 Participant List

name Affiliation emailAdam Smith University College London [email protected] Shenvi Centre for Complexity Science, University of Warwick [email protected] Borowska University of Glasgow [email protected] Nishimura University of California - Los Angeles [email protected] Avalos Pacheco University of Warwick [email protected] Diana University of Kent [email protected] Kreuzer Technische Universitat Munchen [email protected] Alahmadi Monash University /Shaqra University [email protected] Herring Duke University [email protected] Ebert Queensland University of Technology [email protected] Fasano Bocconi University [email protected] Finkenstadt University of Warwick, Department of Statistics [email protected] Karimi Ecole Polytechnique [email protected] Schrinski University of Duisburg-Essen, Faculty of Physics [email protected] Balocchi University of Pennsylvania [email protected] Spire University of Loughborough [email protected] Shu Durham University [email protected] Au The University of Sydney [email protected] Gadd University of Warwick [email protected] Forlani Imperial College London [email protected] Tait University of Edinburgh [email protected] Durante Bocconi University [email protected] Spano’ University of Warwick [email protected] Hosszejni Department of Finance, Accounting and Statistics, WU Vienna [email protected] Ashby Imperial College London [email protected] Isabel Cardoso Rocha Center for R&D in Mathematics and Applications, University of Aveiro, Portugal [email protected] Veen Utrecht University [email protected] Myrvoll-Nilsen UiT - The Arctic University of Norway [email protected] Ryan Warwick Medical School [email protected] Lanzarone CNR-IMATI [email protected] Turner University of Sheffield [email protected] on next page

28

name Affiliation emailFrank Owen University of Szeged, Bolyai Institue [email protected] Riutort-Mayol Universitat Politecnica de Valencia [email protected] Moran University of Pennsylvania [email protected] Stagakis Loughborough University [email protected] Finocchio Leiden University [email protected] Kratzer Department of Mathematics, Zurich University [email protected] Rebaudo Bocconi University [email protected] Le Vu Trinity College Dublin [email protected] Pruenster Bocconi University [email protected] Peneva Mathematics for Real-World Systems CDT, University of Warwick [email protected] Schwabe Tilburg University [email protected] Antoniano Villalobos Bocconi University [email protected] Jewson University of Warwick [email protected] Alberto Fuquene Patino University of Warwick [email protected] Carson University of Warwick [email protected] Smith University of Warwick [email protected] Gamper Warwick Mathematics of Systems [email protected] Huang Imperial College London [email protected] Griffin University of Kent [email protected] Rousseau University of Oxford [email protected] Monterrubio Gomez University of Warwick [email protected] Mulder Utrecht University [email protected] Mengersen Queensland University of Technology [email protected] Brock University of Birmingham [email protected] Fee Schneider Georg-August University Goettingen [email protected] Guzman Rincon University of Warwick [email protected] South Queensland University of Technology [email protected] Steel Warwick [email protected] Catalano Bocconi University [email protected] Crispino Inria Grenoble [email protected] Barons University of Warwick [email protected] Moores University of Wollongong [email protected] Iacopini Ca’ Foscari University of Venice [email protected] Vestrucci University of Texas at Austin [email protected] on next page

29

name Affiliation emailMatthew William Sutton Queensland University of Technology [email protected] Rischard Harvard Statistics [email protected] Guindani University of California, Irvine [email protected] Khaleel University of Bath [email protected] Cunningham University of Warwick [email protected]’ Margaritella University of Edinburgh [email protected] Laitonjam School of Computer Science, University College Dublin [email protected] Stevenson University of Auckland [email protected] Argiento University of Torino [email protected] Badawy Aston University [email protected] Deshpande Wharton [email protected] Power University of Cambridge [email protected] Wade University of Warwick [email protected] Homer School of Psychology, Cardiff University [email protected] Vollmer University of Warwick [email protected] Haque Queensland University of Technology [email protected] Joy Li Queen Mary, University of London [email protected] Rizzelli Bocconi University [email protected] Senn Luxembourg Institute of Health [email protected] Yamamura Queen Mary University of London [email protected] Rigon Bocconi University [email protected] Wiecek Analytica Laser / LSE [email protected] Changye University Paris Dauphine [email protected] Xi University of Warwick [email protected] Whye Teh University of Oxford [email protected] P Raykov Aston University [email protected] Zhao University of Oslo [email protected] Yan Department of Economics, Warwick [email protected]

30

A45, M42, M

5, M6, BIRM

INGHAM

M6, M

69, M1 LO

NDON

A46

, M40

, WA

RWIC

K,

KEN

ILW

ORT

H, L

EAM

ING

TON

SPA