2017 Seminars

The R package rstpm2 includes implementations for two classes of smooth survival models. First, we have implemented generalized survival models, where (S(t|x))=eta(t,x) for a link function g, survival S at time t with covariates x and a linear predictor eta(t,x). We allow for penalized smoothers from the 'mgcv' package. These models include left truncation, right censoring, interval censoring, time-varying effects, gamma frailties and normal random effects. The models allow for the estimation of a variety of parameters, including time-dependent hazard ratios, survival differences, standardised survival and attributable fractions. Second, we have recently implemented smooth accelerated failure time models, such that g(S(t|x))= eta0(log(t)-eta(t,x)) for a baseline linear predictor eta0. This model includes time-dependent acceleration factors and a variety of estimators.

Functional complex networks are a powerful way of representing connections between the element composing a (complex) system, and to map the propagation of information between them; as such, they have successful been used to improve our understanding of, for instance, the human brain. While usually not taken into account, links are characterised by some degree of uncertainty, which can affect the final structure we observe. In order to quantify such effect, this study introduces a Bayesian reconstruction framework and validates it with real electroencephalography brain data. We show that disregarding such uncertainty introduces a bias that results in an overestimation of all topological structures, especially when only short time series are available.

Authors:

Massimiliano Zanin, Seddik Belkoura, Javier Gomez, Cesar Alfaro, Javier Cano

In two dimensions, many self-interacting processes are described by the Schramm-Loewner Evolution SLE(kappa), a family of random fractal path joining two boundary points of an underlying domain D. These continuous paths arise as the scaling limits of various discrete self-interacting paths, such as loop-erased random walk.

A self-interacting process has the locality property if it does not "feel" the boundary of its domain D until it hits the boundary. Among the two-dimensional processes known as Schramm-Loewner Evolution SLE(kappa), it is known that only one, SLE(6), satisfies the locality property. In this talk, I will describe the key properties that identify SLE(6) - the Domain Markov Property, conformal invariance, and the (classical) Locality Property - and introduce a "non-local" form of locality also satisfied by SLE(6), describing the behaviour of the process when it first encloses a target set.

The need for predictions of high-dimensional time series arises with data from many applications, for instance, air pollution. To construct an overall predictor we use the matching pursuit algorithm (MPA), which selects a subset from a dictionary of possible predictors. Although there are theoretical results on the performance of MPA, there is no widely accepted stopping rule for the algorithm. We consider stopping rules using a new information theoretic (IT) criterion based on the degrees of freedom given by the trace of the hat matrix found at each MPA iteration. We compare the performance of IT criteria for different time series models using a simulation study.

We will also apply our IT criteria to model choice for air pollution data provided by National Institute of Water and Atmospheric Research, (NIWA).

Making predictions of the impacts of stressors on ecological systems generally requires smart study designs and a range of statistical analyses. In particular analyses need to be able to take into account information available in spatial (and temporal) variability and the likelihood of non-linear responses. I illustrate this with a study on how the ability of a system to deal with nutrients may change with increasing nutrient concentrations. The study design nesting a manipulative experiment within a large scale spatial survey. The analyses included multiple regression, spatial pattern recognition and kriging to extrapolate results across an extensive area. The results demonstrate patchiness across a landscape in performance and the potential for changes in the location of hotspots with increasing nutrients.

Spatial capture-recapture (SCR) methods emerged just over a decade ago, and quickly filled a niche in ecological statistics. SCR's versatility underlies its success: recordings of whale song across stretches of the Pacific Ocean, video images of leopards roaming the African plains, and genetic traces of passing rodents are all staple ingredients that SCR converts into estimates of animal density and distribution. In this talk, I outline some of my contributions to SCR methodology, with a particular focus on acoustic surveys.

The Nested Sampling algorithm, invented in the mid-2000s by John Skilling, represented a major advance in Bayesian computation. Whereas Markov Chain Monte Carlo (MCMC) methods are usually effective for sampling posterior distributions, Nested Sampling also calculates the marginal likelihood integral used for model comparison, which is a computationally demanding task. However, there are other kinds of integrals that we might want to compute. Specifically, the entropy, relative entropy, and mutual information, which quantify uncertainty and relevance, are all integrals whose form is inconvenient in most practical applications. I will present my technique, based on Nested Sampling, for estimating these quantities for probability distributions that are only accessible via MCMC sampling. This includes posterior distributions, marginal distributions, and distributions of derived quantities. I will present an example from experimental design, where one wants to optimise the relevance of the data for inference of a parameter.

The academic research on queues deals mostly with waiting. Yet, the externalities , namely the added waiting time an arrival inflicts on others, are of no less, if not of more, importance. The talk will deal mostly with how the analysis of the externalities leads to the socially optimal behavior, while solving queueing dilemmas such as whether or not to join a queue, when to arrive to a queue, or from which server to seek service at. Customers, being selfish, do not mind the externalities they impose on others. We show how in queues too, internalizing the externalities leads to self regulation. In this setting selecting the service regime is one of the tools for regulation.(Joint with Binyamin Oz)

This talk will be in two parts, the first will discuss recent initiatives to improve statistics education across Australia. The second will discuss collaborative research on health-care standards and improving health-care systems aided by Bayesian hierarchical modelling.

Peter Howley (https://www.newcastle.edu.au/profile/peter-howley) is Chair of the Statistical Society of Australia’s Statistical Education Section and Associate Professor in Statistics at Newcastle. He will describe recently developed national statistical initiatives and resources aimed to increase access to and support within higher education. One of the initiatives recently was awarded the ISI’s 2017 Best Cooperative Project Award. The resources comprise short animated videos, interactive exercises and extension documents developing statistical threshold concepts, and tools to assist primary and secondary school teachers and students engage with statistics via a national schools poster competition, including industry expert and ‘how to deliver’ videos. These aim to enable students and teachers to feel the interdisciplinary and pervasive nature and value of statistics, and make the field of statistics accessible. The exponentially increasing annual numbers of students participating and positive feedback received is very promising. The teaming up of Sustainability, Statistics and STEM for a road trip to remote and rural NSW schools will be discussed, as will recent developments in teaching at The University of Newcastle.

Most of the medical work is done with the Australian Council on Healthcare Standards and Taipei Medical University.

https://www.newcastle.edu.au/profile/peter-howley

In several application problems, the phenomena under study are multidimensional. Therefore, these phenomena are represented by multivariate variables. In multivariate inferential problems, such as tests of hypotheses for comparing two or more populations, where data are assumed to be determinations of random variables, standard parametric methods (e.g. likelihood ratio test, Hotelling T2 test, ...), when applicable, require stringent assumptions that make them non robust and often inappropriate.

The main limits of these methods are:

1) the assumed underlying distribution is not always plausible or cannot be tested (especially for small sample sizes);

2) the dependence structure (apart from the infrequent case of independent variables) must be formally defined and estimated. For example, in the case of normal multivariate distributions, it is necessary to estimate the covariance matrix or correlation matrix.

The proposed combined nonparametric test, is based on the breakdown of the problem into as many sub-problems as many variables, and on the application of a univariate permutation test for each subproblem. The combination of the permutation significance level functions of each test provides a unique test statistic (and a unique p-value) to solve the multivariate problem.

The test is therefore distribution-free and the dependence of partial tests doesn

Coupling two random objects means constructing them out of a shared source of randomness, in order to find out something interesting about one or both of them. This talk will describe (with not much prior knowledge assumed) some of the interesting things we can discover in this way: from classical extreme value theory, to some new results about random graphs, shortest distances and exploration processes.

Many practical problems are related to prediction, where the main interest is at subject (e.g., personalized medicine) or (small) sub-population (e.g., small community) level. In such cases, it is possible to make substantial gains in prediction accuracy by identifying a class that a new subject belongs to. This way, the new subject is potentially associated with a random effect corresponding to the same class in the training data, so that method of mixed model prediction can be used to make the best prediction. We propose a new method, called classified mixed model prediction (CMMP), to achieve this goal. We develop CMMP for both prediction of mixed effects and prediction of future observations, and consider different scenarios where there may or may not be a “match” of the new subject among the training-data subjects. Theoretical and empirical studies are carried out to study the properties of CMMP and its comparison with existing methods. In particular, we show that, even if the actual match does not exist between the class of the new observations and those of the training data, CMMP still helps in improving prediction accuracy. Some examples will be presented including making predictions from breast cancer genomic data samples. Additionally, some delineation of the extension to the unknown grouping structure problem will be provided. This is joint work with Jiming Jiang of UC-Davis, Jie Fan of the University of Miami and Thuan Nguyen of Oregon Health and Science University.

We consider a single server M/G/1 queue in which customers accumulate priority linearly while waiting. There are a number of priority classes, each of which accumulates priority at a different rate. Upon arrival, each customer pays to enter the queue without knowing the current state of the system. The rate at which they accumulate priority depends on the priority class they have entered. When the server becomes idle, the customer with the greatest accumulated priority is chosen for service. Accumulating priority queues have been proposed in healthcare settings since they permit priority to increase with time spent in the queue. In this talk, we focus on the existence of Nash equilibrium and stability of this model.

Many calls have been made for an increased emphasis on prediction in statistics teaching (see for example Harville 2014). Bayesian prediction has been increasingly popularised through the posterior predictive distribution. This talk raises questions about the interpretation of this distribution, and the need for posterior sampling procedures to provide the uncertainty assessment of predictions.

Harville, D.A. (2014). The need for more emphasis on prediction: a "non-denominational" model-based approach (with discussion). The American Statistician 68 (2), 71-92.

http://www.ms.unimelb.edu.au/~maitkin@unimelb/

It is a truth universally acknowledged that logistic regression gives consistent and fully efficient estimates of the regression parameter under case-control sampling, so we can often ignore the distinction between retrospective and prospective sampling in epidemiology. I will talk about two issues that are more complicated than this. First, the behaviour of pseudo-r^2 statistics under case-control sampling: most of these are not consistently estimated. Second, the question of when and why unweighted logistic regression is much more efficient than survey-weighted logistic regression: the traditional answers of 'always' and 'because of variation in weights' are wrong.

6pm NZT - Foyer area, Ground floor, of Building 302, 23 Symonds Street. Refreshments will be available here before each lecture at 6pm.

6.30pm NZT - Lectures commence at 6.30pm (to 7.30pm), Wednesdays, MLT1 Lecture Theatre, Ground Floor, Building 303, 38 Princes Street.

Or

Live Stream to a screen from 630pm NZT onwards

https://www.stat.auckland.ac.nz/en/about/news-and-events-5/events/events-2017/03/ihaka-ross-ihaka.html

Or join a local group screening in your city.(Auckland, Brisbane, Christchurch, Sydney, Wellington...)

The estimation of covariance or precision (inverse covariance) matrices plays a prominent role in multivariate analysis. The usual estimator, the sample covariance matrix, is known to be unstable and ill-conditioned in high-dimensional setting. In the past two decades, various methods have been developed to give a stable and well-conditioned estimator and they have their own advantages and disadvantages. For example, thresholding methods carry almost no computational burden but their estimators can not guarantee to be positive-definite. In this talk, we will review some of the most popular methods and describe a new method to estimate the correlation matrix using the empirical Bayes method. To our best knowledge, we have not yet found any method in the literature using the empirical Bayes method to estimate correlation matrices to date. We use the fact that the elements in the sample correlation matrix can be approximated by the same one-parameter normal distribution with unknown means, along with the non-parametric maximum likelihood estimation proposed by Wang (2007) to give a new estimator of the correlation matrix. Preliminary simulation results show that the new estimator has some advantages over various thresholding methods in estimating sparse covariance matrices.

There is growing interest among survey methodologists in collecting and using auxiliary variables related to the data quality, often called 'paradata'. I shall talk about binary paradata related to measurement error in a survey variable, indicating whether the variable is measured with error. This may be used in the design of the survey instrument. Here, the talk focusses on its use in survey data analysis to correct estimation for the potential biasing effects of measurement error.

http://www.lse.ac.uk/RESEARCHANDEXPERTISE/EXPERTS/profile.aspx?KeyValue=c.j.skinner%40lse.ac.uk

6pm NZT - Foyer area, Ground floor, of Building 302, 23 Symonds Street. Refreshments will be available here before each lecture at 6pm.

6.30pm NZT - Lectures commence at 6.30pm (to 7.30pm), Wednesdays, MLT1 Lecture Theatre, Ground Floor, Building 303, 38 Princes Street.

Or

Live Stream to a screen from 630pm NZT onwards

https://www.stat.auckland.ac.nz/en/about/news-and-events-5/events/events-2017/03/ihaka-genevera-allen.html

Or join a local group screening in your city.(Auckland, Brisbane, Christchurch, Sydney, Wellington...)

All welcome - Drinks and nibbles to follow.

We have developed an app for policy makers which allows them to test policy scenarios around improving child wellbeing. Designed in the R web application, SHINY, the app allows policy makers and analysts to run realistic simulations in which the effects of changes in children

6pm NZT - Refreshments - Foyer area, Ground floor, of Building 302, 23 Symonds Street. Refreshments will be available here before each lecture at 6pm.

6.30pm NZT - Lectures commence at 6.30pm (to 7.30pm), Wednesdays, MLT1 Lecture Theatre, Ground Floor, Building 303, 38 Princes Street.

Or

Live Stream to a screen from 630pm NZT onwards

https://www.stat.auckland.ac.nz/en/about/news-and-events-5/events/events-2017/03/ihaka-harkanwal-singh.html

Or join a local group screening in your city.(Auckland, Brisbane, Christchurch, Sydney, Wellington...)

Some researchers advocate a statistical modelling approach to inference that draws on students' intuitions about factors influencing phenomena and that requires students to build models. Such a modelling approach to inference became possible with the creation of TinkerPlots Sampler technology. However, little is known about what modelling competencies students need to acquire. Drawing and building on previous research including mathematical modelling research, this study aims to uncover the statistical modelling competencies students need to develop. A design-based research methodology was used. Model Eliciting Activities were developed with a focus on natural variation in an authentic context. Six 11-year-old students working in pairs participated. Camtasia was used to capture students' verbalizations and interactions with TinkerPlots. Pivotal moments in their reasoning were transcribed and analysed alongside written and screen artefacts. The focus of this presentation is on one pair of students as they engaged with a schoolbag weight task. Findings indicate these students seem to be developing the ability to build models, investigate and posit factors, take variation into account and make decisions based on simulated data. From the analysis an initial statistical modelling framework and statistical modelling competency framework are proposed. Implications of the findings and future research plans are discussed.

Honor/Master Students' talk.

Lectures commence at 6.30pm (to 7.30pm), Wednesdays, MLT1 Lecture Theatre, Ground Floor, Building 303, 38 Princes Street.

Refreshments will be available before each lecture at 6pm in the foyer area of Building 302, 23 Symonds Street.

More Details:

https://www.stat.auckland.ac.nz/en/about/news-and-events-5/events/events-2017/03/ihaka-hadley-wickham.html

Local groups

Wellington

The Wellington Young Statisticians group has arranged screenings of the Ihaka Seminar Series. Screenings will be at PHARMAC, 40 Mercer Street. Contact person - Amanda Hughes | https://www.facebook.com/events/231424817265602/

Christchurch, Canterbury, New Zealand

The NZSA Local group has arranged screenings of the ihaka Lecture Series. Screenings will be at Room DA02 at the Dovedale Campus of the University of Canterbury. At 6 PM, you can park virtually everywhere. If you want directions contact Richard Penny (richard.penny@stats.govt.nz) | Map: https://goo.gl/maps/EHUApStsW2s

Live Streaming

https://www.stat.auckland.ac.nz/en/about/news-and-events-5/events/events-2017/03/ihaka-hadley-wickham.html

ihaka Lecture Series:

https://www.stat.auckland.ac.nz/en/about/our-department/ihaka-lectures.html

Sparse representations have seen a huge development in the latest 20 years, due to their ability to capture parsimoniously the features of a signal. The overcomplete basis used for sparse representations, called also dictionary, can be fixed or adapted to the class of signals at hand. The presentation focuses on the latter case and the associated problem of training the dictionary. There are several classes of methods, based on different ideas, all of them simplifications of the underlying optimization problem, which is NP-hard. The huge number of variables is also a basic difficulty. Besides the standard problem of dictionary learning, some modifications suited for classification will be presented, because classification is one of the applications with significant results. Other variations of the problem include regularization and adaptations to incoherent and structured dictionaries.

Further details:

https://www.stat.auckland.ac.nz/seminar/