Adam Gee
Assessing the Role of Momentum in Online Chess Using Bayesian Hierarchical Modelling
The "winner effect" is a phenomenon in animal behaviour that states that an animal that has just won a fight is more likely to win their next encounter. There has been great debate whether similar effects also occur in humans. Using Bayesian hierarchical modelling, we aim to uncover whether winner or loser effects can be observed in online chess. Compared to other sports and games, studying online chess provides a unique framework where there is a huge amount of data available, and players play many back-to-back games in a very controlled environment. Our results show little evidence for momentum effects at a global level, but moderate variability at the player level reveals some players may possess such effects. These results seem to be consistent across amateur and professional chess players. We also discuss several model checking procedures to assess the suitability of our model to the data.
Nikolas Krstic
Statistical Consultation: Working with ASDa
The Applied Statistics and Data Science (ASDa) group is a statistical consultation unit within the UBC Department of Statistics that has existed since the department's founding back in 1983. Our group regularly supports UBC students, staff and faculty in their research, and also offers its consultation services to off-campus clients. This talk will cover what it's like to be employed as a statistical consultant, discussing both my positive experiences and challenging scenarios. Topics will include consultant-client communication, methodology selection, personal skill development, time management, and teaching non-statisticians about statistics.
Yulia Kozhevnikova
BLOST: Bayesian Longitudinally Ordinal Sequential Trial Design for Evaluating Respiratory Disease Treatments
Effective treatment evaluation for respiratory diseases, where symptom severity and time-to-recovery are key outcomes, demands a flexible and efficient trial design. We introduce the Bayesian Longitudinally Ordinal Sequential Trial (BLOST) framework, designed to optimize clinical trial processes in evaluating respiratory illness treatments. This Bayesian model incorporates longitudinally observed ordinal outcomes, accounts for patient heterogeneity, and facilitates information borrowing across time points. The sequential framework compares an experimental treatment with a standard one through extensive simulations. We consider three analytical approaches: a standard Bayesian method based on Hamiltonian Monte Carlo, an enhanced version applying Bayesian Model Selection, and a conventional frequentist approach. Frequentist Type I and Type II error rates are maintained through parameter calibration. Our simulations demonstrate the BLOST framework's efficiency and flexibility, highlighting its potential for evaluating respiratory disease treatments and enhancing preparedness for future health challenges.
Son Luu
Is Gibbs sampling faster than Hamiltonian Monte Carlo for GLMs?
The Hamiltonian Monte Carlo (HMC) algorithm is often lauded for its ability to effectively sample from high-dimensional distributions. In this talk I challenge the presumed domination of HMC for the Bayesian analysis of GLMs. By utilizing the structure of the compute graph rather than the graphical model, we can reduce the time per sweep of a full-scan Gibbs sampler from $O(d^2)$ to $O(d)$, where d is the number of GLM parameters. This simple changes to the implementation of the Gibbs sampler allow us to perform Bayesian inference on high-dimensional GLMs that are practically infeasible with traditional Gibbs sampler implementations. I will empirically demonstrate a substantial increase in effective sample size per time when comparing our Gibbs algorithms to state-of-the-art HMC algorithms. While Gibbs is superior in terms of dimension scaling, neither Gibbs nor HMC dominate the other: HMC retains an edge in certain circumstances thanks to its advantageous condition number scaling.
Renny Doig
Auxiliary-try Metropolis: Incorporating auxiliary variables into multiple-try Metropolis
Multiple-try Metropolis (MTM) is an extension of the Metropolis-Hastings algorithm which aims to improve exploration of the state space by drawing multiple candidate proposals at each iteration. This offers potential improvements to both computational and sampling efficiency. We propose a further extension of MTM that draws candidates both from a standard random walk proposal as well as from a proposal distribution conditional on a variable that is auxiliary to the chain being generated by MTM. There is flexibility in the choice of auxiliary variable, allowing the algorithm to be designed to accommodate difficulties presented by a particular target distribution or shortcomings of a simpler algorithm. We make a broad distinction between uninformed and informed auxiliary variables, representing two potential philosophies in the choice of auxiliary. Specific examples of such auxiliary variables are given and the relative performance is demonstrated on selected examples, based on both simulated and real data.
Johnny Xi
Propensity Score Matching for Unpaired Multimodal Data
Multimodal representation learning techniques typically rely on paired samples to learn common representations, but paired samples are challenging to collect in fields such as biology where measurement devices often destroy the samples. We present an approach to address the challenge of aligning unpaired samples across disparate modalities. We draw an analogy between potential outcomes in causal inference and potential multimodal views, which allows us to use Rubin's framework to estimate a common space in which to match samples. Our approach assumes we collect samples that are experimentally perturbed by treatments, and uses this to estimate a propensity score from each modality, which encapsulates all shared information between a latent state and treatment and can be used to define a distance between samples. We experiment with two alignment techniques that leverage this distance -- shared nearest neighbours (SNN) and optimal transport (OT) matching -- and find that OT matching results in significant improvements over state-of-the-art alignment approaches in both a synthetic and real world settings.
Dr. Himchan Jeong
Dynamic random effects for collective risk models
In this talk, I would like to share my experience as a graduate student and some remarks that can be potentially helpful for those who pursue their career in academia. Further, I introduce one of my research pipelines, the collective risk models (CRMs) under the longitudinal context, which can potentially be used to reward or penalize a driver based on their actual and expected claims history. It begins with the simplest case where we assume no unobserved heterogeneity in the driving behaviors (equivalently no random effects), and extends further to the following two cases where there is unknown heterogeneity in the driving behaviors i) that does not change over time (static random effects), and ii) that could evolve over time (dynamic random effects).