Please join us in person or online for our November Queensland Branch Meeting. The seminar will start at 5:00 pm. Details for the seminar are provided below.
TITLE: ASC Scholarship Showcase
TIME: 5:00 - 6:30 pm (AEST), 16th November 2023
VENUE: 511 Teaching Suite, 88 Creek Street, Brisbane and online (Zoom details will be sent with registration).
Special instructions for in-person venue: Enter through the main door at 308 Queen Street and pass through the Atrium. Speak to the concierge at the elevator located at the back of the room, and notify them that you are attending the Statistical Society of Australia event.
Please note that the seminar will be recorded and might be put on YouTube or similar platform.
SPEAKERS' ABSTRACT AND BIO:
Lessons from post-publication statistical reviews of linear regression in health research
This study explores current statistical practice and identifies common statistical misconceptions and errors researchers make when using linear regression. Statistical practices were assessed in the health and biomedical field by randomly sampling 100 published papers from PLOS ONE in 2019. Forty statisticians were recruited to review the papers, with papers randomly allocated to statisticians ensuring that two independent statisticians rated each paper.
The results show that the average researcher tends to over-rely on p-values and significance rather than the contextual importance and robustness of conclusions drawn, with an estimated 69% of authors of papers not discussing the scientific importance of parameter estimates and only 23% directly interpreting the size of regression coefficients. Only 37% of authors reported checking any of the linear regression assumptions; the most frequently assessed assumption was normality, with most authors incorrectly checking the outcome ("Y”) variable rather than model residuals.
Recommendations for improving this interpretation gap include teaching statistics holistically, where most statistics can be seen in a regression framework rather than a series of unconnected and rote-learned tests. Practical recommendations from this study include greater reporting transparency, with journals providing researchers with template papers reporting common statistical methods. To help them assess statistical methods, reviewers should receive basic statistical training and potentially access automated tools which guide statistical feedback.
Lee Jones
An accredited statistician (AStat) and past president of the Statistical Society of Australia Queensland branch, Lee is a Senior Biostatistician at QIMR Berghofer. She contributes to a broad range of projects, including cancer research, clinical trials, and quality improvement, and has gained over $14 million in research funding and 50 publications throughout her career. Lee is completing a PhD in Meta-Research focusing on statistical quality and reproducibility in health research.
Teach a student how to debug, and they can code for a lifetime: Open-source interactive self-paced R learning website
Learning and teaching statistics can be significantly enhanced by the use of statistical software, such as the R programming language. The current focus of teaching statistical software is to teach what is relevant to the subject and assume students know the basics of writing the code for statistical software. Students are thus often thrusted to learn programming while juggling the core statistical concepts. The challenge is that students come from all walks of life – some have no programming experience while others may have learnt different programming languages. In light of this challenge, we developed an online, self-paced interactive learning website called Learn R as an attempt to uniform students’ knowledge. The Learn R website covers all the basics including R and RStudio installation, basic R syntax, data import, linear regression, data visualisation using ggplot2 and writing reproducible reports with R Markdown and Quarto.
One chapter of the Learn R website focuses on how students can debug and ask for help using a minimal reproducible example. We promote the idea of autonomous learning: “teach a student how to debug, and they can code for a lifetime” as one of the teaching goals of the Learn R website. By teaching students how to seek help, they would be better able to self-identify the issue themselves. In this presentation, I will share the design of the Learn R website and the key concepts of developing a self-guided interactive learning tool for statistical programming. This presentation aims to emphasise teaching students the best practices for debugging and seeking help, and provides potential for statistics educators to focus on core statistical materials by directing students to the self-guided Learn R website for self-study.
Danyang Dai
Danyang Dai (Daidai) is a PhD student at the Centre for Health Services Research (CHSR) under the supervision of Professor Jason Pole, Associate Professor Sally Shrapnel, and Doctor Pedro Franca Gois. Her primary research focus is at the intersection of COVID-19 and Acute Kidney Injury, a critical area given the significant impact of this pandemic on global health.
Danyang has a background in biostatistics and econometrics, and her current research interests are in epidemiology and health service research. Additionally, she currently serves as the President of R-Ladies Melbourne, where she actively promotes gender diversity within the R community.
Performance of multivariate time series models over univariate approaches in modelling correlated count data
Time series modelling is used for analysis and forecasting in every field. Most traditional time series models are not able to well model series of discrete data. There is also a lack of tools to deal with collections of count-valued time series. Analysis and forecasting of count responses from correlated entities becomes a tedious task because of interdependent structure among entities.
There is a need for modelling tools that can handle interdependent count time series with options for estimating contemporaneously correlated errors, moving average processes and lag interdependences. We showcase how multivariate count time series can be analysed in a Bayesian framework. Our models can quantify nonlinear covariate relationships while capturing unobserved (and possibly multivariate) temporal dynamics in a joint probabilistic framework with hierarchical impacts using Hamiltonian Monte Carlo (HMC) simulation with Stan language in R environment.
It is essential to understand how key user decisions impact model estimation, inferences on parameters, interpretation process and forecasting performance. We explore consequences of model misspecifications (models, order, structure), distributional assumptions (including common transformation routines that assume Gaussian observations), modelling assumptions on parameters (time invariant mean and variance-covariance structures) and prior specifications.
Using both simulations and empirical data, we show that a wide variety of latent multivariate trend models can be fit with HMC. Compared to univariate modelling approaches, Bayesian based multivariate time series models tend to perform better in forecasting tasks and can uncover important insights about temporal dependence. However, we find that overall performances of modelling depend on goals of analysis and decisions on models’ specifications. Among these specifications, trend dimension (univariate-multivariate), number of series to be modelled jointly, compositional structures of models, distributional assumptions and prior choices are vital to consider.
Kalu Arachchilage Karunarathna
I am a Ph.D. candidate at School of Veterinary Science, Faculty of Science, The University of Queensland. This research is being carried out under supervision of Dr. Nicholas J. Clark in his DECRA project “Modelling Ecological Responses to Climate Change” with PhD thesis title “Statistical Multivariate modelling for ecological forecasting”.
I hold a B.Sc. (Hons) in Statistics from University of Sri Jayewardenepura, Sri Lanka, a M.Sc. in Applied Statistics and M.Phil. in Statistics. Both M.Sc. and M.Phil. degrees are from University of Peradeniya, Sri Lanka. I have been working at Department of Mathematics, Faculty of Science, Eastern University, Sri Lanka since 2011. Prior to this, I rendered my service as an instructor and teaching staff at the University of Sri Jayewardenepura and University of Peradeniya, respectively for a couple of years.
My research fields of interest are statistical modelling, inferences, and designs of experiments. Most of my research are in the fields of health, environment, industrial, education, and social science.