Workshops

Machine learning models often achieve strong predictive performance, but this comes at the cost of transparency: complex models behave as black boxes, making it difficult to understand why they arrive at a given prediction or which features drive their behaviour. Interpretable machine learning (IML) addresses this gap with a toolbox of methods that make black-box models understandable, and this workshop provides a practical, application-oriented introduction to that toolbox.

The workshop covers a broad range of interpretability methods, including:

• Feature effects, such as partial dependence plots and accumulated local effects
• Feature importance measures, including permutation importance
• Local explanation methods for individual predictions
• Shapley values and their use in attributing predictions to input features

Sessions combine short theoretical introductions with hands-on practical exercises in R, allowing participants to apply each method directly and build intuition for when and how to use it.

Climate change is generating new demands on statistical modelling, particularly in regions like Africa, where historical climate records are sparse, irregular, or entirely absent at the spatial and temporal resolution needed for robust risk assessment. This workshop introduces participants to the statistical and machine learning methods being developed to address this challenge, with a focus on probabilistic forecasting, uncertainty quantification, and climate risk index construction.

Drawing on real-world applications, the session will cover:

• Statistical foundations of probabilistic weather and climate forecasting
• Methods for uncertainty quantification and decision-making with limited or noisy data
• Techniques for constructing climate risk indices and early-warning indicators
• Case studies applying these methods across African National Meteorological Services
• Opportunities and open challenges at the intersection of statistics, machine learning, and climate science

Artificial intelligence is beginning to change all knowledge domains, including statistical research, from a sequence of manually intensive activities into an integrated, iterative process spanning problem formulation, mathematical reasoning, simulation, software development, validation and scientific communication. Faster production of code and mathematical text, however, does not necessarily produce more reliable science. Without an explicit research lifecycle, AI can accelerate incorrect assumptions, generate plausible but invalid arguments, conceal implementation errors and create results that are difficult to reproduce.

This workshop introduces an AI-native research lifecycle designed specifically for statisticians. Participants will learn how to translate a research question into explicit assumptions, estimands, conjectures and falsifiable claims; use AI coding agents to construct modular simulation studies in R/Python; and build tests that distinguish mathematical, statistical, numerical and software failure modes. The workflow emphasises invariants, limiting cases, analytically tractable benchmarks, sensitivity analysis, adversarial counterexamples and independent critique.

The rapid emergence of Generative AI is transforming teaching, learning, research, and postgraduate supervision within higher education. While AI tools offer opportunities to enhance productivity and learning, they simultaneously challenge traditional approaches to assessment, academic integrity, research training, and the development of critical thinking. These developments have created growing uncertainty among academics and supervisors regarding how AI should be responsibly integrated into higher education practice without undermining intellectual engagement and scholarly ownership.

This half-day workshop, hosted by the StatSNetSA core group, provides a practical and reflective platform for academics, supervisors, and postgraduate students to engage with the evolving role of AI in academia. It introduces an adaptation of an existing guiding rubric for doctoral research, expanded to address the opportunities and challenges of GenAI — moving beyond simplistic detection-or-prohibition approaches towards AI literacy, transparency, verification, ethical engagement, and independent scholarly thinking.

While the teaching portfolio might not be compulsory at every institution, the momentum is heading that way. This workshop helps the academic statistician improve their own teaching and prepare for subsequent promotion applications, by assisting participants in completing their teaching portfolio.

The workshop is tailor-made for academic statisticians, presented by an academic statistician — and not an education specialist, with terminology and ideas beyond understanding. The basic structure of a teaching portfolio is provided as part of the session.

• Structure of a teaching portfolio
• The purpose of a teaching portfolio
• How to write up a Teaching & Learning philosophy
• The relationships with research and engaged scholarship (or academic citizenship)

Spatial data are now pervasive in many fields of science and social science, and are used extensively by businesses, government agencies and community groups. This workshop is intended for beginners who would like to get started in Bayesian modelling of spatial and spatio-temporal data using the statistical software package, R.

The workshop will be a blend of theory, practical sessions and case study discussions. Specific topics that will be covered in the workshop include the following:

• Overview of spatial analysis and why it is important
• Spatial data basics
• Overview of Bayesian spatial statistical models
• Using R for spatial analysis
• Practical session: Bayesian spatial modelling
• Overview of Bayesian spatio-temporal modelling
• Practical session: Bayesian spatio-temporal modelling
• Visualisation
• Case studies in Bayesian spatial and spatio-temporal analysis

Practical sessions will include a mix of hands-on computer exercises and group-based discussion. The computer exercises will be provided for participants who bring laptops with R installed, and analogous exercises will be available for participants without laptops.

This workshop covers three diverse case studies of multivariate data. Each speaker provides guidance on the procurement of specific public multivariate data, followed by applications to illustrate the characteristics of the data — equipping participants to source a wide variety of multivariate data to support teaching and research.

Topic 1 Sourcing and Combining Publicly Available Data in R for Evidence Based Analysis

Dr Şebnem Er University of Cape Town

Publicly available data sources offer many opportunities for applied statistical analysis. This workshop provides a practical introduction to sourcing, combining, and analysing open statistical and spatial data, in particular using traffic safety as a motivating case study. Drawing on datasets from Statistics South Africa, City of Cape Town, and other open platforms such as the Spatial Tax Database, participants will learn how to align non-georeferenced administrative data, such as traffic crash records, with spatial units (e.g. wards), and enrich these with socio-economic indicators including income, employment, and deprivation proxies.

The session emphasises the full analytical pipeline: identifying relevant data sources, preparing and merging datasets across differing spatial resolutions, and conducting exploratory analyses. Particular attention will be given to key statistical considerations, such as the modifiable areal unit problem (MAUP), and issues of exposure and measurement bias. Through hands-on examples in R, participants will gain insight into how integrated datasets can be used to uncover spatial patterns of risk, spatial inequalities in road safety, and to support evidence based decision making in urban and policy contexts such as transportation.

By the end of the workshop, participants will be equipped with both the technical tools and the statistical understanding needed to critically engage with publicly available data and to translate it into meaningful insights.

Prerequisites: R and RStudio installed, with some R experience — ideally reading data and using basic dplyr functions such as filter(), mutate() and summarise(), without which participants may struggle to follow live coding. A conceptual understanding of aggregation is also essential, for example what it means to analyse data at ward level or police precinct level.

Topic 2 Sourcing Complex Environmental Data

Dr Arno Otto University of Pretoria

Jason Pillay PhD researcher

Environmental decision-making is increasingly data-driven, with the expectation that reliance on data promotes objectivity and reduces human bias. While this appears sound in principle, inconsistencies in data coverage introduce a different form of bias, potentially skewing conclusions and influencing policy in ways that disproportionately affect developing regions.

This workshop introduces publicly available multivariate environmental data repositories, such as EDGAR, USGS, AQS, GEOROC, and ESGF to name a few. After sourcing the data, preprocessing strategies will be demonstrated that focus on transforming irregularities and extreme observations into informative features rather than noise to be removed. A range of statistical modelling techniques designed to accommodate heterogeneity, skewness, missing values, and outliers will be presented, along with methods for assessing data quality and model validity.

Many conventional model-based approaches rely on distributions chosen for algebraic and computational convenience, often implicitly assuming symmetry, thin tails, or both — yet environmental data frequently exhibit skewness, heavy tails, incompleteness, and group-dependent outliers. Ignoring these characteristics can lead to misleading inferences, including overestimation of the number of clusters present and distorted representations of environmental patterns. The workshop explores more flexible modelling approaches that better accommodate this complex structure, aiming to provide statistically sound tools that empower environmental researchers, policymakers, and stakeholders to make more informed decisions.

Topic 3 Sourcing Multivariate Climate Data

Dr Shawn Liebenberg North-West University

Access to multivariate climate data is important for modelling and assessing climate-related risks in financial applications. This session provides a practical introduction to sourcing, processing, and working with two widely used climate datasets: GSOD (Global Surface Summary of the Day) and ERA5 reanalysis data, beginning with an overview of their structure, coverage, and key differences, alongside alternative observed and reanalysis datasets for context.

The workshop then moves to the practical side, showing how to download GSOD and ERA5 data in a reproducible way, common challenges when working with large climate datasets, and basic approaches for visualising spatial climate data. It concludes with a simple modelling example showing how these datasets can be applied in statistical or machine learning contexts.

The Principal Component Analysis biplot was introduced by Gabriel in 1971 as a tool for visualisation of multi-dimensional data in two dimensions. While the underlying linear algebra was elegant, interpreting biplots purely through inner products initially limited their practical use. Gower later revolutionised the field by reframing the biplot as a multivariate scatter plot with calibrated variable axes, transforming it into a powerful visualisation tool. This workshop explores that evolution and its future in two parts.

Part I focuses on the fundamental concepts underlying biplots and the linear algebra associated with the plots — the objectives of a biplot, how matrix decompositions provide a framework for representing observations and variables in a single display, the major classes of biplots, and a range of applications across different disciplines.

This half-day workshop, hosted by SEPIMOD at the SASA Conference, focuses on the statistical formulation and analysis of compartmental models for infectious diseases. It introduces deterministic and stochastic compartmental models (e.g. SIR/SEIR) from a statistical perspective, with emphasis on model specification, identifiability, and inference. Participants will explore how stochasticity is incorporated through discrete and continuous-time processes, and how spatial structure can be introduced via metapopulation and hierarchical modelling frameworks.

A central COVID-19 case study demonstrates likelihood-based and simulation-based inference, model fitting to surveillance data, and short-term forecasting, alongside uncertainty quantification and sensitivity analysis.

• Statistical formulation of compartmental models
• Stochastic processes for infectious disease dynamics
• Spatial modelling frameworks and dependence structures
• Parameter estimation and model fitting (likelihood-based and simulation-based approaches)
• Case study: COVID-19 modelling and inference
• Prediction, forecasting, and uncertainty quantification
• Sensitivity analysis and model robustness

This workshop provides a hands-on, practical introduction to predictive model development, with a strong focus on real-world credit risk applications. Using predominantly retail credit examples, participants will gain experience in data preparation, variable selection, and the development and validation of predictive models, comparing logistic regression alongside machine learning techniques including random forests, gradient boosting, and neural networks.

All examples are demonstrated using the cloud-native SAS Viya for Learners platform, with particular emphasis on its intuitive, visual Model Studio pipelines. Participants will learn how to upload data, build, train, compare, and evaluate models step by step, turning raw data into actionable insights. The workshop also illustrates automated pipelines showcasing AutoML capabilities that automate data preparation, feature engineering, model training, tuning, and comparison within a defined time limit.