This workshop will explain how to use dynamic plots constructed from low-dimensional linear projections, called tours, to examine multivariate data spaces. The tour projections are read similarly to a biplot. There are several tour types, grand, guided, manual, local, and slice, that are useful and you will learn about. These can be helpful when conducting analyses involving non-linear dimension reduction, like t-SNE, and machine learning models, both supervised and unsupervised classification. We will include working with high-dimension, low-sample size data.

• 

  Dianne Cook (MDAG Workshop Presenter)

  Professor at Monash University

This workshop will explain how to use dynamic plots constructed from low-dimensional linear projections, called tours, to examine multivariate data spaces. The tour projections are read similarly to a biplot. There are several tour types, grand, guided, manual, local, and slice, that are useful and you will learn about. These can be helpful when conducting analyses involving non-linear dimension reduction, like t-SNE, and machine learning models, both supervised and unsupervised classification. We will include working with high-dimension, low-sample size data.

• 

  Dianne Cook (MDAG Workshop Presenter)

  Professor at Monash University

Components analysis, broadly speaking, refers to the problem of modelling the primary sources of information contained in a set of data. In the linear context,

which forms the framework for the present discussion, these sources are represented by linear combinations of the variables on which the observations in the data have been measured. By far the most well known is the Principal Components Analysis (PCA) model, where the objective is to retain as much variation from the data as possible, by minimising the squared residuals between the original observations and their projection onto the principal components subspace.

Although already decades old, and very popular in many fields, the Independent Components Analysis (ICA) model is less well known among the "traditional statistics" community. The problem of ICA is to identify the (statistically) independent sources of information in the data. Perhaps surprisingly, given the seemingly very difficult objective in the abstract, when the data generating distribution can be described through a linear "mixing" of statistically independent source variables, even fairly simple objectives lead to consistent estimation; with the only proviso being that the sources are non-Gaussian.

In this talk I will introduce the ICA problem in greater detail, and discuss some important applications where it has had tremendous impact. I will then introduce some of the more common methods for the estimation of independent components (ICs), as well as some work of my own on enhancing the computational aspects of a more direct approach based on non-parametric pseudo-likelihood maximisation. I will also touch briefly on the problem of estimating ICs in the online context, where estimation needs to be conducted in real time with the receipt of a stream of data seen only one at a time.

• 

  David Hofmeyr (Online MDAG Workshop Presenter)

  Senior Lecturer at Stellenbosch University

The workshop invites presentations by participants on research or education in Data Science. Please, contact the organisers via Berthold Lausen, past president International Federation of Classification Societies (IFCS), email blausen@essex.ac.uk , if you are interested to present in the first part of the workshop (1.30 pm to 3.30 pm) and suggest a title and brief abstract of your presentation by 30 September 2022 and if you are planning to attend the workshop.

• 

  Berthold Lausen

  Professor at University of Essex

The workshop invites presentations by participants on research or education in Data Science. Please, contact the organisers via Berthold Lausen, past president International Federation of Classification Societies (IFCS), email blausen@essex.ac.uk , if you are interested to present in the first part of the workshop (1.30 pm to 3.30 pm) and suggest a title and brief abstract of your presentation by 30 September 2022 and if you are planning to attend the workshop.

• 

  Berthold Lausen

  Professor at University of Essex

Stream: General
Chair: Zani Ludick

Authors: A. Neethling, F. Neethling
Abstract: We all preach the importance of learning theory and applying it in practice. But there is always a tension between theory and practice. In fact, it sometimes appears that the practice of statistics and theory in a scholarly context are poles apart. Thus, the divide between theory and practice remains an enduring challenge – particularly in statistics. So, the need to bridge the gap between theory and practice is instructive.

The challenge for a statistician is to find a connection point in between so that the results are statistically sound, given the need for a solid theoretical grounding on the one hand, and the practical circumstances, limitations and acceptable assumptions on the other. Often creative initiatives are required and this further leads to new research problems or the use of existing techniques in alternative ways. This implies that theory is abstracted practice, and practice is applied theory, and with this in mind in this presentation, various practical challenges experienced in sampling and survey data will be discussed.

Stream: General
Chair: Zani Ludick

Authors: Z. Ludick, W. Pretorius, T. Strauss
Abstract: When composing music, a careful balance must be struck between coherent temporal structure, musical variety and novelty. Generative machine learning models, and other statistical models, vary in their ability to balance these features. Additionally, assessments of the musical structure of automatically generated music often rely on subjective evaluations, such as time-consuming listening tests. Although the listener is the ultimate end-user, and subjective evaluations should not be replaced, automatic assessment techniques for identifying elements of structure in compositions can provide an objective alternative which may enhance the training of machine learning and other models. These measures may also provide musically interesting and useful material for the human composer to manipulate.

Self-similarity matrices (SSMs) can be used to capture and represent structural patterns in music. Identified blocks, paths and corners in SSMs correspond respectively with areas of homogeneity, repetition and novelty. Statistical linguistic measures, such as measures of repetitiveness and complexity also lend themselves to the description of musical structures.

We investigate the links between statistical linguistic measures, such as the moving-average type-token ratio (MATTR) and self-similarity matrices in the assessment of musical structure and demonstrate how they can assist in automatic music generation.

Stream: Data Science
Chair: Stefan Britz

Author: S. Pillay
Abstract: The South African manufacturing industry is rapidly growing and provides numerous opportunities to investors both locally and abroad. This present study analyses the financial ratios of large manufacturing enterprises by using the technique of principal component analysis. The research presents an analysis of the reduction of the many financial ratios to fewer unbiased ratios that will assist stakeholders in assessing the performance of the manufacturing industry. The resulting ratios can be used to analyse the financial performance of manufacturing enterprises with a minimal loss of information.

Stream: Data Science
Chair: Stefan Britz

Author: Stefan Britz
Abstract: As with many sports in the professional era, tennis provides a surfeit of data that can be used to both analyse past performances and predict future match outcomes. However, in its raw form information from past matches provides limited benefit to statistical learning models aimed at predicting the winner of a match. To fully harness the power of predictive algorithms, predictor variables (features) must be created and curated (engineered) from the available data using domain knowledge – in this case the understanding of tennis matches and tournaments – in such a way that predictive patterns are possibly generated.
The focus of this talk is on engineering features from ATP tour data from 2010 onwards in order to predict the outcomes of matches in a statistical learning framework. Predictive models popular in the modern literature are applied: elastic-net regularized logistic regression, random forests, xgBoost, feed-forward neural networks, and support vector machines. Classification accuracy is used to assess model performance, whilst variable importance measures appropriate to each model are used to determine the influence of the engineered features on prediction.

Stream: Extreme Value Theory
Chair: Andréhette Verster

Authors: J. Beirlant, S. Bhattacharya, F. Kamper
Abstract: Whether an extreme observation is an outlier or not depends strongly on the corresponding tail behaviour of the underlying distribution. We develop an automatic, data-driven method rooted in the mathematical theory of extremes to identify observations that deviate from the intermediate and central characteristics. The proposed algorithm is an extension of the method proposed in Bhattacharya et al. (2019) for the specific case of heavy tailed Pareto-type distributions to all max-domains of attraction. We propose some applications such as a tail-adjusted boxplot which yields a more accurate representation of possible outliers, and the identification of outliers in a multivariate context through an analysis of associated random variables such as local outlier factors. Examples and simulation results illustrate the finite sample behaviour of the algorithm and its applications.

Stream: Extreme Value Theory
Chair: Andréhette Verster

Authors: D. Maposa, A.M. Seimela
Abstract: A common problem that arises in extreme value theory (EVT) when dealing with several variables is to find an appropriate method to assess their joint or conditional multivariate extremal dependence behaviour. The method for choosing an appropriate threshold in peaks-over threshold approach is also another problem of endless debate. In this era of climate change and global warming, extreme temperatures accompanied by heat waves and cold waves pose serious economic and health challenges particularly in small economies or developing countries like South Africa. The present study attempts to address these problems, in particular, to deal and capture dependencies in extreme values of two variables, by applying bivariate conditional extremes modelling with a time-varying threshold to Limpopo province's monthly maximum temperature series. The present study is carried out in the Limpopo province of South Africa for the period 1994-2009. With the aim to model extremal dependence of maximum temperature at four meteorological stations, two modelling approaches are applied: bivariate conditional extremes model and time-varying threshold. The latter approach was used to capture the climate change effects in the data. The main contribution of this paper is in combining these two approaches in bivariate extremal dependence modelling of maximum temperature extremes in the Limpopo province. The findings of the study revealed both significant positive and negative extremal dependence in some pairs of meteorological stations. Among the major findings were the significant strong positive extremal dependence of Thabazimbi on high temperature values at Mara and the strong negative extremal dependence of Polokwane on high temperature values at Messina. The findings of this study play an important role in revealing information useful to meteorologists, climatologists, agriculturalists, planners in the energy sector among others.

Stream: Extreme Value Theory
Chair: Andréhette Verster

Authors: T. Ravele, C. Sigauke, L. Jhamba
Abstract: Solar power poses challenges to the management of grid energy due to its intermittency. To have an optimal integration of solar power on the electricity grid it is important to have accurate forecasts. This study discusses the comparative analysis of semi-parametric extremal mixture (SPEM), generalised additive extreme value (GAEV) or quantile regression via asymmetric Laplace distribution (QR-ALD), additive quantile regression (AQR-1), additive quantile regression with temperature variable (AQR-2) and penalised cubic regression smoothing spline (benchmark) models for probabilistic forecasting of hourly global horizontal irradiance (GHI) at extremely high quantiles (τ = 0.95, 0.97, 0.99, 0.999 and 0.9999). The data used are from the University of Venda radiometric in South Africa and are from the period 1 January 2020 to 31 December 2020. Empirical results from the study showed that the AQR-2 is the best fitting model and gives the most accurate prediction of quantiles at τ = 0.95, 0.97, 0.99 and 0.999, while at 0.9999-quantile the GAEV model has the most accurate predictions. Based on these results it is recommended that the AQR-2 and GAEV models be used for predicting extremely high quantiles of hourly GHI in South-Africa. The predictions from this study are valuable to power utility decision-makers and system operators when making high risk decisions and regulatory frameworks that require high-security levels. This is the first application to conduct a comparative analysis of the proposed models using South African solar irradiance data, to the best of our knowledge.

Stream: Spatial Statistics
Chair: Inger Fabris-Rotelli

Authors: A. Smit, P. Makena, C. Drotsky, I. Fabris-Rotelli, L. Dyson, K. Mahloromela, C. Kraamwinkel, A. Kijko
Abstract: Hail is an extreme meteorological event causing global, yearly cumulative insured losses up to US\$1 billion. During the austral summer rainfall season, convective thunderstorms are almost a daily occurrence in South Africa and can be accompanied by large hail events. Yet, observational hail data remains sparse, difficult to collect, and are affected by low spatial and temporal resolution, incompleteness and heterogeneous observational systems. This interdisciplinary research project focuses on investigating various aspects of the South African hail climatology and probabilistic hail hazard analyses. Thus far, the project yielded a hail climatology using pseudo-soundings from ERA-Interim reanalysis data and the HAILCAST (EIH) model, predicting the average hail stone diameter expected per day. The estimated hail day frequency compares well to historical climatologies, but the horizontal resolution of ERA-Interim are problematic in areas of steep topography. The project now focus on developing hazard estimates using EIH predictions and include area-characteristic maximum event size, probabilities of exceedance and return periods. The methodology accounts for aleatory and epistemic uncertainty associated with the predicted event sizes and the applied distributions. Maximum likelihood and Bayesian estimates are calculated by employing likelihood functions from convolution and mixture distributions. Another research stream investigates if the spatial dependency in hail event translates from EIH to the probabilistic hail hazard parameters. The spatial autocorrelation between the frequency of hail events from EIH and the spatial autocorrelation between the mean activity rate when uncertainty was taken into consideration are investigated using Spatial Autoregressive models (SAR). The interdisciplinary methodologies applied and preliminary results achieved from the project will be discussed.

Stream: Spatial Statistics
Chair: Inger Fabris-Rotelli

Authors: A. Azeez, R. Mutambayi, A. Odeyemi, J. Ndege
Abstract: Intimate partner violence (IPV) is a major public health issue that affects millions of women around the world. It has enormous global health burden and negative socioeconomic impact on affected individuals of different social classes, religious, and cultural groups. The purpose of this study was to identify disparities in physical or sexual IPV against women using Bayesian structured addictive regression (STAR) models to determine the factors associated with IPV in Nigeria. We used 2018 Nigerian Demographic and Health Survey (DHS) data, which is a national cross-sectional survey database. The sample size used in this study was 6387 women aged 14-59 years. We applied Bayesian Geo-Additive models in this analysis. Of the 6387 women included in the study, about 24% were aged 25-29 years, 69.5% were living in urban area, 96.6% were married and have working partners, 76.2% had partners with no alcohol history and about 39% had secondary education. The risk of women experiencing IPV was 25.2% higher (POR = 1.252, CrI = 1.218–1.290) among women whose partner drinks alcohol, 11.3% higher (POR = 1.113, CrI = 1.023–1.217) among women who had no education, and 11.5% higher (POR = 1.115, CrI = 1.062–1.172) among women with only primary education than among those women who had up to higher education. There was 8.6% risk higher among working partners than among those partners were not working. The prevalence of IPV among older women in this study demonstrates that it is a serious issue with significant variation across and within states showing IPV vulnerability. This cut across socioeconomic status and there is a need for multi-sectoral approaches to preventing and responding to IPV.

Stream: Spatial Statistics
Chair: Inger Fabris-Rotelli

Authors: F.H. Khanye, J. Keddie, R. Blignaut
Abstract: According to studies, global positioning systems are more trustworthy than other techniques for gathering location data and have opened up new research opportunities, particularly in Big Data. The purpose of this study is to investigate mobileDNA users' behaviour in terms of where and when they utilized their smartphone on a daily basis, including the applications they access. We designed an interactive dashboard to study mobileDNA user mobility and mobile application activity in order to gain insights that will help understand location data in a social science setting. We discovered that the application usage sequence of users on a daily path varies from day to day, and that most users using the mobileDNA application travel to more than one city or town in a day . Our data has a significant limitation in the collection of GPS records. There are gaps in the recorded location sequence when a user is inactive on their mobile device at the time the data is recorded, which is every 15 minutes. On this basis, it is recommended to collect the GPS location data irrespective of the user being active or not at the time of data collection. This will allow for more flexibility and accuracy in the analysis process. Further research is needed to cluster mobileDNA users based on the amount of time spent on their smartphones and then discover whether there are any commonalities in the application use sequences of users in that cluster.

Stream: Nonparametric Statistics
Chair: Jean-Claude Malela-Majika

Authors: M.A. Graham, J-C. Malela-Majika
Abstract: Statistical process control (SPC) refers to a collection of statistical techniques that are widely employed to monitor and enhance the quality of processes. SPC primarily entails adopting monitoring schemes which detect changes in a process that could impact the output quality. These monitoring schemes were initially employed primarily on manufacturing processes; their applications have since expanded to include the food industry, education, engineering, environmental research, biology, genetics, epidemiology, medical, finance, law enforcement, and sports, amongst others. The problem is not the exponential growth of these schemes but rather that recent research still uses outdated performance measures, such as the mean of the run-length distribution (ARL). In addition, many recent publications do not verify the necessary assumptions when applying well-known parametric monitoring schemes (which have underlying assumptions, such as normality, that are rarely met in practice), nor do they comply with the required Phase-I sample sizes for these schemes to operate effectively. In general, when standards are unknown (Case U), the unconditional ARL is considered during Phase-II monitoring. The impact of bias in the Phase-I sample may result in remarkably high rates of early false alarms. We explore the idea of restricting the probability of unconditional early false. This new method is referred to as the “lower percentile-based design”. Consideration is given to the design and implementation of several nonparametric techniques employing a prefixed value of some lower percentile point of the in-control run-length distribution. The optimal scheme has the lowest value for a specific higher percentile point of the out-of-control run-length distribution. We illustrate the new design and implementation methodologies with actual data, provide a summary and conclusion, and make recommendations for future research.

Stream: Biostatistics
Chair: Nontembeko Dudeni-Tlhone

Authors: Q. Ndlangamandla, H. Mwambi, S. Ramroop, N. Yende
Abstract: Anaemia is one of the major causes of morbidity and mortality in children aged five or less in Africa, affecting 25% of the world’s population. In developing countries, it accounts for more than 89% of the disease burden. Although it affects all population groups, but children under five years of age and women of reproductive age are more valnurable. This study aims to determine the factors associated with childhood anaemia in Tanzania and Angola. A survey logistic regression (SLR) and GAMM (generalized additive mixed models) were fitted to identify factors associated with childhood anaemia, both of which consider the survey weights, stratification, and clustering within primary sampling unit. The GAMM model offered more flexibility than the standard survey logistic model and provided a better fit to the data. Hence, the results presented are based only on GAMM. According to GAMM factors that are highly associated with childhood anaemia in both countries are: Child age, parents or guardians’ age, parents or guardians’ education level, stunting, standard of living and the gender of the child. In Tanzania, Male children were found to be more likely to be anemic compared to females (p-value =0.042, OR = 1.115, 95%CI (0.809;0.991)). Children with mothers’ who had secondary education had reduced chances of having anaemia compared to children who’s their mothers had no education ((p-value < 0.001, OR = 0.647, 95% CI (0493;0.8450). Children who were not suffering from stunting were less likely to be anemic in contrast to those who were suffering from severe stunting (P-value <0.001, OR=0.628, 95% CI (0.534;0.738)). The models also reveals that poor kids were more likely to suffer from anaemia compared to kids from middle class (P-value = 0.012, OR = 0.920, 95% CI (0.795, 1.064). The child and parent age were fitted as non-parametric terms in the model, both have p-values < 0.001 and have the effective degrees of freedom (EDF) of 5.11 and 1.00 respectively.

Stream: Biostatistics
Chair: Nontembeko Dudeni-Tlhone

Authors: N. Dudeni-Tlhone, J. Holloway
Abstract: As the COVID-19 pandemic began, nations worldwide focused on non-pharmaceutical tactics to slow the spread of SARS-CoV-2 virus, whilst scientists began to find ways to better understand, treat and control the spread of COVID-19 through pharmaceutical means.
When the COVID-19 vaccines became available to the market, their safety and effectiveness were questioned. Also, vaccination resistance among some segments of the global population began to emerge on social media, as a result, post-market surveillance gained attention. Meanwhile, health regulatory bodies such as the South African Health Products Regulatory Authority (SAHPRA) were progressing with post-market surveillance of COVID-19 vaccines licenced for use in South Africa.
Potential adverse drug effects are investigated at various stages of clinical trials. Trials of this kind, however, usually recruit a small number of participants, eliminating potential future drug candidates. Also, these trials may be conducted over short periods of time, which may not be sufficient to identify effects with prolonged latency. As a result, post-market surveillance is carried-out by health regulatory bodies to ensure public safety.
SAHPRA conducts medical product safety testing on a regular basis, but during the COVID-19 pandemic, they required additional capacity to provide timely evaluation of the safety aspects of these vaccines. SAHPRA then collaborated with the CSIR for analytics support and safety monitoring of the COVID-9 vaccines including Janssen COVID-19 Vaccine and Comirnaty (Pfizer-BioNTech). This involved analysing the characteristics of reports of adverse events following immunisation (AEFIs, implementing procedures for safety signal detection and responding to other additional analytic demands. This paper highlights some of the pharmacovigilance actions carried out to aid SAHPRA in responding to the safety concerns generated by the COVID-19 vaccinations and provides a snapshot of the outcomes.

Stream: Econometrics and Business Statistics
Chair: Stefan Janse van Rensburg

Author: S. Janse van Rensburg
Abstract: Asymmetric conditional volatility models abound. Few models can, however, accommodate proper leverage effects whereby the condition volatility decreases in response to positive shocks. As an alternative, local leverage represents a partial form of leverage. We propose a quasi score-driven volatility model that accommodates local leverage. The model retains the robustness of existing score-driven volatility models. We discuss preliminary issues surrounding estimation and inference. A simple empirical application demonstrates the viability of the model.

Theme: Multivariate Statistics
Chair: Tanita Botha

Authors: N. van Biljon, F. Little, H. Zar, M. Lake
Abstract: Conventional methods for modeling longitudinal growth data focus on the analysis of mean longitudinal trends or the identification of abnormal growth based on standardized z-scores. Latent Class Mixed Modelling (LCMM) takes into account the underlying heterogeneity in growth profiles and allows for the identification of groups of subjects that follow similar longitudinal trends. LCMM is based on a combination of linear mixed-effect-, structural equation- and multinomial logistic modelling. We used LCMM to identify underlying latent profiles of growth for a multivariate response of height and weight measurements taken from birth until the age of five years for a sample of 1143 children from the Drakenstein Child Health Study (DCHS). Instead of focusing on an optimal number of latent classes, we examined the added structural information provided by additional classes. When considering the presence of two (k=2), three (k=3) or up to six latent classes (k=6) specific features within the profiles were preserved while new features relating to level and shape of the profiles were also identified. While smaller values of k allowed us to identify large groups of individuals that follow broadly similar growth trajectories, by increasing k we were able to identify slightly more nuanced features that could be used to distinguish subgroups within the larger classes. With the identification of these classes, a better understanding of distinct childhood growth trajectories and their predictors may be gained, informing interventions to promote optimal childhood growth.

Theme: Multivariate Statistics
Chair: Tanita Botha

Authors: T. Botha, J. Ferreira, A. Bekker
Abstract: The Dirichlet distribution is a cornerstone consideration when working with data on the unitary simplex. Several generalizations of the Dirichlet distribution has been developed with more flexible structures which can be applied to data with forms that may illustrate departures from the usual Dirichlet, such as multimodality and models which may exhibit positive correlation. To gain a deeper insight into the impact of applying generalisations of the Dirichlet to data that exhibit structural departures, the Wasserstein distance is considered and investigated between different members of the Dirichlet family. Since this distance gives a natural measure of the distance and difference between two (multivariate) distributions this paper explores the differences between several (competitive) Dirichlet constructions to highlight and examine the effect that these structural changes may theoretically result in.

Theme: Bayesian Statistics
Chair:

Authors: M.A. Sehlabana, D. Maposa, A. Boateng
Abstract: The misuse of alcohol and drugs is a continuous life threat globally, including in South Africa. For that reason, researchers continue to investigate the risk factors associated with alcohol and drugs misuse. Most studies in literature employed the classical logistic regression model to investigate these risk factors. However, some of the issues pertaining to the classical methods are accounted for in the Bayesian framework. Likewise, the Bayesian logistic regression model can also account for problematic issues to the classical logistic regression model. Several studies used the Bayesian logistic regression model to investigate the risk factors associated with alcohol and drugs misuse. Usually, most Bayesian studies utilize default prior probability distributions such as Jeffereys’ prior and Zellner’s informative g-prior distributions. This study aims to evaluate the effectiveness of a modified Zellner’s informative g-prior distribution and subdue separation in modelling the misuse of alcohol and drugs. The model developed through the use of a modified Zellner’s g-prior distribution is compared to the models developed through the use of a hyper g-prior distribution and mixtures of g and n prior distribution. Comparisons are based on precision and average prediction error. Although the models yielded similar results, the modified version of Zellner’s informative g-prior distribution resulted in narrow credible intervals, and a small everage prediction error. Separation is also accounted for in the model. In this study, the modified version of Zellner’s informative g-prior distribution is evidently effective. All models are developed using the Bayesian adaptive sampling (BAS) R package. Further research may include evaluating some of the recommended prior distributions for the Generalised Linear Models (GLM) and comparison of Bayesian binary logistic regression developed in this study with logistic regression in Machine Learning algorithms.

Theme: Bayesian Statistics
Chair:

Author: T Strauss
Abstract: Many parametric models and statistical laws have been suggested to model word frequency distributions, the most famous being Zipf’s law (the inverse proportionality between the rank of a word and its frequency). Zipf’s law has been generalised and extended to model the relationship between the frequency spectrum and its groups, and also to include more parameters. As such, there are currently several proposed models of Zipf-like distributions that seem to describe word frequency distributions relatively well. These models, with their respective advantages and disadvantages in particular settings, have been shown to hold in different contexts: particular ranges of sample size, certain genres of text, etc.

In this study, we implement a Bayesian approach by fitting a class of general models based on Zipfian distributions that encompasses these current propositions for word frequency distributions. We further identify several linguistically relevant features that may be calculated from word frequency distributions of languages and express natural language as a multidimensional array of these measures. Ultimately, through a Bayesian analysis, we compare the theoretical model with empirical data based on 200 languages from 22 language families, to determine the space that these measures occupy within the posterior distributions. Comparing the patterns and distributions of empirical natural languages to what may be expected from these theoretical models allows us to determine whether some of these linguistically relevant measures also behave in a universal manner across languages and we investigate the linguistic implication.

Theme: General
Chair: Praise Obanya

Authors: L. Ndwandwe, J. Allison, M. Smuts, J. Visagie
Abstract: We propose new classes of tests for the Pareto type I distribution. These tests utilise the empirical characteristic function and is based on a lesser-known characterisation of the Pareto distribution. The finite sample performances of the newly proposed tests are evaluated and compared to some of the existing tests, where it is found that the new tests are competitive in terms of empirical powers.

Theme: General
Chair: Praise Obanya

Authors: A. van der Merwe, J.T. Ferreira
Abstract: Analysing time series data remains a relevant and evolving matter of interest, where oftentimes the assumption of normality is made for the error terms. In the case when data are of a discrete nature, the Poisson model may be assumed for the structure of the error terms. In order to address the equidispersion restriction of the Poisson distribution, various alternative considerations have been investigated in an integer environment. This paper, inspired by the integer autoregressive process of order 1, considers a Poisson-negative binomial noncentral Lindley model for the error terms; which emanates from a novel noncentral Lindley type construction. The systematic construction of this model is discussed and juxtaposed against alternate candidates, and thus with a threefold contribution: development of a continuous noncentral type Lindley model, a discrete counterpart, and its competitive nature within a time series framework. The development and obtained insight from this work is meaningful within the broad spectrum of noncentrality, and shows promise as valuable contenders in the distribution theory space.

Theme: General
Chair: Praise Obanya

Authors: P. Obanya, C. Olivier, T. Verster
Abstract: Permutation entropy is a time series analysis tool that captures permutation patterns/ordinal relationships between the individual values of a given time series. Being a pattern recognition tool which provides deeper insights into the underlying driving mechanism of a time series, it can also be used to determine the degree of complexity of an economic system, thus aiding in the identification of chaotic and stochastic behaviours. Permutation entropy has been successfully applied in various fields such as biomedicine and physical sciences.

Due to the complexity of financial markets such as the capital, commodity, derivatives, foreign exchange, futures markets, amongst others, it is difficult to determine the level of predictability of each sector of the market and hence its efficiency level. This problem can be addressed through the application of permutation entropy.

In this research, we present an overview of permutation entropy and its successful application in a variety of research fields. We also investigate the application of permutation entropy in determining the predictability and efficiency of the commodity markets. The results obtained and the interpretation are discussed in details.

Theme: Data Science
Chair: Annegret Muller

Authors: E. Slabber, T. Verster, P. de Jongh
Abstract: Factorisation machines originated from the field of machine learning literature and have gained popularity because of the high accuracy obtained in several prediction problems, in particular in the area of recommender systems. This article will provide a motivation for the use of factorisation machines, discuss the fundamentals of factorisation machines, and provide examples of some applications and the possible gains by using factorisation machines as part of the statistician’s model-building toolkit. Data sets and existing software packages will be used to illustrate how factorisation machines may be fitted and in what context it is worth being used.

Theme: Data Science
Chair: Annegret Muller

Author: N. Khoza
Abstract: Whenever there is large data (i.e., big data) the only feasible way to get an inference about the data is to sample the population. The subsection representing a population that contains all substantial elements of the population is called a sample. Since it is not ideal to evaluate a large population, we sample the population and make an inference about the population from the sample. However, we show that not all samples are representative of the population because of the uncertainty that exists. We evaluate the uncertainty about the sample in estimating the population using variable importance of the HPSPLIT procedure. Instead of randomly selecting a sample from a population without knowing its uncertainty when estimating the population, we select the most representative sample about the population. This allows the study to make accurate inferences about the population.

Theme: Data Science
Chair: Annegret Muller

Authors: T. Rangongo, I Fabris-Rotelli, R. Thiede
Abstract: Big data can result in more than sufficient information if used efficiently and effectively. Big data poses challenges in storage, management, processing, analysis and visualisation. Techniques for handling big data, specifically geospatial data, have advanced over the years. However, most require high computational power and time. The use of metadata is a solution. Metadata provides a descriptive, administrative, structural and statistical summary of data. This paper constructs metadata of a remote sensing image dataset for crop classification, and proposes a novel multivariate stratified sampling algorithm which selects the most informative images to minimise the number of images used for training. The proposed sampling algorithm performs effectively on a big spatial image dataset of crop types. The results are assessed by measuring the number of images sampled and as well as matching the proportionality of the population crop percentages.

Theme: Spatial Statistics
Chair: Renate Thiede

Authors: K. Mahloromela, I. Fabris-Rotelli, C. Kraamwinkel
Abstract: Window selection for spatial point pattern data is complex. Often, the point pattern window is given a priori. Otherwise, the region is chosen using some objective means reflecting that the window is representative of a larger region. Common approaches used are the smallest rectangular bounding window and convex windows. The chosen window should however cover the true domain of the process. Choosing too large a window results in estimation and inference in regions where the possibility of observations has not been confirmed. We propose a new algorithm for selecting a point pattern domain based on spatial covariate information without the restriction of convexity, allowing for a better fit to the true domain. The proposed algorithm is applied in the setting of rural villages in Tanzania. As a spatial covariate, remotely sensed elevation data is used. The algorithm is able to detect and filter out high relief areas and steep slopes; observed characteristics that make the occurrence of a household in these regions improbable. A modified kernel smoothed intensity estimate using the Euclidean shortest path distance is proposed to estimate the intensity on the resultant nonconvex window, producing more representative intensity estimation.

Theme: Computational Statistics
Chair: Arno Otto

Authors: R.J. du Randt, S.M. Millard, F. Kanfer
Abstract: This presentation considers a semi-parametric finite mixture of partial linear models with Gaussian errors. The semi-parametric structure allows for flexible modelling of the expected value of the response variable. These models are used in cases where the regression structure includes both parametric and non-parametric covariate structures. We demonstrate the properties of the profile likelihood expectation maximization algorithm (PL-EM) using a simulation study. The estimation algorithm is also demonstrated on real data.

Theme: Biostatistics
Chair: Isaac Singini

Authors: A. Nqayiya, S. Pazi, G. Sharp
Abstract: Acute kidney injury is a severe renal disease. It occurs when there is a rapid decrease in kidney’s ability to purify blood and is common in critically ill patients. It is associated with adverse complications such as permanent failure in kidney’s function, requirement of dialysis for blood purification, and death. Early prediction of AKI can alert medical caregivers about patients at risk of AKI so that they may timely implement suitable care, and potentially prevent AKI. The purpose of this study was two-fold. First, this study sought to identify risk factors associated with acute kidney injury. The second objective was to develop a machine learning model for prediction of acute kidney injury. The data set used consisted of 849 observations, 19 covariates and one response variable. The 849 patients were adult patients who were admitted to an intensive care unit during a period of one year from the 3rd of January 2017. The response variable was the acute kidney injury status, indicating whether a patient had acute kidney injury or not. Logistic regression was used to identify statistically significant risk factors associated with acute kidney injury. The statistically significant risk factors were then used to develop five machine learning models for prediction of acute kidney injury. Performance metrics such as area under receiver operating characteristic curve, accuracy, sensitivity, specificity, and precision were used to compare the models. The results showed that all the models exhibited good performances, with the random forest model being the superior model.

Theme: Biostatistics
Chair: Isaac Singini

Authors: I. Singini, D. Chen, F. Gumedze
Abstracts: Joint models for longitudinal and survival data are a class of models that jointly analyse an outcome of interest repeatedly observed over time along with the associated event times. There are two main classes of these models, namely; shared parameter and latent class joint models. These models are useful in two practical applications; firstly focusing on survival outcome whilst accounting for time varying covariates measured with error and secondly focusing on the longitudinal outcome while controlling for informative censoring.
Interest on the estimation of these joint models has grown in the past two and half decades with minimal effort directed towards developing influence diagnostics.
In this study we compared Cook's statistics for detecting influential subjects to classes identified by the latent class joint model which in effect would classify influential subjects through population heterogeneity. We used simulation studies to evaluate our approach and we only present one scenario in this paper. We then illustrated this approach using data from a multi-centre clinical trial on TB pericarditis. Simulation studies and the motivating dataset confirmed our hypothesis that latent class joint models can be used as an alternative to diagnostics to identify influential subjects in the shared parameter joint models for longitudinal and survival data. This is done by classifying heterogeneous classes using a latent variable. This is especially useful if there's little focus on the association between the bio-marker and risk for and event in which case the longitudinal and event processes are explained by some latent population heterogeneity. Modeling population heterogeneity has practical applications in personalised medicine by predicting the risk for an event adjusted for the longitudinal process.

Regency Hall Foyer