arXiv subjects

Kerrie Mengersen

Publications and source records attributed to Kerrie Mengersen.

At least 73 records · Page 4Linked to original sources

arXiv2020

Bayesian item response models for citizen science ecological data

So-called 'citizen science' data elicited from crowds has become increasingly popular in many fields including ecology. However, the quality of this information is being frequently debated by many within the scientific community. Therefore, modern citizen science implementations require measures of the users' proficiency that account for the difficulty of the tasks. We introduce a new methodological framework of item response and linear logistic test models with application to citizen science data used in ecology research. This approach accommodates spatial autocorrelation within the item difficulties and produces relevant ecological measures of species and site-related difficulties, discriminatory power and guessing behavior. These, along with estimates of the subject abilities allow better management of these programs and provide deeper insights. This paper also highlights the fit of item response models to big data via divide-and-conquer. We found that the suggested methods outperform the traditional item response models in terms of RMSE, accuracy, and WAIC based on leave-one-out cross-validation on simulated and empirical data. We present a comprehensive implementation using a case study of species identification in the Serengeti, Tanzania. The R and Stan codes are provided for full reproducibility. Multiple statistical illustrations and visualizations are given which allow practitioners the extrapolation to a wide range of citizen science ecological problems.

stat.AP↗

arXiv2020

Bayesian Computation with Intractable Likelihoods

This article surveys computational methods for posterior inference with intractable likelihoods, that is where the likelihood function is unavailable in closed form, or where evaluation of the likelihood is infeasible. We review recent developments in pseudo-marginal methods, approximate Bayesian computation (ABC), the exchange algorithm, thermodynamic integration, and composite likelihood, paying particular attention to advancements in scalability for large datasets. We also mention R and MATLAB source code for implementations of these algorithms, where they are available.

stat.CO↗

arXiv2020

Estimating a novel stochastic model for within-field disease dynamics of banana bunchy top virus via approximate Bayesian computation

The Banana Bunchy Top Virus (BBTV) is one of the most economically important vector-borne banana diseases throughout the Asia-Pacific Basin and presents a significant challenge to the agricultural sector. Current models of BBTV are largely deterministic, limited by an incomplete understanding of interactions in complex natural systems, and the appropriate identification of parameters. A stochastic network-based Susceptible-Infected model has been created which simulates the spread of BBTV across the subsections of a banana plantation, parameterising nodal recovery, neighbouring and distant infectivity across summer and winter. Findings from posterior results achieved through Markov Chain Monte Carlo approach to approximate Bayesian computation suggest seasonality in all parameters, which are influenced by correlated changes in inspection accuracy, temperatures and aphid activity. This paper demonstrates how the model may be used for monitoring and forecasting of various disease management strategies to support policy-level decision making.

stat.AP↗

arXiv2020

The role of intrinsic dimension in high-resolution player tracking data -- Insights in basketball

A new range of statistical analysis has emerged in sports after the introduction of the high-resolution player tracking technology, specifically in basketball. However, this high dimensional data is often challenging for statistical inference and decision making. In this article, we employ Hidalgo, a state-of-the-art Bayesian mixture model that allows the estimation of heterogeneous intrinsic dimensions (ID) within a dataset and propose some theoretical enhancements. ID results can be interpreted as indicators of variability and complexity of basketball plays and games. This technique allows classification and clustering of NBA basketball player's movement and shot charts data. Analyzing movement data, Hidalgo identifies key stages of offensive actions such as creating space for passing, preparation/shooting and following through. We found that the ID value spikes reaching a peak between 4 and 8 seconds in the offensive part of the court after which it declines. In shot charts, we obtained groups of shots that produce substantially higher and lower successes. Overall, game-winners tend to have a larger intrinsic dimension which is an indication of more unpredictability and unique shot placements. Similarly, we found higher ID values in plays when the score margin is small compared to large margin ones. These outcomes could be exploited by coaches to obtain better offensive/defensive results.

stat.AP↗

arXiv2019

Factors associated with injurious from falls in people with early stage Parkinson's disease

Falls are common in people with Parkinson's disease (PD) and have detrimental effects which can lower the quality of life. While studies have been conducted to learn about falling in general, factors distinguishing injurious from non-injurious falls are less clear. We develop a two-stage Bayesian logistic regression model was used to model the association of falls and injurious falls with data measured on patients. The forward stepwise selection procedure was used to determine which patient measures were associated with falls and injurious falls, and Bayesian model averaging (BMA) was used to account for uncertainty in this variable selection procedure. Data on 99 patients for a 12-month time period were considered in this analysis. Fifty five percent of the patients experienced at least one fall, with a total of 335 falls cases; 25% of which were injurious falls. Fearful, Tinetti gait, and previous falls were the risk factors for fall/non-fall, with 77% accuracy, 76% sensitivity, and 76% specificity. Fall time, body mass index, anxiety, balance, gait, and gender were the risk factors associated with injurious falls. Thus, attaining normal body mass index, improving balance and gait could be seen as preventive efforts for injurious falls. There was no significant difference in the risk of falls between males and females, yet if falls occurred, females were more likely to get injured than males.

stat.AP↗

arXiv2019

Profile regression for subgrouping patients with early stage Parkinson's disease

Falls are detrimental to people with Parkinson's Disease (PD) because of the potentially severe consequences to the patients' quality of life. While many studies have attempted to predict falls/non-falls, this study aimed to determine factors related to falls frequency in people with early PD. Ninety nine participants with early stage PD were assessed based on two types of tests. The first type of tests is disease-specific tests, comprised of the Unified Parkinson's Disease Rating Scale (UPDRS) and the Schwab and England activities of daily living scale (SEADL). A measure of postural instability and gait disorder (PIGD) and subtotal scores for subscales I, II, and III were derived from the UPDRS. The second type of tests is functional tests, including Tinetti gait and balance, Berg Balance Scale (BBS), Timed-Up and Go (TUG), Functional Reach (FR), Freezing of Gait (FOG), Mini Mental State Examination (MMSE), and Melbourne Edge Test (MET). Falls were recorded each month for 6 months. Clustering of patients via Finite Mixture Model (FMM) was conducted. Three clusters of patients were found: non-or single-fallers, low frequency fallers, and high frequency fallers. Several factors that are important to clustering PD patients were identified: UPDRS subscales II and III subtotals, PIGD and SE ADL. However these factors could not differentiate PD patients with low frequency fallers from high frequency fallers. While Tinetti,TUG, and BBS turned to be important factors in clustering PD patients, and could differentiate the three clusters. FMM is able to cluster people with PD into three groups. We obtain several factors important to explaining the clusters and also found different role of disease specific measures and functional tests to clustering PD patients. Upon examining these measures, it might be possible to develop new disease treatment to prevent, or to delay, the occurrence of falls.

stat.AP↗

arXiv2019

Assessing the predictive ability of the UPDRS for falls classification in early stage Parkinson's disease

Identification of risk factors associated with falls in people with Parkinson's Disease (PD) is important due to their high risk of falling. In this study, various ways of utilizing the Unified Parkinson's Disease Rating Scale (UPDRS) were assessed for the identification of risk factors and for the prediction of falls. Three statistical methods for classification were considered:decision trees, random forests, and logistic regression. UPDRS measurements on 51 participants with early stage PD, who completed monthly falls diaries for 12 months of follow-up were analyzed. All classification methods applied produced similar results in regards to classification accuracy and the selected important variables. The highest classification rates were obtained from model with individual items of the UPDRS with 80% accuracy (85% sensitivity and 77% specificity), higher than in any previous study. A comparison of the independent performance of the four parts of the UPDRS revealed the comparably high classification rates for Parts II and III of the UPDRS. Similar patterns with slightly different classification rates were observed for the 6- and 12-month of follow-up times. Consistent predictors for falls selected by all classification methods at two follow-up times are: thought disorder for UPDRS I, dressing and falling for UPDRS II, hand pronate/supinate for UPDRS III, and sleep disturbance and symptomatic orthostasis for UPDRS IV. While for the aggregate measures, subtotal 2 (sum of UPDRS II items) and bradykinesia showed high association with fall/non-fall. Fall/non-fall occurrences were more associated with individual items of the UPDRS than with the aggregate measures. UPDRS parts II and III produced comparably high classification rates for fall/non-fall prediction. Similar results were obtained for modelling data at 6-month and 12-month follow-up times.

stat.AP↗

arXiv2019

Monitoring through many eyes: Integrating disparate datasets to improve monitoring of the Great Barrier Reef

Numerous organisations collect data in the Great Barrier Reef (GBR), but they are rarely analysed together due to different program objectives, methods, and data quality. We developed a weighted spatiotemporal Bayesian model and used it to integrate image based hard coral data collected by professional and citizen scientists, who captured and or classified underwater images. We used the model to predict coral cover across the GBR with estimates of uncertainty; thus filling gaps in space and time where no data exist. Additional data increased the models predictive ability by 43 percent, but did not affect model inferences about pressures (e.g. bleaching and cyclone damage). Thus, effective integration of professional and high-volume citizen data could enhance the capacity and cost efficiency of monitoring programs. This general approach is equally viable for other variables collected in the marine environment or other ecosystems; opening up new opportunities to integrate data and provide pathways for community engagement and stewardship.

stat.AP↗

arXiv2019

Likelihood-free parameter estimation for dynamic queueing networks: case study of passenger flow in an international airport terminal

Dynamic queueing networks (DQN) model queueing systems where demand varies strongly with time, such as airport terminals. With rapidly rising global air passenger traffic placing increasing pressure on airport terminals, efficient allocation of resources is more important than ever. Parameter inference and quantification of uncertainty are key challenges for developing decision support tools. The DQN likelihood function is, in general, intractable and current approaches to simulation make likelihood-free parameter inference methods, such as approximate Bayesian computation (ABC), infeasible since simulating from these models is computationally expensive. By leveraging a recent advance in computationally efficient queueing simulation, we develop the first parameter inference approach for DQNs. We demonstrate our approach with data of passenger flows in a real airport terminal, and we show that our model accurately recreates the behaviour of the system and is useful for decision support. Special care must be taken in developing the distance for ABC since any useful output must vary with time. We use maximum mean discrepancy, a metric on probability measures, as the distance function for ABC. Prediction intervals of performance measures for decision support tools are easily constructed using draws from posterior samples, which we demonstrate with a scenario of a delayed flight.

stat.ME↗

arXiv2019

Computationally Efficient Simulation of Queues: The R Package queuecomputer

Large networks of queueing systems model important real-world systems such as MapReduce clusters, web-servers, hospitals, call centers and airport passenger terminals. To model such systems accurately, we must infer queueing parameters from data. Unfortunately, for many queueing networks there is no clear way to proceed with parameter inference from data. Approximate Bayesian computation could offer a straightforward way to infer parameters for such networks if we could simulate data quickly enough. We present a computationally efficient method for simulating from a very general set of queueing networks with the R package queuecomputer. Remarkable speedups of more than 2 orders of magnitude are observed relative to the popular DES packages simmer and simpy. We replicate output from these packages to validate the package. The package is modular and integrates well with the popular R package dplyr. Complex queueing networks with tandem, parallel and fork/join topologies can easily be built with these two packages together. We show how to use this package with two examples: a call center and an airport terminal.

stat.CO↗

arXiv2019

A feature-based framework for detecting technical outliers in water-quality data from in situ sensors

Outliers due to technical errors in water-quality data from in situ sensors can reduce data quality and have a direct impact on inference drawn from subsequent data analysis. However, outlier detection through manual monitoring is unfeasible given the volume and velocity of data the sensors produce. Here, we proposed an automated framework that provides early detection of outliers in water-quality data from in situ sensors caused by technical issues.The framework was used first to identify the data features that differentiate outlying instances from typical behaviours. Then statistical transformations were applied to make the outlying instances stand out in transformed data space. Unsupervised outlier scoring techniques were then applied to the transformed data space and an approach based on extreme value theory was used to calculate a threshold for each potential outlier. Using two data sets obtained from in situ sensors in rivers flowing into the Great Barrier Reef lagoon, Australia, we showed that the proposed framework successfully identified outliers involving abrupt changes in turbidity, conductivity and river level, including sudden spikes, sudden isolated drops and level shifts, while maintaining very low false detection rates. We implemented this framework in the open source R package oddwater.

stat.AP↗

arXiv2019

A framework for automated anomaly detection in high frequency water-quality data from in situ sensors

River water-quality monitoring is increasingly conducted using automated in situ sensors, enabling timelier identification of unexpected values. However, anomalies caused by technical issues confound these data, while the volume and velocity of data prevent manual detection. We present a framework for automated anomaly detection in high-frequency water-quality data from in situ sensors, using turbidity, conductivity and river level data. After identifying end-user needs and defining anomalies, we ranked their importance and selected suitable detection methods. High priority anomalies included sudden isolated spikes and level shifts, most of which were classified correctly by regression-based methods such as autoregressive integrated moving average models. However, using other water-quality variables as covariates reduced performance due to complex relationships among variables. Classification of drift and periods of anomalously low or high variability improved when we applied replaced anomalous measurements with forecasts, but this inflated false positive rates. Feature-based methods also performed well on high priority anomalies, but were also less proficient at detecting lower priority anomalies, resulting in high false negative rates. Unlike regression-based methods, all feature-based methods produced low false positive rates, but did not and require training or optimization. Rule-based methods successfully detected impossible values and missing observations. Thus, we recommend using a combination of methods to improve anomaly detection performance, whilst minimizing false detection rates. Furthermore, our framework emphasizes the importance of communication between end-users and analysts for optimal outcomes with respect to both detection performance and end-user needs. Our framework is applicable to other types of high frequency time-series data and anomaly detection applications.

stat.AP↗

arXiv2018

Testing hypotheses via a mixture estimation model

We consider a novel paradigm for Bayesian testing of hypotheses and Bayesian model comparison. Our alternative to the traditional construction of posterior probabilities that a given hypothesis is true or that the data originates from a specific model is to consider the models under comparison as components of a mixture model. We therefore replace the original testing problem with an estimation one that focus on the probability weight of a given model within a mixture model. We analyze the sensitivity on the resulting posterior distribution on the weights of various prior modeling on the weights. We stress that a major appeal in using this novel perspective is that generic improper priors are acceptable, while not putting convergence in jeopardy. Among other features, this allows for a resolution of the Lindley-Jeffreys paradox. When using a reference Beta B(a,a) prior on the mixture weights, we note that the sensitivity of the posterior estimations of the weights to the choice of a vanishes with the sample size increasing and avocate the default choice a=0.5, derived from Rousseau and Mengersen (2011). Another feature of this easily implemented alternative to the classical Bayesian solution is that the speeds of convergence of the posterior mean of the weight and of the corresponding posterior probability are quite similar.

stat.ME↗

arXiv2018

Predicting Sediment and Nutrient Concentrations in Rivers Using High Frequency Water Quality Surrogates

A particular focus of water-quality monitoring is the concentrations of sediments and nutrients in rivers, constituents that can smother biota and cause eutrophication. However, the physical and economic constraints of manual sampling prohibit data collection at the frequency required to capture adequately the variation in concentrations through time. Here, we developed models to predict total suspended solids (TSS) and oxidized nitrogen (NOx) concentrations based on high-frequency time series of turbidity, conductivity and river level data from low-cost in situ sensors in rivers flowing into the Great Barrier Reef lagoon. We fit generalized least squares linear mixed effects models with a continuous first-order autoregressive correlation to data collected traditionally by manual sampling for subsequent analysis in the laboratory, then used these models to predict TSS or NOx from in situ sensor water-quality surrogate data, at two freshwater sites and one estuarine site. These models accounted for both temporal autocorrelation and unevenly time-spaced observations in the data. Turbidity proved a useful surrogate of TSS, with high predictive ability at both freshwater and estuarine sites. NOx models had much poorer fits, even when additional covariates of conductivity and river level were included along with turbidity. Furthermore, the relative influence of covariates in the NOx models was not consistent across sites. Our findings likely reflect the complexity of dissolved nutrient dynamics in rivers, which are influenced by multiple and interacting factors including physical, chemical and biological processes, and the need for greater and better incorporation of spatial and temporal components within models.

stat.AP↗

arXiv2018

Scalable Bayesian Inference for the Inverse Temperature of a Hidden Potts Model

The inverse temperature parameter of the Potts model governs the strength of spatial cohesion and therefore has a major influence over the resulting model fit. A difficulty arises from the dependence of an intractable normalising constant on the value of this parameter and thus there is no closed-form solution for sampling from the posterior distribution directly. There are a variety of computational approaches for sampling from the posterior without evaluating the normalising constant, including the exchange algorithm and approximate Bayesian computation (ABC). A serious drawback of these algorithms is that they do not scale well for models with a large state space, such as images with a million or more pixels. We introduce a parametric surrogate model, which approximates the score function using an integral curve. Our surrogate model incorporates known properties of the likelihood, such as heteroskedasticity and critical temperature. We demonstrate this method using synthetic data as well as remotely-sensed imagery from the Landsat-8 satellite. We achieve up to a hundredfold improvement in the elapsed runtime, compared to the exchange algorithm or ABC. An open source implementation of our algorithm is available in the R package "bayesImageS."

stat.CO↗

arXiv2018

Approximating the Likelihood in Approximate Bayesian Computation

This chapter will appear in the forthcoming Handbook of Approximate Bayesian Computation (2018). The conceptual and methodological framework that underpins approximate Bayesian computation (ABC) is targetted primarily towards problems in which the likelihood is either challenging or missing. ABC uses a simulation-based non-parametric estimate of the likelihood of a summary statistic and assumes that the generation of data from the model is computationally cheap. This chapter reviews two alternative approaches for estimating the intractable likelihood, with the goal of reducing the necessary model simulations to produce an approximate posterior. The first of these is a Bayesian version of the synthetic likelihood (SL), initially developed by Wood (2010), which uses a multivariate normal approximation to the summary statistic likelihood. Using the parametric approximation as opposed to the non-parametric approximation of ABC, it is possible to reduce the number of model simulations required. The second likelihood approximation method we consider in this chapter is based on the empirical likelihood (EL), which is a non-parametric technique and involves maximising a likelihood constructed empirically under a set of moment constraints. Mengersen et al (2013) adapt the EL framework so that it can be used to form an approximate posterior for problems where ABC can be applied, that is, for models with intractable likelihoods. However, unlike ABC and the Bayesian SL (BSL), the Bayesian EL (BCel) approach can be used to completely avoid model simulations in some cases. The BSL and BCel methods are illustrated on models of varying complexity.

stat.CO↗

arXiv2017

Using history matching for prior choice

It can be important in Bayesian analyses of complex models to construct informative prior distributions which reflect knowledge external to the data at hand. Nevertheless, how much prior information an analyst can elicit from an expert will be limited due to constraints of time, cost and other factors. This paper develops effective numerical methods for exploring reasonable choices of a prior distribution from a parametric class, when prior information is specified in the form of some limited constraints on prior predictive distributions, and where these prior predictive distributions are analytically intractable. The methods developed may be thought of as a novel application of the ideas of history matching, a technique developed in the literature on assessment of computer models. We illustrate the approach in the context of logistic regression and sparse signal shrinkage prior distributions for high-dimensional linear models.

stat.AP↗

arXiv2017

A network flow approach to visualising the roles of covariates in random forests

We propose novel applications of parallel coordinates plots and Sankey diagrams to represent the hierarchies of interacting covariate effects in random forests. Each visualisation summarises the frequencies of all of the paths through all of the trees in a random forest. Visualisations of the roles of covariates in random forests include: ranked bar or dot charts depicting scalar metrics of the contributions of individual covariates to the predictive accuracy of the random forest; line graphs depicting various summaries of the effect of varying a particular covariate on the predictions from the random forest; heatmaps of metrics of the strengths of interactions between all pairs of covariates; and parallel coordinates plots for each response class depicting the distributions of the values of all covariates among the observations most representative of those predicted to belong that class. Together these visualisations facilitate substantial insights into the roles of covariates in a random forest but do not communicate the frequencies of the hierarchies of covariates effects across the random forest or the orders in which covariates occur in these hierarchies. Our visualisations address these gaps. We demonstrate our visualisations using a random forest fitted to publicly available data and provide a software implementation in the form of an R package.

stat.OT↗