ENBIS-14 in Linz

21 – 25 September 2014; Johannes Kepler University, Linz, Austria Abstract submission: 23 January – 22 June 2014

My abstracts

 

The following abstracts have been accepted for this event:

  • Diagnostic Quality Problem Solving

    Authors: Jeroen de Mast (University of Amsterdam)
    Primary area of focus / application: Quality
    Secondary area of focus / application: Six Sigma
    Keywords: Diagnosis, DMAIC, Problem solving, Root-cause analysis
    Submitted at 31-Mar-2014 14:52 by Jeroen de Mast
    Accepted
    23-Sep-2014 10:15 Diagnostic Quality Problem Solving
    Diagnostic problem solving, which is the task of discovering causal explanations for unwanted effects, is an important element of problem solving. This presentation contributes a conceptual framework for the generic process of diagnosis in quality problem solving by identifying its activities and how they are related. It then presents six strategies that structure the diagnostic process by suggesting a certain sequence of actions and techniques. We analyze when each of these strategies is likely to be effective and how they may help in making the diagnostic process more efficient. Finally, we propose and motivate a generic sequence of stages in diagnosing quality problems.
    The framework offers a scientific basis for studying and evaluating problem solving methodologies such as Six Sigma’s DMAIC model, Kepner and Tregoe’s Problem Analysis method, and Shainin’s system. For the practitioner, the framework clarifies the rationale for many problem solving techniques offered in courses and textbooks. We also offer indications and contra-indications when techniques are promising, and demonstrate how they fit together in a coherent strategy.
  • Statistical Evaluation of Binary Tests

    Authors: Thomas Akkerhuis (IBIS UvA), Jeroen de Mast (IBIS UvA)
    Primary area of focus / application: Metrology & measurement systems analysis
    Keywords: Binary measurement, Binary MSA, MSA, Latent measurand, Gold standard unavailable, False acceptance probability, False reject probability
    Submitted at 31-Mar-2014 15:21 by Thomas Akkerhuis
    Accepted
    22-Sep-2014 16:20 Statistical Evaluation of Binary Tests
    The statistical evaluation of the reliability of binary tests and inspections is challenging. Two goals are: quantifying the reliability of a test, and quantifying systematic differences between multiple tests. In this presentation, we propose an approach that deals with three common challenges: the true condition of the inspected items is unobservable (“gold standard unavailable”), the probabilities of false acceptance and false rejection vary across items, and rejections are relatively rare.
    The first challenge is dealt with by latent variable modeling. The varying probabilities of false acceptance and false rejection are accounted for by assigning parts a continuous condition that the inspections aim to reflect. To deal with the low prevalence of rejections, we propose an intricate sampling strategy, which combines items sampled from multiple sources (total items population, stream of rejected items, and historical data about rejections).
  • Appointment Scheduling in Healthcare

    Authors: Alex Kuiper (IBIS UvA)
    Primary area of focus / application: Modelling
    Keywords: Appointment scheduling, Utility function, Stochastic modeling, Healthcare
    Submitted at 31-Mar-2014 16:52 by Alex Kuiper
    Accepted
    22-Sep-2014 15:10 Appointment Scheduling in Healthcare
    In scheduling theory the quality of the schedule is determined by how it balances conflicting interests of the service provider and her clients. The interest of the service provider is to maximize its capacity by fully using her available time. On the other hand, clients want to be served as quick as possible after their arrivals, i.e., no waiting time. Based on statistical research one found that the patients’ service times are stochastic variables. In scheduling theory this problem is often solved under the assumption of exponential service times due to its tractability. However, this assumption is often oversimplifying the problem.

    We study the problem in a wider context to incorporate flexibility in the actual service time distribution. Our approach is particularly valuable for practitioners who experience variation in the service times, e.g., healthcare situations, such as a doctor or dentist seeing patients. We give explicit statistics for the optimal schedule and show how it outperforms other commonly-used scheduling approaches.
  • Sparse Bayesian Modelling with Spike and Slab Priors

    Authors: Helga Wagner (Johannes Kepler University)
    Primary area of focus / application: Modelling
    Keywords: Categorial covariates, Effect fusion, MCMC, Sparsity
    Submitted at 2-Apr-2014 11:03 by Helga Wagner
    Accepted
    22-Sep-2014 10:45 Sparse Bayesian Modelling with Spike and Slab Priors
    Sparse modelling and variable selection is one of the most important issues in regression type models, as in applications often a large number of covariates on damparably few subjects are available. Estimation of regression effects in such large p, small n problems is ill-conditioned: estimated regression effects typically have large standard errors, estimation results are instable and fitted models have no good predictive performance. To identify thos regressors which have a non-negligible effect, many methods have been developped.
    In a Bayesian approach, variable selection relies on appropriate prior distributions on the effects subject to selection. Such priors are usually specified as mixtures of a spike and a slab component: the spike is centered at zero with a very small variance and the slab is flat. The finite mixture structure allows classification of effects as (practically) zero or non-zero.
    Spike and slab priors have been employed to achieve sparsity in more complex models, e.g. random effects or state space models. This talk reviews Bayesian modelling with spike and slab priors in various model classes and discusses extensions to achieve a sparse representation of the effect of categorial covariates. As the effect of a categorial covariate with k+1 categories is modelled by a set of k regression coefficients - one for each covariate level except the baseline category - sparsity is achieved whenever the effect of the categorial predictor can be represented by fewer regression effects. We show, how fusion of levels with essentially the same effect is feasible by choosing appropriate prior distributions. For all models, Bayesian inference relies on MCMC methods. Performance of these methods will be illustrated on simulated as well as real data.
  • I-Optimal Mixture Designs

    Authors: Peter Goos (Antwerp University, Leuven University), Bradley Jones (Antwerp Univeristy, SAS Institut), Utami Syafitri (Antwerp University, Bogor Agricultural Univeristy Indonesia)
    Primary area of focus / application: Design and analysis of experiments
    Secondary area of focus / application: Design and analysis of experiments
    Keywords: Mixture experiment, Continuous mixture-experiment, D-optimality, I-optimality
    Submitted at 2-Apr-2014 14:36 by Utami Syafitri
    Accepted (view paper)
    22-Sep-2014 10:45 I-Optimal Mixture Designs
    A mixture experiment is an experiment in which the experimental factors are ingredients of a mixture, and the response depends only on the relative proportions of the ingredients. This kind of experiment is frequently used in the chemical and food industry for finding the optimal formulations of products. A mixture experiment is intended to estimate a regression model that allows prediction of the response(s) for all possible combinations of proportions of the mixture. This enables the researcher to optimize the proportions for each ingredient. Ideally, when performing a mixture experiment, a good experimental design is used, so that the resulting data set is rich in information. Various criteria exist to measure the quality of an experimental design. One prediction-oriented criterion to select an experimental design is the I-optimality criterion which seeks designs that minimize the average variance of prediction. According to the early literature on the design of mixture experiments, this criterion is the most suitable one. However, most of the literature on mixture experiments deals with the D-optimality criterion which focuses on a precise estimation of the regression model's parameters. In this presentation, we first review the limited literature on the I-optimal design of mixture experiments for Scheff\'e types of regression models, the most commonly used models for data from mixture experiments, and identify several contradictions. Next, we present I-optimal continuous designs for the second-order Scheffe model and for the special cubic Scheffe model and contrast them with published results. Finally, we discuss the construction of I-optimal discrete designs and compare the resulting experimental designs with their continuous counterparts. We also contrast D- and I-optimal designs.
  • Bayesian Effect Fusion for Categorial Predictors

    Authors: Daniela Pauger (Johannes Kepler University)
    Primary area of focus / application: Modelling
    Keywords: Categorial predictors, Spike and slab prior distribution, Sparse modelling, Dummy coding
    Submitted at 10-Apr-2014 14:26 by Daniela Pauger
    Accepted
    22-Sep-2014 11:25 Bayesian Effect Fusion for Categorial Predictors
    In many applications the collected variables are categorial, measured on an ordinal or nominal scale. To include a categorial variable as a covariate in a regression type model, the usual strategy is to define one category as the baseline category and to introduce dummy variables for all other categories: for a categorial covariate taking k + 1 different values, thus k dummy variables have to be defined and the effect of one categorial covariate is not described by a single, but by a group of k regression coefficients. Including categorial variables with many levels as covariates in regression type models therefore can easily lead to a high-dimensional vector of regression effects. Therefore sparse modelling of the effect of a categorial covariate is an important issue. Sparsity can be achieved excluding variables and level effects without influence on the response or fusing categories with (almost) the same effects on the response to one group.

    In a Bayesian approach sparsity can be achieved by choosing appropriate prior distributions, e.g. spike and slab prior distributions. These priors are mixtures of two components: the spike is centred at zero with very small variance and the slab is comparably flat. The finite mixture structure allows classification of effects as (practically) zero and non-zero. These spike and slab prior distributions are very popular for Bayesian variable selection.

    Our aim is to extend the methods to allow variable selection and effect fusion for ordered as well as unordered categorial predictors. We demonstrate the developed methods using extensive simulation studies and real data.