Program

Schedule

All times are in the Central Time Zone.

• 11:00 – 11:10: Welcoming Remark (Chair: David Dahl, Brigham Young University)
  • Ron Wasserstein, Executive Director, American Statistical Association
  • Mine Çetinkaya-Rundel, Professor of the Practice and the Director of Undergraduate Studies at the Department of Statistical Science, Duke University; and 2023 Chair of the ASA Section on Statistical Computing
• 11:10 – 11:50: Keynote Address (Chair: Mingzhao Hu)
  • Speaker: Simon Urbanek, - Executive Editor of R Journal; - Senior Lecturer of Data Science, University of Auckland
  • Title: R in Big Action: Handling large data sets and R as a service
• 11:50 – 13:05: Data Jamboree (Chair: Sam Tyner, DLA Piper) Each language, in alphabetical order, will be allocated 15 minutes to tackle the same problems from the NYC 311 requests data, followed by questions/answers.
  • Python: Shannon Tass, Associate Professor, Department of Statistics, Brigham Young University. Code
  • Julia: HaiYing Wang, Associate Professor, Department of Statistics, University of Connecticut. Code
  • R: Lucy D’Agostino McGowan, Assistant Professor, Department of Statistical Sciences, Wake Forest University. Code
• 13:10 - 14:25: Lightning Session (Chair: Kun Chen, University of Connecticut)
  • Bowei Xi, Purdue University
  Title:Vulnerabilities of Deep Learning and Robust Deep Ensemble

  Although AI is developing rapidly, AI's vulnerability under adversarial attacks remains an extraordinarily difficult problem. We discuss the root cause of adversarial examples through studying the deep neural network's (DNN) classification boundary. The existing attack algorithms can generate from a handful to a few hundred adversarial examples given one clean sample. We show there are a lot more adversarial examples given one clean sample, all within a small neighborhood of the clean sample. We then define DNN uncertainty regions and show the transferability of adversarial examples is not universal. The results lead to two conjectures regarding the size of the DNN uncertainty regions and where DNN function becomes discontinuous. The conjectures offer a potential explanation for why the generalization error bound -- the theoretical guarantee established for DNN -- cannot adequately capture the phenomenon of adversarial examples. We then introduce a deep ensemble with high accuracy over the adversarial examples.
  • Howard Baek, Fred Hutch Data Science Lab
  Introducing Loqui: A Shiny app for Creating Automated Videos

  Loqui is an open source web application that enables the creation of automated videos using ari, an R package for generating videos from text and images. Loqui takes as input either a Google Slides URL or a Microsoft PowerPoint file, extracts the speaker notes from the slides, and converts them into an audio file. Then, it converts the slides to images and ultimately, generates an mp4 video file where each image is presented with its corresponding audio. The functionality of Loqui relies on two R packages, namely ari and text2speech, which run in the background. Although it is certainly possible to go directly to these packages and run their functions for course generation, we realize that not everyone feels comfortable programming in R. This web application offers an intuitive and user-friendly interface allowing individuals to effortlessly create automated videos without the need for programming skills.
  • Kris Sankaran, University of Wisconsin-Madison
  Beyond Black Box Simulation

  Simulation models have come to play a central role both within and beyond statistics, providing the basis for power analysis in molecular biology, theory building in particle physics, and resource allocation in epidemiology, for example. We will highlight points of contact between statistics and simulation, with an emphasis on how careful interface design can make simulators more transparent both to the scientists who build them and the broader audiences that must draw conclusions from them. We overview approaches to iterative, interactive simulation design and visualization of simulation outputs in an information-dense way. We will illustrate these strategies using examples from experimental design, mixture modeling, and agent-based simulation. Further examples can be found in the recent review, [“Generative Models: An Interdisciplinary Perspective”](https://doi.org/10.1146/annurev-statistics-033121-110134) and code notebooks be accessed [online](https://github.com/krisrs1128/generative_review).
  • Zoe Rehnberg and Emily Robinson, California Polytechnic State University
  Enhancing Statistical Computing Education through Game Plans: A Pedagogical Approach

  In statistical computing education, students often grapple with the transition from conceptualizing a data task, such as data wrangling or visualization, to writing the necessary code. While students likely have the original data set and a vision of the desired outcome, we need to teach students how to translate a general task (e.g., add a variable, combine two data sets, summarize groups, create a visualization, etc.) into appropriate (and correctly ordered) lines of code. Further, as tasks get more complicated and datasets get larger, this translation between the data task and code becomes increasingly difficult. Drawing from computer science education literature, which advocates breaking down steps of complex problem-solving tasks and writing about code (Catrambone, 2011), we introduced “game planning” into four sections of introductory statistical computing that focus on the tidyverse in R. Game plans serve as strategic guides that prompt students to map their coding strategies before implementation. Students can create game plans in various formats, such as pen-and-paper or digital tools like the online whiteboard Excalidraw. Our presentation explores the rationale behind game plans, showcases diverse student approaches, and provides practical tools and examples, all aimed at improving students’ proficiency and structured thinking in statistical computing.
  • Jonathan Sidi, Sage Therapeutics
  mmrm: a robust and comprehensive R package for implementing mixed models for repeated measures

  Mixed models for repeated measures (MMRM) analysis has been extensively used in the pharmaceutical industry to analyze longitudinal datasets. SAS PROC MIXED has been the gold standard for this analysis in the past, and so far R packages fall short for one of the following reasons: model convergence issues, unavailability of covariance structures or adjusted degrees of freedom, or numerical results being far from PROC MIXED results. To fill in this important gap in the open-source statistical software landscape, cross-company collaboration via the “Software Engineering Working Group” (SWE WG) has been initiated and developed the new {mmrm} R package. A critical advantage of {mmrm} over existing implementations is that it is faster and converges more reliably. It also provides a comprehensive set of features: users can specify a variety of covariance matrices, weight observations, fit models with restricted or standard maximum likelihood inference, perform hypothesis testing with Satterthwaite or Kenward-Roger adjusted degrees of freedom, extract the least square means estimates using the emmeans package, summarize with the broom package and integrate with the tidymodels framework. We aim to establish {mmrm} as a new standard for fitting MMRM.
  • Shane Sacco, University of Connecticut
  Tips and tricks to improve the speed of your prediction pipelines and other analyses

  Statistical analysis on large datasets may take days or even weeks to complete, costing us valuable time and essentially halting the progress of a project. While in some cases, there is little we can do, in many cases, there are methods available to speed up computationally expensive tasks. Especially when executing a prediction pipeline, there are many of such tasks including cohort processing, marginal screening, model training, and repeating the experiment. In this presentation, I will provide tips and tricks to reduce the time required for tasks across various stages of the prediction pipeline, which also generalize to other types of analyses. I will focus on analyses in R, but will also make suggestions for Python when appropriate.
  • David Corliss, Peace-Work
  Designing Against Bias in Machine Learning and AI

  Bias in machine learning algorithms is one of the most important ethical and operational issues in statistical computing today. This presentation describes common sources of bias and how to design and develop algorithms that mitigate them. Analysis of disparate impact is used to quantify bias in existing and new applications. New open-source packages such as Fairlearn and AI Fairness 360 Toolkit quantify bias by automating the measurement of disparate impact on marginalized groups, offering great promise to advance the mitigation of bias. These design strategies are described in detail with examples and source code in R, Python, and SAS.
  • Yulia Marchenko, StataCorp LLC
  Software reproducibility in a nutshell

  Reproducibility of scientific studies has been gaining increasing attention in recent years. But what exactly is reproducibility? How is it related to replication, repetition, and automation? When can we achieve reproducibility and to what degree? How can we incorporate reproducibility in our data analysis? And what role does software play for reproducibility? In this presentation, I will briefly address these questions and more.
  • Arinjita Bhattacharyya, Merck
  AACTREVEAL – An R Package for Analysis and Aggregation of the Content of ClinicalTrials.gov

  Aggregate Content of ClinicalTrials.gov (AACT) is a publicly available database that contains a variety of trial-level information for every study registered in ClinicalTrials.gov. Content is downloaded from ClinicalTrials.gov daily and loaded into AACT. This research work introduces the open source aactreveal R package, which provides functions for consolidating analysis datasets from AACT and conducting meta-analyses using the analysis datasets. The main function, extract_aact() conducts a comprehensive search of clinical trials by looking for related search terms (e.g., “pembrolizumab” or “breast cancer”) and utilizing fuzzy string searching. Then, it fetches the desired outcome data such as treatment effect estimates, confidence intervals, and variance estimates. An R package is also developed, for public use. Although the demo of the package will focus on oncology trials, the package makes it easier for statisticians to explore AACT in other respective therapeutic areas. It is available publicly at https://github.com/Merck/bards-aactreveal.
• 14:30 - 15:30: Panel Discussion (Moderator: David Dahl)
  • Theme: Open-source software, open data, and open computing
  • Panelist:
    • Tracy Teal
      • Open Source Program Director, Posit PBC
    • Carol Willing
      • Python and Jupyter Core Developer, VP of Engineering, Noteable
    • Achim Zeileis
      • Editor-in-Chief of Journal of Statistical Software
      • Professor, Universität Innsbruck