Below are all of my presentations, including slides and video when available. You might also be interested in my peer-reviewed publications.
Do LLMs write like experts? Quantitative techniques to measure writing style and the effect of LLMs on student writing. Seminar, Department of Statistics, University of Washington. November 4, 2025.
Do LLMs write like experts? Quantitative techniques to measure writing style and the effect of LLMs on student writing. Seminar, Department of Statistics, Brigham Young University. October 16, 2025.
Do LLMs write like experts? The distinctive style of large language models and their role in student learning. JSM 2025, Nashville, TN, August 5, 2025.
Statistical writing is a key part of the undergraduate statistics curriculum, as it both helps students learn to reason about complex ideas and teaches them an important professional skill. As large language models (LLMs) offer increasingly sophisticated tools for automating writing, we must ask what effect they have on the development of students as statisticians and statistical writers. We describe the results of several experiments comparing the grammatical and rhetorical style of students, experts, and LLMs, showing significant differences in how LLMs manage information, express confidence, and structure sentences. The results suggest that LLMs do not effectively adapt their style to the genre and audience of their writing, and we explore the implications this has for teaching. If students offload the work of writing to a tool that does not write like a data scientist, how can we teach them to communicate like an expert data scientist?
Writing to learn or learning to write: Writing in the statistics curriculum and the role of LLMs. ICERM Workshop on Applied Math in Statistics and Data Science Education, Brown University, May 20, 2025.
Statistical writing is a key part of the undergraduate statistics curriculum, both because it helps students reason about complex ideas and because it teaches them an important professional skill. Through writing, students learn to construct arguments about data, present evidence, and communicate results. Large language models (LLMs) thus raise important questions about the role of writing in statistics education: can they serve as useful tools for students still mastering writing, or do they impede development of important skills? We describe recent research comparing the style and structure of student-, expert-, and LLM-written statistical reports, revealing unexpected biases in the writing style of LLMs, and discuss the implications for students still learning to write like experts.
The COVID-19 Trends and Impact Survey: Lessons for pandemic response and for the role of software in statistics. Seminar, Department of Statistics, University of California, Santa Cruz. May 22, 2023.
Delphi’s COVIDcast Project: Lessons Learned Building Statistical Software in Real Time. JSM 2022, Washington, DC, August 8, 2022.
Beginning in April 2020, the Delphi group has produced COVIDcast: a massive, continuously updated database of COVID-19 indicators across the United States. The data features numerous unique signals, such as behaviors reported by Facebook users in the largest public health survey ever conducted, or symptom indicators extracted from aggregated medical claims data. This data is available via a public API and is widely used by forecasters and researchers. In this talk, I’ll discuss the statistical computing challenges this posed: the challenges of delivering timely, accurate, and detailed data on COVID-19 at a large scale, and of performing reliable statistical estimation on a daily basis. Developing fast, accurate, and reliable statistical software requires more than just the right algorithms; the statistician must learn to think like a software engineer, and to focus on delivering products, not simply analyses. I’ll use examples of our own mistakes to illustrate how designing statistical products requires different skills from designing statistical analyses.
Think-Aloud Interviews: A Tool for Exploring Student Statistical Reasoning (with Ciaran Evans and Amanda Luby). CAUSE/Journal of Statistics and Data Science Education webinar series. June 28, 2022.
Think-aloud interviews, in which students narrate their reasoning in real time while solving problems, are a valuable but underused tool for statistics education research. In this webinar, we suggest possible use cases for think-alouds and summarize best practices for designing think-aloud interview studies. We hope that our overview of think-alouds encourages more statistics educators and researchers to begin using this method.
CTIS: The COVID-19 Trends and Impact Survey. Seminar, Department of Mathematics & Statistics, Swarthmore College. December 7, 2021.
Insights on Vaccine Hesitancy and Healthcare Inequity from a Survey of Millions of Individuals in the US and Worldwide (with Frauke Kreuter). COPSS-NISS COVID-19 Data Science webinar series. June 10, 2021.
During the COVID-19 pandemic, epidemiologists and data scientists have relied heavily on publicly reported data about cases, hospitalizations, and deaths. But this data does not give the full picture of the pandemic, particularly when public health reporting systems are not always timely and complete. We present a global survey, developed in consultation with the WHO, CDC, and many other international stakeholders, that used Facebook as a sampling frame to randomly select daily samples to reach a total of over 60 million respondents in a year. The survey collects data about symptoms, social distancing, mental health, vaccines, vaccine hesitancy, vaccine equity and more. The survey’s broad reach and daily sampling permits tracking of trends as the pandemic progresses, and the results suggest that online surveys may have an important role to play in public health.
Delphi’s COVID Surveys: Tracking a Pandemic in Real Time. Seminar, Department of Statistics, Oregon State University. April 26, 2021.
The Delphi Group at Carnegie Mellon University has for years developed tools to forecast epidemics like influenza. In early 2020, it faced a new challenge: How can you forecast a pandemic with only patchy disease surveillance data? Through a unique partnership with Facebook and the University of Maryland, it set out to collect its own.
Delphi’s COVID surveys have since received nearly 20 million responses in the United States – or 50,000 per day for over a year – combined with nearly 50 million responses received by the University of Maryland’s international effort. This is the largest research survey ever conducted outside of a national census. Aggregate data is publicly available, updated daily, and has already been used to develop forecasts and guide national responses.
In this talk, I’ll introduce the survey and describe its expansion from tracking COVID symptoms to collecting information on testing, mental health, social distancing, mask use, and most recently, attitudes to COVID vaccination. I’ll talk about the statistical challenges in running such a large survey, open statistical problems in analyzing its results, and what the data engineering process has taught us about the role of statistics in emergency response.
Computing in the Statistics and Data Science Curriculum. Panel discussion with Mine Çetinkaya-Rundel. CAUSE/Journal of Statistics and Data Science Education webinar series. January 26, 2021.
Delphi COVIDcast: Publishing early indicators of COVID-19, from massive surveys and medical data. Seminar, Department of Biostatistics, University of Pittsburgh. August 20, 2020.
Since the beginning of the COVID-19 pandemic, a crucial problem has been obtaining timely, high-quality data about its spread, ideally at the county level. In this presentation, I’ll give an overview of Delphi’s COVIDcast project, which now maintains over 500 million observations of data – mostly daily, county-level signals. Some of this data is from public sources, such as reports of confirmed cases and deaths, but much of it is unique to Delphi. Using medical data examples, I’ll illustrate how this data gives unprecedented insight into the pandemic, and show how the COVIDcast API makes it possible for any researcher to benefit from the data.
After discussing the project generally, I’ll focus on one data source in particular: an online survey conducted in partnership with Facebook and the University of Maryland. The survey reaches 70,000 respondents daily in the United States and asks about symptoms, COVID testing, social contact, mental health, and many other items – and with 10 million responses in the United States and many more outside it, represents the largest research survey ever conducted outside a national census. Aggregate data from the survey is available publicly and is used in Delphi’s COVID case forecasts; individual data is available to researchers. I’ll discuss insights gained from the data, potential areas for future research, and how researchers can get access to pursue their own research goals using the surveys.
Support for COVID-19 Research through the Symptom Surveys (with Curtiss Cobb and Frauke Kreuter).
UIDP Webinar, June 23, 2020.
The COVID-19 symptom surveys are designed to help researchers better monitor and forecast the spread of COVID-19. In partnership with the University of Maryland and Carnegie Mellon University, Facebook users are invited to take surveys conducted by these two partner universities to self-report COVID-19-related symptoms. The surveys may be used to generate new insights on how to respond to the crisis, including heat maps of self-reported symptoms. This information may help health systems plan where resources are needed and potentially when, where, and how to reopen parts of society.
Using Cognitive Task Analysis to Uncover Misconceptions in Statistical Inference Courses (with Mikaela Meyer and Josue Orellana).
eCOTS 2020, May 19, 2020. Video poster presentation.
When taking an introductory undergraduate statistical inference course, students struggle to approach problems in the ways experts do. However, articulating how experts solve problems is difficult, and instructors might not be able to detect the misconceptions students harbor when solving these problems. To enable research into student learning, we propose combining cognitive task analysis, a research method from cognitive science, and think-aloud interviews with graduate and undergraduate students to better understand the steps students and experts take to solve statistical inference questions. After using cognitive task analysis to break down the discrete cognitive skills needed to solve simple mathematical statistics problems, we used think-aloud interviews to determine how students and experts applied these skills and to identify skills students lack. In this presentation, we will discuss our analysis of coded think-aloud interview transcripts and our cognitive task analysis of problems where students must identify the relevant variable to operate on, then apply mathematical rules to obtain their answer.
Exploring how students reason about correlation and causation (with Ciaran Evans, Philipp Burckhardt, Rebecca Nugent, and Gordon Weinberg).
eCOTS 2020, May 19, 2020. Video poster presentation.
Introductory statistics students find it notoriously difficult to reason about correlation and causation: when does correlation imply causation, and when does it not? Through think-aloud interviews with students and assessment results, we have seen that while they often recognize that correlation does not necessarily imply causation, they usually fail to understand why. To aid student understanding, we introduce a new activity to our introductory statistics labs that explores correlation vs. causation through simple causal diagrams. In this poster, we present preliminary results from this lab activity and some insights from our think-aloud research, to gain insight into how students think about this fundamental statistical concept.
Using think-aloud interviews to assess student understanding of statistics concepts (with P Burckhardt, P W Elliott, C Evans, K Lin, A Luby, M Meyer, J Orellana, R Yurko, G Weinberg, J Wieczorek, and R Nugent).
USCOTS 2019, May 16-18, 2019. Breakout session.
Assessing student understanding of statistics concepts is quite difficult: conceptual questions are difficult to write clearly, students often interpret questions in unexpected ways, and students may choose answers (even the correct answer) for unexpected reasons. This makes it difficult to assess student learning of concepts, but as we continuously improve our introductory statistics courses, we need tools to understand what students understand.
In this breakout session, we will report on a year-long exploratory project to build an assessment using a powerful tool: think-aloud interviews. Audience members will learn to use think-aloud interviews to elicit student misconceptions and revise assessment questions, providing a practical method they can use in their own courses and research to better assess student learning. We will then share surprising misconceptions discovered during our own round of 36 student interviews, and summarize our assessment’s results from several hundred students in several introductory courses.
Identifying misconceptions of introductory data science using a think-aloud protocol (with S Hyun, P Burckhardt, P Elliott, C Evans, K Lin, A Luby, C P Makris, J Orellana, J Wieczorek, R Yurko, G Weinberg, and R Nugent).
eCOTS 2018, May 23, 2018. Video poster presentation.
Think-aloud interviews can provide assessment designers with insights into student thinking that may not be clear from test responses alone. We present results from preliminary rounds of think-aloud interviews with introductory students and describe surprising misconceptions we have identified, along with insights from our experience designing assessments and performing think-aloud interviews.
A Spatio-Temporal Statistical Model of Crime Hotspots (with Daniel S. Nagin).
American Society of Criminology Annual Meeting, Philadelphia, PA, November 17, 2017.
The concentration of crime into small “hotspots” has been widely observed across many different cities and types of crimes. Current tools to understand the causes and dynamics of crime hotspots are limited. A variety of mapping tools, for example, have been proposed to detect hotspots in crime data, but these tools cannot correlate clusters to events or covariates which may have caused them. Separately, methods such as Risk Terrain Modeling attempt to identify spatial features that predict crime rates, such as gang territories, bars, or poverty, but consider only chronic hotspots, not accounting for temporary flare-ups.
We propose a statistical model which accounts for spatial and temporal variation in crime by modeling both spatial features and near-repeat and retaliatory crimes, allowing it to model the birth and death of crime hotspots and the reasons they appear, and to statistically test hypotheses about each predictive variable. We demonstrate the model on a large dataset of crimes in Pittsburgh, Pennsylvania, showing its utility in understanding the dynamics of crime.
A Spatio-Temporal Model of Crime Dynamics.
Workshop on Social Interactions and Crime, University of Chicago, October 20, 2017.
Point process modeling with spatiotemporal covariates for predicting crime (with Joel Greenhouse and Xizhen Cai).
JSM 2016, Chicago, IL, August 3, 2016.
Extensive research has shown that crime tends to be concentrated in hotspots: small pockets with above-average rates of crime. Criminologists and law enforcement agencies want to better predict crime hotspots and understand the factors that cause them, in order to target interventions. Prior research suggests that past crime hotspots, spatial features (like bus stops or bars), and leading indicators (like 911 calls) are all predictive of future crime, but no proposed predictive policing model can account for all of these factors. We have adapted a previous self-exciting point process model to incorporate past crime data, leading indicators, spatial features and spatial covariates (like population density or zoning data), and developed new tools to evaluate the performance of the model and select variables. We show the basic model and demonstrate its application to seven years of Pittsburgh crime data, comparing its fits to previous hotspot models. These results can be used to better guide crime prevention programs and police patrols.
Statistics Done Wrong: Pitfalls in Experimentation.
LASER Workshop, Washington, DC, October 16, 2014. Video.
Most research relies on statistical hypothesis testing to report its conclusions, but the seeming precision of statistical significance actually hides many possible biases and errors. The prevalence of these errors suggests that most published results are exaggerated or false. I will explain some of these errors, such as inadequate sample sizes, multiple comparisons, and the keep-looking bias, and their impact on published results. Finally, I will suggest solutions to these problems, including statistical improvements and changes to scientific funding and publication practices.