STAT 320

At Amherst, I was the TA for an upper-level statistics elective called Statistics Communication, taught by Dr. Pamela Matheson. My duties included holding office hours, answering student questions, and maintaining the class’ GitHub organization.

This course is unusual but rewarding. Its aim is to build models or run statistical tests targeted at specific audiences and to communicate the results effectively. Other colleges offer similar courses under titles like “Statistical Consulting.”

The prerequisite for the class is intermediate statistics (covering logistic regression, ANOVA, model selection techniques, R programming). As a result, the class was essentially tailored to each student’s background. Some students, with just the bare minimum prerequisites, were generally limited to multiple linear regression, while others with extensive knowledge of ML/statistical theory could engage in more advanced methods like gradient boosting and MICE. Therefore, I had to work with students at their level of knowledge to help them build something that could solve the problem they were working on.

An example of an assignment from this class: a lawyer who possesses a dataset and wishes to support his argument in a court case with a statistical analysis. Half of the class was tasked with writing a report using an appropriate model or test based on the dataset to support the lawyer’s argument. The other half wrote a report that countered the lawyer’s argument using the same dataset. Then, in class, the students presented their arguments in groups in a mock trial and voted for the winner.

In this assignment, students learned a great deal about what can be said using data (correlation versus causation, causal inference, types of studies), as well as how to defend their data against an adversarial position and answer some tough questions (p-value of .04? So, 4% of the time we shouldn’t convict? How can you throw someone in jail when you aren’t 100% sure? Black-box model? How can you ethically make a choice when you don’t know why your model is saying what it does?).

In general, students’ reports and presentations had to be effective and clear. Students had to justify their choice of plots, the language they used (less technical if their designated audience was the general public, very technical if their audience were PhDs in statistics), and the format of their reports/presentations.

Dr. Matheson and I placed heavy emphasis on ensuring that students learned version control. There aren’t many courses in the Amherst statistics major that emphasize version control: that was more typical on the CS side of things. Consequently, most students in the class were using version control for the first time and had to learn about the importance of having a well-documented repository, how to handle merge errors, forking, and how to manage projects involving several collaborators.

The students’ final project synthesized everything they had learned. They chose a topic of interest, collected relevant data, and identified an audience to direct their findings toward. Students had the option to work in small teams or individually. They assembled a comprehensive technical appendix that included all their statistical theory, code, and models. Additionally, they prepared a 2-3 page report or presentation that effectively summarized their findings for their chosen audience. They then presented these findings to their classmates, who assumed the role of the target audience. The project was documented in a well-organized GitHub repository, structured so that it could realistically be passed on to someone interested in exploring their work further.