• Small (2400 students) undergraduate only liberal arts college
• Typical class size ~20 students
• Stats within the math dept: 1.5 statisticians
• No major/minor in stats
• Courses:
  • Intro stats
  • Year-long 300-level prob & stats sequence
  • Applied/methods classes in other depts

Various Data Science initiatives at Middlebury:

• Description of New MATH216 Intro to Data Science course
• Proposed Minor in Data Science
• ASA DataFest

Prereqs

Only intro stats and some exposure to R or any other programming language

Syllabus

• 5 biweekly analyses
  • Submitted in R Markdown: reproducible research
  • Feedback delivered via GitHub
• Term project: both written report and 12 min presentation
• In-class participation

14 students: 6 Seniors, 6 Juniors, & 2 Sophomores

Of which:

• Double majors:
  • Environmental Sciences-Econ
  • Econ-Linguistics
  • CS-Econ
• Singe majors:
  • Economics x 4
  • Molecular Bio & Biochem x 3
  • International Politics and Econ, Neuroscience, Bio, CS

• Mixture of lab & lecture: students bring their own laptops to class.
• Use real, messy, complex data.
• Discussions in class
• This class uses R, but is not a class on R. I try to teach things in a language agnostic fashion: transferable ideas & concepts.
• "Minimizing prerequisites to research.", quote by George Cobb.

• Key: Forget Base R
• How? The Hadleyverse.
• In particular
  • dplyr package for data wrangling/manipulation
  • ggplot2 package for data visualization

Data manipulation via the following verbs on tidy data. The command name is the action we want to perform!

1. filter: keep observations matching criteria
2. summarise: reduce many values to one
3. mutate: create new variables from existing ones
4. arrange: reorder rows
5. select: pick columns by name
6. join: join two data sets
7. group_by: group subsets of observations together

A statistical graphic is a mapping of variables in a

• data set to
• aes()thetic attributies of
• geom_etric objects.

ggplot2 allows us to construct graphics in a modular fashion by specifying these components.

6 dimensions of information on a 2 dimensional page:

Domestic flights leaving Houston airport (IAH) in 2011. Four data sets:

• flights: info on all 227,496 flights
• weather: hourly weather info
• planes: information on all 2853 airplanes
• airports: information on all 3376 destination airports

• Sample of 10% of San Francisco OkCupid users in June 2012 (\(n=59946\))
• 40.2% of the sample was female
  
• Use logistic regression to predict gender
• Overfitting, out-of-sample prediction, cross-validation

Best predictors have distinct differences (in gender) in large segments of the population.

Need to normalize to compare proportions, not counts!

All 222,540 songs played on the Reed College pool hall room jukebox from 2003-2009.

quandl.com has a great R interface

Two R packages for interactivity:

• Shiny: A web application framework for creating interactive web applications with no HTML, CSS, or JavaScript knowledge required.
• Leaflet: Embed maps into OpenStreetMaps

Example on Middlebury Shiny Server Pro: VT Census Tracts.

• Targeted at students in a field where facility with data is a valued skill, but not MATH and CS majors.
• Courses
  • 3 computer science
  • 2 stats
  • 2 domain courses in bio, econ, psych, chem, physics, political science, etc.
• Full details are here

DataFest is an internationally coordinated undergraduate data science hackathon run by the ASA.

Biggest ones:

• UCLA
• Duke
• 5 Colleges in Western MA

Prof. Philip Yates at Saint Michael's College and I organized the first inaugural DataFest Vermont 802 the weekend of April 8th-10th at Saint Michael's College.

• Data visualization is a gateway drug to statistics.
• Prez from Season 4 of "The Wire":
• 
• Students got really excited by ggplot, maps, and Shiny apps.

It's like learning a language. Frustration!

• Point-and-click vs command line.
  
• Thinking algorithmically
• Debugging: help files and Google
• Hadley's wisdom

• Learning in class should reflect how we learn in real life
• Experimenting with an open-learning format: students can collaborate completely and see each other's work. Still haven't decided on its merits.
• However, new tools like Datacamp are increasing the ratio: \[\frac{\mbox{Payoff from learning R}}{\mbox{Startup costs}}\]

• Developing skills and intuition takes time. At Middlebury classes are small: attention and good feedback can be given.
• Feedback via GitHub: they won't learn Git, rather just use it. GUI for feedback delivery.
• Like giving feedback on a paper: more art than science.

• Slides: bit.ly/DataScienceBTV
• GitHub Organization for this class