Reproducibility breakout session at USCOTS

Somehow almost an entire academic year went by without a blog post, I must have been busy… It’s time to get back in the saddle! (I’m using the classical definition of this idiom here, “doing something you stopped doing for a period of time”, not the urban dictionary definition, “when you are back to doing what you do best”, as I really don’t think writing blog posts are what I do best…)

One of the exciting things I took part in during the year was the NSF supported Reproducible Science Hackathon held at NESCent in Durham back in December.

I wrote here a while back about making reproducibility a central focus of students’ first introduction to data analysis, which is an ongoing effort in my intro stats course. The hackathon was a great opportunity to think about promoting reproducibility to a much wider audience than intro stat students — wider with respect to statistical background, computational skills, and discipline. The goal of the hackathon was to develop a two day workshop for reproducible research, or more specifically, reproducible data analysis and computation. Materials from the hackathon can be found here and are all CC0 licensed.

If this happened in December, why am I talking about this now? I was at USCOTS these last few days, and lead a breakout session with Nick Horton on reproducibility, building on some of the materials we developed at the hackathon and framing them for a stat ed audience. The main goals of the session were

  1. to introduce statistics educators to RMarkdown via hands on exercises and promote it as a tool for reproducible data analysis and
  2. to demonstrate that with the right exercises and right amount of scaffolding it is possible (and in fact easier!) to teach R through the use of RMarkdown, and hence train new researchers whose only data analysis workflow is a reproducible one.

In the talk I also discussed briefly further tips for documentation and organization as well as for getting started with version control tools like GitHub. Slides from my talk can be found here and all source code for the talk is here.

There was lots of discussion at USCOTS this year about incorporating more analysis of messy and complex data and more research into the undergraduate statistics curriculum. I hope that there will be an effort to not just do “more” with data in the classroom, but also do “better” with it, especially given that tools that easily lend themselves to best practices in reproducible data analysis (RMarkdown being one such example) are now more accessible than ever.

Interpreting Cause and Effect

One big challenge we all face is understanding what’s good and what’s bad for us.  And it’s harder when published research studies conflict. And so thanks to Roger Peng for posting on his Facebook page an article that led me to this article by Emily Oster:  Cellphones Do Not Give You Brain Cancer, from the good folks at the 538 blog. I think this article would make a great classroom discussion, particularly if, before showing your students the article, they themselves brainstormed several possible experimental designs and discussed strengths and weaknesses of the designs. I think it is also interesting to ask why no study similar to the Danish Cohort study was done in the US.  Thinking about this might lead students to think about cultural attitudes towards wide-spread data collection.

Notes and thoughts from JSM 2014: Student projects utilizing student-generated data

Another August, another JSM… This time we’re in Boston, in yet another huge and cold conference center. Even on the first (half) day the conference schedule was packed, and I found myself running between sessions to make the most of it all. This post is on the first session I caught, The statistical classroom: student projects utilizing student-generated data, where I listened to the first three talks before heading off to catch the tail end of another session (I’ll talk about that in another post).

Samuel Wilcock (Messiah College) talked about how while IRBs are not required for data collected by students for class projects, the discussion of ethics of data collection is still necessary. While IRBs are cumbersome, Wilcock suggests that as statistic teachers we ought to be aware of the process of real research and educating our students about the process. Next year he plans to have all of his students go through the IRB process and training, regardless of whether they choose to collect their own data or use existing data (mostly off the web). Wilcock mentioned that, over the years, he moved on from thinking that the IRB process is scary to thinking that it’s an important part of being a stats educator. I like this idea of discussing in the introductory statistics course issues surrounding data ethics and IRB (in a little more depth than I do now), though I’m not sure about requiring all 120 students in my intro course to go through the IRB process just yet. I hope to hear an update on this experiment next year from to see how it went.

Next, Shannon McClintock (Emory University) talked about a project inspired by being involved with the honor council of her university, when she realized that while the council keeps impeccable records of reported cases, they don’t have any information on cases that are not reported. So the idea of collecting student data on academic misconduct was born. A survey was designed, with input from the honor council, and Shannon’s students in her large (n > 200) introductory statistics course took the survey early on in the semester. The survey contains 46 questions which are used to generate 132 variables, providing ample opportunity for data cleaning, new variable creation (for example thinking about how to code “any” academic misconduct based on various questions that ask about whether a student has committed one type of misconduct or another), as well as thinking about discrepant responses. These are all important aspects of working with real data that students who are only exposed to clean textbook data may not get a chance practice. It’s my experience that students love working with data relevant to them (or, even better, about them), and data on personal or confidential information, so this dataset seem to hit both of those notes.

Using data from the survey, students were asked to analyze two academic outcomes: whether or not student has committed any form of academic misconduct and an outcome of own choosing, and presented their findings in n optional (some form of extra credit) research paper. One example that Shannon gave for the latter task was defining a “serious offender”: is it a student who commits a one time bad offense or a student who habitually commits (maybe nor so serious) misconduct? I especially like tasks like this where students first need to come up with their own question (informed by the data) and then use the same data to analyze it. As part of traditional hypothesis testing we always tell students that the hypotheses should not be driven by the data, but reminding them that research questions can indeed be driven by data is important.

As a parting comment Shannon mentioned that the administration at her school was concerned that students finding out about high percentages of academic offense (survey showed that about 60% of students committed a “major” academic offense) might make students think that it’s ok, or maybe even necessary, to commit academic misconduct to be more successful.

For those considering the feasibility of implementing a project like this, students reported spending on average 20 hours on the project over the course of a semester. This reminded me that I should really start collecting data on how much time my students spend on the two projects they work on in my course — it’s pretty useful information to share with future students as well as with colleagues.

The last talk I caught in this session was by Mary Gray and Emmanuel Addo (American University) on a project where students conducted an exit poll asking voters whether they encountered difficulty in voting, due to voter ID restrictions or for other reasons. They’re looking for expanding this project to states beyond Virginia, so if you’re interested in running a similar project at your school you can contact Emmanuel at addo@american.edu. They’re especially looking for participation from states with particularly strict voter ID laws, like Ohio. While it looks like lots of work (though the presenters assured us that it’s not), projects like these that can remind students that data and statistics can be powerful activism tools.

Pie Charts. Are they worth the Fight?

Like Rob, I recently got back from ICOTS. What a great conference. Kudos to everyone who worked hard to organize and pull it off. In one of the sessions I was at, Amelia McNamara (@AmeliaMN) gave a nice presentation about how they were using data and computer science in high schools as a part of the Mobilize Project. At one point in the presentation she had a slide that showed a screenshot of the dashboard used in one of their apps. It looked something like this.

screenshot-app

During the Q&A, one of the critiques of the project was that they had displayed the data as a donut plot. “Pie charts (or any kin thereof) = bad” was the message. I don’t really want to fight about whether they are good, nor bad—the reality is probably in between. (Tufte, the most cited source to the ‘pie charts are bad’ rhetoric, never really said pie charts were bad, only that given the space they took up they were, perhaps less informative than other graphical choices.) Do people have trouble reading radians? Sure. Is the message in the data obscured because of this? Most of the time, no.

plots_1Here, is the bar chart (often the better alternative to the pie chart that is offered) and the donut plot for the data shown in the Mobilize dashboard screenshot? The message is that most of the advertisements were from posters and billboards. If people are interested in the n‘s, that can be easily remedied by including them explicitly on the plot—which neither the bar plot nor donut plot has currently. (The dashboard displays the actual numbers when you hover over the donut slice.)

It seems we are wasting our breath constantly criticizing people for choosing pie charts. Whether we like it or not, the public has adopted pie charts. (As is pointed out in this blog post, Leland Wilkinson even devotes a whole chapter to pie charts in his Grammar of Graphics book.) Maybe people are reasonably good at pulling out the often-not-so-subtle differences that are generally shown in a pie chart. After all, it isn’t hard to understand (even when using a 3-D exploding pie chart) that the message in this pie chart is that the “big 3″ browsers have a strong hold on the market.

The bigger issue to me is that these types of graphs are only reasonable choices when examining simple group differences—the marginals. Isn’t life, and data, more complex than that?Is the distribution of browser type the same for Mac and PC users? For males and females? For different age groups? These are the more interesting questions.

The dashboard addresses this through interactivity between the multiple donut charts. Clicking a slice in the first plot, shows the distribution of product types (the second plot) for those ads that fit the selected slice—the conditional distributions.

So it is my argument, that rather than referring to a graph choice as good or bad, we instead focus on the underlying question prompting the graph in the first place. Mobilize acknowledges that complexity by addressing the need for conditional distributions. Interactivity and computing make the choice of pie charts a reasonable choice to display this.

*If those didn’t persuade you, perhaps you will be swayed by the food argument. Donuts and pies are two of my favorite food groups. Although bars are nice too. For a more tasty version of the donut plot, perhaps somebody should come up with a cronut plot.

**The ggplot2 syntax for the bar and donut plot are provided below. The syntax for the donut plot were adapted from this blog post.

# Input the ad data
ad = data.frame(
	type = c("Poster", "Billboard", "Bus", "Digital"),
	n = c(529, 356, 59, 81)
	)

# Bar plot
library(ggplot2)
ggplot(data = ad, aes(x = type, y = n, fill = type)) +
     geom_bar(stat = "identity", show_guide = FALSE) +
     theme_bw()

# Add addition columns to data, needed for donut plot.
ad$fraction = ad$n / sum(ad$n)
ad$ymax = cumsum(ad$fraction)
ad$ymin = c(0, head(ad$ymax, n = -1))

# Donut plot
ggplot(data = ad, aes(fill = type, ymax = ymax, ymin = ymin, xmax = 4, xmin = 3)) +
     geom_rect(colour = "grey30", show_guide = FALSE) +
     coord_polar(theta = "y") +
     xlim(c(0, 4)) +
     theme_bw() +
     theme(panel.grid=element_blank()) +
     theme(axis.text=element_blank()) +
     theme(axis.ticks=element_blank()) +
     geom_text(aes(x = 3.5, y = ((ymin+ymax)/2), label = type)) +
     xlab("") +
     ylab("")

 

 

Data Privacy (L.A. Times)

The L.A. Times ran an article on data privacy today, which, I think it’s fair to say, puts “Big Data” in approximately the same category as fire. In the right hands, it can do good. But…
http://www.latimes.com/nation/politics/politicsnow/la-pn-white-house-big-data-privacy-report-20140501,0,5624003.story

Conditional probabilities and kitties

I was at the vet yesterday, and just like with any doctor’s visit experience, there was a bit of waiting around — time for re-reading all the posters in the room.

vodka

And this is what caught my eye on the information sheet about feline heartworm (I’ll spare you the images):

cond

The question asks: “My cat is indoor only. Is it still at risk?”

The way I read it, this question is asking about the risk of an indoor only cat being heartworm positive. To answer this question we would want to know P(heartworm positive | indoor only).

However the answer says: “A recent study found that 27% of heartworm positive cats were identified as exclusively indoor by their owners”, which is P(indoor only | heartworm positive) = 0.27.

Sure, this gives us some information, but it doesn’t actually answer the original question. The original question is asking about the reverse of this conditional probability.

When we talk about Bayes’ theorem in my class and work through examples about sensitivity and specificity of medical tests, I always tell my students that doctors are actually pretty bad at these, looks like I’ll need to add vets to my list too!

Ranked Choice Voting

The city of Minneapolis recently elected a new mayor. This is not newsworthy in and of itself, however the method they used was—ranked choice voting. Ranked choice voting is a method of voting allowing voters to rank multiple candidates in order of preference. In the Minneapolis mayoral election, voters ranked up to three candidates.

The interesting part of this whole thing was that it took over two days for the election officials to declare a winner. It turns out that the official procedure for calculating the winner of the ranked-choice vote involved cutting and pasting spreadsheets in Excel.

The technology coordinator at E-Democracy, Bill Bushey, posted the challenge of writing a program to calculate the winner of a ranked-choice election to the Twin Cities Javascript and Python meetup groups. Winston Chang also posted it to the Twin Cities R Meetup group. While not a super difficult problem, it is complicated enough that it can make for a nice project—especially for new R programmers. (In fact, our student R group is doing this.)

The algorithm, described by Bill Bushey, is

  1. Create a data structure that represents a ballot with voters’ 1st, 2nd, and 3rd choices
  2. Count up the number of 1st choice votes for each candidate. If a candidate has 50% + 1 votes, declare that candidate the winner.
  3. Else, select the candidate with the lowest number of 1st choice votes, remove that candidate completely from the data structure, make the 2nd choice of any voter who voted for the removed candidate the new 1st choice (and the old 3rd choice the new 2nd choice).
  4. Goto 2

As an example consider the following sample data:

Voter  Choice1  Choice2  Choice3
    1    James     Fred    Frank
    2    Frank     Fred    James
    3    James    James    James
    4    Laura 
    5    David 
    6    James              Fred
    7    Laura
    8    James
    9    David    Arnie
   10    David

In this data, James has the most 1st choice votes (4) but it is not enough to win the election (a candidate needs 6 votes = 50% of 10 votes cast + 1 to win). So at this point we determine the least voted for candidate…Frank, and delete him from the entire structure:

Voter  Choice1  Choice2  Choice3
    1    James     Fred    <del>Frank</del>
    2    <del>Frank</del>     Fred    James
    3    James    James    James
    4    Laura 
    5    David 
    6    James              Fred
    7    Laura
    8    James
    9    David    Arnie
   10    David

Then, the 2nd choice of any voter who voted for Frank now become the new “1st” choice. This is only Voter #2 in the sample data. Thus Fred would become Voter #2’s 1st choice and James would become Voter #2’s 2nd choice:

Voter  Choice1  Choice2  Choice3
    1    James     Fred
    2     Fred    James
    3    James    James    James
    4    Laura 
    5    David 
    6    James              Fred
    7    Laura
    8    James
    9    David    Laura
   10    David

James still has the most 1st choice votes, but not enough to win (he still needs 6 votes!). Fred has the fewest 1st choice votes, so he is eliminated, and his voter’s 2nd and 3rd choices are moved up:

Voter  Choice1  Choice2  Choice3
    1    James
    2    James
    3    James    James    James
    4    Laura 
    5    David 
    6    James              
    7    Laura
    8    James
    9    David    Laura
   10    David

James now has five 1st choice votes, but still not enough to win. Laura has the fewest 1st choice votes, so she is eliminated, and her voter’s 2nd and 3rd choices are moved up:

Voter  Choice1  Choice2  Choice3
    1    James
    2    James
    3    James    James    James
    4     
    5    David 
    6    James              
    7    
    8    James
    9    David    Laura
   10    David

James retains his lead with five first place votes…but now he is declared the winner. Since Voter #4 and #7 do not have a 2nd or 3rd choice vote, they no longer count in the number of voters. Thus to win, a candidate only needs 5 votes = 50% of the 8 1st choice votes + 1.

The actual data from Minneapolis includes over 80,000 votes for 36 different candidates. There are also ballot issues such as undervoting and overvoting. This occurs when voters give multiple candidates the same ranking (overvoting) or do not select a candidate (undervoting).

The animated GIF below shows the results after each round of elimination for the Minneapolis mayoral election.

2013 Mayoral Race

The Minneapolis mayoral data is available on GitHub as a CSV file (along with some other smaller sample files to hone your programming algorithm). There is also a frequently asked questions webpage available from the City of Minneapolis regarding ranked choice voting.

In addition you can also listen to the Minnesota Public Radio broadcast in which they discussed the problems with the vote counting. The folks at the  R Users Group Meeting were featured and Winston brought the house down when commuting on the R program that computed the winner within a few seconds said, “it took me about an hour and a half to get something usable, but I was watching TV at the time”.

See the R syntax I used here.

 

Personal Data Apps

Fitbit, you know I love you and you’ll always have a special place in my pocket.  But now I have to make room for the Moves app to play a special role in my capture-the-moment-with-data existence.

Moves is an ios7 app that is free.  It eats up some extra battery power and in exchange records your location and merges this with various databases and syncs it up to other databases and produces some very nice “story lines” that remind you about the day you had and, as a bonus, can motivate you to improved your activity levels.  I’ve attached two example storylines that do not make it too embarrassingly clear how little exercise I have been getting. (I have what I can consider legitimate excuses, and once I get the dataset downloaded, maybe I’ll add them as covariates.)  One of the timelines is from a day that included an evening trip to Disneyland. The other is a Saturday spent running errands and capped with dinner at a friend’s.  Its pretty easy to tell which day is which.

movings1movings2

But there’s more.  Moves has an API, thus allowing developers to tap into their datastream to create apps.  There’s an app that exports the data for you (although I haven’t really had success with it yet) and several that create journals based on your Moves data.  You can also merge Foursquare, Twitter, and all the usual suspects.

I think it might be fun to have students discuss how one could go from the data Moves collects to creating the storylines it makes.  For instance, how does it know I’m in a car, and not just a very fast runner?  Actually, given LA traffic, a better question is how it knows I’m stuck in traffic and not just strolling down the freeway at a leisurely pace? (Answering these questions requires another type of inference than what we normally teach in statistics. )  Besides journals, what apps might they create with these data and what additional data would they need?