A review of my JSM 2016 (Part 1 – Sunday and Monday)
JSM 2016 was last week in Chicago (the greatest city in the world). As always, it was awesome.
This year I was delighted to get to put together the first (hopefully annual!) JSM Data Art show. This year’s inaugural show featured 6 artists: Marcus Volz, Alisa Singer, Craig Miller, Jillian Pelto, Elizabeth Pirraglia, and me (Gregory J. Matthews)
The first session I attended was also the session where I gave my talk on Sunday at 4. The title of the session was “For the love of the game” and featured speakers Dan Nettleton, Brandon LeBeau, Douglas VanDerwerken, Thomas Fullerton, and myself.
Dan Nettleton (with Dennis Lock) spoke about whether or not to go for it on 4th down in the NFL using a random forest model. They used this model to make 4th down decisions based on maximizing win probabilities. Details of their model can be found here and the full paper is here. The examples they used in their talk were from the 2016 AFC championship game where Denver beat New England. Based on their model, the Patriots could have increased their win probability by choosing to kick the field goal rather than go for it on 4th down in both instances.
Dan Nettleton with Dennis Lock – To go or not to go. 4th down analysis. Using Random Forests. N = 430168 plays. 12 years of data. Belicheck should have kicked in the Pats Broncos game both times when they went for it on 4th down near the end of the AFC championship game. Interesting point of do you want to win more in the long run or do you want to win THIS game.
Brandon LeBeau (@blebeau11) spoke about hiring decisions in college football using an item response theory model. Based on this model and measuring success of a coach in terms of expectation of team strength, he concludes that Tim Brewster was a terrible hire for Iowa State. Also a terrible hire? Brady Hoke.
Douglas VanDerwerken spoke about soccer suspensions in the English Premier League (EPL). He found that fouls were reduced by 12% and 23% when players were facing a 1 and 2 game suspension, respectively. He also noted that they found evidence that players foul more often when refs give fewer cards, but also when refs give MORE cards. This may seem surprising at first, but I think what’s happening here is that when ref’s give fewer cards players are more likely to foul more because they know they can get away with it. And when a ref gives MORE cards, they are also probably calling more fouls. I think an interesting extension of this would be to treat the ref’s call like the result of a medical test and have some panel of experts decide whether something was actually a foul. Then you could better measure the actual rate of fouling and the actual rate of fouls being called. Right now those two effects are confounded with each other. This might be tedious and I’m not even sure you you could get experts to agree on what is an objective foul.
Tom Fullerton spoke about college football ticket sales at UTEP. He found many interesting things, but I thought it was particularly interesting as a coach coaches more games, ticket sales on average go down.
I spoke about statistical disclosure with respect to Baseball Hall of Fame voting data. You can find my full slides here. My actual slides at JSM were a much more condensed version of this talk and I actually managed to finish my talk early. It is the first time I have ever finished a talk early.
On Monday morning, I showed up late to the session “Advanced Methods for Statistics in Sports“. I got there in time to see Andrew Swift talk about modeling sports outcomes using a ratio based point differential. His slides can be found here. He also used something called the Farey sequence, which I spent some time reading about. Interesting. Scott Powers (@saberpowers) presented some work with all-stars, Trevor Hastie and Robert Tibshirani. His method, nuclear penalized multinomial regression (npmr) (R package here), was applied to outcome in baseball. They view the outcome of a plate appearance as a multinomial outcome and multinomial regression is a natural choice in this setting. The key to their penalty is, rather than using the Froebenius norm, they penalize the rank of the matrix of coefficients. The resulting estimates can then be interpreted as latent “skills”. So really instead of estimating 9 different outcomes, they are really estimating something like three latent tools or skills. For batters, the three skills work out to be basically the three “true” outcomes (i.e. plays where no fielder is involved). These are K, HR, BB, and HBP. The second skill is power, which is characterized by more flyballs and HR. The third batter skill is plate discipline (i.e. more walks and less K). For pitchers, there are also essentially three skill: contact avoidance (More K and BB), trajectory skills (more singles and ground balls) and command (more ground balls). Scott conclude that in the context of baseball, this method is not significantly better than ridge regression, however, the resulting interpretation of the output of npmr is much more interesting than ridge regression.
(Weird story. I met Powers at SABR. Didn’t realize he was the npmr author and used npmr for a completely unrelated application (fossilized tooth classification). I then emailed him a question about the package, and he responded very quickly with the answer. Then like a few weeks later when I was looking at JSM program for talks I wanted to see, I realized that he was the author of npmr and that I had already met him. #smallWorldStats)
Next, Sameer Deshpande spoke about pitch framing using hierarchical Bayesian logistic regression (The slides are here) using PitchFX data from 2011-2015. In aggregate, the best catchers in terms of framing are Montero, Zunino, Lucroy, Rivera, and Rene. However, these results are related to the number of opportunities a catcher gets. When this is removed you get a statistic, which they refer to as SAFE2. The best catchers in terms of SAFE2 are Rivera, Conger, Vazquez, Montero, and Grandal. After than you have players like Zunino, Maldonado, Stewart, Martin, and Butera.
Stephanie Kovalchik (@statsonthet) was supposed to speak in this session, but her flight from Australia (!) was delayed and she wasn’t able to make her talk. However, you can find the slides from her talk here.
Monday night I was supposed to go to the UConn alumni dinner, but I never made it as I went to the Statistics in Sports section meeting before and just couldn’t get my act together.