Can the NFL Ever Predict Injuries?

Knowledge at Wharton

Ask about Knowledge at Wharton

Episode Description

ABOUT THE EPISODE NFL teams have more injury data than ever. But can they actually predict who will get hurt?

Zach Binney, Associate Professor of Epidemiology and Data Science at Oxford College of Emory University, explains why injury prediction remains so difficult and what researchers are learning from player, coaching, and turf data.

Binney discusses why risk varies so much from player to player, why some artificial turf has been linked to higher injury rates, and whether biomechanics could help teams identify potential injuries before they occur.

ABOUT THE PODCAST Sports is a game of numbers. Wharton experts Eric Bradlow, Shane Jensen, Cade Massey, and Adi Wyner team up to tackle the world of sports, from current events to longstanding issues such as: What sports streaks are the most impressive? How do you rank the best players? Can athletes be compared across sports? Moneyball Highlights explains how decision-makers in the game can avoid the common mistakes and embrace the data. Episodes are recorded at the Wharton School.

----- The Wharton School of the University of Pennsylvania is committed to sharing its intellectual capital through Knowledge at Wharton, the school’s online business analysis journal.

Launched in 1999, Knowledge at Wharton offers free access to:

- Articles, podcasts, and videos highlighting Wharton faculty research and analysis of current business trends - Interviews with book authors from Wharton School Press - In-depth series of curated content like Wharton Executive Education’s Nano Tools for Leaders - A searchable database of more than 10,000 articles covering all aspects of business - A weekly newsletter that delivers Knowledge at Wharton insights directly to your inbox

Visit the Knowledge at Wharton homepage: https://whr.tn/3qUpSBE Subscribe to the Knowledge at Wharton YouTube channel: https://whr.tn/3DEAgAU

Read the full transcriptHide the full transcript
  1. Welcome, welcome to Wharton Money Vault. Welcome to a full hour of sports analytics
  2. here on the Wharton Podcast Network. This is Cade Massey hosting this week with the whole crew.
  3. Eric Bradlow is here. Shane Jensen is here. Audie Weiner is here. We are all in Philadelphia or near Philadelphia,
  4. though we're scattered around four different locations this time. Sadly, we are here.
  5. Some combination of us are here almost every week of the year, 48, 49 weeks of the year, have been doing it for more
  6. than 12 years now. We are recording on a Tuesday afternoon. The show will go up on Wednesday, as
  7. it usually does. Also, as usual, typical format this week, and have a guest in the first half of
  8. the show and roll into some open topics, open lines in the second half of the show.
  9. I'm glad to be back. I've missed the last two weeks because of teaching responsibilities.
  10. It's fun to be back in here. Lots to talk about to begin with. To begin with, we have a guest, first
  11. -time guest, Zach Binney, Professor Zach Binney. Zach is an Associate Professor of Epidemiology and
  12. Data Science at Oxford College at Emory University down there just outside of Atlanta.
  13. Zach is also in Atlanta. Zach does work on sports injuries, athlete health, uses stats to figure out what goes into
  14. those things. He's worked with organizations across the NFL, Major League Baseball, NBA, a little soccer, EPL in
  15. the Premier League, but it looks like maybe Atlanta. He's got some loyalties to Atlanta.
  16. He's flashing on his back wall. Glad to hear about all that. Zach, thanks for joining us.
  17. Appreciate the time. Absolutely. I was actually in Philly last week, had the best sandwich in my life, not a
  18. cheesesteak, but Reading Terminal Market, man. I texted my wife. I was like, when I die, I know
  19. it's probably illegal. You can sprinkle them in the toilet or something if you need to, but you got
  20. to leave a little piece of me at Reading Terminal so my soul can just graze on it for eternity.
  21. What was the sandwich? Tell us. No, it was Il Malocchio, the evil eye at Chef Betts.
  22. Shout out to Chef Betts, man. Incredible stuff. Wow. That is an endorsement. All right.
  23. Reading Terminal is one of the two or three big, big things you need to do when you're in Philadelphia.
  24. So good on you for pulling that off. Zach, start with just telling us like right
  25. now, what are you thinking about? Like what's the project on the front of your head, front of your desk?
  26. What's central to your thinking at the moment? We can take you a lot of different
  27. directions. Yeah. I mean, one of the things that I'm thinking about a lot and have been for
  28. years is injury prediction, right? That's the holy grail of what we do, right? Especially in the NFL, predicting who's going to
  29. get hurt and who is not going to get hurt. It turns out that's real hard. Nobody's any good at it.
  30. And so one of the things that I'm on with Ron Yurko over at Carnegie Mellon actually- Never heard of him.
  31. Never heard of him. Never heard of him. Nobody listens to the show. He's been on the show as many times
  32. as we have. Oh, I thought he was persona non grata here at some point. No, Ron and I are working on this
  33. multi-level modeling approach to tease out different levels at which there may be factors influencing
  34. NFL injuries. So kind of trying to figure out like how much of it is variation from player
  35. to player, how much of it is team -level factors, coach-level factors, medical staff-level
  36. factors, and then stadium-level factors as well. The latter being mostly turf issues.
  37. Substation-level factors? No substation-level factors. If anybody heard me on Pablo Torre a
  38. couple of months ago, I think we pretty conclusively put that one to bed. But we can talk about San Francisco if
  39. you want. So Zach, can you tell us, I would think, maybe it's not a, I'll call it
  40. a multi-level model, which all of us here have fit to data before, but are you telling me that there hasn't been a
  41. large-scale analysis to predict injuries as a function of, I'll call it, I don't know,
  42. the type of turf at the stadium or whether it's a team-specific effect or, I don't know, the BMI of the player or
  43. whatever it is. Are you telling me there has not been like a meta-analysis or a massive dataset
  44. constructed where this can be done? There have been plenty of attempts to kind of describe bivariate relationships.
  45. So like, as BMI varies, how does injury risk vary? Of course, you have to take into account
  46. position there as well. Because actually, this is one of the examples of a confounding that I often use in
  47. my lectures is I'll show students, hey, if you actually just look at football player weight
  48. or BMI and injury rate, as BMI goes up, injury rate goes down. But that's because some of the lightest players
  49. are like kickers, punters, quarterbacks. These are not players who are getting hurt
  50. that much. And once you split it out by position, you see a very clear Simpsons paradox example
  51. where the relationship reverses. And within every position, the heavier you are,
  52. the higher your injury risk is. So there's certainly been work done on that, a good amount of work done on turf,
  53. although the best work that I think was actually done by the NFL internally kind of
  54. leaked a few years ago, like a slide or two, but it's never actually been published.
  55. But a former student of mine, we're working on sort of doing the best we can with public data, and hopefully that'll be published
  56. soon. And we actually did replicate the NFL's findings, which are basically, I don't know if you
  57. guys remember the slit film turf controversy from a few years ago, but basically everybody talks
  58. about artificial turf being worse. But what the NFL found and what we found is it's not all artificial turf.
  59. There are two basic kinds of artificial turf. One is monofilament, so that's like individual blades
  60. of grass sticking up. It's not grass, it's plastic, right? But individual blades sticking up.
  61. The other is called slit film. It's like a Y. So you've actually got two sort of threads
  62. coming up from each stem. That is associated with a greater injury risk. The NFL actually got rid of that a
  63. couple of years ago, because there's some resistance to like, well, you can't plant grass everywhere.
  64. I don't know if you can do it in Chicago. I feel like you can figure it out in most places.
  65. But yeah, but there wasn't a lot of pushback to like, okay, let's switch kinds of
  66. artificial turf. So the NFL got rid of that a couple of years ago. And now we don't really see actually that
  67. much of a difference between the remaining artificial turfs and natural grass. The other factors that you mentioned in the
  68. study that you've launched into with Ron, obviously, they're just central to organizations.
  69. It's a little bit like character and makeup in that people know it matters. It's just hard to measure.
  70. And so they're forever trying to figure out better ways to measure it. Two things you mentioned that are seemingly pretty
  71. central and interesting. It's like, whether there are, you said, coaching effects or staff effects, like basically training, kind
  72. of a training regimen that goes up. You know, there's always talk in summer about
  73. coaches that push them too hard versus not preparing them enough for tackling. You know, that seems like a real tradeoff.
  74. Do you know anything about that? And then the other one, really, the Holy Grail, as you said, is, is there anything
  75. in player evaluation as they go to draft a college kid or as they go to sign a pro free agent?
  76. Is there anything at the individual level that we're missing? Those are the two big questions that come
  77. to mind. Yeah. So the coaching stuff is really interesting. And now that we have a bigger data
  78. set, when we initially looked at this, we only had a few years of data. And now we've got about 12 years of
  79. data, which means you have variation in coaches across franchises. Right. So like one of the really interesting things
  80. that we found, John Fox in Chicago, one of the safest coaches to play for John Fox in Denver, one of the most dangerous
  81. coaches to play for. I have wanted for years to get on the phone with John Fox and be like,
  82. what do you make of this? Because it's those kind of splits, right? Where you could maybe find something and I'd
  83. love to hear, you know, what, what was he doing something different? Was it a different roster?
  84. What are we doing? Let us jump in Shane and then Adi. No, yeah. I mean, maybe kind of use that as
  85. a jumping point. Like what are the samples? Like you, I mean, you do say you've got 12 years of data, so you can
  86. kind of look at like, well, what, what are those kind of sample sizes that you're
  87. basing this John Fox distinction on? Yeah. So we've got, I believe, three years in
  88. Chicago and two years in Denver, vice versa. I don't remember exactly. I'd have to go back and look.
  89. So I guess I, I guess the follow up question is, is like, how many years of data do you kind of need to
  90. like say something about it? Given that injury or injuries aren't hard occurring
  91. constantly, like how, what, what kind of data do you need to say something substantial about
  92. it? And related, what's the granularity of the data and the level of analysis? I'm waiting to get in here, guys.
  93. So just let, anytime you're ready. It's the same question, Adi. We're just following up.
  94. A radio PSA. If you could only see the expression on Adi's face for the last 45 seconds, we'll
  95. give you, we'll give it Adi time momentarily. Zach, please continue. Yeah, no. So that's a really good question.
  96. And that's something that Ron and I have thought a lot about actually. Originally we did this kind of bootstrapping approach
  97. to get a handle on the uncertainty, but we never really loved that because the unit
  98. of analysis that we have is the player game, right? So a player nested within a team, nested
  99. within a game, nested within a season, right? And then we had like 12 seasons. And so when you do a bootstrap sample
  100. on something that nested, it's like, what is the proper level to do this at? Anyway, Ron could talk a lot more about
  101. this, but now we're taking a fully Bayesian approach. So the answer to how much data do
  102. we need is, well, you can kind of look at the credible intervals, right? That we're getting for John Fox in Denver
  103. versus John Fox in Chicago. How much overlap do we have? And there are definitely what looked to me
  104. like meaningful differences. And then you've also got some similarities sometimes.
  105. So like Miami, the Dolphins actually threw out multiple coaches. Originally it was Adam Gase and Joe Philbin,
  106. but now we've had a couple more. Their injuries have stayed low. Same thing in Philadelphia.
  107. This year may be an exception, but there have been some freak occurrences. We all know stuff happens.
  108. But Philadelphia across coaches generally stays low. So that kind of cues you in like,
  109. hey, what are Philly and Miami doing that's not tied to the coach level? So it can lead us in some interesting
  110. hypotheses, I think. Zak, you've got a friend that's making an appearance on the show.
  111. Want to introduce your friend? Yes, I do. This is Taco. You will probably see her walk in front.
  112. Sorry about that. Danny is around here somewhere too. Hold on, hold on. Zak just told us that he has a
  113. cat named Taco. Taco cat. That's an amazing- It's a palindrome. Palindrome. There you go.
  114. All right. Audie Weiner has been waiting for a moment. Jump in, Aud. Well, yeah.
  115. So we've been talking around this, and I think maybe you can answer. Maybe we're going to have to bring Ron
  116. to get the answer. If you're being Bayesian, which is reasonable, I would expect to be Bayesian too about these
  117. problems. The way I would ask Ron is what's your tau? What's your sigma? But the way I'll rephrase it is sigma
  118. is going to be your noise, year to year noise, individual. It's the rate at which things are just
  119. happening randomly. Tau is the difference between the factors that you're trying, the actual effects that you're trying
  120. to measure. Those tend to be pretty small. They're big when you're looking at player talent,
  121. so they're very big differences between player talented, and they dwarf the natural variations.
  122. But you're looking at injuries and things related to that. I would guess that sigma dwarfs tau and
  123. that every estimate is therefore very heavily shrunk. Audie, can I try English real quick?
  124. Layman's real quick. Are you saying that, just consider player injuries for example.
  125. You're saying, look, huge variation of what happens to a guy. There's a lot of random noise of what
  126. happens to a guy. There may be individual proclivities towards injury, but those individual differences are relatively small next to
  127. the huge variation of just shit that happens. Is that kind of? Right. When we observe the shift you talk about
  128. with the coach, one team to another, most of that is probably attributable to the natural
  129. variance. Yeah, you can have a few years, but that shrinks your sigma. More and more data shrinks the natural variation
  130. and the uncertainty there. But I just want to get a sense of that, right? Do you have any sense of what are
  131. the effect sizes of the things you think are driving this? What are the natural variations?
  132. What's the relative size of those numbers? I can get some handle on it. Yeah. So the median odds ratio for a player,
  133. for like a difference between players in terms of injury rate is about 1.8. Well,
  134. it's about a double. Real quickly, odds ratio reflects the increase in likelihood of something happening where one is null
  135. essentially. So 1.8 says it's 80% more likely to happen for a person. Yeah. So a median odds ratio is a little
  136. more complicated. So it's basically if you take any two random players, right? And you take the higher risk player and
  137. the lower risk player, right? And you compare them and you do an odds ratio just between those two.
  138. And then you do every possible comparison and you take the median of all of those
  139. comparisons. It's about an 80% higher injury rate across players. When you look at team head coach combinations,
  140. which is the level that we looked at. So we didn't just look at like comparing Miami Dolphins and New York Giants.
  141. We would take a look at comparing Miami Dolphins under Mike McDaniel to New York Giants
  142. under Brian Dable, because when the head coach changes, a lot of other things change, staff
  143. changes, practice patterns change. So we wanted to look at differences there. The median odds ratio is like 1.1,
  144. 1.15. So it's much, much smaller, much smaller. So the biggest differences are definitely from player
  145. to player than they are from team to team, but there might be some small amount of signal from.
  146. So just to rephrase it. So odds ratio. So odds are just the probability of getting
  147. injured divided by the not, probably not getting injured. Right. So it's probably getting injured is small.
  148. You're basically just looking at the ratio. And so one minus the probability is really
  149. big. You're basically just looking at the ratio of the probability. So that's easy.
  150. And so 1.8 means approximately 80% more likely to be injured. So players really are different from player to
  151. player to player. That's true. And I would imagine a lot of that is driven by size, position, things like that.
  152. But the coaching, anything else, those tend to be really tiny, 1.05s, and that's in
  153. the medians. So that gives me exactly the answer that I wanted. So you're looking at so much more variation
  154. among players than among the things you're trying to measure and that drive it. And then the question is, what is it
  155. between players that's actually causing this? Right. That's kind of the next step.
  156. And I think it's very easy to find bivariate associations between various things.
  157. It is a lot harder to take that and try to predict forward what's going to happen, even when you're focused purely on non
  158. -contact injuries. So I'm never going to be able to predict a guy falling on another guy's ankle
  159. in the pile and breaking it. You're literally never going to be able to predict that.
  160. I thought you were going to say Aiden Hutchinson celebrating a sack and tearing his ACL
  161. for the season. Now, Eric, I was about to say that, and then I corrected myself because in theory,
  162. that was a non-contact injury. There might be something in Aiden Hutchinson's history,
  163. in his genetics, in his training patterns, in the way he plays, there might be something
  164. there that you could tease out. So that's actually one that in theory, maybe you could predict, but nobody's been able to.
  165. Let me ask you, do you and Ron have any, I'll call it, I don't call it magic data source, but like, do you
  166. have any other data, whether it's, well, all right, so let me ask you a question
  167. as a someone that's an epidemiologist, not epidemiologist. Yes. Yes. Yes. Yeah. Epidemiologist.
  168. And if you could have any data you wanted, forget the data you have, if you had any data you wanted, would it be
  169. like literally, I'll call it body MRI data? Would it be motion tracking data? Genetic.
  170. Genetic. Genetic data. Maybe it'd be that. It would be all of the above, but it's so like, even if I had all
  171. of that data, even if you gave me genetics, player tracking data, biomechanics, the new biomechanical
  172. tracking data that the NFL is collecting, it's like, what do I even begin to do with it?
  173. It's so hard to figure out what even in there is going to be most useful. And to be honest, I think we still
  174. have no idea. Switching tracks to baseball briefly. One of the things that I have been
  175. kind of fascinated by is there was an article on, I think the hard ball times a little over a decade ago called the
  176. injury zone revisited. It was by Josh Kalk. Some of you guys may know Josh and what he was trying to do.
  177. And then somebody I'm forgetting the name of the guy who followed up on it. Basically trying to do ultra short-term injury
  178. prediction. So I am working with somebody from major league baseball, a former student of mine right
  179. now, to try to figure out basically if we can figure out within a game, is there a signal when a pitcher's elbow or
  180. shoulder is going to give out? Wow. So can we look at like variation over the last 15 fastballs in either VLO or
  181. placement versus target or variance in that? And we're working with some biomechanists and others
  182. to try to dig into that. So I think that biomechanics data, which is a little, I mean, still very difficult to
  183. get your hands around in baseball, but is orders of magnitude more difficult in football.
  184. I think maybe that has some promise, but we'll see what our work turns up. Zach, I was going to ask you about
  185. biomechanical because that's been kind of the frontier for the last few years. And I feel like it's so hard to
  186. figure it out that it's still kind of below the radar. And presumably some of it's because you need
  187. some theory, like you said, where do you start? So you give me a bunch of biomechanical
  188. data, but I need to know where to start. So I'll probably need some theory. It reminds me of some of the horse
  189. racing guys we had on years ago, talking about watching the horses run by and modeling
  190. their gates, for example, and trying to forecast something off of that. Is there not, I understand football is more
  191. complicated because guys are moving around, but for example, if the biomechanical folks just filmed people
  192. running 40s or the three-cone drill or something like that, do they not have some
  193. models and theory on what is kind of good and what is less good? Or is it like a golf swing where,
  194. look, it kind of doesn't matter. There may be an ideal golf swing, a plutonic ideal of some kind, but if it
  195. works for you, it works for you. And every now and then you get somebody with a very unorthodox swing that works just
  196. fine. We certainly have ideas about like the better ways to move and the worse ways to
  197. move. I think linking those to injury risks, at least at the professional level, that has still
  198. proven pretty elusive. And you also use biomechanics to improve performance, which I think is where a little more
  199. of the focus has been. Obviously you want to prevent injury as well, but it's just proven very elusive to really
  200. tie all that together. Maybe there's some work being done internally that I'm not aware of.
  201. Teams are investing in more biomechanists and kinesiologists and movement scientists and things like that.
  202. But at least publicly that I'm aware of, there hasn't been a lot of very successful
  203. work on that front. When you work with teams, what kinds of questions do they ask you that we haven't
  204. been asking you, we haven't been talking about? What have you heard that you've been kind
  205. of surprised that they might be interested in? This probably won't come as a surprise, but
  206. what I always tell my students is you would be surprised at the simplicity of questions
  207. that teams need answered. I'll tell an anecdote. Granted, this was a decade ago now, but
  208. the first NFL team I worked for, which I won't call out here, I went in, I was ready to produce a fancy predictive
  209. model. We're going to figure out who's going to get hurt. You know what the first question they asked
  210. me was? How often do running backs get hurt? That's it. They had no idea, had no clue.
  211. Honestly, when I go in, it's a lot of, how can we use the data that we already have better?
  212. How can we improve our decision-making processes? It's not making these fanciful predictions, but it's
  213. a lot of like, what's a different way that we can think about our rehab data? For example, you may have it in this
  214. messy spreadsheet and maybe I can teach your guys about survival analysis and Kaplan-Meier curves,
  215. and we can learn how to read one of those curves and kind of cut these guys by different factors and see what is
  216. leading to quicker rehab and what isn't. So honestly, it's a lot of more descriptive
  217. and bivariate stuff that just relies on getting a handle on the data that they have
  218. internally already. It tells you something about the state of performance. It's just the kind of nature of the
  219. beast. It will get a lot better over the coming years. Is injury related to usage?
  220. Do people get injured more at later parts of the game versus earlier? Do people get injured more later in the
  221. season versus earlier? Do people get injured more on four days rest or in back-to-back games?
  222. I'm sure these are all questions you've heard before. Oh, I can answer each of those questions
  223. individually, and the answer varies. So the first part was, what was the first part of your question?
  224. It was, do they get... Like later in the game versus earlier in the game. So I had never actually seen this looked
  225. at until a couple of my students agreed to look at it just last year. And we published a blog post.
  226. We haven't actually turned it into a paper yet. But what we did is we went into the play-by-play data for the NFL,
  227. and we pulled out whenever injury is written, right? And we did a little basic natural language
  228. processing, pulled it out, figured when these injuries were occurring. It is about a linear increase over the
  229. course of the game. We all kind of had a sense that it increased with fatigue. Nobody knew how much.
  230. Nobody knew what shape the curve was. Turns out it's linear. It about doubles. So I can tell you that in the
  231. NFL. I can also tell you in the NFL what happens over the course of a season. Injury rates stay flat.
  232. This surprises a lot of people, but I think what we're seeing is basically a healthy
  233. worker effect. So the highest injury rates occur right at the start of training camp.
  234. You lose all these guys, ACL tears, Achilles are popping all over the place because they're
  235. coming off the off season, and they need to work up their load. And that's when they get hurt.
  236. By the time you've gotten into season, by week eight, anybody that's gotten hurt has gotten
  237. hurt. And you're just left with a more robust population. Well, you got that. Yeah.
  238. The low end of the 1.8 is still playing. Exactly. And that balances out. You're left with those guys.
  239. I'm very interested about in the game because I'm wondering whether or not it's just that
  240. they're exhausting themselves, and that's when injuries are much more likely. That's when you don't have the strength to
  241. do the thing. No, you need linear though. It's not exhausting. It's like every unit of energy expended increases
  242. the risk. So what mechanism gives you that linear increase? Well, it's probably hard to detect any up.
  243. You're getting up is what you're getting. A very simple model of injury is you get injured when you've got a body part
  244. and your fatigue exceeds the load that that body part can handle. It's almost like profit equals revenue minus cost.
  245. It's like your muscle has a certain ability to withstand load. As it gets fatigued, that ability to withstand
  246. load goes down. The load continues going up. At some point, these two meet and the
  247. load exceeds what you've got and you get injured. Is there any way to separate types of
  248. injury? Some injuries are more like fluke injury from someone whacking you with a ridiculously hard tackle.
  249. And that's not the same as an injury that happens due to fatigue and load. Yeah.
  250. So usually what we'll try to do is divide contact and non-contact injuries. Forget contact injuries.
  251. You're never going to predict those. You just look at non-contact injuries. That is somewhat tricky to do with public
  252. data. You need to either do a film review and even then if an injury occurs in practice, which most of them don't, a little
  253. over 80% of injuries that occur in the NFL during the season occur in games. You could go and review the film or
  254. you could get access to the NFL's injury data, which can be a tricky process, but
  255. it's definitely possible. Zach, Eric asked you a couple of others high-level questions that you said you had
  256. answers to. So what about within... You had three of them, Eric. You said within game, within season, and...
  257. I don't remember what I said. Maybe, I don't know, over career. I don't think I said age and time,
  258. but I'm sure you could answer that too. Yeah. Well, age is another fun one because you've
  259. got the whole issue with aging curves of guys dropping out. And so you get an aging curve that
  260. doesn't look quite right because at the older ages, you've only got, again, this healthy worker
  261. effect. That's what we call it in epidemiology. I'm sure there's another name in different fields.
  262. Over time, we've seen not that much of an increase over time in injuries in the NFL.
  263. And the problem with looking over a long period of time also is you've got different
  264. reporting practices, especially when it comes to things like head injuries, but just in general, kind
  265. of more recognizing when you need to pull a guy out of a game, sort of the culture changes.
  266. So it can be hard to track those longer-term changes over time. Okay. Zach, we're going to need to wrap up
  267. with you. We wish you the best with the work. It's really important work in general.
  268. It's good for the health of the leagues, but also individual teams, super, super interested in
  269. this stuff. Wish you the best. Glad to have made contact with you. Let's talk again more down the road.
  270. Thank you. I'd be happy to. That's Zach Binney, Dr. Zach Binney, an Associate Professor of Epidemiology and Data Science at Oxford
  271. College at Emory University down there in Atlanta, doing work on all things sports injury and
  272. health across all kinds of sports. And that has been the first half of this week's Wharton Money Ball.
  273. Come back and join us after the break. Welcome back. Welcome back to Wharton Money Ball.
  274. Welcome to the second half of this week's show here on the Wharton Podcast Network.
  275. Cade Massey hosting with the whole crew. Shane Jensen is here. Audie Weiner is here.
  276. Eric Bradlow is here just off the line with Dr. Zach Binney, Epidemiologist and Data Scientist
  277. down there at Emory, doing lots of interesting work. That's a worthy life, right?
  278. Dive into that whole world. Bring epidemiology and data science chops to help people figure that stuff out.
  279. Long way to answers, but man, it sure does seem like we should have answers at some point, right?
  280. It can't be impossible. Description, prediction first, then you move on to prevention and everything else.
  281. So, I mean, unless you have a measurement system, it's pretty hard to know what treatments
  282. are effective towards it, right? Right. I'm optimistic about the biomechanics, obviously genetics, and
  283. we're going to get there eventually. Super interesting. Okay, fellas, what else in the world of
  284. sports? I have been away for a couple of weeks. We are late. We're deep into the college football season.
  285. There are only two weeks left in the regular season. I mean, rivalry week is only one week
  286. away. We'll do conference championships after that. We'll get the playoffs seeded.
  287. We're still in 12-team playoff land last year. Next year goes to 24. I don't know how closely y'all are
  288. watching. I ran into Adi at 7.30 in the morning walking into the building this week.
  289. He's like, hey, Cade, good victory. And I felt like, man, we've done some good. Adi knows about college football weekend results.
  290. Oh, yeah. And I read the article about Arch Manning and how he's been working himself into holes,
  291. but then coming back with these miraculous comebacks. Am I always in view on those things?
  292. They can't really – you can't have miraculous comebacks forever. Seriously. I hear you.
  293. And Texas is not – they get enough press. We don't need to give them more press.
  294. I've been a bit evangelical lately. I've been hot on our Twitter feed about analytics used for the playoff rankings.
  295. And it's purely evangelical. It's purely like, let's just use a couple of good models.
  296. Let's use a simple system. And how far can we get with that? Just real simple. And I've found it helpful.
  297. And I'm hoping it maybe affects the dialogue a little bit. We've got plenty of sharp people out there,
  298. but I'm not sure how many are in the committee's room. So maybe it's probably a multi-year effort.
  299. But I've been really pushing two metrics, some kind of strength of record. Of course, these are the obvious metrics.
  300. Most deserving and best. These are what everybody talks about. Most deserving and best.
  301. Most deserving, strength of record, some kind of reflection of the quality of your accomplishments.
  302. Just wins. Just win, baby. An Al Davis way of thinking about it. And then best, which is easy.
  303. That's just power rankings, market numbers. I thought there'd be more of a debate
  304. within the committee on what the weight is on those things. So far, they're heavily skewed towards most deserving.
  305. So let me talk you through a couple of things real quick. The most interesting thing is that the field
  306. this past week really kind of clarified. We'll see, clarifies probably too grand because the
  307. season's not over. But I don't think there's as much debate about what 12 teams are in this week.
  308. And at the same time, I have no idea how they're going to rank them. I mean, these things are so jumbled up
  309. that there's so much parity. Teams have played very different schedules. I don't have strong positions on how they
  310. rank on one to 12, but I'm going to have strong positions if they start throwing
  311. teams in there that don't seem to be in there. So the new team, the exciting new addition,
  312. the one that people ought to be talking about is Wake Forest. Who had Wake Forest in the playoffs while
  313. Oregon and Ohio State are not in the playoffs? They were just outside the committees 12 last
  314. week. They ought to be squarely inside this week. We'll find out. We'll find out if they do.
  315. But Wake is looking at market probability on Wake Forest making the playoffs is like 77%.
  316. It's one of the great stories of the year so far. Okay, Cashew, and since you've, I know you've
  317. studied this as well. Do the conference championship games matter? Like I don't even know, for example, can
  318. Texas get to the SEC championship game? And if answer is if they lost, that would be a third loss for them.
  319. And so on the one hand, they got to the championship game, but on the other hand, they got a third loss.
  320. How do you see that? Or is that baked into these probabilities as well? The probability I just gave you is a
  321. market probability. That's a call sheet probability. Everybody's just speculating on the committee, of course.
  322. It's a great question, Eric. It's a debated question. It's a philosophical question to some level.
  323. People are like, how can it hurt you to get there? But then others are like, hey, it's a
  324. short season. We don't have a lot of information. We're going to get one more signal of
  325. the quality of your team. And so why wouldn't we use it? So it's an open, debated question.
  326. It has been used in both ways. I think it's been dismissed altogether, and it has been used against teams for sure that's
  327. happened. Texas, presently, something like the second most likely team in the SEC to make it, but
  328. there's a lot still to happen, like Alabama still in that mix, Tennessee's in that mix.
  329. One of the most interesting things that's happened is a few weeks ago, we thought SEC
  330. is trending towards five teams in the playoff, and now they've got three teams that seem
  331. pretty likely, and Tennessee is still fighting for its life. It's got a tough schedule in front of
  332. it. Think about LSU this weekend. They're home dogs against LSU. Four loss LSU is a road favorite against
  333. Tennessee. Tennessee's like number seven in the country or some crazy thing. So Tennessee, SEC still has some work to
  334. do to get a fourth team in the playoffs. There's been a real leveling of the playoff
  335. field. If you look at the call she odds, you can look at the call she odds and I stacked them up by conferences.
  336. And it looks like about 10 of the spots are, they're not locked up is a bit of a strong word a few weeks
  337. out, but 10 of the spots are pretty well accounted for. If you say somebody between BYU and Tech,
  338. Texas Tech is going to get the second big 12 spot. So you've really got three conferences buying for
  339. two spots. It's SEC's fourth spot, the big 10's third spot, and the ACC's second spot.
  340. I was going to ask you, this is the thing I don't understand about, I understand
  341. these are call she odds. Let's imagine Miami plays Wake Forest in the ACC championship game.
  342. Isn't that likely? Let's imagine they play each other in that game and Miami blows them out, which could
  343. also very well happen. Wake Forest is going? I think if Wake remains, they win until
  344. then and they lose, they've had a good season and people are going to be reluctant
  345. to let a conference championship game that they earned knock them out. But it's possible.
  346. I mean, you make it bad enough and it's possible. I don't think it's taken a zero, but
  347. it's an open debate, Eric. It's still an open debate. But it's kind of an interesting balance.
  348. It's like a two-thirds chance that SEC gets the fourth, a two-thirds chance that
  349. big 10 gets its third, and a two -thirds chance that ACC gets its second. So two of those are going to happen.
  350. Now the big 10, it's a motley group. It's Ohio State, a failing, quickly failing, dying
  351. Ohio State because they lost their quarterback. Iowa and Oregon, who's kind of rising from
  352. the ashes. So those are the three teams that are in the mix for a possible third spot
  353. from the big 10. I apologize, who's that group of 16? Boise. Boise's the leader, the 83% chance of
  354. making the playoff. Boise's been the leader there for a little while. I see. I think those are the most important bits
  355. to share with you. The power rankings still the thing that's happened with power rankings is that Georgia's finally snuck
  356. up a little bit. We've been, historically, especially late 2010s, the top power ranked teams were above 30.
  357. Above 30 points versus the average FBS team on a neutral feed. The best teams were reliably above 30.
  358. We haven't smelled 30, haven't sniffed 30 this year. And Georgia's sneaking up there now.
  359. Georgia's up there around 28, 29 and getting some breathing room between the next teams.
  360. And that hasn't happened. So they're really, I think this is, it matches the eye test.
  361. This is what they've been doing on the field. We'll see if they can hold it up. They've got a road trip against South Carolina.
  362. Do you see a world, Cade, where the top four teams don't come from, let's assume
  363. Notre Dame wins out. Let's say Notre Dame's going to be one of the top four, I would think for
  364. sure. But like, it's like, would there ever be a scenario, because I got rid of this,
  365. would there ever be a scenario where I'll make it up? Texas, as the second team from the SEC,
  366. is in the top four over the Big 12 champion. I don't think so. We have to see. I'm actually, at the moment, not remembering how
  367. the committee ranked teams last week. But, you know, like Utah's number two. I think Alabama might be number three.
  368. So Alabama in theory is ranked higher than Indiana right now. So the Big 10 may not get one
  369. of those top spots. It's possible. It's possible. It's possible. It's possible. I mean, the Big 10 just hasn't had
  370. as, Cignetti went on a rant this week about the schedule they've played. About how they beat teams that were good
  371. when they showed up and they haven't done as well since, but they were good when
  372. they showed up and he just went on a rant. But mostly they haven't played the schedule that
  373. justifies it. So right now, the deserving metric that we've been using, I've been claiming it for Word
  374. and Money, pal, by the way. I've been saying, this is our metric guys. Wins above bubble, wins above bubble.
  375. So we take the 12th best team and ask how they would have been expected to do against the team schedule.
  376. And the teams didn't get rated, number of wins, expected wins above or below how that
  377. bubble team would have done. Let me ask you a question, Cade. Let's imagine you use some blended combination of
  378. those three factors to determine who gets in. Would you be okay that once we've determined
  379. which 12 are in, that you rank them or seed them by one of those criterion? Like for example, power rankings after we've got
  380. the 12 in there. We're going to use a blend to determine who gets in, but once that's in, we
  381. got to give the best team the one, the next, the two, the next, the three, or do you not view it that way?
  382. I almost feel the opposite. The opposite. Okay. I'd like to use the metrics to determine
  383. the field and then let them see them however they want to. And that fits actually a little bit the
  384. way I think about drafting players as well. It's kind of silly to try to distinguish
  385. between one through 12 within a tier. You can kind of categorize with tiers, but
  386. then within tiers is kind of silly. So I'm happy to let them say, well, this player was hurt or this head to
  387. head. I'm happy for them to tweak it that way. You know, there's noise in our metrics, Eric.
  388. They're all jumbled and who's to say whether our metrics are exactly right. I think what really matters, I mean, I
  389. think you've said it the right way. I think at least in the 12 team playoffs, there's sort of three bins of four.
  390. I mean, the first four get a game off. Of course, it does sort of matter. The next four at least get home field
  391. in the next or in the next first round of games. And then there's the bottom four.
  392. So really what you mostly care about is which bin of four you're in. Is there any sense, do you think in
  393. the committee's ethos or psyche or whatever, that it's kind of almost a punitive thing that
  394. they take into account strength of schedule? Like they would want to weight it heavily
  395. to try and encourage, you know, I mean, I kind of want them to weight strength of schedule quite heavily because I don't like
  396. these teams setting up these cream puff schedules for themselves. So like, I don't know if there's a
  397. world where the committee almost is trying to kind of through the selection mechanism, try and
  398. change policy or change behavior. People talk about things like that. A related thing, Shane, very related, is will
  399. they punish Ohio State for going to Austin and losing against Texas? That kind of thing.
  400. It's just like they schedule the team. They played a tough one. Maybe they shouldn't take that big of a
  401. hit. So if they're the team with three losses at the end of the year, maybe don't.
  402. That's in the dialogue anyway. But I think first and foremost, they're just trying to, they do, I don't love it,
  403. but they use strength of schedule as an explicit criteria as opposed to something like strength
  404. of record, which translates strength of schedule and how you perform into something like deserving.
  405. Okay, guys, that's- Actually, just one last thing. I'm a fan of strength of wins.
  406. Just like when I read someone's tenure case, I care about their best papers. I want to know how great they can
  407. be. And to me, let's say someone goes in there and beats Georgia in Georgia. I understand they might have three or four
  408. other losses, but I mean, they did go to Georgia and beat Georgia. So I'm a strength of wins guy.
  409. I'm just saying for myself. All right. Certainly looking at transitioning to the NFL, I
  410. like to think about the Patriots season more about the Macs over their games than like,
  411. say, they're mean or like, you know, mean. That's really going to go well. Howdy, you were going to say something.
  412. I hate to- Yeah, I was just going to, this whole idea about great performance versus overall quality, this deserving resume.
  413. Now you're really getting into it with the college football. That's exactly the, every year at the end
  414. of the year in baseball, we always argue about awards and how do you dole them out.
  415. And in the NLL and the rookie league, rookie of the year vote, it was a major blow up because the writers voted for
  416. this guy, Sal Stewart, a young kid from Cincinnati, first baseman who got almost all the
  417. first place votes and just walked away with the award. But all the advanced metrics say he was
  418. the fifth or sixth best rookie in the league. And that's a pretty big gap. And almost all of it is attributed to
  419. the fact that he had 112 RBIs and really drove a shitty offense in Cincinnati to
  420. produce an incredible, basically, that's a big line for a rookie, for anyone in today's environment.
  421. And this is because he hit a lot of his home runs and doubles and singles when there was men on base and he
  422. made a difference. And the traditional stats are all just average. It's just, on average, you're going to do
  423. this. So one of my students has created a metric called RBIs above expected, and he had
  424. almost 30 RBIs above expected, which is really a lot. It's an enormous number. Well, but I thought you said the writers
  425. liked this guy and the analytics didn't. Yeah. So the writers liked it, but the analysts
  426. don't. But I mean, I don't think... So let me just finish saying that. But you're saying that's because the analytics aren't
  427. good. So the analytics is all average. They don't like clutch. Analysts have never liked clutch, never give you
  428. value for clutch, because it doesn't predict year to year. And so when you drop down your home
  429. runs, they're just assumed to happen where they happen. Yeah. And I mean, if the exercise here is
  430. to take the top five or six rookies that got votes and project forward who's actually
  431. going to be better in the future, it's not going to be Sal Stewart, but it's a reward for what they actually did.
  432. So I feel like the fact that he was unusually, not predictively, not prospectively, but retrospectively
  433. clutch should be rewarded. That's exactly my point. Who had the best rookie of the year
  434. season should factor in like, oh, and he got lucky. I mean, I think the luck should almost
  435. be built into kind of his evaluation when you're thinking about an award like this one.
  436. So what we added was a little bit of an extra, because most analysts just discount
  437. RBIs just completely. And because they're so contextual, right? If you come up in the right spots,
  438. you'll get RBIs. If you come up in bad spots, if you're batting late in the lineup or early
  439. in the lineup, you tend to come up without men on base. And so we have a way to sort
  440. of calculate that. So just, it's not complicated. And it just shows you it's, what's that?
  441. I knew you said what the metric is. So all it does is calculate how many RBIs, given your actual situations, how many RBIs
  442. an average player would produce in those situations. I see. Sal Stewart actually had a lot of those.
  443. So he was around 80. That's a lot, given most players do not expect 80 RBIs if you're average.
  444. But he was in a, he got up - It does sound like he was unexpectedly clutch. But he ended up out-beating that by
  445. over 30, which is really extreme. He should have had, and all the other players who are better at fielding and better
  446. at on base and better at the things, and that ended up with wars that were twice as big as his.
  447. Would your metric be correlated? Are better going forward or did better? So no, they have, last year, their wars
  448. are higher, right? Their actual war are substantially higher than Sal Stewart. That's last year, but war is a retrospective
  449. measure, but it removes, it keeps what your actual statistics are, doesn't shrink those down, but
  450. it doesn't, it removes context and it doesn't reward context. Slotty, would your metric of RBIs above expected,
  451. would it be highly correlated with slugging percentage with men on base? Absolutely.
  452. It's incredibly highly correlated with that. It's actually highly correlated with on base time
  453. slugging, which is actually a slightly better statistic. No one likes it because nobody likes to
  454. multiply when you can add, but there's something very interesting and peculiar, which is that there's
  455. a compounding effect of home runs. Home runs are averaged about one point, I think it's like 1.4 runs per home
  456. run is what the weighted on base percentage would evaluate them. But if you're batting third or fourth in
  457. the lineup, home runs have more than that value. And when you bat first or ninth in
  458. the lineup, home runs have much less than that value. And so actually being able to be extreme
  459. in home runs and then being used in the third spot has a compounding effect on production that these metrics don't actually value it.
  460. That's my point. It relates to college football insofar as you see, who do you give the award to?
  461. The guy who's going to be the best next year, who's currently, who currently we think
  462. is the best, or do we give it to the person, the team that's just done it on when it needs to get done?
  463. Like, you know, Arch Manning. And that's, that's, well, that's back, that's wins
  464. above bubble. That's exactly wins above bubble. That's exactly the metric that we built.
  465. And they all are using the same language that we apply, which is just how many wins more do they get than you'd have
  466. expected? If say the 12th best team played it, which is the bubble team. And we, and it's added, it's added to
  467. your body. Just like you said, we don't want to multiply if you can add, it's just wins
  468. added up. Now it's a small schedule. It's, you know, it's 12 games. I mean, that's really noisy.
  469. They move around, but that's the best way I have to play with. Shane, do you want to talk about NFL
  470. for a minute? Your, your boys, my boys, not so pretty, but I don't actually really want to talk
  471. about them, but I've been thinking about them. Let's, let's look for a second. We've got NFL calls.
  472. You talking about the Jets? Anybody? They're not on the picture either. The Jets and Patriots odds to go to
  473. the Superbowl. I've actually equalized a lot in the last week. So not, not, not really any movement on
  474. the Jets part, but. Guys, I mean, I threw this up here. It's, it's call sheet odds.
  475. This is basically mid-season just after mid -season by team organized by division conference and
  476. division. And we have really spread evenly. Like each conference has just about 50%
  477. within each conference divisions are fairly even spread. There's only two divisions in the NFC that
  478. are relevant, but they're spread equally between North and West. And then all four in the AFC are
  479. pretty, is that unusual? There's about 13 teams still in the mix above 1%. It just feels more parody than you.
  480. Yeah. I kind of feel like it's either there, there are no really dominant teams or there
  481. are. A lot of really, you know, there are so many really dominant teams that we don't
  482. really think of them as dominant anymore. I mean, I think I saw somewhere that there's, at least in the NFC, there's an
  483. unusual number of kind of like seven win team. You know, there's like three or four, seven
  484. win teams, like, you know, at this point in the season after only like nine or 10 games.
  485. That's, I feel like we've got, yeah, like I said, either, either no team that's really
  486. standing out from the pack or just like a really a pack of what we think look like really good teams.
  487. I think the two, the things that stand out, and I love this chart, by the way, I think the thing that stands out
  488. to me about the chart is first of all, the Rams were the heavy preseason favorite
  489. to win the Superbowl. And right now this chart has the 49ers almost three and a half times as likely
  490. to win the Superbowl as the Rams, which seems way over bloated to me. I mean, the Rams are six and four.
  491. It's not like they can't make the playoffs. Matter of fact, they're targeted to make the
  492. playoffs. And also the Superbowl champions who are the Seahawks are, I feel like the 49ers are
  493. four times as likely as the Superbowl champions. That seems strange to me. The other team that seems low-ish to
  494. me is the Detroit Lions. You know, they're sitting there at 4%. The Bears are three times as likely almost
  495. as the, as the Lions are. Bears haven't won anything in the playoffs. Caleb Williams has never won anything in the
  496. playoffs so far. So what the hell is it? Why? And also I think they all have the
  497. same record. I think the Bears, Lions, and Vikings are also- But you know, we're still at
  498. the point in NFL season where you can definitely argue strength of schedule. I don't know if that's what's being factored
  499. in here necessarily. But you know, those are examples or like, you know, people kind of just sort of
  500. see the Bears as kind of fundamentally a better team at the, you know, offensive end.
  501. I don't know. I don't know. Okay, two quarterback developments in the last couple
  502. weeks that feel like they've changed the field. One, Cousins to the Vikings with that incredible
  503. defense to watch. It's so much fun to watch that team these days. And then Burrow, Burrow seems like he's peak
  504. Burrow form. And he's got a couple of phenomenal targets and the Bengals are making noise.
  505. I mean, I think it's the most interesting year of a long time. And my team is in the doghouse.
  506. But I think it's a super interesting year across the NFL. Yeah, my team's done too.
  507. But I'm thinking of picking up the Jags, the Jaguars. They're kind of a fun team to play.
  508. You know, I think they, I'm all for the Jaguars. Trevor Lawrence, he's been creeping his way up
  509. the ranks of quarterbacks in all the years. He's been slow rise. I'm going to take my infinite odds on
  510. the NFC South, which has apparently a zero probability to win that division. And what the heck?
  511. Why not? We can do it. Zero on the division. Why don't you pick up the NFC East while you're at it, Eric?
  512. All right, fellas. Why don't we wrap it up there? Many thanks to our listeners. Our thanks to Jacob Grodnick and the Grodd
  513. crew for Grodd squad, the Grodd squad running production. Deep Patel, the boss lady, our guest, and
  514. you guys for listening. Come back and join us next time between now and then. Enjoy your sports.

Brief Summary

Zach Binney explains why NFL injuries are hard to predict, what player and turf data reveal, and whether biomechanics can improve forecasts.

Key points

  • Why injury prediction is so hard. Injury prediction is the holy grail, but even with years of data, researchers struggle to tell who will get hurt. Random variation makes it difficult to turn observed patterns into reliable forecasts.
  • Not all artificial turf is alike. Research found that slit-film turf is associated with greater injury risk, while the remaining artificial turf types show much less difference from natural grass. The NFL has since moved away from slit-film turf.
  • Players matter more than teams. The median injury-risk difference between players is much larger than the difference between team-head-coach combinations. That leaves researchers asking which individual traits actually explain the variation.
  • Could a pitcher's mechanics warn of injury?. Binney and collaborators are exploring whether changes across a pitcher's recent fastballs could signal an elbow or shoulder injury during a game. Biomechanics may hold promise, but the data are difficult to interpret.
  • The NFL team’s first injury question. Binney expected to build a fancy predictive model. Instead, the team asked a much simpler question: how often do running backs get hurt?
  • Why injuries spike in training camp. Injury rates stay flat over the season, but the highest rates come at the start of training camp, when players are building back up after the offseason.

Episode highlights

Tension

18/ 100Calm

Related episodes