Untangling the NFL Injury Web
by Zach Binney
In a sport as chaotic and violent as the NFL, the goal of predicting which players are going to get injured is both a Holy Grail and, to date, wholly unattainable with public data. While we have written many articles that identify risk factors for injuries -- weight, prior injury history, position and age, to name a few -- we have not written about a model that allows us to successfully predict which individual players will stay healthy or get hurt. That's partly because if we had such a model it would be too valuable to share freely, but also because our efforts to date have come up short.
Prediction is difficult because getting hurt in the NFL is the product of a complex web of risk factors. Here's a simplified version of what one such web might look like for an ACL tear:
There are player-level factors that contribute to injuries, such as their age and injury history. But there are also team-level factors, such as the amount of training players are subjected to and the related fatigue players experience. Stadium-level factors, such as weather and the type of turf, also play a role. There's also a role for plain old bad luck, though more optimistic analysts might call luck "factors to be named later."
To begin teasing apart this web, Ron Yurko (a PhD student at Carnegie Mellon and co-author of a computer program for scraping and analyzing NFL play-by-play data) and I wanted to look at how much player-, team-, and stadium-level factors contribute to NFL injury risk.
We used data from the Football Outsiders injury database on 3,694 players and 3,874 game-loss injuries from the 2012 to 2016 regular seasons. A game-loss injury is an injury that causes the player to miss one or more future games. We looked at how strongly each level (player, team, and stadium) impacted a player's risk of suffering such an injury each week. Because "team-level" factors are partly driven by the franchise (owner) and partly by the coaching staff, we looked at how each team/head coach combination impacted injury risk. While the head coach is only a proxy for the team's full staff (for example, the team physicians often carry over from coach to coach), historical data on medical and training staff is harder to find. Due to the uncertainty in a single NFL game, we also generated 1,000 simulations of the model to get a range of effects for player, team/coach, and stadium factors. We're going to gloss over the other technical details of the model, but they are available here if you're curious.
Before we get to the results, a quick word about why you should care: if we can show that coaching staffs and stadiums impact injury risk, we can identify which ones are safest and mimic their practices (for coaches) or designs and maintenance procedures (for stadiums). This may also help us identify new risk factors for injury we had not considered before.
What We Found
Player-level factors -- such as age, injury history, and playing style -- are by far the most important in determining a player's injury risk. The orange curve in Figure 2 shows this: on median, a higher-risk player (for example, Arian Foster) has nearly three times the odds of injury of a lower-risk player (for example, Frank Gore). "On median" simply means that half the time the difference is bigger than that, and half the time it is smaller.
Compare this with team/head coach-level factors (the blue curve in Figure 2), which might include practice load and intensity, roster construction strategies (more young players vs. veterans), in-game player usage, and so forth. Here, a player moving from a low-risk team/coach combination to a high-risk combination can expect, on median, a 21 percent increase in his odds of injury. If I were a player (or owner), that would be a big enough effect to worry me.
Stadium-level effects from things like turf and climate are a bit smaller but still substantial: on median, a player playing in a higher-risk stadium faces a 13 percent increase in his odds of getting hurt. If you're an NFL owner, is a 13 percent increase in injuries enough to make you think carefully about your stadium design and turf?
Let's look at each of these levels in a bit more detail. Here are the five highest- and lowest-risk players at each position:
There's a lot to unpack here, but we can point out a few things. The safest quarterback and running back are Eli Manning and Frank Gore, respectively, whose ironman reputations precede them. Among tight ends, Greg Olsen is considered the second-safest player. Keep in mind we only used data from 2012 to 2016 in this model, a period during which Olsen was as dependable as they come. The odds of each of these players getting injured is 50 percent lower than the average player. Obviously, anything that you can do to be more like these players in terms of health, you should do. But that's not especially illuminating.
Instead, let's move on to teams and coaches:
A few interesting things jump out here. First, Philadelphia and Miami each have two different coaches among the ten safest. That suggests something smart being done at the franchise level that transcends specific coaches. That could be the product of strong investments in sports science and analytics, careful monitoring of player workload, or any number of other factors. If you're looking to keep your team healthier, though, I'd start by talking to the hard-working folks of the Eagles and Dolphins. Playing for one of these low-risk combinations is associated with a 20 percent reduction in your odds of injury versus an average team and coach.
Then we have John Fox. Fox was among the safest coaches to play for in Denver, but one of the most dangerous in Chicago. The reason for this is unclear. It's possible some of Fox's habits that impact a player's injury risk changed drastically between Denver and Chicago, but there are other non-Fox-centric explanations: perhaps Chicago's roster carried a higher injury risk than Denver's, or there were differences in the effectiveness of the medical and training staffs or injury prevention efforts of the two franchises. It would be interesting to talk with Fox to get his thoughts.
Finally, Jeff Fisher was the third-safest coach in our model. So he's got that going for him.
Lastly, let's look at which stadiums are the safest and most dangerous:
Playing a game in a safer stadium, such as San Francisco, versus a more dangerous one, such as Seattle, is worth about a 20 percent decrease in your odds of injury. Turf type may be the single biggest factor at play here; it accounts for about 30 percent of the variation in injury risk between stadiums.
The curves above are colored by stadium turf. The first thing you might notice is that stadiums with natural grass (green curves) tend to be safer than those with artificial turf (other colors). The 10 safest stadiums to play in all use natural grass, while the six most dangerous feature some form of artificial turf. One particular brand, FieldTurf, was installed exclusively in five NFL stadiums from 2012 to 2016; three of them (Seattle, Indianapolis, and Detroit) were among the eight most dangerous to play in. Turf type isn't the whole story, however. How that turf, including artificial turf, is maintained is at least as important. This is all consistent with an earlier analysis of turf and injuries here at FO.
Pittsburgh and Washington are considered by many to be two of the worst surfaces to play on in the NFL, but they are two of the three safest stadiums in our model. Despite being muddier these fields may actually be softer, lessening the forces on players' knees and ankles. That is just a guess without additional data, though.
San Francisco, meanwhile, has maintained a low injury risk across two different stadiums. Other teams might be wise to learn more about their grounds crew and maintenance procedures, though their climate helps, as well.
We noted above that about 30 percent of the variation across stadiums is due to turf type. The other 70 percent is due to other factors such as climate and, potentially, how risky it is to play the stadium's home team.
Is Seattle's stadium so dangerous not because of the turf or the weather, but because of the Legion of Boom (remember, the data are from 2012 to 2016)? The answer is a resounding "probably not." Here are the risks of offensive injuries when each opposing defense is on the road, to separate stadium from defensive effects:
Seattle ranks in bottom third of "injurious" defenses, so the Legion of Boom does not seem to be why Seattle's stadium looks bad. The defense that saw the most offensive injuries across the field was Washington, whose stadium ranked as the safest in our model. Miami and Philadelphia were the next most injurious defenses; their stadiums ranked as about average and the fifth-most dangerous in our model. Overall there does not seem to be a strong correlation between injurious defenses and dangerous stadiums.
NFL injuries are complicated and difficult to predict. We've shown that a player's team, coach, and even the stadium he's playing in combine with his own injury proneness and a healthy dose of luck to determine whether he makes it through the week. Hopefully we have also identified some coaches and stadiums that may be worth mirroring to reduce injury risk. After all, the NFL is always more fun with Aaron Rodgers, Allen Robinson, and Deshaun Watson than without them.
21 comments, Last at 10 Oct 2018, 1:04pm
#1 by Aaron Brooks G… // Oct 03, 2018 - 1:00pm
Any concern that you have more variables than data and are overfitting your model?
\someone needs to explain how Oakland and Candlestick were safe fields with the surface transitions they had.
\\Or ask RGIII's knee how safe Washington's grass was
#2 by Zach Binney // Oct 03, 2018 - 1:38pm
Overfitting: short answer, no. I'll spare you a deep dive into the math behind mixed models, but basically to get every player, team-coach, and stadium effect we only had to estimate 3 parameters. So stack those against nearly 4,000 injuries and over 114,000 player-games, and overfitting isn't a big worry.
Oakland being that low is a surprise, I'll admit. Keep in mind this is only 2012-2016 data, so Candlestick shouldn't have been hosting baseball for years.
As for RGIII, respectfully, I'll trust our model on 114,000+ player-games versus the results of a few cherry-picked bad cases. :)
#4 by Aaron Brooks G… // Oct 03, 2018 - 2:33pm
My understanding was that there wasn't general agreement on how to count degrees of freedom for mixed-models.
Some seem to account it using parameters, some seem to account it using n-2 (k=1), some perform some strange intermediate voodoo.
I'm curious how much movement there was between teams. How many players actually changed teams during this period?
#10 by Zach Binney // Oct 03, 2018 - 5:41pm
1,307 players (35%) logged time with two or more franchises during this time (48% with multiple team-HC combinations).
Even if no players moved, though, keep in mind these mixed models are still able to tease out player vs. team effects by nesting players within teams (and stadiums). We compare the individual variance in risks within the players across different games/seasons, and within teams across different players.
#7 by Jim C. // Oct 03, 2018 - 4:44pm
As you look at Figure 5, it might be helpful to remember that in 2016 (the end of the sample period), Baltimore replaced its atrocious Sportexe Momentum Turf with natural grass.
#13 by Zach Binney // Oct 04, 2018 - 7:46am
Assuming the above comment isn't satire, which it may be...
I'm debating whether or not to engage with this comment because I don't want to feed the trolls, but I'm going to try and twist it positively as a teachable moment.
To judge science, probably the first thing you need to understand is what question the work is trying to answer. Nowhere in this article do I claim this is a model designed or useful for predicting injuries. Indeed, I state that prediction is a.) incredibly difficult and b.) something I and others have failed at time and again. This model was designed to help better understand how much factors at different levels contribute to injuries, as well as some coaches and stadiums we may want to emulate to reduce injuries in the future. Hopefully the model's framework and the results here will help us build better predictive models in the future. But building such a model was not the point of THIS model or THIS study.
If you want to say the model isn't USEFUL in your eyes because it doesn't predict anything, I disagree but that's a valid point. But to call it "truly terrible science" - and imply that any study without an out of sample test is such, regardless of whether such a test is even appropriate for the research question at hand - displays a reckless and profound misunderstanding of what science even is.
#16 by Aaron Brooks G… // Oct 05, 2018 - 9:07am
First, a question -- the injury prone players have very wide odd ratio distributions. Also, for the most injury prone, the distributions tend to be multi-modal. What's going on what with? What is odd ratio actually capturing?
Basically, how are you quantifying "injured"? Is it season-ending, binary game lost (yes/no), count of games lost per season, etc?
What effects do you think are captured in a given variable?
For player, you have their individual tolerance, their play style, their work-environment (teammates, offensive/defensive system, stadium exposure, climate exposure, team player usage, and training/medical staff), and their person work history.
Coach also captures team, stadium exposure, climate, and training/medical staff.
Stadium captures a team-biased exposure (Meadowlands gets two teams; the Chargers don't count as a professional organization, so LA only has one), climate, and turf construction.
It's not surprise player has the tightest distribution. Because we're looking at per-player outcomes, individual players should always be their own best estimator. Coach is filtered across 53+ players and stadium across hundreds. It's no surprise those are noisier distributions. It would be shocking if they were not.
#17 by Zach Binney // Oct 05, 2018 - 11:54am
Great questions. Thanks for being so engaged with this piece.
Weird player distributions - to be honest, Ron and I aren't entirely sure what's going on there and wouldn't encourage a ton of faith in those. It may be a quirk of our bootstrapping procedure. That's why I didn't really zero in on those. We have a lot more confidence in the other results.
As described in the article, "injured" is defined objectively as any injury that causes a player to miss one or more games. This controls for reporting differences across teams. Each observation is a player-week, so each player was either injured or not injured (or not exposed due to not playing) each week.
What effects are captured in a given variable - another great question that I just didn't have the space to get into in the article. I think you've actually done a really good job describing them in your comment! I just have a few additional thoughts/corrections:
Player: this *should* be separate from contextual team-environment effects - that's the beauty of these multi-level models. Imagine yourself with your personal habits, diet, wealth, etc.; and the neighborhood you live in with access or lack thereof to sidewalks, healthy food, etc. They both have effects on your health, but multi-level models with a random effect for you and for your neighborhood separate those out. So here we're really looking at things like playing style, workload history, injury history, genetics, age, and so on.
Team-coach: this is those team-environment variables you mentioned above along with what you listed.
Stadium: pretty much agree with what you wrote but, again, these effects should be *mostly* separate from team-coach effects since our model separately controls for that. That's the whole point.
Lastly, the player effects are actually wider than the coach and team effects, indicating that who you are is more important than for whom or where you play. SEE Figure 2. The different scales in Figures 3-5 may be, admittedly, confusing there.
#18 by Aaron Brooks G… // Oct 05, 2018 - 1:29pm
I was thinking of figure 2, but I admit I am confused by what the changing heights of the y-axis are conveying. I think it's just the tightness of distribution, but the staggered plot level is a new touch.
Some of what I was thinking of as player is separately binned out, but my argument is that it's incompletely captured at the stadium-team-coach level. If you had put Brett Favre on the Cleveland Browns, does he succumb to a staph infection? Does the presence of Favre lower the infection rate? Does it drive up local copper prices?
But also: what if you play a style that increases injury risk for one position group, but lowers it for another? Mike Martz ran a system that probably measurably reduced the life expectancy of his QBs, but the vertical style reduced the number of killshots likely to be delivered to his WRs. That would probably wash at the team/stadium level and appear as an individual hazard, but was really a team-level effect.
Or consider Indianapolis. Was it safer being a Colts lineman when they had Manning versus Luck, and they didn't have to block as long? Was that a team effect, an individual effect, or a consequence of roster construction? That's what I was talking about.
#21 by Zach Binney // Oct 10, 2018 - 1:04pm
So Figure 2 shows two different things related to the point I think you're making: player effects are on average bigger than stadium or team-coach effects, *and* they're more heterogeneous (spread out).
"Some of what I was thinking of as player is separately binned out, but my argument is that it's incompletely captured at the stadium-team-coach level. If you had put Brett Favre on the Cleveland Browns, does he succumb to a staph infection? Does the presence of Favre lower the infection rate? Does it drive up local copper prices?" - My understanding (and others can chime in if I'm wrong) would be it's both. These are *independent* effects, each controlling for the other. Probably the more sensible way to think about it is the former, though - Favre would be less safe on the Browns if the Browns are unsafe because of organization-level issues that make players less safe when they play for them. The Browns could also be thought of as being safer with Favre, though, if Favre were a low injury risk (which would surprise me, but I'm just going with your example). What players a team chooses to sign is *part* of the team-coach effects we're capturing here, along with practice and training regimens, medical staff quality, etc. It's really complicated.
As to your last couple paragraphs: huh, I hadn't thought about effect modification of team-coaches by position (epidemiology-speak for some team-coaches might be safer for some positions but not for others). That's something our model could account for in the future. I agree it's a possibility.
#15 by panthersnbraves // Oct 04, 2018 - 3:50pm
Did this only take into account injuries that took place on the main playing field? Kelvin Benjamin blew out his knee on the Panthers practice field. iirc, there was also a Miami Dolphins' player who was also injured that same day as well, during the joint practice...
nvm - this said regular season.
#19 by JimSteeg // Oct 07, 2018 - 12:29pm
This is amazing research that should serve as a tool to analyze injury trends. What is missing is practice field and training camp fields. Those that use indoor bubbles and focus on the grass or artificial fields is where players spend more time than on the game field. One issue is that the Rams and Chargers wanted their game fields to deteriorate as part of their effort to move.
#20 by Zach Binney // Oct 10, 2018 - 12:55pm
This is a very fair point. One major limitation I didn't note in the article is we don't know when each injury occurred (in-game or in practice), though we do know from prior research that 80% or more of regular season football injuries occur in games rather than practices. We also don't have data on practice field surfaces, so the analysis really makes the assumption that 100% (rather than upwards of 80%) of injuries occurred on the playing fields. Not ideal, but it's the data we have available historically at this time.