Sunday, March 24, 2013

On the injury rates of running QBs: The future is now

The NFL, we are told at length, is on the brink of a new era of dual-threat quarterbacks. These new QBs will take advantage of their rushing ability to put the defense off-balance and change the way the game is played. Stop me if you have heard this before.

The dual-threat QB is hardly a new phenomenon. Quarterbacks were, of course, originally more like running backs in the early days of football. While my command of football history is far short of sufficient for this or any other analysis, such names as Fran Tarkenton, Randall Cunningham, Donovan McNabb, Michael Vick and Steve Young have demonstrated the power of a QB with the ability to run the ball.

That a dual-threat QB could offer an improvement over a traditional drop back passer should not be in serious dispute. Given two QBs of equal passing ability, the one of the two with greater rushing ability will almost certainly have better performance – in passing as well as in rushing. The statuesque, lumbering traditional QB will have far fewer scrambles for first down, more sacks (assuming equal ability and judgment to throw the ball away in a hopeless situation) and more marginal throws forced in.

The dual-threat QB will have a huge impact even before the snap. The defense will have to consider him a threat to run even on a play that looks like a certain pass, pulling one defender away from the passing game and opening up better opportunities for completion. Once the play begins, the dual-threat QB will have the benefit of an additional option should the receivers look covered. Rather than forcing a marginal throw, throwing the ball away or taking a sack, this QB can make something happen on his own. Again assuming equal judgment and passing ability, the dual-threat QB will trade some of the worst throws of his almost-doppelganger for some scrambles that are far less likely to be turnovers and far more likely to yield positive yardage.

Despite all this upside the dual-threat QB seems prone to injury concerns that will keep him out of reach. Redskins fans in particular may be having second thoughts about Robert Griffin III tallying 120 runs – tied for the league lead among quarterbacks.

Shortened Careers

The biggest problem with a QB who can run seems to be that he may end up running. It is well-known that running backs have one of the shortest average careers of any position. The short careers are assumed to be due to the pounding that they take.

The 119 RBs drafted between 1994 and 1999 who actually made a roster for at least a season averaged 5.4 seasons in the league. The 51 QBs who meet the same criteria averaged more than a full season more at 6.6. The difference is even starker when looking at starts. Those 51 QBs averaged 40.1 starts (and Peyton Manning still going) while the 119 RBs averaged almost a full season less with 27.6 starts. QBs have 50% more starts than RBs and it seems likely that injuries account for a significant portion of that difference – if anything I would suspect that QBs are more likely to be removed for sub-par performance because of the visibility of their mistakes but I can’t credibly claim any statistical support here[1].

QB Running

At this point all I’ve shown is that QBs have longer careers than RBs. As I mentioned above: this is well-known. Let’s take a look within the QB population to see if we can detect some differences between dual-threat QBs and their traditional counterparts.

Figure 1: 1993-2011 Seasons, QBs having started at least 50% of their games

Figure 2: 1993-2011 Seasons, QBs having started at least 50% of their games

Quarterbacks with more rushing attempts tend to play more games and have a higher quarterback rating that those with fewer. To get a little more clarity on where that higher rating comes from, we can compare QB rating to passing attempts per game and rushing attempts per game.

Although passing attempts and rushing attempts do not explain a great deal of the passer rating distribution (R^2 = 0.13), they are both significantly positively correlated. The 95% interval for marginal rushing attempts is a 0.82 to 2.25 improvement in QB rating. The 95% interval for marginal passing attempts is 0.70 to 1.03. It makes sense that both would be positively related – better QBs get more attempts, though the attempts themselves aren’t improving the rating – but it is important to note that even controlling for passing attempts, rushing attempts have a strongly positive relationship. This supports the hypothesis that the ability to rush improves a quarterback’s passing.

So dual-threat QBs appear to have a higher QB rating, and there is evidence that the rushing attempts are related to higher QB rating independent of passing attempts. Despite this, the highest bucket of rushing attempts, 4+ per game, is associated with a larger decline in games played from year to year than we might expect.

*Note – These graphs are still for quarterbacks with at least 50% of their games as starts in Yr 0

The 1-2 and 2-3 buckets trend very close to the average QB – 2.09 fewer games from one year to the next. These two buckets also contain 308 and 264 player seasons, respectively, out of 842 total. Since we saw earlier that QBs with more rushing attempts are likely to have higher passer ratings, it is odd that they would see such a large negative change in games played absent some other factor.

The major components driving the change in games played are likely to be age, performance (as captured by QB rating) and injury. Obviously it is injury that we are looking for here with rushing attempts per game standing in as a proxy.

Looking at year to year regressions is a fairly nebulous business. A lot can happen to a player between two seasons that would affect playing time more than injury from running the ball (e.g., riding a motorcycle like an idiot or a similar, but distinct, idiot).

One Layer Deeper

Looking at week to week data, on the other hand, may just be granular enough for us to separate some factors and see if we can separate the impact of rushing attempts – or more broadly, being hit – from other factors causing injuries.

For this data we can look no further than once again, retrieving the game lines for all players who attempted at least one pass in a given game. PFR also has the weekly injury reports for each team. The NFL standardizes injury reporting into the categories of Probable (25% chance to miss game), Questionable (50%), Doubtful (75%) and Out/Injured Reserve/Physically Unable to Perform (100%). While individual teams may vary in application, what we are really looking for is any week where there is an increase. If a QB moves from Doubtful to Out from weeks 6 to 7, that will show up in the data as week 6’s events having increased his level of injury by 25%. On the other hand, if a team puts a player at Probable every week it won’t affect the data because there is no change.

After pulling this for the 2011-12[2] seasons and adding in number of times a QB attempts a rush (subtracting rushing TDs), is sacked or makes a tackle (typically after an interception), we are all set for the big reveal: How much more likely are QBs to be injured for each additional hit?

Not much at all, or maybe a negative amount. Wait, what?

Comparing increase in percent chance of missing a game against rushing attempts, sacks taken and tackles made (all for current week) yields an overall correlation of 0.0206 with an R^2 of 0.00043 – not a good fit. The details are below:

Coefficients Standard Error t Stat P-value Lower 95% Upper 95%
Intercept 0.04579 0.00793 5.77793 0.00000 0.03024 0.06134
Ratt-TD -0.00093 0.00179 -0.51813 0.60446 -0.00443 0.00258
Tkl 0.00179 0.02522 0.07087 0.94351 -0.04769 0.05126
Sacks 0.00139 0.00266 0.52097 0.60249 -0.00383 0.00661

Not only is it not a good fit, but it looks like rushing attempts actually make a QB (very) slightly less likely to be injured in a given week.

If we instead compare these factors and add a tracker for cumulative sacks or rushing attempts – still terrible[3]. The overall correlation improves slightly to 0.0471 with an R^2 of 0.00222.

Coefficients Standard Error t Stat P-value Lower 95% Upper 95%
Intercept 0.040576 0.009017 4.499833 0.000007 0.022884 0.058268
Ratt-TD -0.000045 0.002009 -0.022573 0.981994 -0.003986 0.003895
Tkl 0.002046 0.025219 0.081132 0.935351 -0.047434 0.051527
Sacks 0.001118 0.002678 0.417384 0.676475 -0.004137 0.006373
Cratt -0.000479 0.000408 -1.174507 0.240432 -0.001279 0.000321
Csacks 0.000913 0.000638 1.431200 0.152641 -0.000339 0.002165

The sacks and rushing attempts are much less interesting when we control for cumulative rushing attempts and cumulative sacks. Interestingly, the cumulative rushing attempts show a negative correlation with injury while the rushing attempts themselves flip from negative to positive, albeit at very low significance. Both individual game and cumulative season sacks have a positive relationship.

To put the size of the relationship in perspective a player with 100 rushing attempts entering a game would have his chance of injury reduced by 4.79% while a player with 40 sacks would see his chance of injury increased by 3.65%. These totals are near the league leaders at the end of the season so you can see how small the overall impact is.


Based on the entirety of the data, I have a hard time concluding that running increases the chance of injury for a quarterback. Sacks, on the other hand, are slightly stronger in their relationship to injury and are positively correlated. I could probably back into a story about quarterbacks who are sacked being unprepared while runners brace for a hit, but the weakness of the relationship and the minimal amount of injuries explained speak to other, unobservable factors at work such as a player’s predisposition to injury or the high variance of types of hit ranging from Deion Sanders to James Harrison.

These findings conform to the conclusion of this analysis by Omar Bashir and Chis Oates that was featured on Slate. Their analysis found running quarterbacks – under several different filters – appear slightly less likely to miss games due to injury. I realize this is a long way to go to confirm that another analysis appears sound, but the methods are slightly different so I hope that this is a useful contribution to research on the topic.

Autopsy of Biases

I’ll admit that I went into this one with a pretty strong conviction that there would be a positive and significant relationship between number of hits and injury. In doing so, I have fallen prey to the availability heuristic in a big way. I can think of lots of examples of running quarterbacks getting injured (Michael Vick! Robert Griffin III!) but very few of the circumstances when traditional quarterbacks went down. I was so certain in my intuition that I drafted a 500 word analogy between playing a dual-threat QB and drafting a player with known injury history (sadly, this has since been deleted)

As the NFL reacts to a Super Bowl in which a traditional QB stood statuesque in the pocket and won while the dual-threat QB could not come through, I wonder whether some analysis now is falling into the confirmation bias. Talking heads saying “I told you so” about the ability of running QBs to win are revealing their own going-in position. Dual-threat QBs have a number of advantages over similarly skilled passers without the same rushing ability and while this doesn’t mean the Broncos should have Peyton Manning running the zone-read, it seems like something that should be considered by the many teams looking for their future QB every winter.

[1] Some advanced stats might be able to answer this question where both positions get an all-in-one ranking that is ideally comparable across positions, but at a minimum allows you to see whether those players ranked lowest relative to their peers are more likely to be benched at either position. That’s not what this post is looking at so I’ll leave it there for now.

[2] has available week to week injury reports that, once collected over several more seasons, should allow a better look.

[3] I looked into a few regressions with additional variables but ran up against some collinearity between variables. The cumulative sacks and cumulative games terms have a correlation of 0.90 so we need to be wary including them in a model together. Even though the model appears better with them (correlation of 0.084) this is probably down to overfitting more than true correlation.

No comments:

Post a Comment