Peyton Manning signed a $90 million contract extension that would hypothetically keep him playing through age 40. Nagging questions about his health will almost certainly plague him for the remaining years of his career. To back up the ailing Manning, the Colts brought in an even older veteran passer, Kerry Collins. In Philadelphia, Michael Vick signed a $100 million contract that will supposedly keep him playing until he's 37. What kind of performance should we expect from older passers?
It's a much more difficult question than it first seems. Averaging the performance of all the recent QBs by age doesn't work. A survivor bias ensures that only the successful QBs stay in the league long enough to have their stats in the sample. Another complication is the the steady inflation of passing stats over the years. In this post, I'll try to tackle those problems to better understand how QB performance is affected by age.
For this analysis, I used Adjusted Net Yards Per Attempt. (Other stats such as EPA or WPA don't go back far in enough in history for a usable sample size.) Adjusted Net YPA starts with simple yards per attempt numbers. 'Net' means that sack yardage is included and sacks go in the denominator as pass plays. 'Adjusted' means that a 45-yard penalty is included for each interception. And because the data came from Pro-Football-Reference.com, a bonus of 10 yards is included for each TD pass. I only included seasons where the QB had at least 10 starts in the regular season, and I only included QBs who were drafted in 1978 or later, which is considered the modern passing era in the NFL.
First, let's address the inflation of passing stats over the years. The graph below plots the steady increase in all QB AYPA over the years. This trend has the effect of masking a QB's true decline. A hypothetical QB who throws for 5.0 AYPA every year of his career is actually declining with respect to the rest of the league. To correct for this trend, just like economists adjust for inflation, I chose a reference year and added or subtracted the appropriate amount for each QB's stats in each year. In this case, I chose the approximate mid-point of the period, 1990.
The problem of survivor bias is tougher. Each QB will have his own unique career pattern. Some are quick studies, who learn from their early experience, while others are slow on the take. Some rely on their foot-speed, which declines more with age than other physical attributes, while other QBs rely on their accuracy and decision-making. Some QBs seem to have supernatural, age-defying abilities to heal, while other QBs have bodies that break down quickly. Survivor bias means that we'll only see QBs who perform well enough to remain in the league.
So if we simply average AYPA by year, we get the following graph:
It appears that QBs don't decline with age at all! In fact, they just keep getting better and better! We know this is survivor bias at work, so we need a different way to look at age and performance.
One way to begin is to fit the analysis to the specific question asked. In this case, I wonder what kind of decline can we expect from Peyton Manning as he continues to age? He's had some incredible seasons, and how will his remaining seasons compare to those? So I looked at QBs like Manning, who already have proven themselves by starting several years, specifically 10 or more seasons. I averaged each QB's decline from his peak years, defined as the average of his three top seasons, at each age. For example, Chris Chandler peaked at 7.0 AYPA at age 27, 30 and 31. At age 32 he put up 5.5 AYPA, which is a decline from his peak of 1.5 AYPA. Averaging franchise QBs' improvements and declines with respect to their own peak years helps overcome the selection bias, and produces the following aging pattern.
Here we see a pattern consistent with what we'd expect. Young QBs gain from experience and possibly from physical development, until a peak age at which the effects of age overcome the added experience. For established 10+ year starters, the peak is at about age 29, and the decline appears relatively shallow, with performance dropping about 0.5 AYPA over a 7 or 8 year period.
A similar approach is called the delta method. This method looks at consecutive year-pairs, and it measures the improvement and decline from one season to the next. By averaging all the QBs' change in performance by age, we can see the effect of age without the survivor bias. For established 10+ year starters ("franchise QBs"), the results are fairly linear:
The graph above can be read like this: Values above zero indicate improvement, and values below zero indicate decline. For example, a QB aging from 21 to 22 will tend to improve by 1.0 AYPA. By age 30, the decline has begun, consistent with the previous 'peak difference' method. Another way to look at the same data is as a cumulative improvement/decline by age, as shown below.
In the above graph we see rapid improvement until a plateau beginning around age 26, and then a shallow decline beginning at age 29.
So far we've only looked at QBs who have 10 or more years as an established starter, something we already know about Manning. Let's open the aperture and look at QBs with 5 or more years as an established starter. I chose 5 years because I'm interested in real-world QB decline, and not just theoretical physiological effects of age. Guys with fewer than 5 years probably aren't good enough to survive past their peak.
The delta method for 5+ year starters produces the following aging pattern:
The aging pattern for 5+ year starters looks very different than for the 10+ year guys. The peak is not as high and the decline goes further. The magnitude of the improving years for the 5+ years group does not peak as high as for the 10+ years group, which makes the decline appear stronger. If we took a QB who has started for 5 years and no longer, and knew nothing else about him, this might be the aging curve we'd expect.
One of the lessons here is that there is no one true aging pattern for the league as a whole. It depends on the type of guy and the comparison data used in the analysis. For example, if you perform the delta method for all QBs, regardless of career length, you see a flat plateau from age 22 through 28, with almost no improvement. This is because the the many young QBs who never catch on weigh down the averages for everyone., much like we saw in the first graph of simple averages by age, which was one big plateau.
One of the more interesting things in the numbers was that the final year of a QB's career, regardless of age, is usually pretty bad, but not necessarily worse than the usual year-to-year variation in any individual QB's resume. In fact, the final year of a QB's career, on average, represents a decline of -0.75 AYPA. This is far worse than any one year of average decline due to age--actually equivalent to about 6 years of decline. To me, this suggests that natural variance is helping end many QB careers.
In fact, if you simply remove the very last season of each QB's career from the data, age-related decline virtually disappears. Here is what the aging pattern would look like:
A lot goes into the decision to retire, and it's not always completely the player's choice. Older QBs are checking out of the league after a down year, but there's no guarantee that the downward trend would continue. Although it's unlikely they'll reach the highs of their peak years, regression to the mean says that the following season is more likely going to be an up-year, at least relative to the previous one. At some point, a QB has absorbed enough sacks, had enough surgeries, made enough money, won enough thrillers, and lost enough heart-breakers for a lifetime. If the prospects for future success aren't very good, it's time to hang up the cleats, even if those prospects are somewhat of a statistical illusion.
The bottom line is that very successful quarterbacks like Manning aren't going to become bad slowly. All of sudden one year, they'll have significant drop-off in performance. If they were 26 and had the same kind of season or had a similar injury, they'd no doubt be back at camp the following July. But at 36, that job in the broadcast booth will seem quite enticing. Successful, established QBs will generally continue to be successful until one day they're not. We won't see it coming. But of course, everyone will pretend they did.
How Quarterbacks Age
By
Brian Burke
published on 8/30/2011
in
analysis,
basic,
player analysis,
quarterbacks,
research
Subscribe to:
Post Comments (Atom)
A few comments:
1) Don't you still have survivor bias when you look at the stats of QBs who have started for at least 5 years? In other words, aren't some of those QBs dropping out of the league before they turn 39? The same would be true even for the 10 year starters.
2) I understand that an average yards per attempt metric works well for your win probability modeling, but perhaps a more appropriate metric of productivity in this case would be total yards (perhaps again adjusted for INTs and TDs). This would take into account injuries and ability to throw a large number of passes per game. Certainly the Colts wouldn't be happy if Manning had one game with a very good 12 Adjusted YPA, just to spend the rest of the season injured and unable to play. This would be a good season by your metric though. Seems reasonable that the risk of injury increases with age, at least beyond a certain point.
3) If you used total yards as your measurement, you could completely get rid of survivor bias by assigning 0 yards to QBs who drop out of the league for the ages in which they are retired. You could also account for yards gained on the ground by the QBs. If you stick with Adjusted YPA, using 0 for retired players makes less sense because of division by 0 attempts. The problem with total yards is that different offenses will use passing more or less often, so you might want to look at each QBs performance as a % differential from the previous season. There are probably other problems with it, but they are likely minor compared to the survivor bias problem.
Looking at 5 and 10 year starters for comparison is a decent way of doing conditional comparisons, which is appropriate for this type of analysis. You could also project Manning by looking at other QBs who had been starters at age 34, or who had started as many games as he has.
The analysis on QB's last seasons is interesting and makes a lot of sense.
Sorry for the second comment, but just to hit the survivor bias point home:
Let's say you're trying to project Manning through the age of 39. In what you've done here, you are assuming there is 0 probability that he will exit the league before then, either due to an injury or degradation in skills. But the probability is certainly higher than that. That's why you need to look at the conditional scenarios of QBs who have played as many games as he has or are as old as he is, and include the fact that some of them put up 0 stats before the age of 40.
would be interesting to see how kurt warners career went in relation to these stats, iirc he didnt have such a bad final season.
the survivor bias is overstated here. and, by following a qb through his career, eliminates that bias that would have been apparent in a simple analysis.
the thing that strikes me, is that obviously qbs at an odd age perform much better than qbs of an even age :)
Elway retired on top after a strong season. There are exceptions to the theory advanced in this article.
Does this explain Brett Favre? Cause pretty much everyone wanted him to retire after 2006 (a randomly poor year?). But then he bounced back in 2007 in a near MVP season (if not for Brady). He was again great in 2008 with the Jets for more than half the year (until a shoulder injury hurt him). He had possibly his best season ever with Minn. So maybe 2010 was another randomly poor season?
Another point which could "spoil" the numbers is that old QB´s (even those with 5 years) only survived long enough b/c they were on teams which were good enough. A Manning in Detroit would be out of the league since long. Same goes with Brady, Montana and so on.
Bad teams have revolving doors at the QB-Position, ending careers of many talented young QB´s in a very short time.
Karl from Germany
Another point which could "spoil" the numbers is that old QB´s (even those with 5 years) only survived long enough b/c they were on teams which were good enough.
True enough. Look at the record, and it is very easy to concieve e.g. that if Brady had been drafted in the fifth round by the Lions nobody would remember him today (except maybe a Lions trivia geek.) For instance *if* he gets off the bench to start in his second year, and he improves the Lions from their actual 2-14 to maybe 3 or 4 wins, everyone says "that sucks, but what can you expect with a 5th rd pick?". The next year they draft Harrington, etc.
The general problem here nobody has solved is separating the QB's own influence on his numbers from the rest of the team's influence on them. PFR.com's "approximate value" metric is about the best attempt I've seen (concluding iirc that the QB's numbers are only about 30% the QB's, which has a lot of implications) but that is useful only in retrospect.
I think for the purpose of "QB aging" the sample size is sufficient so individual cases don't matter so much, and the conclusions are viable and interesting.
Peyton's suddenly aging at an accelerating rate.
His contract pays him $55 million even if he never plays another down.
If Peyton is human, the thought has to enter his mind, "$55 million for nothing ... what if this pain in my neck never goes away? ... Hmmm..."
I know he is a super-human competitor, but he's got to have some mere human components.
BTW, the Colts gave him those contract terms after he was injured. I want my next job to be with the Colts!
Late to the party...
You may also look at "aging" as a factor of downs played as opposed to pure age. A QB who gets a late start to his career, whose body doesn't have the wear and tear, may be able to play later in age than someone who began as a starter at an earlier age.