Catcher Aging is a Curve, Not a Cliff
The Phillies just gave Carlos Ruiz a guaranteed three year contract that covers his age-35 to age-37 seasons. In both my defense of the deal here and on Twitter, I have invoked Russell Martin’s name as a comparison, but noting that Martin was a huge bargain last winter has met with some resistance because Martin is significantly younger than Ruiz. His two year deal with the Pirates covered his age-30/31 seasons, so Pittsburgh wasn’t committing guaranteed money deep into his mid-30s, as the Phillies are doing with Ruiz.
There simply aren’t that many examples of highly productive 35 year old catchers, as most catchers are on their last legs at this point in their career, and that has helped fuel the belief that catchers age very poorly. The belief that catchers age in dog years is a prevalent one, and is one of the main reasons that the Ruiz deal has been poorly received. However, I think the evidence mostly points to this idea being incorrect.
To illustrate this point, I asked Jeff Zimmerman to send me his aging curve chart for catchers over the last 30 years. He was kind enough to do so, and I’ll present it below.
This chart is in batting runs relative to league average, so we’re only talking about offense here, since quantifying catcher defense is a bit of a tricky animal anyway. Since it’s a counting stat, it also factors in playing time, so we don’t have to make separate adjustments for performance and health. As we can see, peak offensive levels are from 25 to 28, as is pretty commonly accepted, and then there’s a gradual decline as a player gets beyond those years. But notice the trend of the blue line (catchers) compared to all position players (the red line); they move basically in lock-step from 27 to 32, after which point catchers actually age better than the norms for non-catchers.
Here’s the table of the actual numbers represented by the chart, for those of you who are more numbers driven than graphics driven.
| Age | Catchers | All |
|---|---|---|
| 21 | -8.5 | -14.6 |
| 21 to 22 | -4.2 | -8.7 |
| 22 to 23 | -8.0 | -4.3 |
| 23 to 24 | -4.1 | -1.5 |
| 24 to 25 | -3.0 | -0.2 |
| 25 to 26 | 0.0 | 0.0 |
| 26 to 27 | -4.7 | -2.5 |
| 27 to 28 | -1.5 | -3.3 |
| 28 to 29 | -7.2 | -4.8 |
| 29 to 30 | -9.9 | -9.8 |
| 30 to 31 | -14.0 | -11.5 |
| 31 to 32 | -17.9 | -17.1 |
| 32 to 33 | -22.8 | -21.8 |
| 33 to 34 | -22.4 | -25.3 |
| 34 to 35 | -28.9 | -30.8 |
| 35 to 36 | -32.5 | -35.8 |
| 36 to 37 | -35.9 | -39.7 |
| 37 to 38 | -45.0 | -49.7 |
| 38 to 39 | -51.7 | -57.1 |
| 39 to 40 | -61.4 | -68.5 |
| 40 to 41 | -80.2 | -79.7 |
There simply isn’t this huge early-30s drop-off that is widely accepted as a fact of life for catchers. The existence of a huge cliff at which productive catchers simply turn into useless backups isn’t supported by the data. Like players at all other positions, catcher aging is a mostly linear downwards trend, and there just aren’t certain ages at which player performance gets exponentially worse. Skills decay over time; they don’t evaporate over night.
This isn’t just a lesson that should be applied to catchers. Aging curves for all players should be viewed as fairly linear trends. Jeremy Greenhouse (now working for the Chicago Cubs) created this aging curve by WAR a few years ago at The Baseball Analysts:
That’s basically a straight line down from age 30, with no significant increases in performance decay at different stages. If you have a +3 WAR player heading into his age-35 season, he shouldn’t be expected to be demonstrably worse in the next season than a +3 WAR player headed into his age-31 season. The more important variable in a player’s expected future production is that player’s estimated production for the upcoming season, not his age in that season. Significantly decaying a player’s future performance because of his age is putting too much confidence in the effects of aging and not enough confidence in the player’s actual performance.
Good 34 year olds don’t regularly become bad 35 year olds. In general, you should expect players to decline at something like +0.5 per season. If they’re especially injury prone and their bodies are breaking down, limiting future playing time, maybe you knock off +0.75 WAR per season. A reasonable aging curve should return a decay of something in that range, for nearly any set of years. Whether a player is 30 or 35 should not drastically change the amount of aging you expect in the future.
And whether a player is a catcher or not should also not greatly affect our expectations for his future performance. In general, catchers age about how we’d expect any other position player to age. They get worse, but they don’t fall apart without notice. Expect decline, not collapse.
Dave is the Managing Editor of FanGraphs.
Great deal.
That’s a bold statement.
Great analysis and point well taken.
My only question is how the aging curves account for catchers who move to a less physically demanding position. In other words, would Joe Mauer’s career numbers count towards the catchers’ aging curve throughout his whole career, or just up until 2013?
my guess is this treats each player-year as independent of any years for that same player, and looks at averages? in which case this either overestimates (for not including productive years of non-catching) or underestimates (by counting people like 2013-mike napoli as catchers) the true rate of decline. this was an incredibly unhelpful reply.
Don’t be so hard on yourself.
Didn’t answer my question, but correctly stated the issue behind my question.
Which gives your comment positive Value Over Replacement Reply….
This was my immediate reaction as well. I think we’re seeing a selection bias in that when catchers get injured or simply show signs of aging, they are switched to another position if they can still hit well.
This would account for the blue line actually being above the red line at advanced ages.
However, this does not negate the idea that catchers don’t age as poorly as is generally thought.
Not sure if it matters, but Ruiz only started in pro ball at the age of 21 after the Phillies converted him to catcher. This may mean less wear and tear than is normal for a typical 34 year old catcher.
How much are these results impacted by the nature of the offensive game of catcher? The later years probably includes more defense focused catchers who never had high offensive production to begin. I’d be curious to see this broken out more by offensive peaks or even league average hitting catchers vs non-league average hitting catchers (perhaps defined as production in 27 year old season or something like that).
Aging Catcher < Cliff Lee Curve
I’ve discussed aging curves with Jeff before. I wonder if he includes an attrition/survivor bias adjustment, and if not, if it might be particularly relevant for catchers (if lots of older catchers drop out from one year to the next, then it could be a significant problem that needs to be corrected).
Yes!
Here’s the number of catchers with >0 PA for each age, going back to 1980 (according to FanGraphs leaderboards):
34: 123
35: 104
36: 69
37: 47
You can see there’s a huge dropoff at Age 35 -> Age 36; 35 catchers lost.
I should create the survivor disclaimer for all these curves (only for players who don’t play in both seasons). Both curves are created with the same bias and Dave’s point that they age the same as the general population still stands.
Maybe I’m dumb (probably), but it seems like players of different positions don’t show the same survivorship bias over age.
For example, 1B lose only 23% of their players going from age 35 to age 36.
DHs lose 27% of their players.
whereas…
2B lose 34% of the players, similar to catchers.
SS lose ~40% of the players.
Those seem like big differences. What am I missing…?
My guess? An aging SS will likely be moved to a lesser defensive position, like 2B, while 1B and DH have nowhere else to move too, assuming moving off the position counts in the % (Catcher and SS are harder than 2B are harder than 1B/DH)
Ruki: sure, I get that notion. But doesn’t that contribute to a significant position-based survivorship bias?
That is to say, not all positions seem to lose players at equal rates, meaning that comparing catchers to a “general population” doesn’t seem meaningful if catchers are especially prone to drop out of the league.
This feels a bit like drowing while wading across the river of an average depth of only 3 feet. For the aggregate, the curve runs smooth, but it may well be that each player drops off a cliff, and drops off at different points, and averaged out it looks rather smooth. I could test this actually somewhat easily, if there are paramaters that could be mutually agreed upon.
As a phillies fan and general fan of Ruiz, i hope that the cliff doesnt approach soon, but I’m sort of dreading it will.
I argue this point all to frequently. People seem to think that catchers age in some some rapid bubble compared to the rest of the MLB.
If survivor bias is accounted for here?
Please do answer “Just Wondering’s” question.
How would you properly account for survivor bias?
I can’t think of a way without significant flaws… is there a standard methodology I don’t know about? Any thoughts?
In terms of survival analysis, Cox proportional hazards models are the classical option, since you can censor when a player drops out, although you need some kind of binary outcome which doesn’t really fit the question that well. GEE or random effects models might be an option, since they’re a little more robust to missing data.
Another option might be multiple imputation, but there’s some reasons why you might not want to go down that road, either.
The problem is the missingness is strongly associated with your outcome (performance). There’s only so much statistical magic you can use in that situation. Missing Not At Random (MNAR) is basically a statistician’s nightmare.
yeah–thanks for your answer. That’s what I was trying to ask about, without knowing the proper terminology.
In this case, I just wouldn’t look at the aging curves in aggregate at all, but try to model the shape of the average decline for each catcher. A solid way to do that would be to normalize each players’ annual performance versus their mean performance and stdDev, then use a model that can optimize over the age that the declines occur (i.e., shift the curves until the variation in normalized performance for each year is minimized).
If we did that, we’d have a great idea about how quick the actual declines happen for a class of players, independent of the age of the player or their raw performance. It would not be prone to the issues we’re seeing above, where if each catcher immediately “hit a cliff” but did it at different ages, we’d still see a smooth-looking curve.
So are you optimistic on A.J. Pierzynski for this year?
“Good 34 year olds don’t regularly become bad 35 year olds. In general, you should expect players to decline at something like +0.5 per season. If they’re especially injury prone and their bodies are breaking down, limiting future playing time, maybe you knock off +0.75 WAR per season. A reasonable aging curve should return a decay of something in that range, for nearly any set of years. Whether a player is 30 or 35 should not drastically change the amount of aging you expect in the future. ”
Miguel Cabrera is 30.
He put up about 7.5 WAR this season.
Assuming a rate of loss of WAR of .5/season, Miguel Cabrera will continue to produce positive WAR for 15 more seasons, until he is 45.
Votto is also 30, and put up ~6 WAR, so he’s good for 12 more years.
Shin-soo Choo is 31, and put up ~5 WAR, so he’ll be playing until he’s 41.
Jose Bautista is 33, 4 WAR, so he’s also sticking around until he’s 41.
Finally, rejoice Red Sox fans, Ortiz is going to be there until he’s 45 or 46! Oh, and he ought to put up 4+3.5+3+2.5+2+1.5+1+.5 = 18 more WAR, giving him a pretty reasonable HOF case.
QED?
At this rate, MLB is going to be very, very old in ~2020.
So your argument is…what exactly? I see lots of snark but not much fire.
Aging is a curve and a cliff.
If everyone aged at the clocklike rate DC says, all superstars would stick around until their 40s.
They don’t, because there are frequently abrupt, rapid downward shifts in performance, which don’t get accounted for in the numbers. This could be because of survivor bias; it could be because of position-shifting through aging; it could be because “star” players age at a different rate than non-star players.
The numbers in the curve represent the “mean” aging curve, but as TheUnrepetantGunner notes, every individual aging curve could have a cliff, whereas the aggregate is smooth.
Bottom line: the simple formula of -.5WAR/year is obviously broken, because if you tried to extrapolate forward performance it would predict unrealistic results, namely every 30-year old star playing until they are in their 40s.
This.
The average numbers show just that–averages among age. The problem is, the players need to GET to that “playing age” to be able to be recorded in the numbers.
So we need to be graphing the standard deviation of the change in war, and the number of players dropping out year to year, as well as the change in war.
If you use a decent projection instead of just 2013 WAR, most of these examples disappear or are heavily altered. Ortiz is projected for 2.8 WAR next year, and I doubt seriously he could put up any positive value in 2019 (or really 2018 and probably 2017).
I’m confused. You say that we should use the projections–by which Ortiz is set for ~3 WAR. But even using that standard, Ortiz is set to play for 6 more years.
But then you say, you don’t believe he’ll put up positive value any time after 2017.
Which is kind of my point; the straightforward linear decline model is clearly failing–I don’t think the most optimistic fans expect Ortiz to play another 6-8 years.
Also, I’m not sure what you mean that using projections causes the problem to “disappear”. It does for Shin-Soo Choo; he’s only projected to play 6 more years at this rate. Using Steamer projections instead, Votto sticks around for 10 years, instead of 12; Miguel Cabrera is around for 14, instead of 15; and Bautista actually gets an extra year, 9 instead of 8.
So the projections might help a little, but not much.
Do you know how many position players older than 40 there were in all of baseball last year (according to FanGraphs)? 3. None of them put up any positive WAR.
Does that not sink the Dave Cameron system, which already projects at least 5 players to stick around into their 40s (check Chase Utley, as well)?
Well, straight linear decline obviously doesn’t fit every case, but for prediction purposes, linear fits are generally the default. What function would you suggest? Exponential?
It doesn’t seem to fit for many good players in their 30s, which is a problem (these are the ones for which future performance is most important, since they are getting mega-contracts). What’s more, while linear decline seems to work alright for the early part of the age curve, the latter part is way out of whack, as I showed above: very few players are even playing in their upper 30s, nevermind accruing positive WAR.
An exponential fit is a good place to start, sure. There’s all sorts of complex regression models one could build, taking into account position or body-type or skills or history. I’ve been learning about machine learning recently, and I’d love to build an SVM-based predictor of future performance.
But that’s not the point; when your model is flawed (which this is), and flawed in a particular fashion or at a particular time, you stop trusting the output of the model at that particular time.
We know most players don’t last into their 40s, perhaps because of sudden, rapid declines. Rather than say “Sudden, rapid declines don’t happen” (effectively the headline of this entry), we SABRists ought to be more honest: our models are unable to accurately predict into a player’s late 30s, so decline could be rapid or it could be smooth. A number of outstanding issues like position-shifting and survivor bias affect the results such that we cannot come to a firm conclusion, and won’t be able to until smart people with a good grasp of the complexities of statistics can attack the problem in earnest.
Anyway, I’m not trying to be snarky, I just don’t trust this -.5WAR/year model and I think you shouldn’t either.
Well, all models are flawed. They’re models! That’s why we have mean squared error and stuff. I’m guessing you’re arguing that this one is a particularly bad fit. I’m not as convinced without seeing some data, but a possible solution could be to say that the decline is close to linear, but the SEs blow up and the confidence intervals widen to a huge degree such that talking about means isn’t helpful.
One thing that doesn’t generally get emphasized as much on fangraphs is variation in the data and standard errors. It’s generally practice in stats when presenting summary measures to present a mean and measure of variation, but that doesn’t happen much in the baseballing stat community.
In one last apologia for linearity, I’ll say that it’s a hell of a lot easier to understand and talk about than fractional polynomials or whatever. A quick and easy think might be just a 0.5 WAR decline 30-33 and 0.7 decline thereafter. Like a spline or something. I mean, it’s not rigorous or anything, but that sounds like what you’re after. There’s so much variation the predicts probably aren’t worth much, but if you’re trying to figure out what guys are worth, you gotta start somewhere.
the problem is you don’t understand that single season WAR does not indicate the true talent level of a player.
how much of this curve is ruined by PED era? More than anything PEDs distorted the aging curve.
I was wondering this myself. How do PEDs affect the aging curve? Does it smooth things out or cause a faster drop-off? Does a player getting caught mid-career (Melky Cabrera) vs. nearer the end (Ruiz, ARod) significantly affect the aging curve? It would not be surprising, given that the length of treatment potentially would be longer.
Great news guys! I checked out this chart (http://epress.anu.edu.au/apps/bookworm/view/Demographic+and+Socioeconomic+Outcomes+Across+the+Indigenous+Australian+Lifecourse/5111/images/Fig8.2_fmt.jpeg) and have concluded that each person (indigenous or not) dies gradually over time, rather just dying all at once! I know that for a long time we thought that death was just like a “cliff” where you were alive or dead, but it turns out that when you average out the data, people just seem to get less and less alive over time.
However, this does have some unsettling implications for those worried about zombies.
————————————————–
In general, you should expect players to decline at something like +0.5 per season.
————————————————–
I don’t understand this statement. Based on the final graph, it appears that the average player peaks at around 1.5 WAR at around age 25 and then decays to around -1.5 WAR at around age 40. That’s about 15 years to decline by 3 WAR, which would be a loss of around 0.2 WAR per season.
A Cameron piece where the analysis can be summed up as, “a win is worth $5-6mm and players decline by .5/season.”
SHOCKING!