Recently, we overhauled our approach to how we evaluate passed balls and wild pitches here at Baseball Prospectus. It started innocently enough, as an attempt to make our data betterbehaved, but progressed to a gradual recognition that we—and as far as we can tell, plenty of others—have been taking the wrong approach to these events for quite some time. Today, we’ll talk about what we’ve learned, and how our models are much the better for it.
Background
It’s no secret that some catchers are better at blocking pitches than others. Yadier Molina seems to be pretty good at it, and Mike Zunino does not. But raw wild pitch and passed ball numbers can be unfair. The catcher, after all, is not the one throwing the pitch, and some pitching staffs are wilder than others, particularly if those pitchers like to throw certain pitches in certain places. The sabermetric community’s longstanding skepticism of official scoring has also led to the practice of combining passed balls and wild pitches for modeling purposes, even though the former are judged by the scorer to be the catcher’s fault, and the latter to be the fault of the pitcher.
As with all things sabermetric, the means of adjustment for these factors have become more sophisticated over time. At the simplest level, we could simply trust the official scorer, and assume the other factors largely balance out. A more sophisticated approach is the “With or Without You” method, which grades a catcher based on how he does without certain pitchers, or how pitchers do without various catchers. Going one step beyond, researchers have tried to identify relevant factors driving passed balls and wild pitches, incorporated them into models of “likely” passed ball/wild pitches versus “actual” such events, and then grading a catcher on the difference. FanGraphs has adopted a model created by Bojan Koprivica as the basis for its Runs per Passed Pitches (RPP) metric. (The parameters of that model appear to be proprietary, although Bojan does describe the relevant aspects). Finally, we unveiled our own blocking model last year, called “RPM WOWY”: a combination of PBWP likelihood, as determined by PitchInfo, followed by a WOWY assignment of credit among catchers and pitchers.
When we began incorporating mixed models into our catcher metrics earlier this year, we converted our catcher blocking model over as well, since it made sense to have all our catcher metrics on the same basic method. And so, throughout the 2015 season, we combined all of what we now call “errant pitches” into a linear mixed model, specified as follows in the R programming environment:
glmer(PBWP ~ log(prob) + (1pitcher) + (1catcher), family=binomial(link=’probit’)
The model was, frankly, a bit of an afterthought. We were focused more on converting RPM WOWY into our new framework rather than thinking it over from scratch. The log transformation of errant pitch probability (the former “RPM,” and the “prob” in the new blocking model) was added as much to assist convergence as anything else.
And yet, the model continued to behave somewhat erratically over the course of the season. The values were not unreasonable, but convergence continued to be an occasional problem, finally culminating in us taking it down a few weeks ago. Around the same time, we worked with August Fagerstrom, who was looking for data that would help investigate the role of R.A. Dickey in Russell Martin’s pitchblocking this year. I proceeded to run the model both with and without knuckleballs, and August wrote about the findings. The answer: Martin was terrible at blocking pitches this year even when not catching knuckleballs.
All of this, combined with the conclusion of the regular season, provided good reason to dig a little deeper and start questioning not only how our models could be better specified, but also some of our core assumptions. When we did so, we were surprised at what we learned. We’ll now share those lessons with you.
Trusting the Official Scorer
As mentioned above, our model (and other recent efforts) treat passed balls and wild pitches as part of one overall population. The underlying concept is reasonable, in that a good blocking catcher can affect both categories: passed balls by not having them, and wild pitches by getting in front of balls that would otherwise head to the backstop.
But it turns out that treating passed balls and wild pitches as identical is wrong. This became increasingly clear as we studied the probability metric (prob) that we were using for errant pitches. Although the combination of pitch type and pitch location was helpful in predicting wild pitches, it was essentially useless in fitting passed balls, with a correlation of about 1 percent. In other words, passed balls have almost nothing to do with the type of pitch thrown or the location where the pitcher throws it. This makes sense, in that a pitch is generally graded as a passed ball only when the catcher has little excuse for not catching it. This also suggests that official scorers are doing a good job — or at least being conservative — in assigning passed balls to catchers.
In fact, when we started separating out the various components of the errant pitch probability metric, we found there to be only one factor, aside from the identity of the catcher, that made any difference in predicting passed balls: the knuckleball. With knuckleballs, the usual ratio of passed balls to wild pitches for other pitch types is reversed, with passed balls being the more common error. In fact, the importance of the knuckleball was significant at a level less than 2e^{16}. This, in a way, also makes sense. Passed balls are assigned when the pitch arrives at a location where the official scorer thinks an average catcher with reasonable effort could have caught the pitch. The knuckleball, of course, is unusually unpredictable, making it more likely a catcher will be fooled even by a knuckleball at “catchable” height. The significance of the knuckleball in our new passedball model confirms that this pitch remains the one “blind spot” in official scoring of passed balls. By adjusting for it in our new model, we help protect catchers from the unfairness of having to catch a knuckleballer.
The irrelevance of pitchers to passed balls can be confirmed statistically. We test this hypothesis by adding the category of “pitcher” to the categories of “knuckleball” (yes / no) and “catcher” and seeing if the identity of the pitcher, once we control for knuckleballs and catcher, has any remaining effect on the probability of a passed ball. For 2015, the model outputs the following:
In mixed models, we are interested in the variance of our groups, a/k/a “random effects”: a variance of 0, or close to it, confirms that a particular group of players does not affect the outcome of whatever is being modeled. Here, since we are controlling for knuckleballs, the pitcher, with variance of 0, is otherwise irrelevant in the causation of passed balls. We therefore removed pitchers from our passed ball model entirely, and let knuckleballs and the identity of the catcher be the sole predictors of interest.
The Polynomial in our Midst
With respect to wild pitches, our existing probability formula remains quite useful: The 2015 correlation between wild pitch probability and an actual wild pitch is .4. Not bad at all. But the most startling thing about this relationship is evident when we plot those data points against each other, here for the 2015 season:
Our mathematically inclined readers will spot what is going on here right away: The predictor, now rechristened wp_prob, certainly has a strong connection with wild pitches. But more importantly, wild pitch probability operates on an exponential curve. It turns out that the best fit of wp_prob to wild pitches is a fourth (!) – degree polynomial function, and the improvement made by the addition of that polynomial is clear when we use generallyaccepted indications of model quality. Here is the likelihood ratio comparison between each degree of the polynomial, as indicated in the model titles:
The fourthdegree polynomial is clearly superior.[i] This was something we missed before, and have now incorporated.
The Catcher’s Role in Wild Pitches
The converse of passed balls being correctly classified by official scorers might seem to be that wild pitches are being correctly classified as well. This is largely true, although, as we will see, catchers still play a role.
Incorporating our fourthdegree polynomial, and using our typical pitcher and catcher categories, we get this summary:
Well now: This fit suggests that the catcher plays virtually no role in the likelihood of a wild pitch. This seems odd, given that we are pretty confident some catchers are much better at preventing wild pitches than others. It’s as if the model is saying: “The effect is there, just not in the way you are asking for me to look for it.”
So, let’s think about this. We believe that catchers play a role in stopping wild pitches from happening, but the model is reporting that this hypothesized skill does not manifest itself on a consistent basis. Why would that be? Our hypothesis focuses on the fact that majorleague pitchers have decent command and stick with pitches they can consistently locate. For these pitchers, having extraordinary blocking skills isn’t that important. On the other hand, if a lot of team innings are being eaten by a pitcher or pitchers that tend to throw balls in the dirt, those blocking skills would be very useful indeed. Essentially, what we need to do is relate the blocking skill of the catcher to particular pitchers. Let’s try it and compare the results:
As you can see, the results are very similar, but the likelihood ratio test confirms that the pitchercatcher version of the model is the better one. So, let’s now rerun the model, but with only pitcher and pitchercatcher as our groups:
And . . . oh dear. The pitchercatcher interaction wasn’t just the key to understanding the catcher’s contribution; it covers the pitcher’s contribution too. Which is to say, wild pitches really are all about the relationship between the skill sets / tendencies of particular pitchers with particular catchers. With a variance now of 0, “pitcher” needs to be removed too. That brings us to our final, and the best, wild pitch model:
This turns out to be the answer. Our final wild pitch model has the best AIC, the BIC, and was confirmed by a likelihood ratio test to have the best explanatory power (p<2e^{16}) among all models considered.
The Change in Distribution
We’ve learned a lot about passed balls and wild pitches, but we haven’t yet discussed the one thing that truly makes these models work: the change in assumed distribution.
Previously, our model treated the occurrence of a passed ball or wild pitch (combined, of course) as part of a binomial distribution. On the surface, this seemed like a natural choice: You either have a PBWP or you don’t, which makes it 1 or 0. Moreover, most statistics education begins with binomials and the classic coin flip as a means of explaining probability, so the binomial distribution is a natural “default” of sorts for this type of data.
Except that here this assumption is wrong. The idea behind the binomial distribution is that each event has approximately an equal chance of occurring at each unit of measurement (here, a PA). That’s not even remotely true of passed balls or wild pitches, which are very rare events: Out of 203,183 opportunities during the 2015 season, there were only 1,549 wild pitches and 285 passed balls. These events also have other interesting qualities. Their respective mean likelihoods and the variance are almost identical. Furthermore, the distribution of passed balls/wild pitches for any pitchercatcher tandem can never be less than zero, and can only climb by one integer at a time. In other words, you can have three wild pitches, but you can’t have 4 or 7.25 wild pitches.
There is, of course, a wellrecognized distribution that is designed for rare events, which can never be negative, increases only by integer, and which assumes substantially similar mean and variance: the Poisson distribution. Switching from a binomial to a Poisson distribution is actually the key to the whole thing. Not surprisingly, when you choose a distribution that more accurately characterizes your data, your results get better and your model becomes more stable.
And so, our final mixed models for passed balls and wild pitches, as specified through the lme4 package of the R programming environment, are as follows:
glmer(passed_ball ~ KN + (1catcher), family='poisson', nAGQ=0)
glmer(wild_pitch ~ wp_prob + I(wp_prob^2) + I(wp_prob^3) + I(wp_prob^4) + (1pitcher:catcher), family='poisson', nAGQ=0)
If you calculate the two models, predict both of them onto the original data set, and then summarize by catcher and pitcher the model’s total expectation over all of their opportunities for both events, you get the models’ composite prediction of what we call Errant Pitches fairly attributed to each pitcher and catcher.
Our final rate statistic for Errant Pitches is Errant Pitches Above Average, or EPAA.[ii]
Out of Sample Testing
So, we’ve found excellent descriptive fits to the data. But, of equal interest is how consistently these models identify the same abilities in these same individuals. In other words, does a catcher’s passedball rate in the firsthalf of the season suggest to us what that rate will be in the secondhalf of the season? And similarly, should we expect a strong pitchercatcher wildpitch combo to retain its effectiveness down the stretch?
The answer to both questions is yes. In fact, these models show some of the highest reliability I have seen in baseball statistics. To generate them, I randomly split each season from 2010 through 2015 into two equal halves, ran the models separately on each half, repeated this for each season 20 times, and then took the average correlation for each model between the two halves over all 20 splits for each season. For simplicity’s sake, I limited the predictions to passed balls for catchers and wild pitches for pitchers.
Here are the results:
Season 
Wild Pitch 
Passed Ball 
2010 
0.87 
0.91 
2011 
0.88 
0.97 
2012 
0.87 
0.91 
2013 
0.87 
0.95 
2014 
0.89 
0.94 
2015 
0.88 
0.93 
As you can see, our new models not only fit the data, but they do so consistently as well.[iii] If you want to know the true talent of a particular catcher at blocking pitches, these models are an excellent place to start.
The Effect of the Models
Finally, we’ll provide the catchers who benefitted and suffered the most from their batterymates during the 2015 season:
Name 
Pred PBWP 
Actual PBWP 
PBWP Saved 
Runs Added 
32.2 
22 
10.2 
2.9 

51.3 
42 
9.3 
2.6 

28.5 
20 
8.5 
2.4 

29.2 
21 
8.2 
2.3 

14.3 
7 
7.3 
2 

29 
23 
6 
1.7 

28.1 
22 
6.1 
1.7 

41.3 
36 
5.3 
1.5 

41.4 
36 
5.4 
1.5 

Yadier Molina 
37.9 
33 
4.9 
1.4 
13 
17 
4 
1.1 

28.6 
33 
4.4 
1.2 

12.6 
17 
4.4 
1.2 

32 
37 
5 
1.4 

Mike Zunino 
41.4 
47 
5.6 
1.6 
38.9 
45 
6.1 
1.7 

21.7 
28 
6.3 
1.8 

39.7 
48 
8.3 
2.3 

40 
52 
12 
3.4 

Russell Martin 
49.6 
64 
14.4 
4 
And here are the top and bottom pitchers:
Name 
Pred PBWP 
Actual PBWP 
PBWP Saved 
Runs Added 
12.3 
6 
6.3 
1.8 

11.6 
6 
5.6 
1.6 

7.3 
2 
5.3 
1.5 

8.2 
3 
5.2 
1.5 

8.0 
3 
5.0 
1.4 

6.9 
2 
4.9 
1.4 

9.5 
5 
4.5 
1.3 

5.1 
1 
4.1 
1.2 

10.9 
7 
3.9 
1.1 

3.7 
0 
3.7 
1.0 

7.7 
12 
4.3 
1.2 

1.6 
6 
4.4 
1.2 

2.3 
7 
4.7 
1.3 

2.2 
7 
4.8 
1.4 

12.9 
18 
5.1 
1.4 

1.7 
7 
5.3 
1.5 

5.6 
11 
5.4 
1.5 

1.5 
7 
5.5 
1.5 

7.5 
13 
5.5 
1.5 

4.0 
11 
7.0 
2.0 
That will do it for now. As always, we welcome your comments.
[i] Polynomials of any higher degree begin to degrade the fit, and would be problematic for prediction in any event.
[ii] To get the Errant Pitches Above Average (EPAA), we take the predicted number of Errant Pitches, subtract the actual number of Errant Pitches, and divide them by the number of opportunities for each player.
[iii] It’s important to note that “prediction” here is not necessarily the same as an actual projection, although these numbers should help with that as well when used appropriately.
Thank you for reading
This is a free article. If you enjoyed it, consider subscribing to Baseball Prospectus. Subscriptions support ongoing public baseball research and analysis in an increasingly proprietary environment.
Subscribe now
Maybe it doesn't even matter for the purposes of predicting future events, but it seems like a different category from your runofthemill passed balls.
Meanwhile, we have posted the data for 20082015 in the sortables http://www.baseballprospectus.com/sortable/index.php?cid=1892377 which you can download and explore.
As groups:
C's w/20+ FRRA (total 1,890): +53.4 BLRA
C's w/pos FRRA (2,213): +43.5 BLRA
C's w/neg FRRA (1,053): 43.9 BLRA
C's w/10+ FRRA (773): 18.3 BLRA (see Ruiz)
Best: McCann 167.9 FRRA/17.8 BLRA (Martin 170.1/2)
Worst: Doumit 84.1/5.5
Other Notables:
Carlos Ruiz 33.1/21.0
Matt Wieters 24.7/16.0
J Molina 144.4/11.9, B Molina .8/10.8
Miguel Olivo 20.9/16.5
Wilin Rosario 14.9/15.9
With no runners on (and not a dropped third strike), a pitch will never be called WP/PB. Many/most will be the result of the pitcher having really no reason to worry and therefore being willing to bury it in the dirt, and/or the catcher not caring whether he stops it. *Some* will be the result of a breakdown in control which could reinforce the observation of the particular pitcher's predilection for throwing the ball away.
The actual runners involved also affect the issue. There are events that will advance Terrence Dyson that will not budge Billy Butler. The former get called, the latter do not.
So do either of these matter, or are the effect too small? In the former (no runners), the indifference of both the pitcher and catcher probably mask any possible useful observation. In the latter, the battery is on its "best behavior", but the identity of the (lead) runner changes the margin for error they may have.
The other thing I find interesting is that while Zunino is listed midway on the poorer scale, I don't see any Mariners pitchers (of what I remember) on the poor side of the pitcher's stats. Shouldn't there be a corresponding ratio?
Just like more strikes are called balls before balls are called strikes,more passed balls are called wild pitches than wild pitches are called passed balls.A big part of the catcher's job is to prevent runners from advancing 90'
(Blocking balls in the dirt & controlling the running game.)Just because the ball is in the dirt does not mean that it is a wild pitch. Catchers are EXPECTED to block many of those pitches & keep them close enough so that runners are not able to advance.(Good point previously re: Butler vs. Gore running)I think that catchers need to be evaluated on every ball in the dirt relative to the speed of the pitch & the distance it hits relative to the catcher(Front & side). I realize that this is a tedious endeavor, but we are only talking about 10 pitches on average per game per catcher. One thing that muddies the waters is the fastball in the dirt.The perception in today's game is that you can't block fastballs in the dirt & that you have to pick them.(Hard to accept)In many of those cases you are evaluating pickers & not blockers. IMO,& where it gets somewhat subjective, if the catcher tries to pick,what you think is a blockable fastball, then it becomes a misblock.
Sorry to get away from the gory math but I'm not sure the #'s really indicate how good a blocker the catcher is.Maybe we need a new stat,blocking %.(% of blockable pitches blocked based on speed & distance to remove some of the subjectivity.)
JW
Speed, pitch type, where it hit the ground http://www.baseballprospectus.com/article.php?articleid=22934