THE BOOK--Playing The Percentages In Baseball

Isn’t it true that hi-lev situations are more likely to have base-out states that favor contact hitting vs. someone who assembles his slash line through a more “feast-or-famine” power hitting approach? 

If true, wouldn’t that naturally lead to a list on the plus side that is tilted towards contact hitting, and on the negative side, people who whiff a lot?  And if the nature of the output list is a natural consquence of the question being asked, might that not necessarily be a point in favor of clutch skill existing?

Didn’t Nate Silver claim that clutch hitting as a skill is really more a function of situational hitting skill? 
I interpreted that as meaning that slappy line drive singles/doubles hitters are better able to place the ball through the hole or into a gap when they need to as compared to strong power-hitters who basically try to launch long fly balls no matter what the situation is.

The list above would support this theory.

"And if the nature of the output list is a natural consquence of the question being asked, might that not necessarily be a point in favor of clutch skill existing?”

I would not be a point in favor of the idea of clutch hitting being some extension of a person’s character. Clutch guys being heroic and non-clutch hitters being frauds.

But a far more important point of evidence that certain types of hitters are undervalued and overvalued.

Clutch per PA (using only the 200 listed in the link)

Rk Player Clutch PA
192 Jim Beauchamp 2.7 730 0.00369863
95 Danny Sheaffer 3.8 1032 0.003682171
44 Duffy Dyer 5.2 2266 0.002294793
61 Rodney Scott 4.7 2487 0.001889827
135 Scott Bradley 3.2 1801 0.001776791
178 Jimmy Stewart 2.8 1602 0.001747815
140 Timo Perez 3.1 1825 0.00169863
14 Jose Lind 6.7 4000 0.001675
54 Jim Leyritz 4.9 2961 0.001654846
136 Ed Ott 3.2 1961 0.00163182

If one were using this to evaluate Sosa’s HOF candidacy, it would be a mistake to look at his 59 WAR and then subtract 16 from it.

His batting wins are 30.3, very close to what I have on my site, and the basis for his WAR.  The clutch is figured by looking at his WPA/LI, which comes out to 40.8, and subtracting that from his 24.8 WPA.

If you replace his batting wins with the 24.8 WPA, then that would be justifiable.  But if you dock him 16.8 without giving credit for the WPA/LI being higher than his batting wins, you are making a big mistake.

So as the metric stands, Sosa is 59 WAR.  Knock him down to 53 and you’ve got a good case.  Knock him down to Jim Rice territory and you are wrong.

What does the high WPA/LI say about him?  Does that mean Sammy was very efficient at turning run scoring situations into actual runs, but did so at times that meant the least to his team wins?

His high WPA/LI tells you that he had alot of batting outcomes that had more value for that inning, score, base, out than they would at a random point in the game.

His poor clutch means that he took that great skill and dropped it down substantially because he applied that particular skill when it didn’t matter.

***

You’ve given me something to think about then.  While the clutch metric is valid, it definitely needs to be understood in the context of WPA/LI, and not his basic LWTS (which is really how I was thinking it).

So, thanks for keeping my eyes open to that point.

Chris: word to the wise: clutch per PA with all those decimal places doesn’t convey much.  Turn it into clutch per 700 PA.

How does clutch hitting skill correlate with handedness and platoon splits?  My best guess is that lefties have relatively larger increaes in BABIP w/ runners on base (1B covers the bag, bigger hole), but I would also guess that bigger platoon splits would lead to same-handed relievers being used especially often in high leverage situations.  Does the data match my intuition?  Is that easy to check?  This is interesting stuff.  Thanks for posting.

Good observation Matt.  Of the 8 clutch hitters listed, 7 were switch hitters or lefties, Taylor the exception.

Of the 11 non-clutch hitters Tango listed, Thome bats left and the others right.

Cool, makes sense.  Thome makes great sense because he’s got a HUGE platoon split .294/.429/.615 vs. .238/.339/.420.  Teams must have wised up to that and thrown LOOGYs at him in clutch situations.  I’m curious how well this theory works on a larger scale.  Is it just really extreme hitters who hook it in the hole?  Pull hitters?  I found once that the top 8 career lefty sluggers at the time who get shifted on have bigger BABIP spikes with men on than the top 12 career righty sluggers at the time.  Those guys were obviously pull hitters.  But maybe the clue about this is platoon splits might have a countering effect on clutchness.

Handedness—that makes a lot of sense as a factor. I would also think most of the slap hitters who do well in the clutch don’t have big splits and therefore are harder to neutralize. 

Isn’t it true that power hitters usually have larger platoon splits than slap hitters? I’m not certain about that.
If so, that would put them at a disadvantage in high leverage situations where a LOOGY or ROOGY could more easily neutralize them.

Could we get some sort of composite of the hitters who are good in WPA Clutch and those who are not? It should be easy for someone with a good database around here. Maybe pitchers are less likely to give up a home run with a game tied or something, and being able to identify players who will help a little there would be worth knowing.

I started half my own question. Per 600 PAs, the list of 200 hitters Tom posted (good WPA clutch numbers) hit .275/.339/.399, 1B/2B/3B/HR breakdown of 108/25/4/11, K/W of 67/51, SB/CS of 13/6. Really, decent batting average, possibly good strikeout rate (but this is going back to 1950, when folks struck out less) but otherwise, unremarkable regulars. If there’s a link to the bottom 200, their composite stats would be interesting for contrast.

And, I forgot to add, they added about 0.4 clutch wins per 600 PA.

I know I’m in the minority on this, but I don’t truly understand this definition of “clutch.” I mean, I’ve read all the explanations, but I don’t truly get it.

I prefer the “clutch” approach James uses in Win Shares: an extra run for each hit above expected with runners in scoring position and one run for each home run above expected with runners on. Straightforward and valid within the context of what it’s trying to do.

No kidding.  I submit that not one fan in 1,000 could make sense of this distinction:

“His high WPA/LI tells you that he had alot of batting outcomes that had more value for that inning, score, base, out than they would at a random point in the game.....His poor clutch means that he took that great skill and dropped it down substantially because he applied that particular skill when it didn’t matter.”

And if clutch performance is delivered disproportionately by lefties and by hitters whose BA makes up a large share of their offensive value, it seems at least as likely that you’ve discovered a problem with the metric as a new insight into hitting.

I prefer the “clutch” approach James uses in Win Shares: an extra run for each hit above expected with runners in scoring position and one run for each home run above expected with runners on. Straightforward and valid within the context of what it’s trying to do.

Studes, in that case, you would compare RE24 to LWTS.  You (and James) don’t care about the inning and score.

No kidding.  I submit that not one fan in 1,000 could make sense of this distinction:

Guy, I consider it my biggest failure that I can’t explain what I think is the most important concept.

Wow, we’ve discovered that left-handed singles hitters have more success against high-strikeout, right-handed relievers than right-handed sluggers do.  I never could’ve guessed.

Nellie Fox

http://www.baseball-reference.com/players/split.cgi?n1=foxne01&year=Career&t=b

Under clutch stats late and close (1574 PA) he has a tOPS+ of 95, which I think means he did worse than usual in those situations.

High leverage, 1766 PA, he has tOPS+ of 119.

So, why wouldn’t late and close be better?

Studes, in that case, you would compare RE24 to LWTS.  You (and James) don’t care about the inning and score.

Sure, but that’s not really my point.  Put it this way: I’d rather look at the BRef High/Med/Low Leverage breakouts than look at someone’s clutch score.  I understand those.

Interesting example, dq.  Maybe Nellie Fox had a better innate understanding of key situations!

dq

That tOPS of 95 for Fox in late and close means, I think, that his OPS was lower then than it was during other times. But falling only 5% might be alot less of a fall than the average player sees, so that could mean he was actually “clutch.” But I’m not sure.

Cy

Also, when I looked at the clutch hitting leaders from 1987-2001, switch hitters generally did a better job of maintaining their OPS in late and close than other guys. Maybe not being susceptible to the handedness issue helped

If we take everyone from the list who had 4000+ PAs and ranked them by cluth per 700 PA, this would be the top 10

Jose Lind 1.17
Sandy Alomar 1.09
Dick Green 1.00
Nellie Fox 0.91
Tony Taylor 0.86
Joe Cunningham 0.86
Scott Fletcher 0.84
Tommy Helms 0.80
Dan Meyer 0.78
Orlando Merced 0.77

22/you missed the point. Nellie Fox is the greatest clutch hitter in the retrosheer era, but he did worse than average in late and close games.

I’m confused.

I sorted the list by Clutch Wins/600 PA, and looked at the top 100 versus the bottom 100—the guys who were really good versus the guys who were pretty good. The pretty good guys went .280/.348/.413—not a bad group of hitters. The really good guys weren’t anywhere as good a group of hitters: .267/.324/.374. Per 600, the pretty good guys hit more doubles (25 vs. 23), home rums (13 vs. 9), drew more walks (55 vs. 45) and were faster (14 SB, 6 CS and 11 GDP vs. 11 SB, 5 CS and 12 GDP). The really good guys had more sacrifice bunts (6 vs. 4). The other categories held more or less constant.

I’m still inclined to think this is an adjustment on the part of the pitchers. What would be interesting to see is the difference in WPA for each event for a good clutch hitter versus a bad—are the outs constant and the hits less valuable for the bad clutch guys, or vice versa, or something else?

"I consider it my biggest failure that I can’t explain what I think is the most important concept.”

Sometimes a concept is not successfully communicated because it’s complex, and it just takes time for people to get it.  And sometimes because the communicator does a poor job.  But you’re a good communicator and you’ve given it your best shot, over a long period of time.  I think you now have to conclude that the factors that seem very important to you just don’t comport with what most people believe are the relevant factors/categories.  You need to reconsider your metrics, rather than trying new ways to explain it.  I’m not saying you’re “wrong,” just that you seem to have an idiosyncratic view on this that most others aren’t buying into.

I think people want two, or at most three, metrics.  In descending order of interest:

1) a measure of “clutch” which tells us whether a hitter performed better or worse (relevant to his own talent) in situations where the performance matters more in terms of winning the game.

2) a “Clutch run production” measure that does the same thing as #1, but ignores the game score.  Does a hitter deliver more when it will produce more runs for his team?

3) Situational hitting.  Does the hitter change his approach based on game situation such that his offensive output is more efficient than average (hitting singles with runners on 2nd and 3rd, hitting homers with 2 outs and bases empty).  Personally, I’m not sure this 3rd metric is really very useful, and I’m not convinced it can be defined in a way that is clearly distinct from “clutch” and can be understood by most fans.  But I’m willing to listen. 

Does either “Clutch” or WPA/LI tells us any of those 3 things?

dq: there is some overlap with late & close and high-leverage situations.

So, studes is suggesting that Fox’s great performance came in the parts where the two did not intersect.

Fox batted .273 with none on, .317 with runners on and .307 with runners in scoring position. Maybe that accounts for part of it. And, as some said above, he may have had a hole to shoot for with runners on that led to alot of hits.

I think if you subtract RISP stats from runners on stats, you get what a guy did with a man on first only. I came up with Fox hitting .331 with a man on first only

Another aspect about this is that there were not such big differentials between close and late hitting and other situations back in the 1950s and 1960s. This could be due to the increased use of relief pitching. See

http://cybermetric.blogspot.com/2009/10/did-increased-use-of-relief-pitching.html

Guy/26: WPA/LI answers question #3.  At it’s most basic:

a) bases-loaded, bottom 9th situation:  walk = HR. 

b) bases-loaded, top of 1st situation: walk not = HR.

WPA/LI and WPA are the *only* metrics that ensures this fact.

TO THE EXTENT that a pitcher and batter are aware of the impact of a HR and BB that they would change their approach to that game state, then we need a metric to reflect that.

Clearly, the batter and pitcher change their approach based on the plate count (3-0, 0-2).  Presumably, they change their approach with runner on 3B and less than 2 outs (pitchers try to get the K, batter tries to contact the ball; possibly it cancels out overall; likely it doesn’t cancel out for certain types of pitchers and/or batters).  Possibly, they change their approach with runner on 3B, less than 2 outs, and game either a blowout or tied.

So, this is the question on the table.  This is the reality that we’re trying to model.

And WPA/LI steps up to the plate and answers that, like nothing else does.

And that’s what I’ve been trying to explain for years.

I took all the guys from the list who had 4000+ PAs. Then subtracted IBBs from PAs. Call that PA*. Then I found clutch/PA* and HR/PA*. The correlation between those two is -.296. So does this mean that HR hitters have a harder time being clutch?

Tom, can you generate a list of all the guys with 4000+ PAs in the Retrosheet era? That way we could find this correlation for all the guys who had 4000+ PAs.

I also want to raise a question for Tom or anyone else who has worked for a MLB team. Does or has clutch hitting ever affected any personnel decisions?

I don’t discuss any specific issue regarding any MLB or NHL teams.  My answer would be the same if you asked me if they look at performance on Tuesday at 15:34 as it would if you asked if they look at their career performance from 2007-2010.

I called up all the players who had 4000+ PAs from 1950-2009 (758 guys). The sum of their clutch stats is -446.6. What does that mean? That the average player is below average in the clutch? Shouldn’t it all sum to zero? Maybe being negative is okay and I don’t completely understand it.

I took all 758 of the guys who had 4000+ PAs from 1950-2009. I did not subtract IBBs from PAs this time since my guess is that IBBs count in WPA calculations. I then found clutch/PA and HR/PA and the correlation between the two is -.49. So it looks like HR hitters don’t come out so good.

For all guys who had 4000+ PAs from 1950-2009 here are ten worst in clutch/700PA

Javy Lopez -0.870
Joe Adcock -0.882
Jim Thome -0.888
Frank Robinson -0.948
Jermaine Dye -0.970
Cecil Fielder -0.978
Casey Blake -0.981
Mike Lieberthal -1.014
Lance Parrish -1.140
Sammy Sosa -1.188

27/ Nellie Fox’s difference in clutch is due to his batting average, basically hitting singles. He hit a few more triples, but less doubles, virtually same homeruns and walks.

In EVERY year, 1950-1964, his batting average was worse in Late & Close versus High Leverage. For his career, he was .051 worse.

In every year, his differential was more than the league differential - the odds should be 32,767 to 1 against that.

In 1950, his L&C ba was .171; his high leverage .362

I would guess that there is some amount of overlap in the 2 categories.

I would like to reconcile the two. I don’t know if anyone has the data for one of his years, preferably 1959 where there is also a big gap (and a Sox pennant).

I called up only the lefty hitters who had 4000+ PAs from 1950-2009. Then I found the sum of their clutch stats and it was 93.9. So lefties are positive and every one else is negative. It looks like two things that highly impact this stat are handedness and power hitting.

There were 233 lefties

There were 420 righties. The clutch stat sum for righties was -621.3. You might notice that 93.9 + (-621.3) = -527.4, not the -446.6 I mention above. There must be about 100 switch hitters and they must be positive.

I took all the righties and lefties (no switch hitters) from 1950-2009 with 4000+ PAs and ran a regression with clutch/PA being the dependent variable and HR/PA & a dummy for righties being the independent variables. Here is the equation

Clutch/PA = .000586 - .00026*Righty - .0198*HR/PA

The t-values for the two independent variables, in order, were 6.8 & 15. The r-squared was .303. The standard error works out to about .33 wins per 700 PAs.

Being a righty costs a player about .18 wins per 700 PA. A one standard deviation increase in HR rate (.014), costs a player about .19 wins per 700 PA.

I took the top 15 among lefties in clutch/PA. Then I found their averages with a runner on first (which you can get directly from Baseball Reference without any calculations like I mentioned above). Here they are. The numbers are their runner on first average, overall career avg and then the difference. They averaged about 20 points better than overall with a man on first. I don’t know for sure what the normal difference is and I did not check for any righties. Maybe I will get to that later. But in the 1990s, the average with runners on base for all hitters was about 6 points higher than the overall average.

Nellie Fox 0.331 *** 0.290 *** 0.041
Joe Cunningham 0.278 *** 0.291 *** -0.013
Dan Meyer 0.272 *** 0.253 *** 0.019
Dan Driessen 0.283 *** 0.267 *** 0.016
Craig Reynolds 0.281 *** 0.256 *** 0.025
Vic Wertz 0.283 *** 0.276 *** 0.007
Lance Johnson 0.285 *** 0.291 *** -0.006
Billy Goodman 0.330 *** 0.299 *** 0.031
Dave May 0.307 *** 0.251 *** 0.056
Tony Gwynn 0.355 *** 0.338 *** 0.017
Ozzie Guillen 0.276 *** 0.264 *** 0.012
Bobby Tolan 0.311 *** 0.265 *** 0.046
Tony Womack 0.276 *** 0.273 *** 0.003
Rusty Greer 0.333 *** 0.305 *** 0.028
Jerry Lumpe 0.282 *** 0.268 *** 0.014

I’m thinking a better way to handle biases would to do a WOWY, same batter, same pitcher, same bases/outs, varying only the leverage

The LHH / Runner on 1B issue was also reported in The Book.

Can you give me a page number? It has been awhile since I read it

Maybe another factor:
I think (no research) that clutch plate appearances are more likely to occur against an opponent’s strongest relief pitchers.  These are the guys that strike out a lot of hitters.  So a player more suceptible to K’s should fare worse than one that puts the ball in play.  At least in comparison to all plate appearances.

Excellent piece, Tangotiger!

Does any of this analysis take into consideration outcomes odds of different types of players hitting in clutch situations?  The data show that Gwynn and Rose were mega-clutch, but those hits were probably singles.  Meanwhile, Sosa, Thome, Schmidt, et al., may have gotten significantly fewer clutch hits, but couldn’t many of them been late-inning three-run homers?

Or, again, is this factored in somehow?

Thanks for the great article!

Bobby

The correlation between clutch/pa and so/pa is -.375. So yes, players who do strikeout alot fare worse.

Ham Man

It would be interesting to see if high strikeout players see an even bigger increase in their strikeout rate in clutch situations than low strikeout guys. I will look into that when I get a chance.

Cy

Bobby, this is based on WPA, and a clutch 3 run homer usually does more for win probability than a clutch single.  So that should already be accounted for.

Bobby: it’s all factored in, since the currency is in wins, not hits.

***

Cy, p. 323, here’s my Arthur Conan Doyle description:

Do certain families of batters have a particular advantage in the situation of runner on first and less than two outs, compared to their overall wOBA in any situation?
...
Does experience matter? We broke up the hitters into young (under 25), old (over 35), and the rest, which we call middle-aged. The middle-aged batters had a 16-point advantage, the old batters had a 10-point advantage, and the young batters had a 1-point advantage! This is a very interesting result. What we have here is a situation where the defense has been greatly disrupted, but that the young batters are giving that entire advantage back, as they’ve been disrupted by the runner just as much!

Thanks. So with a runner on first only, age affects how well you do compared to other times? I guess I don’t recall that fromt the book but I have it handy and there it is!

I did a study for “By The Number” once where I found that experienced players do a better job of maintaining their performance in close and late situations than inexperienced hitters. It is at

http://cyrilmorong.com/Clutch-experience.htm

I happened to have a list of players with 6000+ PAs from 1987-2001 with their OPS in close and late situations (CL) and their OPS in non-CL situations. I took the ratio of CL/nonCL. Tino Martinez did the best, with 1.095, meaning that his OPS in CL situations was 9.5% higher than nonCL. The correlation between CL OPS/nonCL OPS and SO/PA is -.364. So it looks like guys who strikeout alot have a little harder time doing well in the clutch

Also, if you go to the rankings, you can see that 10 of the 12 best players in maintaining their OPS in the CL were lefties or switch hitters

http://cyrilmorong.com/clutch.htm

And it looks like 8 of the bottom twelve are righties

Going back to the data set that had all hitters from 1950-2009 with 4000+ PAs, using only righties and lefties, here is the regression equation where clutch/PA is dependent on SO/PA and being a righty or lefty. The dummy variable is 1 for righties.

Clutch/PA = -0.000226*Righty - 0.0047*SO/PA -0.000657

Multiplying -0.000226*700 = -.158. So for a full season, a righty loses .158 wins. The t-value was -5.41.

The standard deviation for SO/PA was about .0449. So multiplying that by -.0047 leaves -0.00021. That times 700 is .148. So a one standard deviation increase in strikeout rate will cost you .148 wins.

The r-squared was .199 and the standard error was 0.0005. Times 700 leaves about .35. So the standard error of the regression is .35 wins.

But like I said in #40, it looks like righties and high strikeout guys will not do so well in the clutch.

The t-value on SO/PA was -10.55

Do strikeouts cancel out the adjustment for home runs, or do we still need something here?

I really am not sure but I don’t think so. I think the regression with HRs gave better results. I could put both SOs and HRs in the regression but I think that would not help since they are probably very highly correlated with each other. I only did the SO version because someone raised the point about guys who strikeout alot not doing as well.

You’d have to do a multiple regression, obviously. But it gives us two avenues. First, that pitchers are probably painting the corners a bit more with a power hitter in a clutch situation (the home run difference). Second, the guys who strike out all the time are not bumping up as many runners with outs, and that costs more when the game is on the line. But since the first one has a change, I’d gather that it’s the more fruitful avenue.

The correlation between HR/PA and SO/PA is .59

The correlation between SO/PA and Clutch/PA is -.40

The correlation between HR/PA and Clutch/PA is -.50

I think I read once that a rule of thumb is that if two independent variables have a higher absolute correlation with each other than either of them has with the dependent variable, collinearity might be a problem. But the numbers are close so here is the regression.

Clutch/PA = -.00025*Righty - .0169*HR/PA - .00157*SO/PA

All three variables seem to be significant

Righty -6.31
HR/PA -10.49
SO/PA -3.06

R-squared is .314 and the standard error per 700 PAs is .33. So the coefficients on HR/PA and Righty don’t change all that much from #40 and #55 but a big drop for SO/PA although it is still significant

The constant was 0.0007

Good stuff, Cyril. About what I expected, but good stuff. I wonder what else is in there, but I would think power would be the big one.

Thanks. Glad you liked it.

Sorry. In #60 I forgot to say that

Righty -6.31
HR/PA -10.49
SO/PA -3.06

were the t-values

I know we talked about how to deal with Sosa’s “clutch” and WAR, but what about pitchers?  How would a 60 WAR pitcher, for example, be affected by a -3 “clutch” for example?  IS there anyway to determine that?