Pitching to the Score

[spiritof66]

In a recent article on the subscription portion of his website, Bill James evaluated how efficiently pitchers pitch to the score of a game -- that is, "how well the pitcher has matched his effort to what was required to win the game" by distributing the runs he allows so that they are just below their run support in each game.

James looked at how well pitchers from 1952 through 2013 pitched to the score. His conclusion: "the #1 guy by far is Mike Flanagan." And the single-season leader is Mike Cuellar in 1970. Nice to see, and consistent with my own not-too-objective opinion that Flanny was an exceptionally smart pitcher who did what he had to do to put his team in a position to win.

[DrungoHazewood]

If you read through the comments under the article (unfortunately for linking purposes it's subscriber-only), even James himself says:

Not all analysis is designed to measure whether a pitcher, overall, is good or bad. This system is absolutely NOT designed to measure, overall, whether a pitcher is good or bad. Bob Gibson in 1968 comes out a little bit below average, since he did not do a particularly good job of matching his best games with his offensive support. There are a billion systems to measure whether a pitcher is good or bad. This system is not intended to make that a billion and one.

This isn't designed to tell us whether a pitcher has/had a skill in matching his runs allowed to run support. It's observing which pitchers ended up matching their runs allowed to their run support, whether or not they intended that at all.

TangoTiger has several comments regarding this:

tangotiger

Bill's method is a clever way to see how much the matching affected the outcome, regardless as to whether there's a skill there. We are treating as a given the actual RS and RA distributions, and simply rearranging them.

The question is what to do with the results, the "efficiency". As Bill has shown, we get +/- 6 wins at the extremes for one season.

Let's take a similar situation, and look at it from a catcher's perspective. Suppose we gave out W/L records not by pitchers, but by catchers. We'd look at runs allowed based on the catchers, and repeat the exercise. And we'd of course find "efficiency" numbers for them as well. So, what do we do with these results?

The point is that there's a danger that when you've controlled for as much as you can to then just give the "leftover" to the last standing variable, all the random variation that's also part of any metric will also go to that variable.

We could for example first adjust the catcher W/L record based on the pitcher. We can make sure for example that if one catcher has 120 starts and another has 40 starts, that each catcher didn't disproportionately catch one pitcher more than another. This would be a slight adjustment overall.

We'd still get to the point that the catcher, as the last variable standing, would absorb all the random variation inherent here.

Bill is effectively just giving out 10% of the net impact to the pitcher, as shown in the Lee/Hamels comparison. It's reasonable enough. But, if it can be shown that it should be 5% or 2% or 1%, then would we agree that that's the weight to use?

Ok, so this is what I did. I created a "standard" runs scored and runs allowed distribution. The 36 games were split as follows:

0 runs: 3 games

1 run: 5 games

2 runs: 5 games

3 runs: 5 games

4 runs: 4 games

5 runs: 4 games

6 runs: 3 games

7 runs: 2 games

8 runs: 2 games

9 runs: 1 game

10 runs: 1 game

11 runs: 1 game

That gives me a total of 144 runs scored in 36 games, or 4.00 runs per game. I used the identical distribution for runs allowed.

I then randomly sorted the two groups, so we have 36 random games. If you scored more than you allowed, you get a win. If it was a tie, then I just generated a random number to flip a coin to get a win. That's one pitcher. I did the same for 1000 pitchers.

Since these are all .500 pitchers, I subtracted 18 wins from each "season" to get the same differential that you got. I then got the standard deviation for these 1000 pitchers.

The first time I ran this, I got one SD = 1.59.

The second time was 1.85. Then 1.69, 1.65, 1.94, 1.84, 1.67, 1.77. 1.78, 1.89.

The average of these 10 runs of 1000 pitchers was one standard deviation = 1.77.

So, just by pure luck, we should expect "pitching to score" to have one SD = 1.77 if there was no such thing as pitching to the score.

In your case, we only get to 1.60 standard deviation.

The second quote basically says that the "pitching to the score" effect is basically what we'd expect from random variation.

This article most definitely doesn't say or prove that some pitchers are better than their records because they pitched to the score on purpose.

[Frobby]

Wish I could read the article, sounds interesting.

This article most definitely doesn't say or prove that some pitchers are better than their records because they pitched to the score on purpose.

Nobody pitches to the score on purpose. But perhaps there are guys who are better able to bear down in a close game, and/or who tend to get a little cavalier with a big lead. In a sense, there is no way to really know the answer. All you know is the result, not whether it was caused by some trait of the player. It would be interesting to do an experiment like Tango's involving RISP hitting and see what the outcomes were.

[TakebackOPACY]

When you say, "Nobody pitches to the score on purpose.", my first reaction is "Yes, they do." Then I stop to think what you're really saying and think we probably agree.

The pitchers are always trying to get the other team out and aren't trying to let the other team score. So, nobody pitches to the score on purpose I guess.

But as soon as you take a different strategic approach on how to go about that based on the score (and other variables), aren't you pitching to the score on purpose?

[DrungoHazewood]

When you say, "Nobody pitches to the score on purpose.", my first reaction is "Yes, they do." Then I stop to think what you're really saying and think we probably agree.

The pitchers are always trying to get the other team out and aren't trying to let the other team score. So, nobody pitches to the score on purpose I guess.

But as soon as you take a different strategic approach on how to go about that based on the score (and other variables), aren't you pitching to the score on purpose?

The way I look at this is that it's not trivial to retire even the worst MLB hitters, so you have to pitch near similarly no matter the situation. This is especially true in modern baseball where there is less spread in talent than in the past. It may be true that the Pitching to the Score myth has its roots in the same deadball era tactics that included a lot more pacing than is possible today. I assume that pitchers would let up significantly in a 7-0 game in a era where you could lead the league in homers with 11. Today if you let your foot off the gas even a little that 7-0 lead can quickly evaporate when an average 9th-place AL hitter has 11 homers per 600 PAs.

[Ohfan67]

Fun topic. Thanks to spirit for sharing.

Many factors contribute to things like "pitching to the score" effects. For example, a manager is much more likely to leave a struggling pitcher in a game with a big lead than when the game is low-scoring and tight. Managers may trust older pitchers more than younger pitchers when the opposing team is closing the gap. The depth of the bullpen may affect how long the manager sticks with the pitcher. I am not a subscriber so I didn't read the article. Did James control for these variables?

[Pheasants]

I think there also is a fielding to the score factor inthat with a five run lead, you won't care so much about a guy tagging on a sacrifice fly or scoring on a grounder when the easier play is to first base. The outfielders may be positioned differently to prevent the big inning, etc.

[Frobby]

Fun topic. Thanks to spirit for sharing.

Many factors contribute to things like "pitching to the score" effects. For example, a manager is much more likely to leave a struggling pitcher in a game with a big lead than when the game is low-scoring and tight. Managers may trust older pitchers more than younger pitchers when the opposing team is closing the gap. The depth of the bullpen may affect how long the manager sticks with the pitcher. I am not a subscriber so I didn't read the article. Did James control for these variables?

I think there also is a fielding to the score factor inthat with a five run lead, you won't care so much about a guy tagging on a sacrifice fly or scoring on a grounder when the easier play is to first base. The outfielders may be positioned differently to prevent the big inning, etc.

Great points.

[Ohfan67]

If you read through the comments under the article (unfortunately for linking purposes it's subscriber-only), even James himself says:

This isn't designed to tell us whether a pitcher has/had a skill in matching his runs allowed to run support. It's observing which pitchers ended up matching their runs allowed to their run support, whether or not they intended that at all.

TangoTiger has several comments regarding this:

The second quote basically says that the "pitching to the score" effect is basically what we'd expect from random variation.

This article most definitely doesn't say or prove that some pitchers are better than their records because they pitched to the score on purpose.

I don't have time right now to dig into the stats, but I think tangotiger's statistical rationale is really wrong. I think his random number generation is not equivalent to James' analysis. Comparing standard deviations of randomly generated distributions and distributions of the real data doesn't tell you anything (except they are equally variable and even then comparing SD is not the valid test for comparing variation). I think that guy's comments are incorrect. I have to stop thinking about this now because I have to help a graduate student with some statistical analyses.

[DrungoHazewood]

I don't have time right now to dig into the stats, but I think tangotiger's statistical rationale is really wrong. I think his random number generation is not equivalent to James' analysis. Comparing standard deviations of randomly generated distributions and distributions of the real data doesn't tell you anything (except they are equally variable and even then comparing SD is not the valid test for comparing variation). I think that guy's comments are incorrect. I have to stop thinking about this now because I have to help a graduate student with some statistical analyses.

I think it's a good test of any hypothesis to compare it to what you'd get with purely randomly distributed results. It doesn't definitively prove there is no such effect, but it casts doubt when you can model a supposed skill with a random number generator. For example, a baseball simulator with no clutch factors built in whatsoever (maybe OOTP or Diamond Mind) will generate players who have great clutch stats over a 20-year career at basically the same frequency as in real life, and that is purely random. That indicates to me that clutch hitting is largely just random variation.

I think it is incumbent upon the people suggesting a phenomenon to produce evidence that the thing exists outside of randomness. If MLB pitchers do pitch to the score someone needs to explain why you get the same results from an emotionless, clutchless, human factor-less simulation.

[AdamK]

I learn way more about baseball in the offseason. Or I should say, I'm willing to digest more about baseball when baseball doesn't get it the way.

Thanks for the content.

[SteveA]

The way I look at this is that it's not trivial to retire even the worst MLB hitters, so you have to pitch near similarly no matter the situation. This is especially true in modern baseball where there is less spread in talent than in the past. It may be true that the Pitching to the Score myth has its roots in the same deadball era tactics that included a lot more pacing than is possible today. I assume that pitchers would let up significantly in a 7-0 game in a era where you could lead the league in homers with 11. Today if you let your foot off the gas even a little that 7-0 lead can quickly evaporate when an average 9th-place AL hitter has 11 homers per 600 PAs.

True.

But you do hear a lot of talk about things like starting pitchers learning not to throw all out in the early innings so that they can still pull a 98 MPH heater out of their pocket in the late innings, a la Justin Verlander (and we've seen Gausman do it a few times).

So the idea that a pitcher is always at max effort 100% of the time does not necessarily ring true to me either.

[DrungoHazewood]

True.

But you do hear a lot of talk about things like starting pitchers learning not to throw all out in the early innings so that they can still pull a 98 MPH heater out of their pocket in the late innings, a la Justin Verlander (and we've seen Gausman do it a few times).

So the idea that a pitcher is always at max effort 100% of the time does not necessarily ring true to me either.

Starters may not go fully 100% all the time. But I think that over time, especially in last 20 years or so, pacing has pretty radically changed. Used to be you could back things off and throw in the low 80s and you'd be about on par with all the Scott McGregors. Now if you were to throw 82 for a few innings Fangraphs would have an article on it in two days speculating that you tore something. I don't recall ever seeing data that showed starters throwing with less velocity depending on game condition, and we have the data.

[Ohfan67]

I think it's a good test of any hypothesis to compare it to what you'd get with purely randomly distributed results. n.

I totally agree, but I don't think that's what he really did. At all. He didn't compare the runs given up by pitchers with a large versus small lead to a random distribution of potential runs given up or something similar.