THE BOOK--Playing The Percentages In Baseball

I tried to write someting to Rob using either the comments section or the mailbad and it didn’t work; in lieu of that I’ll post it here so at least someone reads it:

*****

Dear Mr. Neyer,

My name is Colin Wyers and I’m writing in response to your blog entry on Tango’s wOBA, specifically:

“Meanwhile, the Baseball Prospectus metrics are—last I checked, anyway—completely opaque. Because BP is a for-profit enterprise that holds its intellectual property closely, we’re just supposed to trust them.

Which I do, generally. But that science-vs.-enterprise dynamic can be tricky. The methodology behind BP’s metrics is not, to my knowledge, peer-reviewed. If one or two people make a big mistake, would anyone else know? Now, let’s jump ahead and say that two or three years down the line, the big mistake was discovered internally. Would BP announce to the world that all those numbers over the previous three years had been wrong? Or would the guys running the show decide that the loss of credibility (and potentially, revenues) isn’t balanced by the loss of integrity?”

This is not - precisely - true, and your hypothetical scenario isn’t quite hypothetical. Essentially, everything that makes up EqA has been published by Baseball Prospectus in some form or another at some point in time - I can and have figured it (and MLVr, the rate stat that forms the basis for VORP) on my own using the published explanations by BP writers. Certainly it’s a bit obfuscated (I once spent an entire afternoon crawling through the old rec.arts.baseball USENET archives for Woolner’s original writings on VORP) but if the construction of BP’s offensive metrics is supposed to be a secret at some point they slipped up.

The problem is that there are issues with the methodology behind both of them, MLVr moreso than EqA. The issues with EqA are a bit difficult to explain without some pretty hairy math that I’m not the best person to explain, and aren’t a major issue. The problems with MLVr are more deep-rooted and unsettling (MLVr criminally underrates the walk, for instance). None of this is exactly secret - I know Tom Tango, Patriot and myself have written about several of these issues in the past, and I haven’t gotten a response from anyone at BP on the matters. Clay Davenport seems to be responsive to some of these concerns - Tango says he’s talked to him before about it. You’d have to ask Tom to be sure, but I don’t think anyone’s said anything substantial about fixing MLV/VORP’s issues.

I don’t want to try and impugn anyone’s motives here - I don’t pretend to know why BP isn’t responding to these sorts of concerns. (If I had to hazard a guess, I’d say it’s because a lot of the people best qualified to deal with these questions have matriculated - either to MLB teams or in Silver’s case to his career as a political analyst.)

In the end, I’d say the big benefit of wOBA over EqA is that the math is a lot simpler - calculating wOBA is simpler than EqA, and figuring a player’s runs from wOBA is much, much simpler (to do it the “correct” way with EqA requires using exponents and is a bit sticky). You do lose some of the nice benefits of the Davenport Translations, like park factors and league adjustments.

As far as non SB/CS baserunning metrics, there are quite a few out there - BP has one (probably the best one out there), although it’s not very well integrated with their other offerings. Bill James Online has one. Lee Panas publishes some on his Tigers blog. I’ve dabbled in it before myself. I think this is one of those things that simply hasn’t percolated all the way yet, and will be pretty common in one or two years - the data and the methods are both available.

Cheers,
--CW

Someone with an ESPN account, might want to tell Rob that the 15 run difference between Pujols’ wOBA runs and his Batting runs, is due to the fact that Batting runs treats IBB as regular walks by weighing them at .33 runs, while wOBA doesn’t assign them any run-value.

Pujols was intentionally walked 34 times. Right there, is an 11 runs difference between batting runs and wOBA.

Listen to this nimrod criticize wOBA on Rob’s blog. How do you respond to this kind of stupidity?

jslevin00

“I agree with the parts about wOBA not being very intuitive. At this point, I think more fans are attuned the value norms of OPS than OBA, so scaling this stat to OBA just seems like a foolish effort. I can’t think why they didn’t scale it to batting average, except perhaps to avoid comparison to EqA, which would not be a good reason.

I look at OPS+, which has the virtue of being easily understood. wOBA just doesn’t have that, and while I like the fact that it translates easily into runs, how many will actually make use of that, even among diehard fans? Very few, I think.”

Another nimrod. He calls wOBA a refurbished EqA, and then praises Marginal Lineup Value which uses basic Runs Created:

“wOPA sounds like an interesting diversion, and i bet it ranks players in pretty much the right order. My problem is that it doesn’t correlate to anything actually happening on the field. It’s a refurbished Eqa.

One stat BP doesn’t give nearly enough attention to is MLVr. It measures the number of offensive runs a player would add to an average lineup per game—Pujols led the majors last year at .621 and Chipper was second at .522. (There’s also PMLVr, which is adjusted for position.) A stat that tells you that Pujols is 6/10 of a run better per game than an average hitter is one that has tangible meaning and that a casual fan can understand.”

When I was a kid in math class I used to get my work done very quickly, sometimes in my head.  I used to get annoyed at teachers who wouldn’t give full credit unless you show your work.

Now I can understand what they were doing.  Outside of Pete Palmer and Bill James, nobody has a right to claim “intellectual property” over something as simple as a run estimator.  James released all his stuff to the public, I don’t know the circumstance with Palmer’s methods but he doesn’t seem to be bothered by others using them.

Oh yeah, David Smyth too.  Baseruns is a legit advancement.  Everything else is tweaking.

I thought Rob was referring to Palmer’s Batting Runs. Equivalent Runs over-values the walk even more than Palmer’s Batting Runs does. That would explain most of the 15 run-difference. When Patriot applied the partial derivative approach to EqR, the implicit LW value for the BB was .347 from 1990 to 2005.

The pecuilar thing about EqR is that the batting events are linear, but base-stealing is dynamic. It is like some sort of mutant-hybrid. Patriot has shown that prolific basestealers distort the linear portion of the formula. The more stolen-base attempts a player has, the more the linear weights for the batting events go down.

Piggybacking on Colin/1, of all the BP stats, EqA is the one that is most out there in the public eye.  We know exactly how it’s figured, at least prior to any sort of league adjustment/"translation".

While I have no beef with wOBA, and I understand and am on board with its use for applications like the studies in The Book, personally I would still rather see RAA/PA, as I prefer to keep the runs unit whenever possible.  Of course, wOBA is directly related to RAA/PA, and I in no means want to impose my preference on anyone else.  Better wOBA than EqA, OPS, OPS+, etc., and always good to see recognition for Tango.

I just put up a piece on StatSpeak on the subject of this issue around intellectual property.  I think this is a bigger issue than the minutiae of what BP has released and what they haven’t.  Let’s say that in my mom’s basement, I come up with the Pizza Cutter super-secret formula to predict baseball performance.  I release my numbers, but not my method, and the numbers end up doing a pretty good job.  Not perfect, but pretty good.  I start charging for access to it.  Do I have a right to do that?  In the free market, I sure do!

Now suppose that eventually, I released my method and let others critique it.  Someone here could make it better.  That’s the beauty of peer review.  I can’t charge for it any more, for the twin reasons that it will become a new creation with other’s ideas incorporated (so assigning credit will be impossible) and it’s open source anyway, so anyone could just run it off for free.  It would be a better system, but I’d lose my meal ticket.

So are we as a field OK with the thought of people making money at the cost of possibly stagnating the progress that could be made if I just released my creation into the public?

I respond in its own thread here:
http://www.insidethebook.com/ee/index.php/site/article/the_history_of_the_woba_part_1/

Also, I usually give a pass (once) to those who disagree without basis (like some of the comments that are being quoted here).  I do this on the idea that the commented is unintentionally misinformed or ignorant of the facts.

On the other hand, if faced with the facts and the commenter is now making foolish statements, then a punch (above the belt) is acceptable.

There is nothing that Sean Forman does that can’t be replicated.  Yet, there it stands, a monument of hard work and inspiration, with people contributing ideas to make it even better.

There is nothing that Davide Appelman does that can’t be replicated.  Again, hard work and inspiration and contributions of readers and fans.

Why in the world would BP think that releasing the innards of their work will spell its downfall, considering that people are not paying for the stats to begin with?  (PECOTA is a separate issue.) If anything, having nuts like me suggesting improvements for their work and advocating the results would be a plus.

So, I disagree that the natural outcome of BP releasing their work is that a competitor would simply eat into BP’s profits.

This is Pizza’s post:
http://statspeak.net/2008/11/neyer-discusses-woba-and-some-sabermetric-philosophy.html

He also says that he trusts others at BP double-checks the work.  This is not the case, certainly not all the time.  It’s more like a federation than a corporation.

PECOTA is a separate matter.  I am talking about metrics that they are currently not charging money for, and yet, are not completely open about how they calculate those metrics.

What is it about EQA or MLVr that is not open source?  I don’t have much use for those methods, as they are overly complicated ways to get inferior results, compared to LW.  But the formuli have been written about in their books.

It’s even in their dictionary:
http://www.baseballprospectus.com/glossary/index.php?ltr=E&context=alpha

So is the issue just something like not knowing what park adjustments they are using?

That’s a good point. 

I think it’s a combination of things.  One, for sure, is the park (and league) adjustments.  Two (where it applies), the replacement level.  Three, and likely the most important, is that it is fairly complex to code in a couple of minutes.

A step-by-step calculation would be nice, so that we can give it the fair scrutiny it deserves, and that everyone is looking at the same thing at the same time.

Woolner has published the specifics of MLV, and Davenport EqR. In order to compute the exact same numbers as BP, then yes, it would be important to know their park factors.

But from where I sit, I don’t think it’s important to tie one specific offensive metric to one specific park factor. If you compute wOBA with no park factors and I compute wOBA with run park factors and someone else computes wOBA with component park factors, they’re all still wOBA.

In this case, I think it’s important to distinguish between method and implementation. The methods for MLV and EqR are publicly available (if not exactly well-advertised; MLV and VORP are available on Woolner’s old website, which BP still links to). Everything I need to calculate VORP the way BP does is published (run estimator, replacement baseline, etc.) It’s up to me to roll up my sleeves and implement it myself if I want to.

And here’s a start - my full code for computing MLV and EqR, according to BP’s published accounts:

http://basql.wikidot.com/baseball-prospectus-run-estimators

I have no doubt this could be improved upon. But at that point you’re halfway to computing VORP.

At a sufficient level of complexity, in order to get method you need implementation. I’m not asking for the full details of PECOTA - without a crash course in VBA macros and STATA and so on I’m not sure I’d be qualified to comprehend all of it anyway. I expect a certain level of transparency and some benchmarking and I’m satisfied. Same with the Davenport Translations.

The IB issue is not handled real well by many of the various metrics.  It is pretty simple how it should be handled.  It should NEVER be treated as a regular BB (or HP) for obvious reasons.  At least I think they are obvious.

We don’t exactly know the value of the IBB in win value, which is really the only thing we care about, especially for such a context-dependent event.  We do know, however, that whether to IBB a batter or pitch to him is close to a toss-up in just about every situation.  Therefore we know that the win value of the IBB is right around the same value as an average PA for that batter.  IOW, the value of the IBB is completely different for each batter.  But the nice thing is that since for every batter it is about equivalent to an average PA, we can ignore it in a rate stat, but include it in a counting stat.  So, when computing wOBA, we can definitely ignore it.  However, when turning wOBA into runs, as long as we use the number of PA that includes the IBB, we are OK.  So you just have to make sure that when computing a counting stat, even total linear weights runs, you have to use the number of PA that includes the IBB.  If you must set a value to the IBB in your calculations, simply use the average value of a PA for that batter.  That would even apply to weak batters in the #8 slot in the NL who get a lot of IBB only because they are hitting in front of the pitcher.  Using a positive value for those IBB’s, especially the value of a regular walk (around .31 runs) is ridiculous of course.

Therefore we know that the win value of the IBB is right around the same value as an average PA for that batter.  IOW, the value of the IBB is completely different for each batter.  But the nice thing is that since for every batter it is about equivalent to an average PA, we can ignore it in a rate stat, but include it in a counting stat.  So, when computing wOBA, we can definitely ignore it.  However, when turning wOBA into runs, as long as we use the number of PA that includes the IBB, we are OK.

EXACTLY.  This is how I handle it in wOBA and in the conversion of wOBA to runs.

What is wrong with including intentional walks in a run estimator at their correct marginal run-value, usually between .16 to .18 runs?

“But the nice thing is that since for every batter it is about equivalent to an average PA.”

No, an intentional walk is worth .16 to .18 runs more than an average plate-appearance. If we reconcile to runs created by adding the average run-value of a plate appearance to all PA events, then the IBB is worth .28 to .30 absolute runs.

It depends what you are after.

If you only care about run creation at the team level, then the run value of the IBB is around .18 runs above the average PA.  End of story.

But, if you care about the IBB as it relates to player evaluation, then you need more.  You need to know how to links to wins.  And, we know that the average IBB adds around +.01 wins per PA (compared to the average NIBB which adds +.03 wins per PA), above average.

Furthermore, if the IBB is only given to players who in the same situation would have gotten +.01 wins per PA, then the IBB is really win-neutral.  That is, if the IBB is only given to great hitters, and if great hitters are worth +.01 wins per PA, then the win value of the IBB is really +.00 wins above the average PA for that player. 

Even furthermore, if MGL’s thesis holds, then the IBB is issued “smartly” in such a way that it adds more than +.01 wins to the really good hitter and it adds less than +.01 wins to the bad hitter, such that the idea as to whether to issue an IBB is really such a break-even call, RELATIVE TO THE HITTER IN QUESTION, then again, it’s a win-neutral event.

Think about it…

We only care about win value, not run value.  We use run value as a proxy for win value.  In the case of the IBB, we don’t want to do that (use run value).  While sometimes the IBB raises the win value of the batting team, sometimes it lowers it.  The presumption is that if managers use the IBB in some reasonable fashion, that the marginal win value is exactly zero.  You definitely do not want to use .16 or .18 for an IBB.  Intuitively, you should realize that that is incorrect.  If I have a player who is around league average per PA, why would I give him extra value for any extra IBB’s he gets beyond league average?  Presumably, if an average player were to get a lot more or less than the league average IBB, it is because he happens to come up a lot or a little with bases open, or he happens to bat in front of a very good or very bad player.  We don’t want a player’s context-neutral lwts to be heavily tied in to context if we can at all help it.

Regardless of what the run value of an IBB is, we don’t want to include that in a player’s lwts.  The simple reason is that we really want the currency of lwts to be win value, but we only use run value because it is convenient.  And the marginal win value of an IBB is around zero.

It is the same idea, or at least similar, with the sac bunt.  The run value is defimnitely negative, but we don’t want to penalize a player for two reasons:  One, the win value is not necessarily negative (and players must sometimes give up run and win value with a sac bunt attempt to gain run and win value when they are not bunting in a bunt situation - game theory), and two, it is an elective strategy decided by the manager, and has nothing to do with the player other than bad hitters tend to bunt more often (we would actually be penalizing good bunters if we included the sac bunt in a player’s lwts).

You must always keep in mind what you are trying to capture in a metric.  For lwts, we are trying to capture a player’s theoretical contribution to winning based on the voluntary outcome of his PA, independent of all context, so that we can use this information to determine which of more than one player is the most valuable (in terms of theoretical wins or marginal wins), assuming an average team in an average park and league, with an average manager, surrounded by average players, in an average lineup slot, etc.

I will admit that I have not read “The Book.” I will head over to Amazon and read the chapters about the intentional walk and sacrifice bunt.

I don’t know that there is anything in The Book to shed any more light on the subject of how to handle the IBB in a lwts (or other) formula.  We have to presume that the marginal win value of the IBB is zero at best, otherwise why would a manager issue one.  If the marginal win value is zero, then the lwts value has to be zero, since, as I said, we are really trying to capture win value and not run value in a lwts (or similar) formula.  The reason we use run value and not win value is that it is more practicable and that in almost all cases (other than the IBB and maybe sac bunt), on the average they are one and the same.

Of course you still need to read The Book!  Especially as a quality poster on this site, which you are!

The easiest way to Park adjust wOBA is to multiply it by the square root of 1 divided by the park adjustment:

wOBA+ = wOBA * SQRT(1/PADJ)

The “Park Adjustment” is the Home + Road Park Factor. If you correct the initial Park Factor calculation for the offset of other parks in the league,

Corrected PF = (PF * N_TEAMS) / ((N_TEAMS - 1) + PF)

then the Park Adjustment is simply:

Park Adjustment = (Corrected PF + 1) / 2

mgl/19 --

I wonder, however, if the distinction between intentional and non-intentional walks is as black-and-white as might first appear.  Batters, especially power hitters, are apt to be pitched around with men on and first base open.

Last year, excluding times when only second and/or third base was occupied, Ryan Howard had 571 plate appearances (ex-IBB).  He walked (again, ex-IBB) 48 times—8.41%.

When there was a man on second and/or third only, Howard had 16 NIBBs in 112 plate appearances (ex-IBB)—14.29%.  Had he walked at the same rate as the other base situations, he would have had 9.4 walks.

So, is it fair to say Howard had approximately 7 sort of intentional walks last year?  Should they be backed out?  (I know that NIBBs receive a lower weight than HBPs precisely because of their non-random nature, but this treats Howard the same as Chin-Lung Hu, which doesn’t seem right.)

MGL,

I appreciate you, Tango, Andy, and your publishers letting us search through “The Book” in it’s entirety at Amazon.com.

If I am only interested in measuring “run-value”, not “win-value”, how would this change your position on not including the IBB and SH? Of course, even from a “run-value” perspective, including the IBB and SH does make the formula less context-neutral.

Informative, fascinating discussion, gentlemen.  Thank you very much.

I’m glad you enjoyed the discussion.

Most of the complaints about wOBA are that it isn’t park adjusted. It is not difficult to park-adjust wOBA. You would park-adjust wOBA the same way you park-adjust OPS, using the square root of the park adjustment.

I’m still trying to digest MGL’s comments about the intentional walk and sacrifice hit. I can’t get it out of my warped mind that a run-estimator should include all plate-appearance events, or at least include all the available ones. This way everything comes as close as possible to adding-up.

Actually, wOBA is easier to park-adjust than OPS. With OPS, you need park-adjust OBP and SLG seperately.

Regardless of what the run value of an IBB is, we don’t want to include that in a player’s lwts.  The simple reason is that we really want the currency of lwts to be win value, but we only use run value because it is convenient.  And the marginal win value of an IBB is around zero.

MGL, I don’t think this is correct.  While each individual decision to intentional walk has an added win value of near zero and therefore strategy wise the decision to intentional walk is nearly a wash, the win value for the batter’s team does increase, but at a rate that is nearly equal to the value that would be expected BY THAT BATTER, not by the average batter.  The team giving up an intentional walk is willing to give up win value and run value by giving up the intentional walk becuase they believe that that batter, if allowed to bat, would cost more win and run value. It is the batter’s skill as a hitter that is being acknowledged by the intentional walk and that skill should be included in linear weights at a value of .10 runs above average, which is the actual average rate that an intentional walk increases scoring.  If you give him no run value for the intentional walk then for that at bat you are treating him as an average batter, which he is not.

Peter, yes, I may have worded it incorrectly, but I already explained that the value of an IBB is approx. equal to the value of that player’s average PA, not an average player’s PA.  Which is why you want to include them in a “counting” run estimator like total lwts (or BaseRuns) or simply converting wOBA to runs or marginal runs.  We are in complete agreement.

Terps, there is simply no reason to want to use “run value” for the IBB or SH, unless for some reason you literally want to estimate how many “runs” (and not wins) a player has theoretically added to his team by virtue of all of his PA (which would be misleading since most people assume that there is a linear relationship between runs and wins and most people also assume that a player’s “value” is a function of how he helps his team to WIN GAMES, not create runs simply for the sake of creating runs).  If that floats your boat, then by all means use whatever the run value of an IBB (and SH) is. Since it adds a runner, the run value usally goes up.  Then again, if the next few batters are poor enough hitters as compared to the IBB’d batter, then the run value may NOT go up, so as long as you are NOT being context-neutral, you might as well figure out the run value of each IBB given the number of outs and the offensive level of the following batters (and the baserunners).  Then of course you are moving towards RPA (run probability added), which no one uses, since no one really cares about runs unless it is a good proxy for wins.

I don’t know what else to say about the IBB.  The way we typically use a lwts formula requires us to treat the IBB as a neutral event, relative to that batter’s average PA, which it is.  A neutral event will not change a batter’s rate value, but will change his total value, which is what I have been trying to say.  It is no more or less complex than that.  Not to mention the fact that the batter himself is not doing anything at all when he gets IBB’d. It is not really part of his skill set. Even if an IBB typically had a positive or negative marginal win value over and above an average PA for that batter, it would still be somewhat debatable whether we would want to include it in our “assessment” of that batter.  The reason being that one year a batter might be IBB’d 100 times (like Bonds) and the next year, even though his offensive talent may not have changed that much, he might be IBB’d only 30 times.  Anyway, that is a different story altogether.  If an IBB were positive or negative, as opposed to neutral, I would have no problem including it in a rate stat, even though the player himself is not doing anything to create that event, other than the fact that his offensive talent, or at least other managers’ perception of it, is a catalyst for that event.  Heck, if you could reduce a very good batter’s marginal win value by IBB’ing him a lot (which you can’t really), you might as well include that reduction in his lwts, no matter what you were using those lwts for.

Bottom line is that you must view everything within the context of its utility - IOW, what you are using it for.  Without that context, a stat like lwts means nothing.  You can’t really discuss a stat without an understanding of its purpose.  And that purpose must usually have some practical value, or it is not very interesting (like a RPA or a run estimator that is not congruent with wins).

MGL - As long as the rate stats that you are using for the player have the number of IBB’s subtracted from the number of PA’s in the denominator as well as 0 value for the IBB in the numerator then we are in complete agreement.  Is that what Tango does for wOBA?

Peter, right we are in agreement.  IBB is treated as a non-event in wOBA.  So, Pujols is a .450 wOBA, without considering how many IBB he has in either the numerator, or denominator.

Then, to convert to runs above average, we mulitiply by PA, which includes IBB and SH.

***

Terps: you can calculate the RUNS, while including IBB however you want.  But to convert to WINS, now each player will have his OWN runs to win converter (because now each run is not worth the same).  Do you really want to do that?

If we treat the IBB as a non-event in the marginal run total, then there is no way to account for it in the runs created total if the runs created are reconciled using R/O. Most people, including me, still use R/O runs created. I still use R/O RC because I don’t think it is appropriate to use R/PA RC with Baseruns.

As for each player having his own custom runs to wins converter, isn’t this the case with Theoretical Team Baseruns since each player has his own set of custom linear weights.

I will admit that I know very little about the relationship between runs and wins. I know what a Pythagorean record is. That’s about all I know. So I can’t answer Tango’s question: Do you really want to use a different runs to wins converter for each player?

The reconciliation between LWTS and Total Runs is described in the “how to calculate wOBA” thread.

I’m not really concerned about how “most people” use RC/out.  I’ve laid out my case as to why I do it the way I do it.

Tango,
I know the reconciliation between LWTS and Total Runs was described in the “how to calculate wOBA thread”. Remember, I was there arguing with you the whole time

If you want to use Baseruns, then you need to use R/O RC. If you are using Baseruns to generate Custom LW, you can subtract out the R/O using a stat that I devised called the R/O rate, and get marginal run-values. The R/O rate is:

Event R/O Rate = Event R/O / Average R/O of all events in the dataset

When you calculate your Custom LW, the R/O rate is multiplied by the Average R/O of all events in the new dataset.

New Event R/O = Event R/O Rate * New Average R/O

Multiply the above number by each events “c” coefficient, and that is the number you subtract from the RC run-value for each event to arrive at the LW run-value for each event.

We do not know the exact proportion of NIBB that are intentional. I would bet that this number is pretty close to the total number of IBB. Why not COMBINE the two values, and use the same LW value for all walks. Say a team pitches around Manny Ramirez in the 4th inning with runners on 2nd and 3rd and 2 outs. The opposing team is leading by 2 runs. They walk Manny on 4 straight pitches. It so happens that the next game, he steps-up to the plate in the exact same situation, with his team trailing by 2 runs and runners on 2nd and 3rd with 2 outs. This time the opposing team decides to intentionally walk him. There is absolutely no difference between the 2 walks, other than how they are accounted by the official statistics. The first walk was an implicit IBB, the second walk was an explicit IBB.  Over the course of a 162 game season, Manny’s “intentional NIBB’s” are probably close his number of IBB’s.

The same things can be said about the SH. There is no difference between the batter advancing the runner on a ground-ball in one plate appearance, and then doing it with a sac-bunt in another plate appearance. Assuming the base-out states are equal, they accomplish the same thing. Why not treat the SH as a regular batting out in a LW formula?

terps - The reason SH and IBB are singled out is that the batter has no choice whether he gets to be in control of his at bat or not.  In the case of the SH his manager takes that away from him and in the case of the IBB the pitcher’s manager takes it away from him.  In the case of pitching around the intention is not to walk the batter.  If the batter swings at an out of zone pitch or the pitcher gets a strike call on a pitch in the black then the pitcher may use his advantage to get the batter out rather than walk him.  The IBB strategy and the pitching around walk strategy are used in measurably different game situations and should not be considered equivalent in run value or win value.

Tango is handling the SH and the IBB in exactly the correct manner.  It is apparent that you are very interested in the technical details of how different metrics work and how baseball decisions are made.  It would really be to your benefit to hold your critcisms until you receive your own copy of The Book and have read it over thouroughly.  Although there are things in The Book that can use your critical point of view, you must take the time to understand the author’s methodology first.

The only thing I want to add is to make sure you understand what you are trying to accomplish, and what *I* am trying to accomplish.  And that while we may do things that overlap, there are (probably legitimate) reasons that I do things the way I do.

There are of course some unintentional-intentional walks, or PAs where the batter may have a strike on him, and the manager then calls for the IBB.  There’s only so much I can do with those.

Terps, I’ve always thought that the win value of the average NIBB for the better batters is probably less than that of the worse batters, for the reason you mention (they tend to be pitched around when the win value of the BB is lower).  I’ve always wanted to see how much lower it is.  I suspect that we need to use different run values for NIBB’s based on something like a player’s SLG average or even just his overall lwts (without the BB’s).

The SH is really a different entity.  As Peter says, it has almost nothing to do with the batter and is very context dependent.  No matter what the outcome, it really should be ignored.  I actually ignore all sac bunt attempts, and not just official SH’s.  The reason is that some hits and outs are the result of sac attempts.  Even though the IBB is also at the opposing manager’s discretion, it is OK to include it in a player’s “skill set” because how many IBB’s a batter gets is a function of how good he is (at least opposing managers’ perception of that), at least in the slugging department.  That is why we include it in a batter’s counting RC or lwts. But, the bottom line is that it is around a neutral event relative to his average PA.

Again, keep in mind that we generally are using lwts or other RC formula in order to estimate batter skill in terms of their win value to an average team in an average context. And then we often use that to establish a context-neutral projection. Keep that in mind and go from there when making decisions about how to treat the various events.  If you do that, things will become fairly evident.  If you start with the premise that all “runs created” have to be accounted for, etc., your decisions will not be congruent with trying to estimate batter skill and in constructing projections and being able to compare batter “worth.” Again, the ONLY thing that matters in terms of batter value (assuming that by “value” we mean helping their team to win games) is wins - not runs.

Right, what MGL said.

I think it’s ok to want to account for all runs that a team has scored.  But, that, by itself, doesn’t really answer the question we are usually interested in.  Almost always, the question, at its core, relates to winning. 

While almost always, winning and scoring and preventing runs are interchangeable, the key word is “almost”.

(Funny: the CAPTCHA word is “always”.)

You have at least convinced me to exclude sacrifice bunts but from a player evaluation formula. In a way, the sacrifice and intentional walk are polar opposites. The decision to sacrifice is entirely at the discretion of the batting team’s manager, while the decision to intentionally walk is entirely at the discretion of the opposing team’s manager. What does CAPTCHA mean?

I shouldn’t of referred to the posters in #3 and #4 as nimrods. But I am an asshole and really couldn’t help it.

Responding to the poster I quoted in #4, I too like the concept of MLV. However, I do not care for the run-estimator that MLV uses.

MGL - I looked into dividing unintentional walks into two categories.  I think I was using discretionary walk to refer to walks that I thought were the result of pitching around.  The problem was that they were not easy to define.  The two most significant variables seemed to be SLG and batting order position.  But game score differential was also a factor and so was the inning.  Leadoff and #2 hitters got almost no discretionary walks because they were almost always followed by batters that were more likely to drive in runs.  #5 and #6 batters seemed to get quite a few because they were often followed by batters that were quite a bit worse at driving in runs.  I think I was figuring that discretionary walks were something like 20 to 25% of all unintentional walks and had a run value around .25 to .27 runs and the non-discretionary unintentional walk was around .33 to .35 runs.  Leadoff guys gained a couple of runs a year and #4 and #5 batters lost a couple.  Seemed hardly worth the effort so I shelved the research.  With each run worth $400,000 to $500,000 a year it probably would be worth the effort if you were helping a player negotiate a contract.

Peter, makes sense.  What about extreme players like Bonds?  What do you think the average value of his NIBB’s are?  I suppose you could group all NIBB into two categories, but I would not do that.  It is a continuum I think, or at least there are several categories.  I think the first thing I would do is to establish the average value of a walk to a poor batter and/or one at the top of the order and the average value for a high slugging batter who is often IBB’s and/or who bats in the middle or the bottom of the order.  Then I would go from there.  If the difference is not that great, then I would probably not worry about it.

MGL - Yes, to establish a baseline I looked at walks to pitchers and walks to lead off batters in situations that I knew the pitcher didn’t want to walk the batter.  One of the valuation methods that I considered was to give the batter the average value for a walk for the particular base out situation he faced and the particular lineup position he was hitting in calculated by the number of runs that actually scored from that situation for the league.  That would have been a continuum that roughly approximated what the batter’s walks were actually worth, but might not have represented accurately a batter’s true talent, i.e. a repeatable skill.

One of the valuation methods that I considered was to give the batter the average value for a walk for the particular base out situation he faced and the particular lineup position he was hitting in calculated by the number of runs that actually scored from that situation for the league.

That is a nice idea.  While I agree that there is a lot of noise and context in that (as you say, it does not necessarily represent a batter’s skill), no sample data represents batter skill exactly anyway, which is why we regress.  The trick is to use sample data and metrics that strip as much as the noise and context out as possible while still preserving most of the integrity and transparency.

Do you have any numbers as to the run or win value of the average NIBB for a player like Bonds or Pujols or A-Rod versus that of a really weak hitter?  As I said, and you said as well, if it is not all the great at the extremes, it is probably not worth worrying about.

MGL - I will look in my old files tonight.  I should mention that it is worth studying this more.  Not so much for individual player values, but differentiating a player’s true skill at descriminating balls from strikes from the situational factor of descretionary walks is very important in fine tuning any game simulation program.  A player is not going to get walked at his average walk rate in either a second and third or a bases loaded situation, and it is important for any simulation to make the correct distinction.

Here are the linear weights values of the unintentional walk by lineup position and number of outs.

Lineup---Outs---BB/(PA-IBB)---Linear Weight

1---------0---------.072------------.505
1---------1---------.076------------.393
1---------2---------.090------------.299
2---------0---------.069------------.552
2---------1---------.074------------.356
2---------2---------.082------------.283
3---------0---------.082------------.585
3---------1---------.092------------.296
3---------2---------.103------------.232
4---------0---------.080------------.471
4---------1---------.093------------.375
4---------2---------.109------------.201
5---------0---------.072------------.440
5---------1---------.080------------.325
5---------2---------.093------------.192
6---------0---------.065------------.499
6---------1---------.072------------.320
6---------2---------.082------------.230
7---------0---------.059------------.471
7---------1---------.068------------.387
7---------2---------.077------------.195
8---------0---------.058------------.515
8---------1---------.067------------.384
8---------2---------.075------------.181
9---------0---------.051------------.534
9---------1---------.056------------.543
9---------2---------.065------------.321

I also looked at Bonds 2004 season where he had 232 walks.  120 of these were intentional.  The 112 unintentional walks break down like this:

Lineup-Outs-Walks-BB/(PAs-IBBs)-Value-Runs

4-------0-----46----.24---------.471---21.7
4-------1-----34----.23---------.375---12.7
4-------2-----31----.20---------.201----6.2
5-------0------1----------------.440----0.4

Total--------112-----------------------41.0

Bonds average walk value ---.367 Runs

So Bonds actually GAINS about 4.6 runs over what he would get if linear weights were calculated the normal way.  I think that this is probably a peculiar case.  I would still expect the normal gain or loss for most players to be within +- 2 runs.

What may be most important is the win value of each walk and not necessarily the run value.  It is interesting though that Bonds’ average walk value is higher than that expected if the walks were random.

The arguement is that a pitcher can choose to issue walks more often in situations where the walk has less RE/WE, right? Well perhaps (at least in extreme cases) the same is true for hitters as well; a player is more likely to take a walk when the walk has a higher RE and is more likely to swing away in situations where there is a greater difference between a base hit and a walk in RE.

That’s one thing I’d really like to study more indepth - to what extent do pitchers and hitters control various aspects of the batter/pitcher matchup. If a low-walk pitcher faces a high-walk hitter, is the walk rate what we would expect using Odds Ratio? Or does the high-walk hitter have more control over that than the hitter?

Colin - A hitter is pretty much at the mercy of the pitcher.  If the pitcher doesn’t throw the ball in the strike zone he shouldn’t be swinging at it, if he does throw it in the strike zone he puts himself in a hole if he doesn’t swing at it.  Some hitters can take a strike and still not be at too much of a disadvantage but few batters can take two strikes without limiting their options.

MGL - Yes, what is important is win value as in all stragic decisions.  But the run value parallels win value pretty closely.  What is happening with Bonds was unexpected by me but when I thought about it I shouldn’t really have been surprised.  Because of his very high production per PA (in 2004) pitchers were more willing to walk him in higher value situations (like with 0 outs) because the cost was still less than having him hit.  That raised his walk value until it was slightly higher than the average value of .36 runs when calculated by the above method.

MGL: “If you must set a value to the IBB in your calculations, simply use the average value of a PA for that batter.”

If a batter has negative runs-above-average, would you assign a negative marginal run value for his intentional walks?

Tango: “I’m not really concerned about how “most people” use RC/out.  I’ve laid out my case as to why I do it the way I do it.”

I agree with Tango. When working with Linear Weights, reconciling to runs created using R/PA is the best method for individual batters.

But what about Baseruns? Baseruns is a team model. It utilizes intrinsic linear weights that are based on R/O.

If you simply ignore IBBs in your linear weight formula, when you reconcile to runs created using R/O, the runs created value of the IBB will still be zero. If you ignore SHs in your linear weight formula, the runs created value of the SH will be around .20 runs, since the R/O for a SH is higher than the average R/O.

Ultimately, I was hoping that I could find a way to handle the IBB and SH that would be compatible with Linear Weights, R/PA Runs Created, and R/O Runs Created. By compatible, I mean having 3 sets of equations where there would be no inconsistencies. An example of an inconsistency that would probably arise is the LW value of a 1B being higher than the R/O RC value, since the LW value of a 1B had to be adjusted upwards, in oder to reconicle the LW to zero without the IBB.

Peter, out of curiosity what was the average run value of Bonds’ intentional walks in 2004, by this method?

Peter, I still don’t see why Bonds’ walks would have a higher value than if they were randomly distributed (or at least distributed as they are for the average batter).  Even if pitchers are willing to walk Bonds when the run value is high, as with 0 outs, they should still be MORE willing to walk him when the run value is low.  Unless perhaps they “run out of” times to walk him when the run value is low, I guess.

Terps, yes, it does not matter what the average value of a PA is for the batter.  An IBB is merely a substitute for an average PA, right?  If I told you that it does not matter if you issue an IBB or let a batter hit in any particular situation/PA, in terms of win value (the win value is exactly the same either way), then you would know that the value of the IBB in that situation was exactly the same as a generic PA.  Well, that is essentially what is happening on an average IBB.  It is the same as letting the batter hit.  Which is the same as just giving him an average PA.  If anything, you can make an argument that an average IBB reduces win value a little (as opposed to letting a batter swing away), which would be an argument for the IBB having a value a little less than an average PA for that batter.  But, no matter what, we can’t use the same value for an IBB for every batter, since clearly they are used differently for weak and strong batters.  If a #8 hitter gets IBB’s 10 times a year in the NL, what is the point of giving those IBB any fixed value as opposed to a generic PA for that batter, or even ignoring it completely?  What about an IBB to Bonds who gets walked almost every time there is a base open? Can we give those IBB’s the same value as those to the #8 hitter in the NL?  Of course not.  And BTW, the fact that a batter has a “negative” run value per PA or whatever, as compared to average means nothing.  It is obviously positive in terms of runs created.  It is only negative because we happen to be using a baseline of average.  If we use replacement as a baseline, then most batters would have a positive run value per PA or per whatever.

Obviously if you are using BaseRuns to actually estimate runs scored, it is a different story.  But who uses BaseRuns for batters? You shouldn’t.  And if you do, you still should not be assigning a fixed value to the IBB.  BaseRuns for individual batters means nothing anyway.  So you don’t need to account for actual runs, since there is no such thing as actual BaseRuns for individual batters.  For teams, when BaseRuns is appropriate, if all you want to do is account for runs scored, then by all means assign a static value to SH and IBB.  If you do that, though, you won’t be able to reconcile teams runs and expected wins.  For example, if team A has a lot of IBB’s, their runs/win will be higher than a team tat has fewer IBB. If you want to reconcile expected wins for a team, and are using runs as a proxy for expected wins, then you have to do the same thing with SH and IBB that you would do for lwts and batters - ignore them and/or assign them a value of an average PA.

Let’s put this IBB and SH to rest.  Once you think about and understand the relationship between runs and wins and once you know what happens to run and win value with an IBB and a SH, which should be pretty obvious, everything falls into place, and you can do whatever you want (I should say whatever is appropriate and correct) with those IBB and SH, depending on what question you are trying to answer.

I will be glad to put this issue to rest. You are probably tired of discussing it, and I am tired of thinking about it.

Now, I have no problem with ignoring IBB and SH when calculating runs-above average. What I do have a problem with is including them in the runs created calculation. Runs created should be calculated with the same categories that were used in the runs above average calculation. This means that IBB and SH should not be counted as plate appearances when figuring runs created using the average runs per plate appearance.

“But who uses BaseRuns for batters? You shouldn’t.”

You’re right. You don’t want to use BaseRuns for batters. However, there is nothing wrong with using Theoretical Team Baseruns for batters. It is even better if you make an adjustment for the number of runs gained/lost based on the batter’s rate of making outs. I like the Theoretical Team method because it tells you how many runs a player would add to a lineup of eight average hitters, or whatever baseline you choose to make the theoretical team.

BTW, I hate getting-up this early (4:30) in the morning. But I have a job interview at 7:00. 7 AM does seem like a strange time to be interviewing someone.

terps - There is really no reason to use BaseRuns for anything.  It really provides you with no more information than the formula:

Runs = (Men on base - LOB) + HR

which is to say no information at all.

MGL - Bonds’ 120 intentional walks pretty much included all the situations with a low walk value so there weren’t many to include in his “pitch around walks”.

I think using BaseRuns for Pedro or Jason Marquis is more than fine.  For hitters, LWTS is more than adequate.

***

I think we can make a distinction between IBB to everyone and IBB to Bonds, as clearly, the IBB given to Bonds were often irrational, or at the least, inconsistent. 

From 2001-04, his BONDS years (his wOBA was over .500 in each year, and he had around 600 PA each year), his IBB totals were: 35, 68, 61, 120.  The 2001 season was fairly consistent for his career, while his 2002-03 seasons responded to his jump in production of previous year(s).  2004 was simply an enormous outlier here. His 2006-07 seasons had a wOBA below his career average, and yet his IBB rate was above his career average.

The analysis of IBB Bonds or not must follow the scheme presented here (based on Win Expectancy, WE):
http://www.tangotiger.net/walkbonds.html
http://www.tangotiger.net/walkbonds2.html

I would guess that his 2004 season was simply managers being overly afraid.

For the record, I am now in full agreement with Tango, MGL, and Peter about how to handle IBB and SH in run Linear Weights and Runs Created. Re-reading the wOBA year-by-year thread convinced me that everything is being handled properly. Sorry for being a dick. You need to keep in mind that I am a rookie when it comes to sabermetrics. I still don’t know how the runs get converted to wins. Is there any introductory material on runs to wins converters?

Look for all teams with a Runs Scored / Runs Allowed differential of 1 run per game, +/- 0.1.

Report back the average run differential, and report back their winning percentage.

You will find the answer is close to 1.00 and .600 (or +.100 above .500).

And 1/.1 = 10.

I used an exact cutoff point of +/- 0.1 runs for teams with a RPG - RAPG differential of 1 RPG. From 1900-2008 there were 97 teams that matched this criteria.

Win%: .603
RPG: 4.99
RAPG: 3.99

The difference is exactly 1 RPG.

“And 1/.1 = 10”

10 Runs per win. Doesn’t this number change according to the run environment?

Yes, the number does change based on the run environment (not much though, plus, if you are talking about players in the same season, it cancels out). 

For purposes of illustration, this is how you can figure out the conversion of runs to wins.  It’s that simple.

(You can do similarly where the differential is -1 runs per game, and you will get similar results.)

Here’s some useful research I did a few years ago to illustrate the relationship between Quality and Run Environment:
http://www.tangotiger.net/winactuals.html

The same methods that Tango, MGL, and Peter recommend for calculating Linear Weights and Linear Runs Created, can also be applied to Theoretical Team Baseruns. There is one catch: you need to incorporate a PAR adjustment into the Theoretical team calculation for it to work cleanly. If you do not apply PAR, then your Baseruns estimate has to be expressed as R/O. I guess you could still use R/O, but then you would have to use the player’s “out-rate” to determine how many outs he would of made in his plate appearances where he was intentionally walked or executed a sacrifice bunt. It is much easier to work with plate-appearances for batters than it is to work with outs. 

I used the same categories in the Linear and Baseruns formulas. I did not set a run value for IBBs and SHs in all the formulas I used. The linear weights formula was based on 1993-2008 data. I found b-values for the Baseruns equation using runs created values that were reconciled with R/O. I am using Albert Pujols’ 2008 seasons in my example.

First, I applied the LW formula to Pujols’ statistics exactly as Tango, MGL, and Peter recommended. The RC total was also figured using the method they recommended.

I then proceeded to apply the Theoretical Team process to Pujols using Baseruns. Again, I left IBB and SH out of the Baseruns equation. PAR was incorporated into the theoretical team calculation. After I finished applying the TT process, I figured Pujols Baseruns per plate appearance (not including his IBB and SH). The average Baseruns value of a Pujols PA was applied to his 34 IBB. I then added this number to his theoretical team baseruns total. Here is what the numbers look like:

Linear Weights: 66.8 RAA, 146.4 RC
TT BsR w/PAR: 68.4 RAA, 149.3 RC

Just for kicks, I applied the Runs Created values that were reconciled by R/O (the values I used in the BsR equations) to Pujols’ statistics, along with a PAR adjustment.

R/O RC w/PAR: 147.6 RC

It does appear that all three methods (LW, R/O RC, adn R/PA RC) are compatible with each other, when SH’s and IBB’s are not included in the run estimator formulas. At least this appears to be the case for individual players. For teams, I still do not know the best way to handle IBB’s and SH’s. Just remember that R/O RC are only compatible with Tango, MGL,and Peter’s method if a PAR adjustment is applied. Since TT Baseruns utilizes intrinsic LW that are based on R/O, you need to incorporate the PAR adjustment, in order to treat the SH and IBB like Tango et. all treat them.

Even with extreme players like Pujols, the difference between the Theoretical Team runs estimate and the linear runs estimate is only a few runs.

And the linear method is 100x easier.

The biggest difference I could find between Linear runs and TT Baseruns was Bonds 04 season and Neifi Perez’ 02 season. Bonds created 8 additional runs when he was added to the theoretical team. Neifi Perez created 7 fewer runs when he was added to the theoretical team.