THE BOOK--Playing The Percentages In Baseball

On further examination of the data, Tino looks pretty bad without Knoblauch.  Chuckie is only repsonaible for -8 of Tino’s 2B rating, Joey Cora is responsible for more.

Part of the problem is that most of Chucky’s Yankee throwing errors came in 1999, which retrosheet is missing.  He was still a decent 2B in 1998, played only 82 games (15 E) in 2000, and didn’t play 2B after that.

Nice work.  Beat me to it as well (and saved me some time).

I’d like to know more about the methodology - e.g. what exactly did you control for?

Also,

It is not that hard to figure out, but I would certainly like to see a column of “per 150 (or 162) games.” Other than knowing or looking up how many years each player played, I have NO idea how many runs each player is saving/costing per season. Without knowing that, I have NO idea how important this skill is.

And of course, we’d like to know how much of this IS a skill and how much is “luck.”

Let’s not get carried away with these WAWO. We ALWAYS need to know how much something is a skill before we conclude ANYTHING or make any assumptions about the data. Of course, the spread and magnitude (as a function of the sample size or opportunities) of the numbers gives us an idea as to how much of these things are skill/luck, but we still can’t really tell JUST by looking at the numbers. It would be nice to have some kind of “r” or something like that to help us do a mental regrssion toward the mean on the numbers. Or perhaps a column with regressed numbers (that actually reflect true talent estimates).

I can do a WAWO on anything I want and get lots of different numbers. That does not, however, necessary mean that one player is “better” (in skill) than another, though. I still have to determine how much of what I am measuring is skill/luck.

Sure, but you can say that about career ERA, OBP, SLG, and whatnot.  All this shows is a player’s sample rates.

Yes, I agree it should be “seasonalized”.  I’m not sure what “1000 throws” corresponds to.  Don Mattingly for example had almost exactly 1 putout per inning.  So, I would prefer something like per 1300 throws or so to denote full season.

I just lost a nice long post responding to #3.  I’ll write another, shorter, version.

Sure, but you can say that about career ERA, OBP, SLG, and whatnot.  All this shows is a player’s sample rates.

Yes, I know, but we have enough of an intuitive idea of how much luck/skill is involved in these metrics, and many people have already given us y-t-y correlations and the proper “regression toward the mean” formulas for them.

For new metrics like this, it is imperative that we have SOME idea as to how much skill/luck there is in the spread we see.  What does Helton’s 57 runs during his career (I assume it is his career) mean in terms of true talent?  50 runs?  30 runs?  I have no idea.  My best guess from eyeballing the numbers is that one SD of talent per season is around 2-3 runs, which is a lot, BTW.

I would like to see it be standard procedure eventually, when giving sample data, to give a y-t-y “r” or something like that.  It really is necessary for us to draw any conclusions about the data, other than what has already been done (by skill, luck, or some combo).

I am not criticizing Rally for this.  This is great work.  First time I have EVER seen this quantified even though we have been talking about it for a long time.

No one gives that kind of info, when presenting data like this, including myself.  I would just like to see it done.  When I see data like this, one of the first things that comes to my mind, is how much of the spread I see is skill and how much is luck.

I have two steps in running this:  Run a query from my retrosheet databases and export only the fields I need to text.  Then create a new “scooping” database and import all those files into it.  Run my queries from there to get results for all the years.

Along the way, I didn’t keep track of innings so I kind of eyeballed it to get 1 season =~ 1000 throws.  I first baseman may get 1300-1400 putouts per 1400 innings, but some of those are on throws from the pitcher or catcher (not included here) and some are unassisted putouts.

I didn’t control for anything, just looking at whether each infielder made an assist to first, or a throwing error.

What I meant by control for things was, let’s say that a certain first basemen was on the field with SS A, B, and C, for 2000 IP with SS A, 1000 with SS B, and 500 IP with SS C, did you look at what SS A did with all other 1B, SS B with all other 1B, and SS C with all other 1B, and then weight by the IP for each SS (2000, 1000, and 500 IP)?  And did you do that for each IF separately?

Then, if you did it that way, what about the other 1B’s with all of these infielders?  Did you control for them?  What if a 1B’s primary backup was particularly good or bad?

So, it is not that simple a job.  What was your exact methodology for the “with and without”?

I controlled for none of that.

Here’s all I did:  Say Jose Reyes made 45 errors in 2000 throws to first, to all first basemen.

Say he made 1500 to Carlos Delgado, and 30 errors.  That makes 15 errors in 500 throws to other 1B, or 45 per 1500.  Delgado gets +15 errors saved for 1500 throws.

I can see why you might want to weight that be the lower of the w/wo innings.

MGL’s first paragraph in post 6 is what Rally is saying in post 7.

MGL’s second paragraph isn’t really needed, since the “other 1B” will come out pretty close to league average.  Later in MGL’s second paragraph, he’s talking about “backup 1B”, but Rally (and I) look at all “others” in their careers, not on a year-by-year basis.

So, in Rally/7’s example, it works out like this:
Reyes, with Delgado: 30E, 1500throws
Reyes, w/o Delgado: 15E, 500throws
prorated, w/o Delgado: 45E, 1500throws

While Reyes has been a lifelong Met, this would be Reyes’s entire career, Mets or otherwise.

Then you repeat with other Delgado shortstops.  I dunno, say Cesar Izturis, with Delgado, and Izturis (in his entire career) without Delgado.

You end up with two sets of numbers: all of Delgado’s career, unadjusted, which is purely a count of all the errors of his SS and all the throws they made to him.

Then the second number is the prorated numbers of his shortstops when Delgado was not their 1B.  Add em up, and compare.

Now, Rally is correct that you would likely prefer to always prorate down to the minimum, rather than prorate to the actual Delgado number of opps with each SS.

When I do this with C/P or SS/P, this isn’t really an issue, since Jeter will have plenty of “other pitchers” to compare Clemens and Mussina against. 

But, in the case where you’ve got players practically married to each other, you do end up with problems here, and so, it’s probably worth it to prorate downwards.

I agree with MGL. I am not all that confident that “"the other” firstbaseman will come out pretty close to league average” when weighted by their PA’s.  I think you need to go back at least another step or maybe more.  Or perhaps the better methodology is, instead of assuming league average skills of “the other” first baseman, you do a second iteration substituting the skill you computed for those first baseman in the first iteration.  It might take a few more iterations but I bet this process will give you better numbers.

I did this step for the catchers in THT08 and here:
http://www.tangotiger.net/catchers.html

And the end-result was that the catchers were mostly being compared to an average catcher.  (i.e., I did go through the second iteration, as Peter/9 is talking about).

However, as noted, catchers automatically don’t play 150+ games every year.  Plus, they have enough pitchers who move around, that it’s an automatic that the breadth of “other catchers” that each of his pitchers was throwing too would be far and wide.

In THT08, the SS/P was similarly even.  There’s enough pitchers moving around, that when you compare Jeter to other shorstops of his pitchers, there’s enough of them.  Just with Clemens, he had a good cross-section.  Throw in all the other pitchers he’s had, so, again, it’s easy to see that you don’t need to go for another iteration.

In some cases though, you have to.  For example, Cal Ripken is a problem.  He played every inning.  So, for Orioles pitchers who did not have a career outside of Baltimore and were in the bigs in the mid to late 80s, this process will fail you.

The “correct” thing to do is do it both ways, so that we can see the impact.  But then, you’ll want to go further.  How about controlling for the handedness of the batter/pitcher?  The GB tendency of the batter/pitcher?

So, there are additional considerations as well, each of which if not handled is a source of bias.  Whether the “other 1B” provides a bigger source of bias, I don’t know.  I doubt it though.

Also note that the beauty of WOWY is the one-variable aspect.  Once you start including more and more parameters, the elegance of WOWY is gone, and you are left with a series of adjustments.  And that makes the results less real.

Tango - I agree with trying it both ways and seeing if it makes a difference.  And a lot depends on how the results are intended to be used.  If you are looking for a career rankings list of players than the one iteration methos is probably close enough to get the rankings 99% accurate is is certainly more elegant.

But if numbers like those generated by Rally for 1B errors prevented or the ones you are contemplating for SS/2B double plays added end up being used as actual values representing ability in some fielding metric then the extra accuracy generated by a multi iteration process may be worth the effort.  There is no elegance in simplicity if the results aren’t accurate for the purposes intended.

There is no elegance in simplicity if the results aren’t accurate for the purposes intended.

I really like that statement!

I may be all wet, but I would think that when doing WOWY (what does that stand for?  I have been using WAWO) with first basemen, that the “without” is dominated by a player’s backup, moreso than with other positions.  I could be wrong though.

I think the correct thing to do is always use the iterative (recursive) process, just to be sure.

With
Or
Without
You

Pronounced: Wow-wee

Not sure why you keep bringing up backups.  We are looking at who Omar Vizquel threw to 1B.  That would include Alvin Davis in Seattle, Paul Sorrento in Cleveland, Eddie Murray, JT Snow, Ryan Klesko, and so on.  These 1B may or may not have played on the same team, but it’s irrelevant.

We are looking at Omar’s performance with Alvin Davis, and without Davis, in his career, not just in 1989.  So, Alvin Davis is being compared to the guys I listed above plus another dozen or two 1B.  Then, you repeat with some other SS that Alvin Davis had. 

Once you go through it, you’ll end up with quite a large “others” to which Alvin Davis is compared against.

As long as the player has had a long enough career, and has not been married to his mate, then you really don’t have to worry about the “other” not being average.  He almost certainly will be.

The general point about the recursive check stands, and I agree with.

Yeah, the backups is no big deal.

Not using the lower of the “with” or “w/o” playing time could be a big deal.

Say a SS makes 3000 throws to a 1B, with 30 errors.  That SS plays his whole career with that one 1B, and only makes 2 throws to another 1B, 1 of them being an error.  My system would look at that as the 1B being +1470 for that SS.

I checked my results, and there aren’t any extreme values skewing to that extent, but there could be.  The worst I found was Steve Balboni and George Brett.  I’m missing most of Brett’s career, starting in 1985, and in this span he did not throw to many other 1B than Balboni.  Brett made 2 errors in 30 throws to others, and 3 in 438 to Balboni, for a +26 rating.

At 2B, Todd Helton and Brent Butler - 2/184 with, 1/7 without, giving Todd a +24 rating.  This needs to be revised.

One solution is to add a number, maybe 100 throws at an average error rate, to the without, as a regression.  Of course doing stuff like that you lose the elegance and simplicity.

Starting on this path is a gateway drug that will lead you into multiple iterations, and sorry guys, I’d rather move on to something else instead of going there.

I just added 100 throws of regression to all the without- makes a huge difference.

Todd Helton drops to +18, 15th overall.  Kent Hrbek, Keith Hernandez, and Wally Joyner top the list at +42, +39, +39.

Next on the list is to see how the variation compares to what you’d expect by chance.

I just followed the steps to calculate a regression point from that zone rating post back in October 2006.

For 1B with 500 or more throws, I get SD of Z-scores at 1.54, regression = .422, average chances = 1333.

So you can regress this to the mean 50% at 975 throws.

"One solution is to add a number, maybe 100 throws at an average error rate, to the without, as a regression.  Of course doing stuff like that you lose the elegance and simplicity.”

This is very close to how I was revising my career SS/pitcher WOWY!  I realized I need to do this because of Cal Ripken, and the mid-80s to mid-90s Oriole pitchers who never played for another team, I was trying to figure out the best approach to this.

1.  Use the lesser of the two opps (with Ripken, without Ripken), that is standard in all these kinds of matched pairs studies.  What I don’t like about this is that the career total for Ripken won’t be his career total, but some smaller number.  For example:
http://www.tangotiger.net/catchers.html
I *want* to keep Gary Carter’s career WP and PB totals with Steve Rogers as 51, 7.  And his career totals with all pitchers as 464, 85.

Then, I *want* to have some accompanying baseline that is prorated (up or down) to Carter’s opps with each of his pitchers.  In that chart, you will see that the most egegious one is Charlie Lea (3061 PA with Carter, and 815 without).

If that’s the worst of it, I’m fine with it.  So, a straight prorating works for me (with catchers/pitchers anyway).

2. Add a league average total to the “without” so that it comes up to the “with”, in the Steve Rogers, Charlie Lea cases.  While we are ok with prorating down, we don’t like to prorate up, because of some crazy multiplier potential (especially with say Cal Ripken). So, the solution is to add enough league average PA, so that it matches the “with”.  In the Carter/Lea case, that means adding about 2200 league-average PA to the “without”.

In the other cases, like Gooden, Darling, etc, you continue to pro-rate down.

Well, once you do this, you realize that it’s exactly like not adding the 2200 to Charlie Lea, but simply adding it at the end.  That is, you know you need to add 4500 for Steve Rogers, 2200 for Charlier Lea, so you might as well just add 6700 league average PA at the end. 

So, it comes down to how many league-average PA to add.

3. Allow pro-rates down, but not pro-rates up.  That is, don’t add in those 6700 league average PA.  What this means is that you will simply let the “without” dictate the baseline, without having a regression component.

The more I work with this stuff, the more I lean towards the 3rd option.  First, it’s easier to program.  Trying to come up with a proper league average baseline is problematic to begin with.  Plus, it loses its elegance when I try to fudge like that.

Basically, what you end up doing is that if Carter’s “with PA” is 72,000, his “prorated-down without PA” might be 65,500.  In order to align it back up, you prorate up *all* of his pitchers by 10%.

While I’d LIKE to say that I’ve got exactly the same number of PA with and without Gary Carter for Steve Rogers and Charlie Lea, the reality is that I don’t. So, I’d rather not fool anyone, and simply go with option 3.  I could also report on how much “prorating up” I had to do, so that the reader knows.

RAlly/17: good job.

So, basically, your regression is:
1000/(1000+throws)

Or
1/(1+seasons)

Pretty straightforward. So, for a guy with 9 seasons, you regress his rate by only 10%.

Given Rally’s regression equation 1/(1+seasons), and again, eyeballing the chart and making a couple of mental calculations, it looks like 1 SD of talent is around 2-2.5 runs a year (everyone is plus or minus 6-8 per year).  Does that sound about right?