Now that I see this chart again ... it looks like the value for an error is low. The value should be around 0.60 runs, but this looks like it is around 0.10 runs.

Also, the putout run rate looks high; it should be half the assist rate.

I don't mean to imply that *every* SS should be rated on the DP based on the rate at which GB4 and GB6 are turned into DPs - all I was doing here was pointing out that (a) the Yankees are particularly awful at it and (b) they're outstanding on the 5-4-3 DP, so it's probably not primarily the 2B's fault. I didn't look at the Red Sox when I did this table, but I see now that they were also very bad at the 5-4-3 DP, so I have no doubt the primary DP problem there was Offerman. (Be interesting to see how Nomar did w/ Sanchez there this year...)

-- MWE

Found 'em both, Charlie - math errors on my part. I'll repost the table later tonight.

-- MWE

When you do repost it, Mike, could you rerun the correlations with DFTs, FRs and chances.

BTW, what was the correlation between DFTs and chances again? You said it was something around r=0.50 while AIM chatting the other night, but I don't remember offhand.

Mike --

Althopugh their numbers do not correspond exactly to yours, STATS 2001 fielding numbers show the Yanks once again finishing last in DP per DP opportunity: 105 out of 203, or 51.7%, vs AL average of 58.4% (NL was at 59.1%). This is for all infield DPs, not just those started GB6 or GB4.

Any idea what the deal is with Mike Bordick? He appears to have gone from awful to great to awful over these 3 years. Unless he had some sort of arm/leg injury in the 2 bad years, something must be skewing his results that didn't get normalized out.

I sent the revisions, with the correlations redone, to Dan S; I'm not sure how to post 'em myself and rather than screw things up royally I'll let someone do it who knows what he's doing :)

The correlation between DFT and SS fieldable opportunities is r=0.526.

Bob Allen, thanks for the 2001 update on DPs. How does STATS determine a DP opportunity? The opps counts seem a bit low to me.

-- MWE

Re STATS DP and opportunity numbers -- They don't provide a definition in the Handbook. I assume it's an infield grounder with less than two outs and a runner on first.

The math error I made in the table is part of the problem; when that's fixed, Bordick goes from average to great to below average (which mirrors what happens in the DFTs). There was also a pitching staff change in 2000; the Orioles were an extreme groundball staff in both 1998 and 1999 and became a fly ball staff in 2000. There is a tendency (which I'll discuss in the next article) for flyball pitchers to allow a higher percentage of hits on groundballs than groundball pitchers do, and that probably drags Bordick's rating down in 2000.

Apropos of the MVP discussion thread, FWIW, ARod and Tejada are fairly close to each other in 1998 and 1999, and ARod comes out *much* better in 2000. Of course, the $22 million man was still with Seattle then; I haven't looked at his Texas numbers at all.

-- MWE

The play-by-play data I use here is licensed by me from Gary Gillette and Pete Palmer.

-- MWE

I'll have the fixed table posted when I get home in a few hours; I'm not at a computer that provides me access to my e-mail.

The DP opps still seem low. In the 2000 AL, using the criteria of any ground ball that was (a) fielded by an infielder, (b) not a sacrifice bunt, and (c) occurred with a runner on 1B and less than two outs, I come up with a range of between 275 and 359 DP opps per team.

WRT Bordick: Davenport's fielding numbers for 2002 show Bordick with FRAA of 32 runs. His last five seasons (based on the V.2003 of DFTs, which has some changes from V.2002 which is what I evaluated in the article) are 11, 34 -12 (-11 BAL/-1 NYM), -1, and 32.

-- MWE

Still on the DP opps question -- in the AL 2001, the STATS figures ranged between 157 and 263 per team (avg 214). NL was only about 202 per team. I can't account for the difference from your numbers.

Some questions for the guys on this board:

For my Superlwts, for DP defensive value, I also use DP's per DP opp, where a DP opp is runner on first and less than 2 outs. I can't remember if I except any situtaions (like bags loaded) from DP opps. If a Dp is turned I give equal credit to the fielder and to the pivot man. This 50/50 credit is arbitrary. My "feeling" is that the pivot man has more influence over whether the DP is turned or not. I've never looked into this though. And for DP opps, I don't think that I make a distinction between ground balls that are initially thrown to second and ground balls that are inititally thrown to some other base.

Anyway, my questions are:

1) What credit split would you use for the fielder and the pivot man when the throw goes to second and either the DP is turned or it isn't?

2) If a fielder throws to first or some other base (and not to second) in a DP situation, should we not consider this an opp for the fielder? IOW, if a fielder does not even attempt a DP, is it becuase no one would have had a chance for a DP (because of where and how hard the ball was hit and/or becuase the runner was in motion or very fast, etc.), or because that particular fielder could not at least get the force at second?

BTW, I think that my UZR runs should have been included in the charts for comparison purposes! ;)

For my Superlwts, for DP defensive value, I also use DP's per DP opp, where a DP opp is runner on first and less than 2 outs.

You have to exclude SH from this - also failed sac attempts (which might be part of the difference, now that I think of it, between the opps I report and the opps that STATS reports; I got the SH but not the failed sac attempts).

Anyway, my questions are:

1) What credit split would you use for the fielder and the pivot man when the throw goes to second and either the DP is turned or it isn't?

For me, this is the same situation as for the ball in the hole that either the SS or 3B could field. I don't know who should have fielded it, so they both get full blame for not fielding it. Same thing here - both the SS and the 2B get full blame for not turning the DP. My subjective feel, as I noted earlier, it's likely that in the Yankees' case the problem is Jeter and in Boston's case it's the 2Bs.

2) If a fielder throws to first or some other base (and not to second) in a DP situation, should we not consider this an opp for the fielder?

I count it, again because I generally don't *know* whether or not the DP was possible.

BTW, I think that my UZR runs should have been included in the charts for comparison purposes! ;)

That's in part 6. Patience, grasshopper - one method at a time :)

-- MWE

One thing I'm trying to sort out:

My understanding of Win Shares, CAD and DFTs is that they place defensive contribution in a team context. Thus, just as the sum of offensive accomplishments of players on a good hitting team will be greater than the sum of offensive accomplishments of a bad team, so will defenses be rated.

As I read these articles, the point is not that team defense has been incorrectly measured. At least, that argument has yet to be advanced. If so, then measuring individual defense within one team becomes a zero-sum game. In other words, if Derek Jeter is underrated by CAD, DFT or WS, then Brosius or Bernie or somebody must be overrated, since the overall team has been measured reasonably well.

Is this correct, or are you arguing that team defense has been measured as suboptimally as individual defense?

Mikael -- more or less, yes. What we're trying to do is to spread out the success based on the number of outs each player makes over or under the league rate. We know how groundball/flyball and left/right skews change a team's out spread -- usually. Every so often, a team like the Yankees will have an unusual out spread due to odd factors.

A good challenge will be learning how to spot these teams. No one knows how to do this yet.

I agree (and concede) that there are some things where objective evaluation can be significantly helped by competent and sustained observation. One of those might be "turning the double play". As it seems (to me) to be difficult to nail down when a DP could or could not have occurred, but for the skill of the fielder or pivot man (as in the questions I asked - which I would like to get more answers to, BTW), based upon PBP data, AND the sample sizes of DP 's and DP opps are relatively small, I think that observation (and scouting reports can be a big help. In fact, it is probably failry easy for a competent person to look at a pivot man's footwork, quickness of hands, etc., over some period of time (probably not too long, as there is not much "fluctuation" in that skill) and determine how good he is at turning the DP. An good example of where "subjective" evaluation is much better than an objective measure would be the following: We want to evaluate a player's speed. Which is better, looking at his baserunning totals (triples, SB/CS, advances on the beses, etc., or either timing him (that's not really objective - it is another objective measure) or watching him run (a trained human eye can prettty much tell how fast a runner is)? Another one like the pivot man situation is arm strength in the OF.

Here's another interesting point about subjective evaluation. I've thought about this before but it just popped into my head again after reading Slappy's post. We (sabers) generally eschew (debunk, poo-poo, disdain, mock, etc.) subjective evaluation when it comes to things like offense and even defense, and rightfully so - to some extent. However...

Since regression is an important part of an objective evaluation process (for ability or projections that is), it is critical that we estimate the population that a players comes from well (sorry about the awful syntax). (See the discussions about regression and population on some of the Clutch Hits [argh!] threads.) Without getting into all of the mathematical details, the reason we need to know about that population is that that determines what number we regress a sample stat towards. If we are given a player with a .300 BA in 100 AB's and we know nothing else about that player, we regress the .300 towards the mean BA of all players. If we find out that our player is a LHB, then we regress the .300 towards the mean BA of all LHB's. If we find out that our player is 6'2", 185 lbs., then we uyse the BA of all such players, etc.

Now here's where subjective evaluation comes into play. Subjective evaluation can give us a better idea of what population a player comes from (in order to determine what number to regress his sample stats towards) - no here's the important part - as long as that subjective evaluation is completely independent of any sample data (the subjective evaluation must either be NOT subject to much if any sampling error - like watching a player run several times is not subject to much sampling error - some but not much OR the sampling error in the subjective evaluation must be independent of the sampling error in the objective evaluation - i.e., the sample stats.

For example, let's say we have a normalized UZR or ZR for a particular OF, say Torri Hunter, that is 1.02 (2% above average) for one season. Normally, we would regress the 1.02 towards 1.00 to estimate his true number. Since our sample is only 1 year, we regress say 50%, to give him a true number of 1.01.

Now what if in observing Hunter, we notice that he gets good jumps on the ball (we think), he runs like the wind, and he makes tough catches look easy (like Andruw Jones). Even if we are "wrong" about some of these observations, doesn't it change the population that Hunter is from? Before, without any observation, he is just a center fielder, so we use the mean UZR of a (starting) CF (around 1.00) to determine what to regress his 1.02 towards. But if we notice all of the above things, we now have to say that the population he comes from is not only all CF's, but CF's who run fast, get good jumps on the ball, etc. So what is the mean (true) UZR of that population? We don't know for sure, but it might be 1.01 or 1.02!. So instead of regressing Hunter's sample UZR of 1.02 towards 1.00, to get an estimate of 1.01 for his true UZR, we might regress the 1.02 towards 1.01 or even 1.02 to get an estimate of 1.15 or even 1.02 for Hunter's true UZR!

This is basically the mathematical explanation for how and why we can indeed use subjective evaluations ALONG WITH objective, sample measures to better estimate players' true abilities (i.e.projections), especially for abilities that are hard to measure through objective means (for various reasons - e.g., small sample sizes, lots of sampling error), and especially when the subjective evaluation is reliable enough to make some significant inferences about the population that a player comes from!

You can even do it with hitting! For example, let's say that you ave the same hitter as above - .300 in 100 AB's. You know nothing else about him, so you regress the .300 towards the mean BA of all players, say .260. Since it is only 100 AB's, you regress maybe 80%, to get an estimate of his true BA of around .268. But what if you watch the guy (let's forget about his handedness, hieght and weight, etc.) and say, "Boy this guy "looks" like a great hitter! He has a sweet swing, etc." Now you may have a different population! Assuming that your observations are not based upon (completely independent of) the fact that he hit .300 in those AB's - which is hard to do of course - you no longer have a hitter who comes from a population of all hitters - he now comes from a population of all hitters who "look" like great hitters! So instead of regressing the .300 towards .260, you might regress it towards .280 or something like that. The tricky part, of course, is to somehow make the subjective evaluations independent of the sample data, in order for the above to be valid. That is tough in hitting, which is why we generally don't like to use much if any subjective evaluation when it comes to hitting, fo rour projections. For example, if we are looking at a player's sample HR total and trying to determine his true HR rate (i.e. HR projection), how can we separate observation and sample performance? If he hits 50 HR in 500 AB's, he is going to "look" like a great HR hitter, right?

But since in order to use subjective evaluation (observation) to change our population, and hence, the number we regress to, the subjective evlaution must be INDEPENDENT of the sample data (the 50 HR's), we cannot change our population in this case even though we observed that this guy "looks" like a great HR hitter, because that observation is obviously going to be influenced by the fact that he hit 50 HR's. Now if we gave a tape of the guy to a person who didn't know that he hit 50 HR's, and on the tape were this player's swing and misses, and our observer (presumably a competent scout), and he remarked "Wow, this guy has a powerful swing! I think he is a great HR hitter," THEN we could legitimately change the population from whence we think this player came!

Anyway, just something (important, IMO) to think about...

Excellent post MGL!

I just want to repeat the most important part again, for those who skim long posts - The tricky part, of course, is to somehow make the subjective evaluations independent of the sample data, in order for the above to be valid.

Now what if in observing Hunter, we notice that he gets good jumps on the ball (we think), he runs like the wind, and he makes tough catches look easy (like Andruw Jones). Even if we are "wrong" about some of these observations, doesn't it change the population that Hunter is from?

Not necessarily. We are watching Hunter, and comparing him to some baseline for an outfielder that exists *in our mind*. If we watched Andruw Jones for a while, or Darin Erstad, or Jim Edmonds, or Doug Glanville, we could easily draw the same conclusion about them as we do about Hunter. It might very well be a requirement for *every* center fielder that he get a good jump on the ball, run like the wind, and make tough chances look easy, in which case the base population hasn't changed at all.

-- MWE

Mike, of course if ALL CF'ers were characterized by "good jump, swift, etc.", then of course, THAT would be the only population of CF'ers. However, if there were ANY (even 1) CF'ers who do not fill this bill (when observing them), then by definition we now have 2 populations, even though the overwhelming majority of all CF'ers may belong to one of those populations.

In practice, what we are doing when we regress a player like Hunter's (whom we observe to have certain positive characterisitics that not ALL Cf'ers possess) sample defensive stats, is saying "Even though the average, say ZR, of a starting center fielder is .73, that includes a few "lumbering" CF'ers, perhaps a few CF'ers whom we "know" (independent of their ZR - hopefully) should not even be playing CF (perhaps the team has no other CF'er), so since we "know" that Hunter is probably not one of those types (lumbering, doesn't really belong there), we cannot use .73 as the "mean of the populaton" ZR to regress his sample ZR to. We would need to estimate the mean ZR of all CF'ers who are not lumbering, etc., which would presumably be greater than .73...

Mike, I apologize if you covered this already, but...

When it comes to the DP, shouldn't the LH/RH thing also apply? For example, what is Jeter's rate at turning the DP with a RH hitter and LH hitter? What is the league rate? Breaking those down with him as the pivot man and not, what are the breakdowns?

If Jeter's ball distribution is really skewed one way, isn't it possible that the DPs are also skewed?

Futhermore, the more balls hit up the middle or directly at him would make it a little easier to turn the DP, than a ball in the hole. So, break it down even further: what is Jeter's rate at turning the DP, compared to the league when
1 - LH/RH
2 - pivot / not-pivot
3 - hole / direct / up the middle

( I realize we lose sample size along the way.)

In practice, what we are doing when we regress a player like Hunter's (whom we observe to have certain positive characterisitics that not ALL Cf'ers possess) sample defensive stats, is saying "Even though the average, say ZR, of a starting center fielder is .73, that includes a few "lumbering" CF'ers, perhaps a few CF'ers whom we "know" (independent of their ZR - hopefully) should not even be playing CF (perhaps the team has no other CF'er), so since we "know" that Hunter is probably not one of those types (lumbering, doesn't really belong there), we cannot use .73 as the "mean of the populaton" ZR to regress his sample ZR to. We would need to estimate the mean ZR of all CF'ers who are not lumbering, etc., which would presumably be greater than .73...

And the gain in doing this is...what? How more accurate are we likely to be if we try to split a population of 30 or so players - all of whom are in the top 1% if not in the top 0.5% of the entire population of baseball players (which would include minor leaguers, amateur players, beer leaguers, etc.) - into two or more groups? Aren't we at the point where we're trying to adjust for differences that are likely to be "smaller" than random year-to-year performance variation? Brock Hanke, in the 1998 BBBA, called this the Law of Diminishing Returns as applied to baseball analysis, and I think he's right.

-- MWE

Mike, whoah! Some things are stated as fact - pure and simple - and some things are stated in the context of practicality and usefulness. The sentences you quoted from me were clearly the former! No offense, but it sounds a little like you got caught stating something that turned out to be false and you are trying to save face!

Funny thing is, I'm actually conceding what non-sabers often say - namely that subjective observation/evaluation is an important part of a proper analysis of a player's ability. I'm just giving a mathematical explantion as to why they may be right...

Mike, whoah! Some things are stated as fact - pure and simple - and some things are stated in the context of practicality and usefulness. The sentences you quoted from me were clearly the former! No offense, but it sounds a little like you got caught stating something that turned out to be false and you are trying to save face!

My answer is that the distinction between CFs based on the skills that MGL has cited has already been made before the players every get to CF - players who don't have the requisite skills aren't playing CF. The variation in those skills within the group is almost certain to be *far* smaller than the variation between those inside the group and those outside the group, to the extent that it does not matter whether you account for it or not, because you cannot see the differences between members in the group from outside the group.

-- MWE

While what Mike is saying about the breadth of skills being narrow for guys already playing CF (so essentially, we can look at the stats in context, and not go further subjectively), I wouldn't do that for say 3B or SS. Agility, quickness, speed, positioning, arm strength, etc would be some attributes that you would need. Again, as long as you can keep the scout from looking at the stats, then you could gain a good regression number based on the scouts report on the various attributed of a fielder.