In the words of the great C. Montgomery Burns: excellent!!

I can't say how grateful I am that the fine-toothed combing of fielding has finally reached a new level.

I see that really only the Colorado shortstop is heavily affected, while the others are slightly affected (3 or less runs).

One note regarding PF: note that a 1.0 in 1993 might not be the same as 1.0 in 2002, since the parks themselves have changed, or even created. A park in 1993 might be a 1.01, but become a .98 in 2002, because a PF is relative to other parks in the league. To solve this, if you wanted to, is one of two ways: 1) the delta approach, as we've discussed many times for other things like age adjustments, or 2) look at only the parks that have not changed from 93 to 02, and make that the baseline 1.00, and make all parks relative to that baseline. But, that's probably just quibbling.

Very very enjoyable.

Jason, good point. Extreme GB or FB pitchers like Sid Fernandez or Tommy John are about 2:1 in GB/FB or FB/GB.

Since MGL said the relationship is about .001 outs made per .1 change in ratio (not the best way to do it, since your ratio will be .50 or 2.00 depending on what your denominator is. I would choose "rate" as in FB/(FB+GB) ), then we should see something like an extra .005 to .010 outs made per BIP. At 500 BIP, that's 2.5 to 5 outs, or about 2 to 4 runs.

This information will be much much more valuable for a pitcher than for an individual fielder.

Whether that difference is due to mistake pitch or because of positioning or other factors makes it clear though that we are talking about a "little thing" effect.

Not that MGL doesn't have enough to do, but I think it would be interesting to try to see the new UZR to try to look at specific *pitchers* - looking at their starts and adjusting for the defensive help they received...

Warren, if you followed the discussion at fanhome, this is exactly what I'm talking about with PZR.

Essentially, DER = UZR + PZR on a team level.

Now, all you have to do is apply the same process, but switch fielders for pitchers. Solving for this will lead you into proof for DIPS.

All these numbers - isn't this the same system that said Mo Vaughn was better then JT Snow last year?

If it is - then just save it.

MGL: what's the error range for your computations, after doing all the adjustments?

Bordick: I assume that those numbers are /162 GP, and therefore, Bordick's performance needs be pared down.

Jim wrote:

"All these numbers - isn't this the same system that said Mo Vaughn was better then JT Snow last year? If it is - then just save it."

Yes it is! Jim, what would the point of an objective "system" be if it only reinforced what it is you think you already know? I think it was Tango who said that a good system (offense, defense, whatever) should coincide with what you think you know 80% of the time and you should be surprised 20% of the time. I don't know if I am doing justice to his statement and obviously no one knows what the numbers are (80/20, 85/15, etc.), but I agree with the general concept. On the flip side, just because a system follows that patter, doesn't make it a good one of course. That is just a quick and dirty "check" on the sytem right off the bat. Implicit in that 20% (or whatever percentage), is that some smaller percentage will be REALLY surprising (like Snow, T. Hunter, maybe N. Perez). Whether that means that the system is "wrong" with regards to those players, I have no idea. I doubt it. I think it either means that these players' true defensive abilitites are somewhere in between (the objective rating and the subjective consensus), or that these players are the ones, for whatever reasons, LOOK good but really aren't. I suspect it is a little of both, but I lean towards the latter theory (since it's my objective rating!). Seriously, I lean towards the latter, since after all, that is the whole point (or one of them) of these objective ratings - to identify those players whose defenewive abilitites we CAN'T, again, for whatever reason, nail down by observation.

For the record, when the smoke cleared, Snow's 02 adjusted UZR runs was -9 (-11 per 162 "games"). Mo Vaughn's was -15, or -25 per 162 "games". So for you Snow fans, while Snow's performance was bad (hey, I don't make up these numbers - I just report them - that's how many ground balls Snow did or didn't catch in the various zones, etc.) in 02, Vaughn was over twice as bad! Vaughn is indeed a statue at first and was the worst first baseman in the NL last year (McGriff, who used to be good, was second at -22 per 162), and should, of course, be a DH (like F. Thomas, who has -27 UZR runs per 162 at first in 59 games since 1999).

As far as Snow, someone suggested that he was hurt last year which could easily have affetced his range, although he has no great UZR years going back to 1999 (he was a +3 in 01, -123 in 00, and a -1 in 99. The subjective consensus also seems to be that he is a wizard at catching bad throws. The low number of errors by the SF infield supports this view. UZR does not measure that skill at all.

BTW, the numbers in the article are absolute runs and not runs per 162, so they have to be taken in the context of the number of games or the number of chances. Also, the "games played" column in the article is the actual number of games played which I took off of the ESPN defensive stats website (the same place I got the range factor (RF) and the STATS ZR from in Part I). Normally, I put down in the "games" column the number of a player's chances divided by the league average chances per game at that position. So for example, Izturis' 9 UZR runs in 290 chances is a much better performance (rate-wise) then Hernandez' 10 runs in 469 chances.

As far as Bordick being "better" than A-Rod, apparently last year he was (again, I don't make up the numbers - just report them), at least in terms of fielding ground balls in the various zones and not making errors, which is what UZR measures. Keep in mind that we have fairly large sample error especially in those one-year samples. For example, if I were to use a (full-time) one-year sample to estimate a player's "ability" I would probably regress on the order of 50% (at least). So Bordick's one year UZR runs of 19 might correspond to a "true" UZR runs (ability) of 9 or 10 (in 116 "games), which is around 12-14 per 162. Of course, if you want to know more about who is "better" in terms of ability or projection for this year, you would look at multiple years to increase sample size. In 01, Bordick was +5 per 162 and A-Rod was +6. In 00, Bordick was +8 and A-Rod was +14. Finally, in 99, Bordick was +21 and A-Rod was +1. So it looks like Bordick is indeed a great defensive shortstop, which is amazing for someone at that age (although good hands will remain more stable with age than good range, I assume). A-Rod looks (from his UZR runs) like a great defensive SS as well, or at least a very good one. As far as who would be better this year (I think Bordick retired, did he not?), it's probably a toss-up. Given Bordick's and A-Rod's age, I'd probably give the edge to A-Rod.

That brings up Tango's question about error range. I have no idea (am I supposed to?). I guess we can take the UZR rate and assume that we have a binomila distribution and go from there. For example, Bordick had 347 chances in 02 with a UZR rate of .815. Using the binomial formula for standard deviation (SD), we get a SD of 21 points on the UZR rate, which is around 7 balls (.021 * 347), which is around 5.6 runs. Since UZR is essentially the sum of the individual ZR's, in the various zones, and becuase of all the adjustments, we probably have a SD which is higher than that, maybe 7 or 8 runs (in those 347 chances). So at 2 SD's (95% confidence interval), that's an "error range" of around 15 runs! That's another reason for not "worrying" about unusual looking numbers like Snow's. It is entirely plausible that Snow's 02 numbers or even his 99-02 numbers make him look bad entirely by chance. Becuase of this, it is completely acceptable to "narrow" that error range by using other independent measures, like observation and scouting.

While I have always insisted that objective measures for offense and defense are MUCH more accurate than scouting and observation, your objective measures are always "handcuffed" by error range (random variation). Observation and scoutng can help to narrow that range. In fact, the larger your sample (for your objective measure), the less important scouting and observation are (and, the more dangerous they can be - I'm actually with Mike Gimble on this - for a veteran player, on whom I have lots of data - I don't even want to look at him - case in point is Tex - they have to LOOK at him in Spring Training to see if they want him to be their everyday player (according to GM Hart)? - God help the Rangers if he happens to get lucky or "look good" during the Spring). Most of us know this intuitively, but it is a very important concept nonethesless - in fact, it is one of MGL's rules!

BTW, I changed my UZR program to use 4 years worth of data, both leagues, for the baseline (league) numbers in the various zones to see how that would change each player's ratings. The idea was to increase the sample size for each zone, since remember I am splitting the data in each zone into 6 pieces (and no one complained about that?)!

IOW, the way the program works right now, Bordick's 02 "outs/BIP" in the "56" zone is compared to the league (for the AL in 02) average "outs/BIP" for all SS's in zone "56". The other method is to compare Bordick's "outs/BIP" to 4 years' and both league's worth of SS "outs/BIP" in zone "56". At the end, I do an across the board adjustment to make sure that all SS's UZR runs add up to 0. USing the old method this happens automatically. Using the new method it does not. For some reason I am afraid to "rock the boat" and use the new method even though it should be much more accurate (basically it should reduce that "error range" for each player's UZR runs, by reducing the range of sample league ZR's in each zone, by using more (8 times more) league data).

BTW, I appreciate that Tango and the other "ratio adjustment guys" didn't scream at me for the way I applied the adjustments. Doing it the "right" way woulkd have been a programming nightmare. ALso, yes I really fudged the pitcher G/F adjustments. At the very least, as Tango suggested, GB/(GB+FB) is better than using G/F ratios...

I really mean no offense, but there isn't an NL SS within 12 runs of Juan Uribe except Jose Hernandez?

Mike Bordick is Ozzie Smith?

Yeah they are numbers and it took a million crazy calculations to figure them out - but they look pretty worthless.

I know you guys love to pat each other on the back but come on - trying to apply numbers like these to the 'real world' and speak about them as truths is why so many mainstream baseball types are so skeptical of statheads.

Jim: your skepticism is noted. Rather than dismissing the results, why not try to explain them. Do you think there is a bias in how the calculations are done, or do you think that there is too much noise in fielding stats, even after accounting for all the variables MGL considered?

On the other hand, if you do have a high year-to-year correlation, do you think that means anything?

There's a reason that Bordick converted all those plays into outs. MGL has considered where the ball was hit (close to him or not), whether the batter was lefty/righty, whether the pitcher was gb or fb, the base/out situations, and the speed of the ball. Is there something else that we should look at? If the results are that far out of whack from our perceptions, then either our perceptions are wrong, or there's alot of noise in the data. Then again, if there is alot of noise, we'd expect low year-to-year correlation. Let's wait for the results of that.

***
Shorty: I suspect that the low out-conversion rates for those short soft flyballs is because they are low flyballs. What we really want, more than anything, is hang time. Not sure why only football keeps track of this.

***
MGL: the error range is more than basing it on the binomial distribution. Every time you apply an adjustment factor, that factor itself, based on the sample size, is subject to error as well. As you noted, it's a balancing act. I suspect that the more factors you introduce, the greater the range of error, even if the overall error is reduced. Not sure if I'm saying it right, nor how to figure it out.

***
The 80/20 rule belongs to Bill James, but I subscribe to it.

Tango, this is a little of the "blind leading the blind" (neither one of us is a statistician per se - there are sabers with way more statistical acument than both of use combined - like Vinay, I think), but by introducing the adjustments, we are increasing the accuracy but decreasing the reliability (or maybe it's the other way around, or neither).

In any case, I think that the binomial model gives us a pretty good idea as to the standard error (the measurement error). The fact of the adjustments increases the standard error somewhat, but I don't think by much. I could be wrong though. For example, the difference in the standard errors of BA and OPS are great. It is more than twice as large for OPS than for BA.

However, even the measurement error does not tell us the "error range" in terms of estimating defensive ability from the UZR results. Those are two different (although related) things (accuracy versus reliability?). For example, we could come up with a totally worthless system (even more worthless than Jim thinks this system is), which would have a finite standard error, even very small, for a large enough sample, yet the "error range" in terms of how well it estimates a fielder's actual ability would be totally different. If that makes sense...

Anyway, I am going to try and address some of Shorty's, er MAH's, comments, and then I'm going back to work (on all my other baseball stuff). I sure wish that his comments were like his (previous) moniker! BTW, can we call you "The poster formerly known as Shorty"? I do like MAH better though. It gives you a little more credibility. That's why I chose MGL, rather than the "the baseball guy" or something like that. Even if my stuff is totally worthless, as Jim thinks, at least my name gives me SOME credibility.

As far as regression, I do them by hand (I'm like a fine craftsman when it comes to regression)! Seriously, I do, when they are not that important. There are many ways to come up with regression formuals. One way is to do a regression analysis and use the resultant regression formula. You want to use, say one year sample data (say, PF's) as one set of variables and as many other years as possible combined as your other set of variables. If you wanted to know how much to regress, say 10 years of data, you would use those 10-year, say PF's, as one set of variables, and another 10 (or some other time frame) years as your second set of variables, and again use the resultant regression formula. One problem (there are many) with this method (at least when it comes to park factors) is that you are assuming that all parks get regressed equally - not the case! For example, if Yankee stadium had a 1-year PF of 1.08 and so did Coors Field, which one would you regress the most? Actually it is a trick question because it is not that you would regress each one differently (you woudn't), it is that you would regress each one towards a different number. Using regression formulas from regression analyses doesn't allow you to do this (unless you plugged other things into your regression and did a mutliple regression analysis). Another way is to take all the 1.03's (for example) for 1 year and look at what they are the next year (or next 5 years) as a group. If they are 1.02 the next 5 years, then that is your regression for all 1 year 1.05's. Again, this doesn't distinguish between parks. Anyway, a good rule of thumb is that you should better regress too much than too little ("It is better to see too little than to see something that isn't there!") If you look at my chart, you will see that all of the PF's are regressed aggresively toward 1.0, unless I have reason to think otherwise (like Coors, Fenway, etc.). Another good rule of thumb is that as long as you think that there is some "true" differences among your samples, you are ALWAYS better off using them with an aggressive regression, then not using them at all. That was way too much about regression!

As far as the differences between GB and FB pitchers, your theory of defensive platooning is interesting. Perhaps that is the only explanation for the differences in out percentages. I hope not! If it is, then I shouldn't be adjusting at all! Perhaps, as someone suggested, infielders are more "heads up" when playing behind GB pitchers, and the same for the OF'ers and FB pitchers (is the number of GB or FB chances per game correlated with GB or FB out %?).

As far as why "soft" fly balls are harder to catch in the short zones, it is simply that the out percentages in the various zones determine how soft, medium, and hard is defined apparently. Apparently a "soft" fly ball or line drive is one with a low trajectory. As I said, the definitions of soft, medium, and hard are not so obvious for fly balls and line drives amd pop-ups. As I also said, it doesn't really matter how they derine any category, as long as all the "stringers" are consistent.

As far as real sow balls down the line being harder to turn into an out, I imagine that there is a subset of "soft" ground balls that become more difficult to field, but this subset must be small, and without having another category (like super-soft) there's nothing we can do about it. Also keep in mind that since I use league averages for every zone and for all 3 categories of batted ball speed, a slow ball in a short infield zone might in fact have a low out percentage.

As far as using GO/FO ratios rather than G/F ratios, I don't think there is going to be much difference (although I always wondered why the "official stat" was GO/FO ratio and not G/F ratio), even for extreme ratios. Let's say a pitcher has, say 200 GB and 100 FB, for a G/F ratio of 2.0. Because he is an extreme GB pitcher, he will have a .748 GB out% and a .695 FB out% (see the chart in the article). So his GB outs will be 149.6 and his FB outs will be 69.5. If a non-PBP databse uses these FO and GO numbers, they will think that the pitcher had 149.6/.735, or 203.5 GB's and 69.5/.702, or 99 FB's (based on an average GB and FB out% of .735 and .702), not too far off of the real 200 and 100. I'm not sure if that's what you meant.

That's all I can think of right now...

I guess I just think there is to much 'noise' when trying to create end all be all defensive metrics.

Let me first state that I am open to new ideas, I was blown away by DIPS and I am realtively openminded?

The first problem is that the information you are using is subjective. You might not think that there is a bias in this information but there inherently is. To draw a parallel from my own experience. I officate high school football. We have rule interpretation meetings where we watch NCAA sanctioned videos on everything from pass interference to unsportsmanlike conduct. They show us 15 plays and tell us what the call should be by NCAA ruling. We all watch the videos disagree with each other and then go out and call the games the way that we feel the rules should be interpreted. These types of biases will go into the PBP data, no matter what Stats tells the stringers to enter. I am assuming (I used to know someone that did it) that they still handle things the same way. The stringers are oftentimes at home watching on TV and many times do the same teams games over and over. Not only do you end up with just a difference of opinion on an event you get the bias of a person who has prejudices towards certain players. Oh, Mike Bordick is the best SS in the AL if he coudln't get to that ball. . . I'll enter my data accordingly.

The second great problem is that defense is not an individual aspect of the game. Unlike offense, on defense the players are completely intertwined - it seems impossible to take them out of a team context into an individual context. Is JT Snow really so bad at first, or does Dusty Baker just like to guard the lines? Is Jeter really that bad at short or did Brosius and Ventura 'steal' outs from him? What about injuries. Nagging injuries that won't keep a player out of the lineup may make a big difference as far as things like range, throwing ability and overall 'hustle'.

Just judging by the overall results - I don't know if they pass the 'Smell Test'. Is the best shortstop in the league really only worth 1 run a week or 1 run every 10 days over the worst shortstop in a huge offensive era? What do the results look like overall at the other positions?

Will there be correlation in future seasons? That might depend on how you define correlation I guess.

I do appreciate the effort and look forward to seeing the rest of the results, but at first look it doesn't look to me as though UZR is the holy grail that you are searching for.

Jim, most of your comments are valid. Some are even enlightening. I don't think that any (or all) of them are fatal to the system, however. I don't think that UZR is the "holy grail" of defensive metrics (whatever that means). I think that: 1) it is a good measure of a player's defensive performance, vis-a-vis his defensive value to his team, 2) it can be used as a good estimate of a player's defensive ability, and thus, his future defensive performnce, and 3) it is by far and away, the best defensive system out there (which does not matter much, of course, if the system does not hold up on its own merits).

Then again, I am biased. As I said on another thread, one of the biggest obstacles to seeing and utilizing the truth is human ego. So you may be right...

One would hope that "coders" are randomly assigned to games so that whatever biases they have are averaged out; e.g., nobody is coding a significantly disproportionate number of games for any particular team.


It strikes me that if this is not so it could be a devastating problem. We have, for example data that show significant deviations in umpires on balls and strikes, etc. However, they are rotated pretty fairly if not perfectly. But I personally doubt this is so for stringers, but it would seem to me that this is very much worth serious investigation and testing by the proponents of ultimate defensive metric systems.

About 6 years ago I worked with a guy who entered info for Stats - he did Cubs games. I was interested in getting involved and all you had to do was watch games on TV and enter what you saw. It only paid 5$ a game, so I wasn't really interested once I heard that. They did ask what teams you could see on TV where you lived and that would effect who they could assign to you.

Who knows, maybe the info is done by 3 people actually in the stadium, I don't know. I do know that people used to do it at home, and therefore would be watching on TV -which in itself is a huge problem.

Do I appreciate the work you've done and think that there are some great ideas there. Yes. However when you mention that there might a 15 run error range then it's hard to put a lot of stock into what I'm looking at.

If defense matters as little as these numbers show - that doesn't really jive with what DIPS says - that makes defense look awfully important, UZR tells me that there really is no difference in major league shortstops outside of the extreme outliers.

Do you have the rest of the positions somewhere on the web or have you not published them yet?

MGL asked me to format the 1999-2002 revised UZR. I'm hoping to do that on Monday, and I'll post them. Darin Erstad really stands out.

When I've looked at fielding in the past, I can pretty much tell you that the range should be +/- 20 runs for each position. And while one player doesn't have that much impact, you could have a team, like the 2002 Angels, that can be worth +100 runs over average.

Well, the above poster is right about bunts. Those are a separate deal. I know 3bs make tons of errors on bunts, so that can really skew things.

One thing I want to do with CAD is do a reconcile with UZR. Until we can improve upon things, we could assume UZR is "right" and then find out where there are differences, and why there are differences. Maybe teams that allow many walks screw with the fielding stats. Maybe teams that strike out many batters have fewer hard-hit balls.

1. How quickly does the SS get rid of the ball & how difficult or far or off balance was the throw? Many quick runners (& some contract dependancy) will have an infield ball to the SS that is often scored a hit. Your numbers only count, can't see the difficulty of the overall play.

To some degree, fixing the numbers for how hard the batter hit the ball lessens this impact.

2. Your infielder plays on a team with poor pitching which has more runners on base more frequently. When holding runners on it shortens the infielders range. How does your rating system account for this?

The base/out table fixes this.

3. You have the wrong shift on because: a. You have a poor manager who doesn't read his reports; b. You have a rookie pitcher that likes to 2nd guess the pitch call or can't hit his spots. Instead of a fastball inside its offspeed over the outside of the plate. Your range factor is now off.

This could affect the results. I doubt it is an issue in the major leagues, or in the high minors, not nowadays. Surely not #b. If #b were true, pitchers would have a significant skill of not allowing balls in play to turn into hits. And thanks to Voros McCracken, we know that, if there is such a skill, it is very small.

4. You play behind one of those lanky kids that falls off to one side or the other of the mound and always screens you from the ball so you have less time to react.

See above.

5. You have a weak fielder (i.e. weak 3b to SS) & they play you deeper and over to help keep balls in the infield.

This is a big issue. At some point when evaulating a fielder, we need to look at the infield as a unit, and the outfield as a unit, or at least look at the fielder with the men next to him.

6. Your team has a knuckeballer.

Clear issue. You know, I wonder on what plays a knuckleballer has an advantage. Does he cause more hard-hit balls? MGL, could you look at Wakefield, Candiotti and Sparks against their teams? Is there more or less of a certain type of play, or are some plays easier?

7. You have a poor fielding or stretching 1B. Your SS, 2B, 3B change the time they take with the throw to get it more accurate.

Clear issue. There is a difference, not a big one but enough of one to alter the numbers.

8. Your team is in a division with more bunters and you are doing your field rating on a 3B.

As I said above, I think MGL should count bunts alone, not with the other plays.

9. You have a poor pitching staff that loads the bases frequently forcing you to draw the infield in.

The base/out adjustment should fix most of this. Of more concern is the manager's tendency to bring in the infield.

For whatever it's worth, I ran a year to year correlation calculation for UZR runs and also for OPS. Here are the results:

For OPS, the data pairs were 1998 and 1999 and 2000 and 2001.

OPS, min 50 games for each year, r=.659, n=748. OPS, min 100 games per year, r=.748, n=324.

For UZR runs, the data points were UZR runs per 162 games, and also for min of 50 and 100 games. Also, if a player had no data pairs for consecutive years (because they didn't have the min number of games in one of those years), I used non-consecutive years. IOW, some of the data pairs in the UZR data are for non-consecutive years, but I never overlapped pairs (every data point is unique).

UZR runs, min 50 games for each year, r=.479, n=331. UZR runs, min 100 games per year, r=.528, n=190.

Here is the breakdown by position:

first base, UZR runs, min 50 games for each year, r=.278, n=56. UZR runs, min 100 games per year, r=.411, n=32.

second base, UZR runs, min 50 games for each year, r=.543, n=48. UZR runs, min 100 games per year, r=.484, n=31.

third base, UZR runs, min 50 games for each year, r=.606, n=48. UZR runs, min 100 games per year, r=.592, n=26.

SS, UZR runs, min 50 games for each year, r=.497, n=50. UZR runs, min 100 games per year, r=.526, n=34.

LF, UZR runs, min 50 games for each year, r=.672, n=39. UZR runs, min 100 games per year, r=.807, n=18.

CF, UZR runs, min 50 games for each year, r=.468, n=44. UZR runs, min 100 games per year, r=.396, n=25.

RF, UZR runs, min 50 games for each year, r=.20, n=46. UZR runs, min 100 games per year, r=.560, n=24.

I don't know what the various confidence intervals (for r) are relative to the sample sizes. Obviously for the indiviual positions, it is large...

My replies are similar to Charlie

1. I suppose MGL needs to add a batter/speed component.
2. MGL has baserunner/outs adjustments.
3. Yes, positioning is part of fielding. The player, not the manager, gets the blame in this system.
4. Screening from ball? Unless you play right behind him, what does this matter?
5. Yes, "stealing plays" and adjacent fielders will have an impact. A very valid point.
6. Knuckles: yes, that's correct. If we had more data, we would correct for that as well. For now, how much impact can 10-15% of the innings have on some of the teams?
7. 1B: yes, valid point. We should classify our 1B as good or poor catchers,and make the adjustmens accordingly.
8. Bunts: how many could there be? I'm not sure if MGL counts or ignores this. I suppose then you need a batter/bunterLevel adjustment too.
9. Bases loaded: covered in #2 above.

Important point: the most impactful situations are covered by MGL's 5 points - park, baserunners/out, pitching staff, batter handedness, speed of ball. And all of that combined causes most players a swing of around 2 or 3 runs (with Colorado an outlier). How much more swing do you thnk you'll get with these lesser variables?

I agree with Tango. All of the suggestions are good ones and someday I may incorporate them into the system (baserunner speed, knucklers, bunts, etc.), however, we can expect that any more "adjustments" will have minimal impact. OTOH, every little bit helps!

One thing to keep in mind. How much impact an adjustment has is dependent not only on the magnitude of the effect that that particular feature has, but on the liklihood that a player will have an anomolous "profle" with regard to that feature. For example, the pitcher's G/F ratio has a minimal overall effect because a pitcher's G/F ratio doesn't impact GB and FB out percentages all that much and it is unlikely that a player will have a pitching staff with a very unusual G/F ratio, even though he has only 4 or 5 pitchers he mainly fields behind (because pitchers in general do not vary all that much in G/F ratio, so 4 or 5 is enough to smooth things out). OTOH, while park effects do not impact GB and FB out percentages tremendously, a player is likely to have a non-average overall PF since he plays half his games in his home park. Anyway, the point I am trying to make is for something like speed of baserunners, while this may impact GB out %'s significantly, the likelihood of an individual player having an unusual group (speed-wise) of batters while is on the field is small. The only reason why the likelihood of a fielder having an unusual group of batters, L/R-wise, is greater is because the handedness of the batters is related to the handedness of the fielder's pitching staff, which is only around 4 or 5 pitchers (starters, at least), so for example, you would expect Oak fielders to see lots more RHB's, since their starting rotation has lots of lefties.

So basically in deciding what adjustmenets to use, you want to factor in not only the impact of that asjustment on a single batter or runner or pitcher basis, but the likelihood that individual fielders will have unusual combinations with regards to those adjutments...

Here's what I have so far:
MGL, great break downs.

Mike Emeigh and I have discussed and disagreed about positioning. MGL's breakdown of LHB adn RHB seem to make MWE right and me wrong. I'd prefer a different set of zones and more specific breakdowns, but I don't think defense should be adjusted for that.

STATS scorers: Jim is right. STATS scorers typically score ofr one team. While the suggestion of bias is fair, I don't think it's correct. I was a scorer just as Jim describes, but he overstates the difficulty *and* the motives. Besides, you get "graded" on your scoring. They would actually pblish a "newsletter' with scorers' "Percent Right". Please note - "Right". With multiple scorers, biases get smoothed over. I can't doctor a player unless all three (+) scorers are doctoring the same player.

Softness of fly balls: well, they are softly hit. Mostly those "loopers" the 2B can't quite get to and the RF is sliding in trying not to get a knee in the face from the retreating 2B. Even GBs, you know what is hard medium and soft. Okay, that isn't always perfect obviuosly, but three scorers will smooth the biaes.

3B plays: Your breakdown of 3B plays is interesting. MAH notes that they are the opposite of what one might expect. I disagree. When a LHB hits one down the 3B line, it is either *hard* or *soft*. much lower percentage of "mediums" (MGL can actually cross-check this). *IN addition* one *very* important aspect to this is MWE's ob re positioning (and a few others around USENET - I think DMN hates this part of it).

I think adjusting UZR based on factors is important. I think "type of play" is one - but I think type of play (zone/speed) mitigates the batter-handedness for the most part. And if not completely, I think unless there is an inordinate number of difficult plays, the adjustments should be very small.

I missed something - how did the COL SS get such a boost on a 3% PF?
Did he just get more "hard 56" balls?

I really like the PFs for COL and BOS LF. THat's some great work. I'm going to apply the PFs to my methodology and see what happens.

Thanks, MGL.

Oh - in my DPI system, I have Bordick at +20 runs prorated to 1430 innings. I need to finish the 02 rankings before 03 starts.. heh.

MGL, you offered your r scores without comment.

Some of the scores looked borderline-OK but I didn't see any really strong numbers. As a whole, pretty underwhelming evidence for UZR as a predictive tool for evaluating individual players. Hate to say this, but the scores reminded me of Voros' r scores for BABIP for indivudual pitchers from year to year (OK, maybe a little better than that, but not a whole lot) - and everyone seemed to regard that as prima facie evidence that BABIP was no way to evaluate a pitcher.

I wonder whether all the attempts to try to isolate a fielder's performance independent of his teammates' performance isn't somehow messing up the data. Even if the adjustment method *seems* logical, if it is too simplistic and fails to take account of the more subtle interactions between how fielders play together as a team, then you might just end up butchering your data beyond recognition, rather than *adjusting* it. Sort of like the cumulative error of every adjustment that Tango talked about.

Here's a reasonableness test. Is there a way to do UZR for the whole team, or the infield or the outfield as a group - basically a team-vs-team measure? If you did this, then you ditch all the adjustments for *sharing* balls between zones (would this be very different than DER?). Then, if a team or infield or outfield showed as being among the best, then you would expect their players to show as at least average to good relative to other fielders. If you didn't see that - in other words, you had a top fielding infield comprised of a bunch of supposedly weak to middling infielders (or vice-versa) - then that would be suggestive, maybe, of some sort of a problem with the sharing adjustments that are happening with UZR. Maybe the White Sox are a team to look at in this regard - it would seem that they're regarded popularly as a team of weak fielders but, if I remember some comments from Part 1, your data suggested otherwise.

Thinking about the team UZR (or DER, if that's what it amounts to), I have a hunch that the *best* defensive teams may just be those with consistent average to above-average fielders, BUT without any real liabilities (and maybe without many real stars - think Minnesota in recent years). In other words, the "greater than the sum of the parts" phenomenon. One liability on defense might significantly counteract the benefit of 3 stars, what with all the various, unnatural adjustments that a team might employ to try to minimize their exposure to their one big liability.

MGL's got the overall r at around .50, which is significant. If I remember Dan Werr's chart, he had BABIP or $H for pitchers at around .20 and for hitters at around .40, though he could correct me on that.

As for the sum of the parts comment, if MGL did the process as I think I read it, it should match exactly. That is, if the Angels individually come out to +100 runs, then as a team, without worrying about sharing, they would also come in at exactly +100.

One of the major reasons that the r is .50 rather than something higher, is that there are alot of interactions going on (pitcher, hitter, fielder). That adds up to alot of noise. And some fielders get alot less plays than others, meaning that their error range will be greater. It may very well be that this is the best that you can get, given the data. That is, given the level of noise, and the abundance of variables to consider, that an r of .50 is as good as it gets. That may be good enough for some, and not good enough for others. And if the quantitative approach isn't good enough, you'll have to rely on "good" scouting to make further adjustments, which is entirely possible, though highly problematic.

Dan Werr wrote the following

http://www.baseballprimer.com/articles/tangotigre_2003-03-20_0.shtml

BABIP shows substantially less consistency for hitters than other stats. Here are year-to-year correlation for hitters' stats 1990-2001... essentially, this describes how good a predictor a hitter's year is of his following year in each area:

BABIP: .476
AVG: .496
BB/(BB+AB): .808
SO/AB: .862
HR/AB: .809

For pitchers, it was:
HR/9: .470
BB/9: .692
SO/9: .776
BABIP: .218

**********

This is Tom. Seems to me that UZR is as reliable an indicator of fielding balls in play, as is batting average to a hitter as an indicator of making contact for a base hit.

Doug,

The individual r's by fielding position are "worthless" (or at least misleading) in the sense that the samples are too small. However, the overall r of .500 and change (for min 100 games or so) seems to suggest that there is considerable predictive value. In fact, if UZR r is roughly comparable to BA r, that would also seem to suggest that there isn't a whole lot of "noise" in UZR, since I don't think we consider there to be a lot of "noise" in BA (in fact, I'm surprised that the BA r is not considerably higher). In any case, you can "like" .500 or "dislike" it, but to "dismiss" it does not seem right...

FJM,

Yes, you are probably right that adjustments could be made for type of pitcher (power/finnesse, etc.). Maybe some time in the future I will add them. It could be that certain players like Womack get "punished" for having some unusual situation that UZR does not address. Who knows?

As far as Counsell, he may be a SS naturally which suggests that he would have better than average numbers at third, even if he were an average SS. I don't think he hits well enough to be a full-time 3B'man (of course neither does M. Williams). It may well be that he could (and should) be a full-time SS on many teams. I would find it hard to beleive that he couldn't do a better overall job than Womack, Guzman, Mears, or Halter.

Keep in mind (and I will repeat this until my dying breath) that these are SAMPLE results. They do not necessarily mean that they are indicative of a player's true fielding ability (sample error, sample error, sample error - the statistician's mantra), or if they are "indicative" we certainly don't know to any large degree of certainty, how indicative. In fact, it is almost a certainty that if you look at a whole list of sample UZR results, that you will see several players whose sample results are NOT particularly indicative of their true abilities (the ones that are 2 or 2.5 standard deviations off). Maybe Counsell is one, Snow is one, and T. Hunter is one. Their reputations suggest that they are...

It's very possible that I missed something in the explanation, but is it possible that Uribe (and other COL IFs) is unfairly getting his UZR increased. It seems very likely that there are more "hard" groundballs in Coors, making groundballs in general harder to field, resulting in the PF of .97.
If Uribe is already getting credit for these "hard" grounders, he may be getting double compensation for one factor.

It's very possible that I missed something in the explanation, but is it possible that Uribe (and other COL IFs) is unfairly getting his UZR increased. It seems very likely that there are more "hard" groundballs in Coors, making groundballs in general harder to field, resulting in the PF of .97.
If Uribe is already getting credit for these "hard" grounders, he may be getting double compensation for one factor.

Ryan, where have you been all my life? Yes, I never thought of that! I checked the exact same thing when it came to the G/F ratios of pitchers and the L/R batters, but I forgot to do it for IF park factors! It could be true for OF park factors as well, to a lesser degree. Since IF PF's are probably based on how hard the ground balls go thru the IF, yes there is probably some double counting, maybe lots of double counting. I will check it out and report back! Thanks!

Why didn't anyone else think of that?!

Well, I reran the park factors and I controlled for batted ball speed in both the infield and the outfield. I also controlled for percentage of LD's and FB's in the OF, whereas previously I lumped them all together.

The only big change was in Colorado, as Ryan suspected. The sample IF PF for Colorado after controlling for ball speed was 1.00, rather than .96 (I had regressed the .96 to .97 for the PF chart in the article). Interestingly, the OF PF's changed dramatically. Whereas previously the sample OF PF's in Colorado were .89, .91, and .89 for CF, LF, and RF respectively, after the batted ball speed and LD/FB adjustments, the sample OF PF's for Coors were .99, 1.02, and 1.01, respectively. I guess that the low unadjusted out percentage in the OF was due more to the batted ball speed and more line drives than to the expansive OF. I guess also that the low IF gound ball out percentage was also due to more hard hit ground balls (there are more DP's in Coors as well), whereas in the other parks, when the IF PF's are due to the playing surface it doesn't "show up" in the batted ball speed. IOW, the stringers must rate the batted ball speeds relative to the playing surface or the playing surface affects the batted ball speed enough to affect the out percentage but not enough to affect how the stringers rate the speed. Or something like that.

Anyway, I am going to change the IF and OF park factors in Coors, which should primarily affect the Rockie players, of course. I'll have Tango update the files on his web site. Thanks again to Ryan for pointing out my mistake!

ALso, when I make the adjustments (LD/FB and ball speed) in the OF, the Fenway sample PF in LF is .97 rather than .81. I'm not sure why. I need to do some more research on this.

One thing I'm not too happy about is that since I only have batted ball speed data from 99-02, my sample for the new PF's (with the adjustments) is much smaller than for the old PF's (4 years versus 10 years), especially for the old parks like Fenway. You'd be surprised how much variability there is in the defensive PF's, especially in the OF, when the sample size gets fairly small (less than 5 years or so).

Also, when I control for batted ball speed in the IF, the ARI IF PF is real low (.90), as compared to a previous factor of .95. Does anyone know whether the IF at the BOB is particularly short (fast)? Maybe that is one reason why Womack's UZR rating is so bad.

I am also a little concerned that when I control for batted ball speed at each park, things could get screwed up if we have "stringer bias", since I guess the same stringers tend to score the same teams...

FJM, Thanks for the info about the BOB. You can look up everyone's UZR rating at Tango's web site, linked above. I could separate hits to the OF versus infield hits for edification, but of course we want to include both (as well as bunts) in our IF PF's, as presumably there would be fewer IF hits on a fast IF (and vice versa), thus somewhat balancing things out...

FJM wrote:

"The percentage of ground ball basehits that stay in the infield (excluding bunts) could serve as an indicator of the speed of
the infield, thereby perhaps allowing you to group parks together for the purpose of calculating IF PF's."

That's a good idea!

BTW, there are 2 links for the UZR data on Tom's site. One is the "min 120 games" link and the other is a downloadable .csv file which has every player who had at least one chance in the field in 99-02...

There really isn't much to check since we already know that all of the Ari players in the IF will have considerable fewer (10% less) GB out's at home as compared to on the road...

FJM, when the computer crunches the data, every out made by a fielder is park adjusted according to what park the out occurred in, so it doesn't matter what team a fielder plays for.

As I said, the complete file should include every player who had at least one chance (or maybe 1 out - I forgot) at any position during 1999-2002, regardless of whether they were still playing in 2002.

As far as some players not showing up, I'll check the ones you mentioned and report back. I'm also running "infield hits" park factors right now, which will be based on non-bunt infield hits divided by total ground balls...

Cintron only played 8 games with 7 AB's in 2001 only, so presumably he didn't make any outs in the field. As I said, I think I have all players who played anytime in 99-02.

I'm trying to run those "IF hits" PF's, but it is a little tough to define an infield hit the way the hit locations are coded. The hit location for a GB hit to the outfield is the IF section where it went through the IF and a GB hit to the infield is where the ball is fielded or where it stops rolling, so the only distinction is who fields the ball (an outfielder or an infielder). The criteria I used for an infield hit is a non-bunt hit that was fielded (picked up, knocked down, thrown to first too late, etc.) by an infielder (not an outfielder) that did not land or was not fielded in one of the deep infield zones.

I'm not sure this is going to yield anything fruitful..

Yup, it looks like the "IF hits" PF's are all over the board, with no apparent rhyme or reason (too much fluctuation, too small sample sizes). It's not even worth posting them. It was a good thought. I think I'll just change the Colorado PF's (IF and OF) and redo the UZR ratings. I think I'm also going to redo all of the OF PF's since they changed slightly (should be more accurate) after I adjusted for the FB/LD ratios at home and on the road. In any case, Tango can update the files on his web site in a day or so...

MGL, your file only shows players with at least 40 games between 1999-2002, combined (that's how you gave it to me).

MAH has a good point. I assume MGL did it by IP. Defensive games are 8.75 Innings long. My rating system uses IP/8.75 for Def G, then i divide total chances by Def G to get an average play per game. Based ont he average play per game of each starter, I calculate the average plays per game for the average team and then make my calculations off that (compar to average).

Based on MGL's IF PFs, I would say IF PF is mostly a wash, and helpful, but not needed (but nice to have).

I have DPI numbers for all starters from 1998-2002, adn once the tweaking is done, I'll make a comparison to our two sets.

I have always used a "pitching staff correction" for OF. Perhaps MGL can tell us the difficulty average for teams behind good pitchers. (perhaps number Hard 3, medium 2, soft 1 to create a s/ number) I made mine based on SLG of non-HR hits. MGL, can you generate a "ptching staff" value for each team? Liek the avrage FB hit off Braves pitchers is 2.15, while the average FB hit off TB pitchers is 2.35 (10% higher). That adjusts by pitchers. Maybe you are saying you do this, but I missed that.

Chris, DIPS of course claims that any differences in average "hardness" of either ground balls or non-HR fly balls would be due almost entirely to chance, so before you use any data like that to adjust a fielder's rating, you would want to know whether that assertion were true. Of course, it would be OK to use any given year's "hardness" data to adjust that same year's fielding ratings, regardless of whether variations in a pitcher's or pitching staff's hardness rating was due to chance or pitcher skill (or the percentage of each). Where you would get into trouble would be to use one year's data to "conclude" that Glavine (or the Braves' staff as a whole) allows easier IF or OF batted balls than the average pitcher (say his average OF ball was a 1.8 and the league average was 2.0), and then you used that adjustment for a fielder behind Glavine (or a Braves fielder in general) for another year. In order to do that, you would need to know how much of that sample 1.8 was luck and how much was skill, if anything. IOW, you would need to know how much to regress that 1.8 towards the league average of 2.0 before you used it to adjust any fielder data for another time period. Again, DIPS says that that 1.8 should be regressed almost if not completely to 2.0. Iwould guess that it should be regressed pretty far towards 2.0, but not all the way (of course, how much to regress also depends upon how much sample data the 1.8 refers to - one year, two years, etc.). And we also know from my research on GB amd FB research that depending upon the GB/FB ratio of the pitcher or pitchers, we would regress the sample IF and OF average ball speed to a different number. For example, an extreme GB pitcher may allow more easy GB's and more hard FB's and vice versa (remember I said that a pitcher's G/F ratio and his GB and FO out percentages were somewhat indpendent of the speed of the FB's and GB's but not completely independent. In any case, you would still want to adjust your fielder ratings for the G/F ratio of that fielder's pitching staff, although as I mentioned in my article, the GB and FB out percentages, and thus a fielder's rating, is not that sensitive to the G/F ratio of his pitchers, and you would be hard pressed to find an entire pitching staff with an unusual G/F ratio...

One more thing...

If you look at pitcher or pitching staff average hardness rates for OF or IF balls, I think that "Stringer" bias is going to show up, or at least be a serious problem, assuming that the same stringer or stringers tend to score for one team. I suppose you could look at road games only for all pitchers, but then you are cutting your samples in half. Kind of like only looking at a batter's road stats only to eliminate home aprk adjustment issues (which is not a horrible idea, BTW, for a player with a very long career)...

An interesting finding (I bet you didn't know)....

Even though a fielder playing on a turf field at home makes substantially fewer errors than a fielder on a grass field (a typical infielder makes 6.8 errors per 200 chances at home on a turf field and 7.2 on a grass field), that is completely balanced out by the fact that a fielder who plays at home on a turf field makes more errors on the road on a grass field than a fielder who plays on grass at home. IOW, if a fielder is used to playing on turf (at home), he has particular trouble with grass on the road, as opposed to a fielder who is used to playing on grass at home (he DOESNT have trouble with grass on the road). SO basically, all fielders, at least as far as turf and grass are concerned, have the same error rate on the home and raod combined, regardless of whether their home field is truf or grass.

I thought that was interesting!

In fact, I may scrap my error park factors, since most of it is based on whether a park is grass or turf, and because of the above "syndrome" I applied them incorrectly anyway (there would have to be 2 different error PF's for grass parks - one for the home players and one for the road players).

Well, I did the ground ball through the infield thing the other way around this time. Rather than try and define an IF hit, I looked at the percentage of GB's that make it through to the OF, and are a hit of course. The results are much more interesting this time (I don't know why I tried it the other way around before).

Here are the "OF hits per non-bunt GB" sample park factors for all parks since 1993. I am also including, for your reading pleasure, the "bunt hits per gb" and the "IF hits per non-bunt GB" park factors.

I guess this should give you some idea as to the "speed" of the infield. One of the interesting things is that even though past research has shown that turf and grass parks have around the same GB out percentage (actually turf parks are slightly higher in hit rate), it looks like considerable more GB's make it through the turf parks but that more infield hits and bunt hits occur in the grass parks.

Does this mean that slow GB hitters are favored in turf parks and fast GB hitters are favored in grass parks? IOW, if you remove the bunt hits and infield hits (which for slow runners and/or power hitters are presumably few and far between), then what is important as far as whether a ground ball is a hit or not is the speed of the truf, which seems to vary signficiantly from park to park, even though that is not evident from only looking at overall outs (or hits) per GB.

(I don't know how to do the formatting on these comments, so this is probably going to look horrible!)

The format is park, OF hit PF, IF hit PF, Bunt PF.

ANA, .96, .97, .92.
ANA2, .96, .90. .96.
ARI, 1.39, .69, .73.
ARI2, 1.33, 1.05, 1.03.
ATL, 1.07, .94, .91.
ATL2, .98, .92, .81.
BAL, .94, .91, 1.28.
BOS, 1.08, .93, .85.
CHA, .95, 1.08, 1.14.
CHN, .98, .99, 1.46.
CIN, .93, 1.00, .97.
CIN2, 1.04, 1.14, 1.49.
CLE, .87, 1.08, 1.26.
CLE2, .93, .99, 1.22.
COL, 1.17, .86, .64.
COL2, 1.21, .98, .87.
DET, .98, 1.14, 1.30.
DET2, 1.10, 1.01, .98.
FLO, 1.02, .82, 1.56.
FLO2, .97, .97, 1.05.
HOU, 1.07, .96, .92.
HOU2, 1.09, 1.11, .58.
KCA, 1.15, 1.00, .74.
KCA2, 1.03, .91, 1.11.
LAN, .88, .96, .94.
MIL, 1.04, 1.16, 1.05.
MIL2, .96, .99, 1.71.
MIN, .98, 1.07, .64.
MON, 1.08, .98, .8.
NYA, 1.05, .96, 1.06.
NYN, .90, 1.06, 1.22.
OAK, 1.03, 1.09, 1.09.
OAK2, 1.05, .93, .83.
PHI, 1.00, 1.16, .91.
PHI2, .91, .83, .85.
PIT, 1.03, .99, .90.
PIT2, .91, 1.15, 1.31.
SDN, .93, 1.08, .96.
SEA, 1.04, 1.00, .76.
SEA2, .87, 1.25, .92.
SFN, .89, .94, 1.25.
SFN2, .97, 1.06, .8.
SLN, .98, 1.03, .64.
SLN2, 1.01, .94, 1.07.
TBA, 1.02, .66, .93.
TBA2, 1.10, 1.07, 1.01
TEX, .94, .83, 1.05.
TEX2, .92, 1.00, 1.22.
TOR, 1.00, 1.09, .67.

Lots of fascinating stuff here! I hope some of you take the time to look at it and comment. There seems to be large variation in speed of the IF, as indicated by the OF hits per non-bunt GB and percentage of IF hits. The bunt hit percentage is curious. It seems as if batters do not like to bunt on turf whether or not the turn is seemingly slow like grass or not (cin, phi, SL and tor all have OF hit PF's of 1.00 or less, indicating slow turf, but the bunt PF's are all less than 1.00). I don't know whether it is still tough to bunt even on slow turf or the batters just think that they shouldn't bunt on turf.

The same thing is true on fast grass. Batters like to bunt on grass even if the grass is fast like turf (ARI2, CIN2, FLO, KCA2, MIL, NYA, OAK, SLN2). Not in all cases though. There are some fast grass parks where batters apparently don't like to bunt (ATL, ARI, COL, BOS, DET2, OAK2). I need to have another category for bnt hits per bunt attempt to see if it is indeed much harder to bunt on turf fields than on grass fields, notwithstanding the speed of the turf or grass. Of course, there will probably be large selective sampling there as only the best bunters may bunt on turf versus grass.

I'm fascinated anyway, especially since it's 5:00 AM and I have to be in school at 7:30 and I haven't been to sleep yet!

Yes, they changed the turf (grass) in Ari in 99. Ditto for TB in 00 and PHI in 01. Of course, SL, Cin, and KC went to grass in 96, 01, and 95, respectively. The others were pretty much new stadiums or renovations.

Here is some more interesting data (note: none of the following data includes Coors, as either a turf or grass field):

There are 50% more bunt attempts (for a base hit) on grass than turf, which is not surprising, but the bunt success (bunt hits per bunt attempt) is actually higher on turf than on grass (.484 to .461). There is of course a different subset of players who bunt on grass and turf (I assume) and since turf bunts are less frequent, they probably occur when they are easier on the average. Nevertheless, since the bunt success in general is so great both on turf and on grass, the data suggests that many more bunts could perhaps be attempted on turf. Do you think?

Also, the bunt success rate ON THE ROAD is higher for turf teams and their opponents, which suggests that, unless for some reason turf teams have better bunters, which I doubt, that fielders who regularly play on turf have trouble fielding bunts on grass. In order to test this hypothesis, I need to break down the "turf team road data" into the home and road team, rather than lumping them together like you do when you conpute PF's.

As I mentioned before, higher error rate on grass (around 5%), but it also looks like when turf fielders play on grass, they make even more errors, as compared to fielders used to playing on grass (just like the bunt thing above). Again, I need to split up the road data.

It looks like turf fields do indeed allow around 3.5% more hits thru the IF. The road data is the same so it doesn't look like turf fielders have more trouble on grass as compared to grass fielders, as far as hits are concerned (a hit is not an error - an error is an out).

Perhaps most interestingly, turf fields actually allow 7.5% more IF hits! Is this because of the "high chop" hits? Is it because infielders are forced to play back further on turf? Both? I would have thought it would be the opposite as the grass fields tend to slow down ground balls. Also, for the IF hits, it looks like the same problem for the turf fielders. They appear to allow more IF hits on grass fields than the grass fielders. Is this becuase they play a little too deep on grass fields? Is it because they just have more trouble coming in on slow hit balls - the same trouble they appear to have on bunt attempts?

Because of the problems that turf fielders seem to have on grass, the PF for IF hits is around 1.00 for both grass and turf parks, even though the IF hit rate is higher on turf than grass.

So really, the reason why grass and turf parks have around the same hit rate is because you get more hits on ground balls on turf but fewer bunt hits (because there are fewer bunt attempts). On top of that you have many more errors on grass. On top of that, how you do on grass depends on whether you are the home team and the road team is a turf team or not.

So now I need to figure out the lwts value of a ground ball in a turf park as compared to a grass park, including errors, and with and without bunts included (since some players bunt a lot and others never bunt)...

Now that you mention it, scorer bias may be as much of an influence on individual "GB error" PF's (although I'm sure the difference between the turf and grass parks is real) as the condition of the field.

As far as the bunts, I try and exlcude all sac bunt attempts. The only thing I consider a bunt attempt (success or failure) is when there are 2 outs or the bases are empty. Everything else I "assume" is a sac bunt attempt (or course that is not always the case). I think this is a good idea though, although some "bunt attempts for a hit" are eliminated, all of the sac atempts are eliminated and most of the pitcher bunt attempts in the NL are eliminated so that NL and AL can be fairly compared as far as bunts are concerned. Also, I don't think there is much, if any, bias left in the bunt attempts I do keep track of (it's nice when you can remove a problem area and there isn't much bias or "selective sampling" in what's left)...

FJM,

Yes, it is surprising to me that even if the BOB infield is very hard, that the OF hit PF would we that high. Remember that these are sample PF's, subject to random variation, and that ARI2 is based on only 4 years of data. Your suspicion that visiting players do not field well at the BOB could be correct. I suspect they may play too shallow since it is a grass field but plays like a turf field (even much faster). I'll check.

As far as the low OF PF's for such turf fields as Philly and Min, yes, that is surprising. Again, the "real factor could be 1.01 or 1.02. Given enough parks (there are 52 I think), some of them will have sample PF's far from their true PF's. Expect 2 or 3 (5%) to have sample PF's more than 2 SD's away from their real PF's, right? That's why I regress them all before I use them.

As far as Sea and Ichiro, you are sugesting that his infield hits contributes to the SEA2 IF hit PF, right? He should have just as many IF hits on the road as at home, so that should not affect the SEA2 PF...

I'm curious about the correlations listed much earlier in this thread. Taking only players who played on different teams in year N and N+1, what is the correlation? I will note that I'm rather leary of the number of adjustments (park, pitcher, etc.) and micro zones that are built into this system and worry that we may be seeing correlation not between performance in N and N+1, but between the adjustments being made to the raw data. That is, the adjustments are swallowing the performance.