THE BOOK--Playing The Percentages In Baseball

I briefly looked through the “article” and noticed a couple of errors right away.  He mentions that the value of the HPB is higher than that of a walk because an HBP is indicative of a struggling pitcher.  I don’t know how Studes generated these run values (did he look at the change in RE or did he look at actual runs to end of inning after these events?), but I don’t think that it correct.  For one, the value of the HBP is greater than that of the BB because HBP are issued pretty randomly and BB’s are issued when they hurt the least (bases open, two outs, etc.) For another, even if HBP were indicative of a struggling piticher, I don’t think you want to use that “higher” value when looking at pitcher stats since you are going to include anyway the stats that ensue after an HBP.  It is double counting.  And I am not convinced that an HBP means that a pitcher is struggling.  I am not even sure what he means by that in the context of setting and using these values.

The author also says that there is little (or was it “no") evidence that a pitcher has any control over the percentage of fly balls that are home runs.  That is completely false.  Pitchers very much have control over the distances of their fly balls and hence the percentage of them that go for home runs.  This is easy to verify in a numbers of ways of course.  I don’t know where this idea came from that pitchers have little control over HR/HR, but it seems to be “going around” in some ssabermetric circles.

MGL, the run values I used were taken from Tom Ruane’s analysis on Retrosheet.

The notion that pitchers don’t have a lot of control over the HR/FB notion has been around for awhile, put forward by THT’s site and Annual, as well as Shandler.  The THT Annual had a .28 correlation year to year.

Typically, I’ve seen it ranked above BABIP but below K and BB rates in terms of predictability and persistence among individual pitchers.

I agree with MGL on the first part, and I made a note of that in Dave’s article at the end of that thread.

As for the second part, Dave Cameron said this:

We’ve seen very little evidence that major league pitchers have significant control over how often their flyballs go over the wall, so occassionally you’ll see a wild swing in performance that is not indicative of a players true talent level, simply because a pitcher is having more or less flyballs go over the wall than should be expected. Felix Hernandez in April and May of this year was a great example of a guy who allowed a lot of home runs per flyball, and that rate has steadily dropped as the season wore on. The average major league pitcher gives up home runs in about 11-12% of his outfield flies - significant variation from that is probably not an indicator of talent for a major league quality pitcher.

It’s semantical.  He was talking about “significant control”.  I’m not sure what that means, but he probably means that if the average HR rate is 11-12% per FB, then we shouldn’t expect pitchers to be above 20%.  Just taking a quick look at King Felix, highly touted at USS Mariner, and the example being used: 11 HR in April and May, with 48 FB.  23%.  After that, 7 HR (including a 3 HR game), 73 FB.  10%.

I don’t know what the “True” HR rate is per outfield FB (though easily determined with the data), but I’ll guess that 95% of the pitchers are within 4% of the league mean.

EVERYTHING gets regressed to the mean!  Some more than others.

As Tango pointed out in the thread on “maximum correaltion for projections” when you are doing year to year regressions, the “r” is limited by the sample size even when there is prefect correlation in terms of the underlying talent.

OK, we are talking about two different things.  One is the correlation from one time period to another and the other is the spread of talent among the population.  As long as their is some spread of talent, given a large enough sample that you are regressing to another large enough sample, the “r” will always be one.  However, when you get an “r” from a year to year (or whatever time period you are talking about) regression, you have to be careful about making an inference about the spread of talent from that “r”.  The magnitude of that “r” is a function of two things:  One, the spread of talent (the larger the spread, the larger the “r” for any given sample size), and two, the sample sizes of the two variables.  The larger the sample size, the larger the “r”, as long as their is some spread of talent.  In this case, since you are only looking at a pitcher’s fly balls, the sample size is fairly small, and an “r” of .28 (and we don’t even know that the “real” “r” is unless he was using a grip of pitchers in his regression) by no means indicates “little or no talent” or whatever you want to call it.

In fact, I wish we would get away from this “little control, a lot of control, etc.” EVERYHTING gets regressed given a finite sample of performance. If there is no spread of talent with respect to that performance, the regression is 100%.  If there is tremendous spread of talent (lots of “control") then the regression can be small, but it depends on the sample.  If we regressed batters’ one week BA on another week BA, what do you think the “r” would look like?  Would we conclude that there is little or no control over BA among batters because we found that the “r” was like .15 (for that week to week correlation)?  If the spread of talent is somewhere in between nothing and a lot, then we regress somewhere between a little and a lot, again, depending on the sample size of the data and our knowledge of the spread of talent.

So I don’t think the issue is semantical.  I really don’t.  I think there is a misconception that pitchers have “little or no” control over those HR/FB and that a pitcher’s true HR rate is almost compeltely a function of his FB rate.  That is simply not true.  In fact, in the research I have done, there is an indication that a pitcher’s true HR rate is very much independent of his FB rate.  Or at least as important.  Obviously there is some correlation (between FB rate and HR rate).

As far as the values of the HBP and BB, again I don’t know if Tom used the traditional method, which is to take the before and after RE state for each event.  If he did, and that is probably the case, then the difference between the HBP and the BB values has nothing to do with the HBP being indicative of wildness.  I don’t know where he got that from.

Here’s the link to Tom’s article:

http://www.retrosheet.org/Research/RuaneT/valueadd_art.htm

I think he used what you call the “traditional” method.  He doesn’t give an opinion regarding HBP vs. BB.

I’m someone who has emphasized the “home run totals generally equal flyball rate plus ballpark” point of view in the past (witness xFIP), though I have seen people overemphasize it, too.  In fact, I’ve resisted requests to highlight xFIP in our stats instead of FIP.

And I agree that I do confuse “control” with year-to-year predictability at times.  Thanks for giving me something to think about.  (Again!)

Right, that’s Tom Ruane’s article.  In The Book, I give the run values for the BB and HBP, and they are Markov-based (meaning the pitcher effect is removed).  It’s solely due to their distribution.

THT actually has the HR/FB correlation at .08, not .28, which if correct does seem to be inconsistent with MGL’s conclusions, even taking small sample sizes into account.  However, I’ve always found that .08 result surprising, and it would be great if MGL could share data that demonstrates a non-trivial talent spread.

Using the “spread in talent” process, I calculate the z-Score as 1.12 for HR/FB in 2005.  That is, if it were completely random, we expect to get 1.00.  We get 1.12.

The true HR/FB standard deviation = .02, meaning 95% of pitchers are +/- .04 HR/FB.

I’ll see if I can generate some data later on tonight.

Unfortunately, while the .02 standard deviation does not sound like a lot, what constitutes a lot of spread and what constitues a little, when it comes to HR per FB?  I don’t have a real good feel for that. 

Certainly when you put it in terms of HR/FB it sounds like a little (.02). However if you put in terms of HR per season or something like that, it sounds like a little more than a little.

A starting pitcher gives up around 300 FB per year (including pop ups I think).  Now, that .02 is 6 HR, which sounds like a lot more than .02, even though it is the same.  That means that 95% of all pitchers give up within 12 HR’s per season of each other given exactly the same number of FB’s.  That means that 1 out of 40 pitchers gives up 24 more or less HR’s per season (assuming a starter) than another 1 in 40!  That sounds like a lot to me.  And that is true talent according to Tango.  That does not sound like “little or no control” to me.  But again, that is only semantics. However, when I see the words “little or no” I don’t expect that to mean one standard deviation of 6 HR per season!

I get slightly different #s than MGL, but I think the conclusion is exactly right.  The average HR/FB rate is about .11.  Assuming an average of about 210 OFs for a starter, the difference between a good (.09) and bad (.13) rate is about 9 HRs/yr (or about 18 HRs between the very best and very worst).  Using back-of-envelope calculations, 9 HRs means about a 15 run difference, or about .65 in ERA.  Obviously, that’s an extremely meaningful talent difference.

Tango:  does a real talent SD of .02 seem consistent with a y-t-r correlation of .08?  If so, this seems like a great example of why y-t-y correlation is a limited tool for looking at true talent issues.

Off the top of my head, I think that .08 is not consistent with a .02 true talent SD.  Of course, we need to know the uncertainty of that .08.  In any case, I think the .08 is wrong, but I am not sure.  It wouldn’t take but a few minutes to figure a y-t-y corr. for HR/FB.  You want to somehow park adjust though or look at pitchers who switched teams, as HR/FB is clearly very sensitive to the park.  I would think that the park alone would produce greater than a .08 corr.

A better indication of a pitcher’s “pure” HR/FB rate would be average fly ball distance, adjusted for at least altitude, if not altitude AND temperature, or something like that.

r = var(true)/var(observed)

r = .02^2 / .04^2 = .25

And that’s when FB = 110

This makes the regression equation
= x / (x + FB)

where x = 330

That is, given 330 FB, you regress your HR/FB rate 50% towards the league mean.

In my sample, 110 outfield FB is on 120 IP.

For comparison purposes, you do the same to a hitter’s OBP rate at a level where x = 200 PA. 

All of this is quick back-of-the-envelope calculation.

The point though is that if one were to spend the required time, all the answers would materialize.

"That is, given 330 FB, you regress your HR/FB rate 50% towards the league mean.”

Without belaboring the point, that is too far from “little or no control” for that statement to be interpreted as correct, no matter how liberal you make the definition of “little.”

Some things are purely semantic, some are partially semantic, depending on your definition of terms.  This issue is NOT purely semantical.

And yes, .08 corr. ("r") is not consistent with Tango’s regression coefficient. 

Remember that regression = 1 minus r for the same sample size.  So if we regressed pitchers with 300 FB on themselves and another 300 FB, we would get an “r” of .5.  If the sample of pitchers with the .08 “r” were around 200 FB per pitcher, we expect the “r” to be around .3 and not .08.  So something is wrong with the .08.  It might simply be sample error if he wasn’t using a large sample of pitchers.  I assume he used a min number of FB (or PA) for the pitchers in his samples.

To be clear, Dave did say “not significant”, and not “little or no”.  Regardless, the numbers speak for themselves.

"Not significant” is significantly more ambiguous.  I agree that the numbers speak for themselves.  I also agree with Guy that a reasonable interpretation of the numbers is that there is “quite a bit” of talent or control among pitchers as far as HR/FB is concerned.  Again, somewhere along the line, someone started the erroneous notion that a pitcher’s fly ball rate is “everything” as far as his HR rate (HR per PA) is concerned.  Without verifying that, otherwise good analysts and writers are starting to throw it in, as David did in this article.  Unfortunately, that’s how bad things get started and perpetuated, even in sabermetric circles, such as, “Sac bunt attempts are bad and should be avoided like the plague, especially early in a game.”

Thanks, Guy. I looked at HR rate instead of HR/FB rate.

The results you guys are discussing here are different than what we found at THT, or what Shandler found.  This is more than semantics; it’s an entirely different research conclusion.

I did it off-the-cuff.  If you want to email me the following:
HR, outfield FB

for each pitcher (no names required), I can give you a more definite answer.  No need to limit the number of pitchers in the list, unless you want to.  You can send me 2005 instead as well.

Thanks, Tango.  Let me follow up with David first.  He did the original research, and I think he’s seen MGL’s research, too.

Here is some data on pitchers in 04 and 05.  I ran a regression of HR/FB in 04 on HR/FB in 05 for pitchers with at least 500 TBF per year.  The average TBF per year was 772 and the average FB was 209.  A FB was any air ball, not including line drives (according to STATS).

“r” was .232.

If I only use outfield fly balls, which is defined as all air balls, but not line drives (again, according to STATS), to only the OF locations (an average of 176 OF flies in the 772 TBF), I get an “r” of .222.

The number of pitchers in the sample is 88.

I I decrease the min TBF to only 200 (an average of 530 TBF), I get an “r” of .081 for 229 pitchers.

That sounds like what THT got.  We need a larger sample to decrease the uncertainty of these “r’s”.

If we increase the sample to include 98 on 99, 00 on 01, and 02 on 03, we get 890 pitchers with at least 200 TBF per year with an “r” of .181 (rather than the .08 with a smaller sample).

For pitchers with TBF greater than 499, we have 317 pitchers with and “r” of .190 (rather than the previous .232 with the smaller sample.

Nothing is park adjusted.

The low r in the Gassko/Bradbury calculations may also have been impacted by BIS coding changes.  For example, from 2003 to 2004 the number of FBs increased by about 3,000, while the out percentage on FBs fell from .914 to .855 (according to Pinto).  With sample sizes of over 40,000, those are huge—i.e. implausible—shifts.

Also in the THT annual TBF was 350, which is well below the number that MGL used to get the 0.22 r but marginally ahead of what he used to get .08 r

In MGL’s first study, he did FB=209, r=.232

This gives us an “x” for the equation of x / (x+FB) of 692.

In this case, 692 FB corresponds to about 2500 TBF.

In his next study, he did FB=176, r=.222 .  The “x” value is 617.  TBF corresponds to about 2700.

In his next study, he did r=.081, with an unknown number of FB, but which I will guess is FB=120.  The “x” value is 1361.

Now, there is a danger in how you do a correlation, if you don’t weight each sample appropriately.  This is why with a straight regression, you want your samples to have a similar number of “n”, amongst themselves, and in the paired sample.  There is a better way to do it otherwise, but more complicated.

Looking at the first two studies, we see that r=.50, when TBF is around 2500 or so, meaning about 580 IP.

In my off-the-cuff example earlier, the r=.50 corresponded to about 360 IP.

Yes, when your sample elements all have different N’s (TBF in this case) you get more variation than the average TBF would suggest.  That is one reason why you want to have at least a min number of TBF (or whatever it is), to keep that variability to a minumum.

In any case, as you can see when you are doing a regression with only a couple of hundred players, you have a large uncertainty.  You can look it up in a chart, what the standard deviation of the “r” is, given the number of elements in your sample and the observed “r”.

Plus, as I said in a previous post, we don’t know how much of the correlation is due to the park.

If I increase the data and get 407 pitchers with at least 500 TBF in each season (average of 760 TBF and 196 FB), I get an “r” of .190.

Oh, and I do not include any games in Coors Field in the data.

I re-ran the correlations for players who switched teams from one year to the other, in order to make sure that the correlation was not being significantly influenced by the park HR factor.

For 94-05 (regressing 94 on 95, 96 on 97, etc.) data, there were 85 pitchers who had at least 500 TBF in each of two consecutive years and switched teams.  The “r” was .203, so the suggestion is that the “r’s” that we are getting are NOT due to the parks only or even mainly.

If you can report the average number of FB, we can calculate the regression equation rather easily:

n * (1-r)/r = x

n is FB or BFP or whathave you

r = n/(n+x)

x tells us how much a sample size we need to regress 50%

In may last sample of players who switched teams, the average TBF was 746 and the average FB was 197.  “R” was .205 for 85 pitchers.

In the larger sample of all pitchers, the “r” was .190 and the sample size was 407, the average TBF was 760 and the average fb was 196.

I would use the “r” and the other data from the second sample, since it is much larger and the uncertainty from the “r” is much smaller.

The resultant “x” would be similar in either case, more or less, 800 FB, or about 670 IP.

I’m a few days late to this party, but I’ll add a few thoughts anyways. 

I’m not sure I agree with MGL that the difference in HR/FB rate is indicative of a talent (among major league pitchers, obviously) as much as it is the biproduct of a skillset.  Yes, there is some correlation in HR rate beyond simple FB%, but I think what we’re seeing is a correlation in pitcher type, not HR-prevention-skill. 

Bare with me, as this probably is semantics, but here’s my thinking - groundball pitchers post higher HR/FB rates than flyball pitchers as a group, and always will, because of mistake pitches.  For a groundball pitcher, most flyballs are mistake pitches, often poorly located, and mistakes get jumped on.  A flyball pitcher, however, is more likely to have good location on his pitch when he gets a flyball.  A smaller percentage of the flyball pitchers flyballs will be mistakes, and thus, a smaller percentage will fly over the wall. 

A little evidence, though its nothing close to a thorough study on the issue, using THT’s data (which is park adjusted, by the way):

48 starting pitchers from ‘04-’06 have posted HR/FB% of 15.0% or higher.  Those 48 pitchers, as a group, had a HR/FB rate of 17.0%.  Their collective GB% was 49%, well above league average.  As a group, the guys who gave up the most home runs in a season were groundball heavy.  The guys with the six highest HR/FB rates the past three years are 2005 Derek Lowe, 2004 Matt Morris, 2005 Jake Westbrook, 2005 Brandon Webb, 2004 Greg Maddux, and 2006 Felix Hernandez.  That’s basically a who’s who of the groundball elite.

Conversely, 58 guys have posted a HR/FB% of 10.0% or less during the same time period.  Their collective HR/FB rate was 9.0%, and as a group, their groundball rate was a much lower 44%.  In fact, of the 16 lower HR/FB% seasons of the past three years, only one (2004 Tim Hudson) had a GB% of higher than 50%, while all of the guys with the six highest HR/FB rates in that same time span had a GB% north of 50%. 

My belief is that if we looked at a pitcher by skillset (maybe broken into quintiles, such as extreme groundballer, groundball, neutral, flyballer, and extreme flyballer) and developed expected HR/FB rates for each of the five different categories of pitcher, we’d see different markers for different pitcher types, and very little year to year correlation in HR rates beyond what we’d expect from that pitcher type. 

Again, more non-thorough data, because all I have access to is what THT or fangrahs publishes.  I took the 261 seasons that qualified for the ERA title from ‘04-’06 that THT publishes and broke them out into these quintiles. 

The 53 pitchers with the lowest GB%, extreme flyball guys posted a HR/FB rate of 12.1%

The 52 pitchers with low-to-average GB% had a HR/FB rate of 11.5%.

The 52 pitchers with an average-to-high GB% had a HR/FB rate of 11.9%.

The 52 pitchers with a high GB% had a HR/FB rate of 12.3%.

The 52 pitchers with the highest GB% had a HR/FB rate of 13.6%

You’ll notice that the rate is higher than 11% for all five groups, and thats because this analysis only includes starting pitchers.  Starters post about 1% higher HR/FB rates than relievers do.

I won’t be surprised at all if Felix’s HR/FB rate is always higher than 11%, because I’m not convinced that his skillset’s true HR/FB is 11%.

Interesting, David.  It sure would be nice to know if pitchers varied in their “true” HR/FB rates solely as a function of their GB/FB ratios.

Although whether that is a “skill” or part of a “skillset” is definitely a semantical issue, there is a big difference between the variance being unique to the pitcher himself or being unique solely to his GB/FB ratio.  If the former, in doing a projection, you would regress every pitcher’s HR rate to some common number.  If the latter, you would simply assign a pitcher a fixed HR rate according to his GB/FB ratio, the latter not fluctuating a whole lot, even in a relatively small sample of performance.  My guess is that it is a combination of the two.  Let’s see:

If I look at pitchers with a GB/FB ratio of less than 1.1 only (average=.90), I get 69 pitchers from 94-05 and an “r” of .258.

For the GB pitchers (greater than 1.6, average of 2.20), I get 78 pitchers with an “r” of .176.

I guess that within each group, the correlation could still be variation in GB/FB ratio.  What about if I look at all pitchers who are about neutral in G/F ratio, say 1.2 to 1.5 (average pitcher with > 500 TBF per year is 1.35)?

We get 49 pitchers with an “r” of .142.

For pitchers less than 1 and greater than .7, we get 29 pitchers with an “r” of .371.

For pitchers greater than 1.4 but less than 1.7, we get 33 pitchers with an “r” of .306.

So, to be honest, although we have some very small sample sizes, I don’t see any suggestion that there is NOT variation among pitchers with similar G/F ratios.

I am pretty sure there are better ways of doing the analysis, like doing a multiple regression, but I am not enough of a statistician to know exactly what to do.

What if we regress G/F rate on HR/FB rate?  We get 1357 pitchers with at least 500 TBF and an “r” of .087, suggesting that G/F rate alone is less predictive of a pitcher’s HR/FB rate than a pitcher’s previous HR/FB rate, again, suggesting that there is a “talent” component to a pitcher’s HR/FB rate, independent of his G/F rate.

Now, one of the problems with this kind of analysis is that you have all kinds of selective sampling and cross-correlation issues.  One way to avoid that is to look at two groups of pitchers in one year, both with the same average HR/FB rates.  One group is fly ball pitchers and one group is ground ball pitchers.  If David’s hypothesis is correct, we would expect to see the HR/FB rate of the GB pitches higher in the subsequent year than that of the FB pitchers. Let’s see:

I looked at all pitchers with HR rates around .11 (.10 to .12) in year X.  Among these, I split them up into 2 groups.  One group had a G/F ratio below 1.25 and the other group, greater than 1.30.  So one group are fly ball pitchers and the other are GB pitchers, but they had the same HR rates, collectively, in year X.  If what David is hypothesizing is true, we would expect their HR rates to be different the following year.  We would expect the GB pitchers to have a higher HR/FB rate in year X+1.  It is true that FB pitchers have a lower HR rate than GB pitchers - .106 to .113 (less than 1.25 G/F ratio and greater than 1.30 G/F ratio).  So we would expect the FB pitchers to have a HR rate of around .106 in year X+1 and the GB pitchers, around .113.  That is not what I found.  They had around the same HR/FB rates the following year.  Here are the results:

Fly ball pitchers

HR rate in year X = .110 HR rate in year X+1=.107

Ground ball pitchers

HR rate in year X = .112 HR rate in year X+1=.109

Anyway, I am tired, and my above analysis is far from rigorous and suffers from a lot of sample size problems, but I just don’t see much evidence for David’s proposition.

Certainly is a fascinating topic!

Btw, please stop using ratios, as a variable in a regression.  You must use rates.  If the mean ratio is 1.0, then a 0.5 GB/FB ratio is exactly the same as a 2.0 FB/GB ratio.  A .333 GB rate or .667 FB rate shows this to be true.

As well, David is looking at all contacted balls, and therefore, his GB rate is GB per contact ball.  MGL looks like he’s eliminating LD and IF (though I can’t tell).

I knew you (Tango) wouldn’t like the ratio thing!  I am eliminating LD only, but I don’t think it would make any difference at all if I included them.

Even with small sample sizes, MGL’s results seem pretty conclusive.  At most, there’s a weak relationship btwn GB% and HR/FB.  It would make sense that there would be some correlation even if the two skills are totally independent, simply because a medium-high FB%/high HR per FB pitcher can’t last in the majors—he’ll give up too many HRs—while an extreme GB pitcher can survive even with a relatively high HR/FB.

And pretty clearly, there IS a HR-prevention skill, separate from GB/FB tendency.

While I appreciate the information, I’d stop short of calling MGL’s work “conclusive”.  I think we really need to see a thorough, extensive study on the issue that considers all factors. 

I agree that this HR prevention skill exists, but I think its far more rare than the idea of a “three true outcomes” mantra would have people believe. I think we can see that there are skillsets that lead to lower HR/FB rates than other skillsets, and there are examples within those skillsets who don’t look to be constrained by that relative baseline. 

But having looked at the issue a lot lately, about the only names of starting pitchers (again, relievers are a whole other animal) that I can come up with that consistently post a HR/FB rate far enough away from the park adjusted norm are Roger Clemens and John Lackey. 

And I’m not convinced on Lackey just yet, because I think Edison Field may be doing something funky that we haven’t figured out how to measure yet.  In Chris Dial’s fielding work, he rated Garret Anderson, Darin Erstad, and Tim Salmon as the best at their respective positions in the past 20 years, and Lackey isn’t the only Angel pitcher who sustained lower than average HR/FB rates while pitching in Anaheim (Washburn, Byrd, Escobar, and now Weaver are all examples).  I don’t have the answer, but I think we need to look at Edison Field’s outfield fly park factor a little closer.

FWIW, I took the Hardball times 2004-2006 stats and filtered by BFP (>350).  The R for year to year was .19.  The 2005-2006 had a lower sample size (we’ll see when the seasons over), but that’s consistent w/ what MGL came up with.  Those numbers are park adjusted and are HR/OF.

Take the following with a grain of salt, the sample sizes are too small to mean anything, but it’s interesting.

I took the sample from the Angels (>350 BFP) and found combined for a 10.3% HR/F, but the team as a whole was 11.0%. Interestingly, the Mariners showed no difference based on BFP>350 vs the team as a whole (12.3%-12.3%). 

Keep in mind this wasn’t using PBP data, and the numbers were determined using Excel functions (Batted Balls * GB, Batted Balls * LD, OF * HR/F) to figure out the counting totals, which I used for the >350 BFP vs whole team comparison.  This is park adjusted, so it’s probably not very accurate (wouldn’t it be better to compare Raw data when comparing the team?).  I also compared the team totals (all pitchers vs >350 BFP), not individual players.

That all said, I’m not the best person to be doing this, or to listen to on tis, but I like to confirm the data I’m given, whenever I can, even if I’m not the most qualified person to do so. 

PS - how accurate is the Retrosheet Batted Ball data?  I’ve toyed with the data quite a bit, but I’ve never been sure of the accuracy.

"But having looked at the issue a lot lately, about the only names of starting pitchers (again, relievers are a whole other animal) that I can come up with that consistently post a HR/FB rate far enough away from the park adjusted norm are Roger Clemens and John Lackey.”

Well, having looked at these stats for about 3 minutes, I found three more:  Pedro (.08 over last 4 years), Glavine (.09), and Santana (.09).  The difference between .09 and an average starter (.12) is about 0.5 R/G.  Perhaps your list is so short because you’re looking for “consistently” --I don’t know what you mean by that, but these numbers will of necessity bounce around quite a bit since N is only about 200 for most pitchers. 

In any case, MGL and Tango have both demonstrated, using totally different methods, that there’s a skill here.  The only evidence to the contrary I know of is the small r seen when the threshold is reduced to 350 BFP.  But that includes pitchers with as few as 80 OFs in a season, an absurdly small sample given a true talent SD of .02.  I don’t see why we need a “thorough, extensive study” before concluding this is a skill—we should assume it is a skill until/unless someone can show it isn’t. 

Now, it may be that Ks and BBs are in some sense a “more true” outcome than HRs, in the sense that variations in those talents are more significant.  I’d guess that is true for Ks, probably not for BBs, but I’m not sure.  To know, we’d have to now the SDs for those rates, and then translate that into runs.

If certain pitchers were repeatedly posting better HR/F numbers at home than other pitchers in the same park, wouldn’t that indicate a certain level of control?

I downloaded the 2004/05 THT data, and did the “spread in talent” process, on the combined data.

For pitchers with a total of at least 200 OF over those two years, I get 147 pitchers, with an average of 306 OF.  The z-score is 1.27.

The random SD is .019, meaning the observed is .024, making the true .015.

r = .015^2/.024^2 = .39

If I expand to at least 100 OF, I get 288 pitchers, average of 185 OF, z-score of 1.18.  The random is .024, observed is .029, true is .016.

r = .016^2/.029^2 = .30

The “x” for the 1st equation using n * (1-r)/r = x is 480.  In the 2nd equation, it’s 430.

So, we regress a pitcher’s performance, 50% towards the mean, when we have around 450 OF, which corresponds to also around 450 IP.

Next up, I’ll look at David’s idea of GB and FB pitchers.

Well, having looked at these stats for about 3 minutes, I found three more:  Pedro (.08 over last 4 years), Glavine (.09), and Santana (.09).  The difference between .09 and an average starter (.12) is about 0.5 R/G.  Perhaps your list is so short because you’re looking for “consistently” --I don’t know what you mean by that, but these numbers will of necessity bounce around quite a bit since N is only about 200 for most pitchers.

THT has Glavine’s HR/FB rates the last three years as 11.1%, 6.6%, and 14.4%.  That, to me, doesn’t come out to 9%, and it certainly doesn’t look like a skill.

THT has Santana’s HR/FB rates the last three years as 12.7%, 10.9%, and 12.4%.  Same comment as above.

THT has Martinez’s HR/FB rates the last three years as 12.6%, 9.0%, and 13.6%.  Same comment as above.

I suspect you’re using non-park adjusted numbers.  You can argue with THT’s park adjustments, I guess, but I can’t imagine an argument where non-park adjusted numbers are okay for this kind of analysis.

Looking at all pitchers with at least 100 OF and 100 GB.  277 pitchers.

Just some background info.  The observed SD of the gb rate is a very large 3.46.  All this means is that GB is a true and very noticeable skill.  The average BIP was 580.  With the random SD for this being .021, and the observed .073, we get a true of .070.
r=(.070/.073)^2=.92
x=.08/.92*580=50
This means that after 50 contacted balls, we regress a pitcher’s GB rate 50% towards the mean.  So, we can pretty much instantly tell if a pitcher is a GB pitcher or not.

(BIP = IF + GB + LD + OF)

Anyway, let’s get on with it.  I now have the HR/OF z-scores, and the GB/BIP z-scores.  The regression between the two gives me an r of .08.  So, there is some support here that the higher the GB rate, the higher the HR rate.  If I take the top 20 z-scores in HR/OF, that’s 2.31.  Their GB/BIP z-score is 1.38.  On the flip-side though, the lowest in HR/OF has a z-score of -2.33, but a GB/BIP z-score of .08.

The top 20 in GB/BIP had a z-score of 7.42, and a HR/OF z-score of 0.27 (essentially zero).  The top 3 GB/BIP (Webb, Lowe, Westbrook) all indeed did have very high HR/OF z-scores, of around 2.0.  Mulder was in the top 5, and he too had a high z-score of .91.  Marquis rounds out the high-GB pitchers with high-HR rates.  (Maddux is up there as well.) However, Drese, a GB pitcher, has a low-HR rate.

I think this is a case where in the very extreme GB guys, they will have a higher-than-normal HR rate, but over the whole population of pitchers, the relationship is fairly weak.

Btw, the z-score for LD is 1.0, meaning it’s completely random.  I’ll have to look at it some more, since this doesn’t seem right.

My analysis is not park adjusted at all, but as I indicated, when I look at pitchers who switch teams, thereby eliminating the home park bias, I still get a reasonable “r”.  I probably should have looked at road data only to get rid of the home park bias.  Even with imbalanced schedules, that is always a nice way to do that (get rid of home park biases).

It was in interesting point that there is probably a selection bias in baseball, in that GB pitchers who have a high true HR?FB rate are going to still fare OK, whereas FB pitchers will not.  In fact, that is an excellent point.  There might not be any cause/effect at all (the traditional and logical view that when a sinker-baller makes a mistake, the pitch gets killed).  It might just be that there are more FB pitchers who are generally better at keeping their HR/FB rates down, for whatever reason.  Speaking of that, it would be nice to know “why” some pitchers have a high HR rate and others have a low one.  One of the secrets which will eventually be unlocked using the TLV pitch data.  Is it that the poor ones make more “mistakes?” Is it that overall the poor ones have less velocity on their pitches?  Do they throw more offspeed?  More fastballs in fastballs counts?  Etc.?  I have no idea.

Just out of curiosity, why would you think the best response would be using only road data rather than just park adjusting all the data? THT thinks its possible, anyways, and I’m sure you have enough data due to UZR to build a park adjustment for HR/FB rates. 

And I agree, the point about selection bias is a good one, and that may very well be the answer - it might not be that groundball pitchers’ mistakes get pounded more often, but that I see more groundball pitchers mistakes get pounded because the flyball pitchers who make mistakes get eliminated very quickly.  That’s definitely a possibility that I hadn’t considered. 

And I agree, TLV data is the key to this secret.  When I did the Charting Felix series (which, if I recall, you weren’t a big fan of, incidentally), it became pretty clear to me that Felix wasn’t “unlucky” in how many home runs he was allowing.  He was legitimately throwing bad pitches, and they were getting crushed.  Now, I don’t have any kind of baseline to compare it to, so I don’t know if he was throwing more “bad pitches” than other pitchers, but most of the home runs came on 92-95 MPH fastballs up in the zone, middle of the plate, in fastball counts.  Pretty easy to see coming. 

Anyways, all that said, I think we pretty much all agree.

Re: Pedro, Glavine, and Santana, I was looking at Studes’ Batted Balls Library, which gives HR/OF for 2002, 2003, 2004, and 2005.
Pedro: .07,.04, .11, .08
Glavine .09, .10, .10, .06
Santana .06, .09, .11, .09
Maybe park factors raise those numbers a lot, but I’d be skeptical of park factors that large.  In any case, it’s pretty clear that Pedro (.69 HR/9 IP lifetime) and Glavine (.68) have this skill; time will tell on Santana.

You’re more skeptical of Studes park factors than you are of not using them at all?

I’m sorry, but that just doesn’t make any sense to me.  And saying that it’s clear that Pedro and Glavine have this skill is just making a claim you can’t support.

I never commented on your Felix Hernandez article, if you were referring to me, David, IIRC.

Park factors are ALWAYS imprecise and problematic.  When you don’t mind decreasing your sample size, or you think the tradeoff is worth it, using road data only is ALWAYS a good way to neutralize the possible adverse effects of PF’s.

And yes, a big mistake that many people make is looking at very anomolous performances and saying that it is OBVIOUS that that player or players has that “whatever.” That is not the way it works.  FIRST you determine if there is ANY spread in talent in the population and if yes, you take a player’s performance and do the appropriate regression.  And even then, that is only a BEST ESTIMATE of that player’s skill.  We HAVE NO IDEA whether that player for certain has that skill or not or what the magnitude of that skill is.  We can only estimate it and/or estimate the chances that he has that skill and the chances that he has X of it, Y of it, etc. 

For example, if we say that we cannot find a clutch hitting skill in the population of ML players (which is not true of course, but let’s assume that it is), then no matter how large any particular player’s clutch hitting splits (clutch versus non-clutch) are, we have to conclude that that player has no such skill.  There is no argument, “But what about so-and-so?  Have you seen his numbers?” It does not work that way. So with regard to Galvine and Pedro and any other player, since we are pretty sure that there is indeed a spread of talent among pitchers with regard to HR/Fb rate, we can estimate the level of skill that Pedro and Glavine have based on their .68 and .69 career numbers of whatever they are and the corresponding number of career FB allowed, by appropriately regressing their sample rates.  That does NOT mean, however, that we are certain that their skill is more or less than the major league average.  All we can say is that there is a 90% chance or 99% (I have no idea what it is) that they have a significant skill, etc.  Or we can do the calcs and say that there is a 20% chance that their true rate is less than .7, and a 50% chance it is less than .8, or whatever (numbers for illustration purposes only).

"saying that it’s clear that Pedro and Glavine have this skill is just making a claim you can’t support.”

Why would you say that?  Through 2005, Pedro gave up 194 HR on 6390 BIP.  For 2002-2006 his OF% was .35, and since your own analysis suggests that GB/OF tendency is a stable skillset, let’s assume that’s his career rate.  Even if he was a bit more of a GB pitcher early in his career, it won’t significantly change the conclusion.  That means 194 HR on 2237 FB or .087 HR/FB (I’ll include any HR on LD, just to be generous).  The 95% confidence interval is +/-.013, so we can say there’s a 97.5% chance that Pedro’s true rate is 10.0 or below. 

Repeat for Glavine (300 HR, 12884 BIP, .29 OF%), and we get .080 HR/FB, +/-.01, or a 97.5% chance he’s below .09.  I’m too lazy to figure out the chances that each are truly above average (<.12), but it’s well over 99%.  So, saying it’s “pretty clear” they have this skill actually understates the conclusion we can reach. 

On park factors, I said I’d be skeptical if they raised the rates “a lot”.  On the two years we have in common (2004-2005), it looks like park-adjusting bumps up the HR/FB rates just 1 point, not a lot.  Even if we had park factors for their whole careers, I can’t see park adjustments changing the conclusion that both Pedro and Glavine are good at preventing HRs on FBs, especially since neither pitched most of their career in Shea (which Studes rates as an extreme anti-HR park).

I have to agree with Guy, although I don’t really think there is a disagreement here.

I just downloaded all of the Batted Ball data (for qualified starters) from Fangraphs.com (2002-2006), and I’m getting an r value of .24 for the correlation to GB% and HR/F, and -.25 for FB%, which is in line w/ what Dave was saying.  Also, the Year to Year r is 0.28 for HR/F using that data.  Is that data (HR/F) park adjusted?  That could be skewing the numbers.

That GB to HR/F correlation doesn’t jive at all w/ the numbers (2004-2006) from THT.  I dont have the spread sheet in front of me (I guess I could download it), but when I looked it was more like .05 not .24.

Iget r=.2 using THT data (GB% to HR/F) I only grabbed qualified pitchers, to make it quick.  Must have done something wrong before.  Like I said, I don’t have the spreadsheet I used when I got .05 in front of me.  Maybe I was using the wrong column or something.  Anyway, this was a great discussion.

OK, I was looking at the THT batted ball data, curious about what Dave Cameron was talking about RE: HR/F being relative to GB% of the said pitchers.

For Qualified Pitchers the correlation for GB%BFP to HR/F is r = .20.  Using GB/(Batted Ball) to HR/F, I got r = .19.

Using IP >= 100, the correlation for GB%BFP HR/F is r = .15 and for GB/(Batted Ball) to HR/F the r = .12. 

Next I tried to weight the GB% w/ an exponent to look at Daves suggestion that it is particularly weighted at the extremes.  I tried a series of exponents (1-16) and found (GB/Batted Ball)^5 correlates .22 to HR/F, which was the best correlation I could get.  This could be overfitting, but it’s interesting, it fits Daves suggestion this is particularly pronounced w/ extreme GB pitchers.

Also, Interestingly (OF/Batted Ball)^2 correlates -.17 to HR/F, yet (OF/Batted Ball) is -.19 and (OF/Batted Ball)^5 = -.12, so perhaps high OF pitchers do not have a lower HR/F.  OF/BFP has an even lower corelation (-.09) to HR/F.

Next I wanted to eliminate single year bias, so I took and made an agreggate of the counting stats, and looked at only players who face >1500 BFP since 2004 and did the same thing (GB/BFP)^2, again GB%^5 was the best fit at r=.25. It was only 118 pitchers so it may not be conclusive, but these pitchers did face more than 1500 batters. 

I took a quick look at LD/airball and GB%, and the R = .80, which surprise me.  It then dawned on me, I should look at HR/Airball instead of HR/OF.  I got an r = -.04 for HR/Airball and GB%.  This (the LD% and HR/airball correlations) was using 3 years data aggregate >= 1500 BFP.  I’ll have to look at this further, but it looks like, while High GB% pitchers, and especially extreme ones do give up higher HR per OF, they give up more LD per airball evening the HR numbers out overall.  An exponent to fit the GB% did not seem to affect the relationship to LD/Airball.

OK, I’d have to look at this further, I think I can say w/ fairly certainty, extreme GB pitchers give up more HR per Outfield Fly, however, it appears (and this I cant say w/ as much certainty) that they give up more LD to other airballs, and this may offset the HR numbers.  Obviously giving up more LD isn’t good, but it’s interesting.

Also, looking at IF #’s, it looks like GB pitchers invoke less IF.  I see r = -0.55 for IF/airball and r = -0.66 IF/Flyball to GB/BFP.  Also, for OF/BFP I get r = 0.56 for IF/airball and 0.42 for IF/Flyball.  It looks like when GB pitchers get the ball lifted in the air, it may be being hit much harder.  I’ll have to look at hit tracker data and see if I can get it into the THT data.  I’m not sure I can, but it would be interesting.

So maybe, JUST maybe FB pitchers are getting too much flak for being FB pitchers.  This data doesn’t tell us that, but I find it interesting none-the less.  BTW, I’m SURE there is huge selection bias, as FB pitchers and GB pitchers who don’t put up good #’s aren’t going to get much playing time.

all of the IF correlations were using 1500 BFP over 3 years aggregate data.

Oh, and the R= .25 for HR/OF was supposed to be .35 not .25 for 1500 or more BFP over 3 years, which is fairly signifigant.  these #’s are using THT, and HR values for OF were calculated using HR/F*OF% (which was calculated from the data).  That said the HR/Airball was using STRAIGHT HR, not the calculated HR off OF I calculated from the battedball data, and is thus not park adjusted.  Still, the information is interesting.

This morning I played w/ these numbers, the HR/F has a .28 correlation to actual HR/F, the LD/Airball at .83, and IF/Airball .67.  I always liked David Gassko’s DIPS 3.0 article, but maybe this would affect it, and simply setting LD% and IF% at some league average isn’t the best way.

HR/F = .115+(GB/BattedBall)^12*8
OR .115+(GB/BattedBall)^8
OR .115+(GB/BattedBall)^12*2+0.0885 (regressed)

LD/Airball = GB%*.8
OR GB%*.64+0.0892 (regressed)

IF/Airball = GB%*.177
OR GB%*.0.119+0.025 (regressed)

In all these cases GB% is GB/BattedBall

Chris, Note that “HR/OF” is not accurate.  The HR includes all HR, including those on Line Drives.  Not sure how many of those there are, but maybe you’d want to do something like:
ofHR = HR - LD*.02
or something like that.  Perhaps some of the PBP wizards out there can tell us the % of LD that are HR.

***

I reported various batted ball type regressions here a little while ago.  I’ll see if I can dig it up.  (I really ought to setup the SEARCH on this site.

***

If you talk to studes, and get into the THT fold, you’d probably have good access to the data you want in a more useful form.

Chris, you might find the regressions here interesting:
http://www.insidethebook.com/ee/index.php/site/comments/mark_teahen_and_gb_fb_ratios/

Thanks Tango, that’s something I was wondering about, if THT’s HR/OF was all HR or just the ones off of OF, so yeah, it wouldn’t be accurate, and I believe .02 is close for the % of LD that are HR, and that would be per the 2006 THT annual, but I dont have it handy (I’m at work), to reference. 

On an only somewhat related topic, do you know how accurate, relative to other pbp data is the Retrosheet Batted Ball Data is?  It looked like to me, 2003-2005 had most of the batted ball type data, but Im not 100% on that.  If it’s somewhat or mostly complete, maybe I can just reference that.

It looks like I made a mistake, I BELIEVE I meant IF/Airball as (1-GB%)*.177 and (1-GB%)*.118+.025 repectively, but I dont have the data in front of me to be sure.  I’d love to see the regressions you posted though.

Dave Smith in the Retrolist yahoo group talked about the reliability of batted ball data of the recent seasons, and said it may be questionable, as the source provider was Howe Sports Bureau, I believe.  Your best bet is to check the archives, and see what he said exactly.  IIRC, the 1991-1992 data came from STATS, and that may be the most reliable.

The Fangraphs data I think comes from BIS, so you may want to look into those as well.

I already wrote something similar to what Chris is doing here:

http://www.hardballtimes.com/main/article/the-truth-about-the-grounder/

But keep up the good work.

Thanks David, I forgot about your article on grounders. 

The numbers are spot-on with what I have in front of me, r = -.05 for GB% to HR/Airball, r = .67 for GB% to LD%.  I get r = -0.53 for GB% to IF/Airball.  That was for qualified pitchers 2004-2006.  Doing an aggregate from 2004-2006 >= 1500 BFP I get even stronger correlations which is expected (.82 FOR ld% .67 FOR if%).

FWIW, I did what Tango Suggested and used ofHR/OF as (HR-ld*.02)/OF.  The relationship to ofHR/OF and HR/Airball and straight GB% is very weak, however it looked like the relatioship follows a V shape curve centered on the mean GB% (.445) when I graphed it.  If I use ABS(lgGB%-GB%) I get a much stronger correlation (.28) than just GB% (.07 ).  A V-Shape also has a stronger correlateion to to HR/Airball (0.20).  These R’s are over 3 years data aggregate >=1500 BFP.  Using single Seasons for qualified pitchers the correlations are weaker (0.17 for to ofHR/OF and .07 for HR/AirBball).  Also if I weight the curve slightly to extreme GB% pitchers, (like .445-GB%+GB%^8), I get a stronger correlation for both data sets, but may be the result of overfitting.

Chris, by qualified pitchers you mean those with 162 innings?

Once again I think selective sampling rears its ugly head.  Groundball pitchers who allow a high % of flyballs to be homers get to keep pitching, since they don’t give up many flies to start with.

If you’re a flyball pitcher and you give up a high % of homers on flyballs, you’ll never last long enough to qualify for anything.

In the Mark Teahen thread I linked above, I wrote:

Regression toward the mean of the various rates is an r=.50, at these levels of number of balls contacted:

This was derived using almost all pitchers, something like 30 or 50 BIP.  I see no reason to use any higher cutoffs, and I don’t see a reason to use year-to-year data.  Simply looking at a single year of data of almost all pitchers, and stripping out the random variance for that level of BIP will leave you with the true talent level for the league.

I agree Rally, I do believe it is the result of a selection bias.

Tango, when I first looked at it, I smoothed the data to try and strip some of the randomness out when I was looking at the relationships.  The correlations posted however, were against the raw data.

By I do believe it is the result of a selection bias, I mean the heavier weighting toward extreme GB% than extreme Airball%.

As Tango noted a couple of days ago, the HR/OF fly in Studes’ batted ball library includes line drive home runs in the numerator. In the mlb.com pbp data I have studied for 2005, the 2% HR rate on line drives which Tango suggests would be an almost perfect estimate; also almost exactly 10% of all HR came on line drives [503/5017]. mlb.com did count fewer line drives than STATS or BIS for 2005 [about 25,600 vs 26,100 vs 27,700; I don’t have the exact totals at hand].

I assume it’s chance, but a few pitchers in 2005 gave up relatively large # of line drive HR - Livan Hernandez 8 (out of 25 total HR), Tomo Ohka 7 (out of 22), Jon Lieber 6 (out of 33).

For those who have access to the data, I think it would be preferable to study line drive and outfield fly ball home run rates separately, since the line drive homeruns can create a good deal of noise if the data is left combined.

I looked at the retrosheet data (so take it with a grain of salt) for 2003-2005 and filtered for >=1500 Batted Balls I get the following correlations to GB%:

HR/Airball 0.07
ofHR/OF 0.24
ldHR/LD -0.13

If I use ABS(.445-GB%) instead, I get the following numbers:

HR/Airball 0.21
ofHR/OF 0.38
ldHR/LD -0.23 - When I graphed it, it appears ldHR follows an inversion of the same curve as ofHR/HR.

Overall: ldHR/LD = 2.5%

Single season data Over 400 BFP there is virtually no correlation in the data, although abs(.445-GB%) is still better fit, albiet slightly, at least for ofHR/OF and to a lesser extent HR/Air.

Note, none of this is park adjusted.

I’m just catching up with this thread, but it’s not true that HR/OF at THT includes all home runs.  Where did that come from?

It only includes HR’s from outfield flies.  HRs from Line drives are separate.

That came from me, from my discussions with you the first year you had the data. You had told me, then, that you didn’t have the HR data split up between FB and LD.  It sounds from your comment that this has changed.

Thanks.  I am getting my data sources mixed up, mostly because I’m 100% working on the book right now.

I think you’re right that the HR/OF on THT’s pages include all home runs (have to check with Bryan).  My bad.  However, the HR/OF figures in the Baseball graphs library includes only HR’s from OF.  Also, the HR/OF data that will be in this year’s THT Annual (where we’ll be basically repliating the library data for 2006) includes only HR’s on OF.

Thanks Studes, that’s what I thought originally, and makes more sense.  The #’s for ofHR/OF I pulled out of Retrosheet this morning are very close to what I got using the data from the THT website. 

I think outliers in the extremely low GB% group are strengthening the ABS(.445-GB%) relationship.  From .35 GB%-.55 GB% is fairly stable (~11.3%).  At GB% >= .55 the numbers shoot up.  #’s < .35 have a sharp increase too.  Both groups contain a relatively small sample, especially the < .35 GB% group.  That group particularly may be skewing the data.

I’ll have to look at the data I was using earlier, this is all off of an excel file of the THT data I have available at the moment.

Chris, is there a place on retrosheet where you can see flyballs allowed? 

Or have you downloaded all the game files and run a program to tabulate that?

I’ve got to learn how to do the latter this offseason.

I have the event files in SQL Server, I used BEVENT to do that, and did an SQL query to get total numbers.  As far as I can tell they’re mostly complete for 2003-2005, accuracy I’m not so sure about.  I dont see any batted ball type data when Hit Value = 4 before the 2003 season.

I’ll have to see what Studes says RE: the THT HR/F #’s.  If they include all HR, then it makes sense that it would be skewed more heavily towards pitchers w/ higher GB%, since they give up more LD per batted ball.

Looking at the SQL queries, THT’s data does indeed seem to include all HRs divided by outfield fly balls. Then again, less than 5% of all home runs in 2006 came on line drives.

David, wouldn’t that would skew the numbers higher for groundball pitchers because of the higher LD%?  That said, at least w/ the Retrosheet data, and using ofHR = HR/F*OF-LD*.02 as Tango suggested shows there is still some correlation on ofHR/F and GB%.

Also, there is a slight negative correlation to GB% and ldHR/LD (which could just be noise), and a strong negative correlation to IF%, and it does appear the overall effect is almost a wash, on HR/Airballs, which is what you mention in the articles about grounders.

Belatedly, I apologize for my false statement about the accuracy of studes’ HR/Fly data at baseballgraphs.com; somehow I had it confused with fangraphs.com. Sorry.

Here’s a few more numbers, courtesy of Retrosheet
Using evetnt files 2003-2005
Filter: Batted Balls >=400

Correlations
GB% to ofBABIP r = .23
GB% to ofSLGBIP r= .25
GB% to ofHR/F r = .11
GB% to ldBABIP r = .02
GB% to ldSLGBIP r = .04
GB% to ldHR/LD r = -.07
GB% to gbBAVG r = -.27
GB% to gbSLG r = -.28

Adjusting each pitchers stats relative to their own team:
GB% to ofBABIPr = .30
GB% to ofSLGBIP r = .29
GB% to ofISOBIP r = 0.24
GB% to ldBABIP r = .03
GB% to ldSLGBIP r = -.06
GB% to ldISOBIP r = -.11
GB% to gbAVG r = -.18
GB% to gbSLG r = -.20
GB% to gbISO r = -.16

I’m not sure I did the team adjusments correctly.  I suspect they need to be regressed.  I simply subtracted the team mean from the players average, for each category.  I was trying to adjust for park and defense.

I should have filtered each component (GB, OF, and LD) seperately instead of using Batted Ball >= 400.

I found Dave’s comment on Relief vs Starting HR/F rates interesting.  I took Retrosheet data 2004-2006, and dumped it into excel.  I seperated relief and starting appearances for all players during that time. 

You can view the data here:
http://spreadsheets.google.com/pub?key=pzbfJeBNaMK0pXtsicEm-ZA

Retrosheet batted ball data looks a little different than BIS data, at least based on what’s been posted at Fangraphs and THT, but that’s all I have.

Based on retrosheet relievers post a HR/OF .008 lower than starters, and 0.013 less LWTS per OF, but post higher LWTS per GB (0.004), and LD 0.002, and lower LWTS per IF (0.001).  Taking the above spreadsheet, the average reliever had 115 batted balls.  If you take the averages for relief and starting percetages of each batted ball type, you get a difference of 0.133 LWTS over 115 batted balls, 0.693 over 600 batted balls. 

Another test was to take a theoretical extreme flyball pitcher (GB% 0.25, LD% 0.183, OF% 0.502, IF% 0.065).  Over 225 batted balls, just 0.404 LWTS difference.  It would take 8568 batted balls to get 10 runs using LWTS. 

It’d be really interesting to see this done w/ BIS data, and maybe using baseruns instead of LWTS (I will probably do this w/ my data).

Retrosheet seems to score IF and OF differently than BIS apparently does, recording about double the IF BIS does.  Actually, Retrosheet lists it as POP-UP and Flyball, so I’m not sure their distinction (vs. BIS), is it zone based, someone mentioned POP-UP is all fly-balls fielded by infielders?  I’d have to check that, that doesn’t sound right.  I think overall, the results would be similar though.

Just taking a look now… your “totals” line summed the percentages, which is wrong.  Your GB% is correct, but OF% is wrong, etc.

That’s been fixed.  I think the pop-up #’s are throwing the OF #’s off.  It’s retrosheet, so OF = ‘F’ in the batted ball field, and IF = ‘P’.

I’ll check the queries tonight, but I’ve always found the popup and OF numbers to be out whack with what Fangraphs and THT report.

Retrosheet’s method of scoring fly balls seems pretty straight forward to me.  Almost all balls in the air that are not line drives and that are fielded by infielders are pop ups. For 2004-2006 there are only 16 exceptions to this.  Almost all balls in the air that are in the air that are not line drives and that are fielded by outfielders are fly balls.  For 2004-2006 there are 2 exceptions to this.

This seems like the most logical definition to me.  If an infielder can get to a ball in the air before an outfielder it should be an infield fly; if he can’t, it should be an outfield fly.

Retrosheet’s big problem with hit ball classification is with ground balls.  It lists ground balls being fielded by outfielders and, of course, they are almost all hits (30 exceptions).  But it doesn’t discriminate between ground balls that reach the outfield that are ground balls as they go through the infield as opposed to line drives through the infield or line drives just over the infielders heads.

OK, NOW it’s fixed.  I forgot to re-publish it.  Not used to google spreadsheets yet.

I agree wholeheartedly with Peter’s major point about batted ball classifications. 

The classification system employed was horribly conceived, as it is practically meaningless to know that Tim Raines picked up a groundball rather than knowing it went through Derek Jeter.  While I would prefer that everything be classified

two groundballs hops past Derek Jeter at point xyz, in 2.3 seconds, followed by 5 more hops for 100 feet in 3.6 seconds, followed by rolling on the ground to point abc picked up by Tim Raines, who threw to 2B whereby the ball was received 3 feet off the bag in 2.8 seconds

a “simpler” system was adopted.  But, out of everything I just enumerated, that Tim Raines picked up the ball at point ab is one of the least important aspects to record.  If you have decided that you can only record a limited number of things, record the most important ones.

However, Retrosheet has nothing to do with this.  They are the delivery mechanism of the data, and not the originator of the data.  The got their data from STATS (1991-1992) and Howe Sports Bureau (2000-2005), IIRC.  They probably got the 93-98 from Palmer/Gillette.

In short, Retrosheet has nothing to do with the horrible practice.