Just some quick notes on MLB seasonal averages:
- Run scoring is up slightly from last year (4.33 R per game, from 4.28), but is still the 2nd-lowest value since 1992.
- Strikeouts have reached another all-time high, which has been true for 5 years running. So far in 2012 they stand at 7.5 per game, up a whopping 5% from the record rate last year of 7.1.
- Meanwhile walks are down to 3.06 per game, the lowest value since 1968!
- As you might imagine, the 2.43 K/BB ratio is astronomically high. That’s up nearly 6% from last year’s ratio, which itself was an all-time record.
- Hits per game are at 8.65 per game, the lowest value since 1989.
- Attendance stands at 31,381 per game, the highest since 2008 and a pinch higher than pre-strike level of 31,256 in 1994.
- Intentional walks are down to just 0.21 per game. That’s the lowest level in recorded history, which goes back only to 1955. I presume as run-scoring goes down, managers are increasingly reluctant to put more runners on base. Sabermetrics has come a long way in the last 20 years to show just how much the chances of scoring increase when any batter is walked.
- Interestingly, sacrifice hits have also dropped way down, to just 0.30 per game. That’s also the lowest level since they’ve been recorded, which is back to 1954. Managers have also learned that giving up an out in exchange for a base advance is worth it far less than thought for most of the 20th century.
It’ll be interesting to see how the decreasing offensive environment unfolds in the coming years. My gut is we are in a transition period, with elements of the high-offense “old guard” still present, while elements of the “new guard” begin to come more into play. The age of defensive and speed players may return. Perhaps we’ll have another Charlie Lau guru of hitting who will stress more contact, less swing-for-the fences approach. I’m not predicting a true 1968, Year-of-the-Pitcher year, but I think within in the next five to seven years, it’s possible we’ll approach early 1970s level of hitting.
Mike, I agree…..baseball now, it looks to be the best of all worlds, the sabermteric reality, accompanied by the non-steroid reality. Things are looking up!
ugh, edit button plz- *sabermetric 😉
Sabermetric reality, the ultimate oxymoron.
Oops.
Irony alert.
The Phillies have by far the best K/BB ratio in baseball at 3.23. And yet they’re 11th in the NL in ERA.
Ed:
That “pound the strike zone” mentality can, at times, get a pitcher like Lee, Hamels, or even Halladay in trouble.
When asked, “What’s the worst pitch a pitcher can throw a hitter”?, Sal Maglie supposedly replied, “A strike.”
I posted yesterday that my analyses regarding success in one run (1-r) games, was wrong, but it was a brief message and people may have missed it. Here’s why it was wrong, and also a new analysis to replace it. I wish I could say it was just a minor gaffe, but it wasn’t; it was a major analytical mistake that changes the whole picture. So it needs to be clarified.
The problem is that I forgot to account for the expected deviation from a .500 record (for each team, over the time period of interest) that would be due to chance alone, when assessing the likelihood of chance vs. skill in determining 1-r outcomes over all teams. I had a nagging feeling something was wrong with the extremely low (too low) probabilities I was getting, and also the fairly wild swings in success from year to year for individual teams that both tag and John had mentioned as evidence against the “skill hypothesis”. Teams change from one year to the next, but they are similar enough that you wouldn’t expect such swings.
So let’s try this again, with a (hopefully) better approximation to reality. I’ll focus mostly on the underlying concepts but post some (very brief) R code at the end for those might be interested.
What I did originally was sum the absolute values of the differences in W-L record for each team, and then convert that into an estimate of what that would translate to if you only had say, two teams playing games, as in John’s example of one team that’s 20-10 and another that’s, by definition, 10-20. The probabilities computed for that scenario were correct–it’s identical to flipping a coin 30 times and coming up with 20 heads (or tails), which has a probability of about p = .02. So far so good.
continued
continuing
The problem comes in when you have a larger number of teams. I alluded to this in one of my explanations to John, where I mentioned that I had to take some extra steps to try to approximate a situation in which there is a single “experimental unit” being evaluated–like a single coin being flipped repeatedly, or a two team league in which all games are between those teams only. The problem is, I did this wrong.
When you have say 30 teams, it’s more complicated, in a way that I hope I can convey, but which is tricky. Basically, you have to account for the fact that your probabilities for outcomes for any given pair of teams, are embedded in a larger system in which there are many such comparisons. This creates probabilities on two different (and hierarchical) levels, and they both have to be accounted for. You still use a binomial probability model, which is designed to estimate the probabilities of defined numbers of successes or failures in repeated trials, given a known probability of success in each such trial. That doesn’t change: we still use a p = 0.5 probability of winning (and losing) any individual game, to create the reference expectation over many such games, against which we compare the observed W-L records in 1-r games, both over all teams and for individual teams, within our defined time period of interest.
continued
continuing
What this means in terms of simulating the situation is that you have to estimate how many times *any individual team* would be expected to go over (or under) .500 by a certain amount given a defined number of games played. This is complicated by the fact that the number of 1-r games played varies from team to team, especially as time-span of interest increases, so that factor has to be included also. It also means that you cannot compute the relative probabilities of chance versus skill in being the likely cause of the overall results, using the number of simulated trials that exceed a defined valued, as was done before. Instead you have to compare the distribution of the expected deviations from .500, for each team in a 30 team league league, with the actual distributions thereof, and then look at the results to see which (if any) teams might have results that far exceed expectation, as defined by *that* distribution.
I made a couple of other changes also, but they are unrelated. The most important one is to back 20 years (1992-2011) instead of just 5 as before (18 years would be ideal because then sample sizes for all franchises would be based on the same number of games played, but this is close enough even if the Rays and D-backs have played fewer games). And I included in the simulation the different number of 1-r games played by each team, which varied a lot even among the non-exapansion franchises. I also ran a million simulated trials instead of 100K. That still only takes maybe 15-20 seconds on my laptop.
continued
continuing:
Here is the R code for the analysis:
# simulation and games played vectors:
sims = rep(NA,1000000)
r1 = c(799,938,939,871,974,929,918,894,912,892,860,936,917,896,891,874,866,953,630,
652,904,816,953,958,884,786,858,964,900,874)
# game trials:
for (i in 1:length(sims)){
games = sample (r1,1)
sims[i] = rbinom(1, games, 0.5)/games
}
# expected 10 percent quantiles:
quantile(sims, probs = seq(from=0, to=1, by=0.1))
r1 is the vector of 1-r games played over those two decades (the two low values are the 1994 expansion teams, Dodgers were high at 974 and Tigers were low at 786, for the non-expansion teams).
So what I’m doing there is saying, if I run a million individual games trials, where the outcome of each is determined purely by chance, what will be the expected win percentage for each team in a 30 team league. Then I simply compare the distribution of those percentages against the observed win percentages in such games over that time. I do this in two different ways: (1) for assessment of the overall results, as the number of teams falling into each of 10 bins (each will ideally contain 3 teams if chance is the determining process in W-L record), and (2) examining the p value for individual teams out on the edge of the distribution (both low and high).
And the results are…
First these are the actual results, 1992-2011 (hope this formats right):
Rk Tm G W L W%
1 NYY 799 443 353 0.557
2 SFG 938 495 442 0.528
3 CAL 939 494 445 0.526
4 OAK 871 451 420 0.518
5 LAD 974 503 471 0.516
6 ATL 929 479 449 0.516
7 CIN 918 472 443 0.516
8 MIN 894 460 433 0.515
9 FLA 912 465 447 0.510
10 CHW 892 453 436 0.510
11 BOS 860 439 421 0.510
12 SDP 936 475 461 0.507
13 PHI 917 460 457 0.502
14 HOU 896 449 446 0.502
15 MIL 891 447 443 0.502
16 CLE 874 438 436 0.501
17 TEX 866 433 432 0.501
18 STL 953 475 476 0.499
19 TBD 630 314 316 0.498
20 ARI 652 324 328 0.497
21 TOR 904 448 455 0.496
22 COL 816 405 411 0.496
23 MON 953 469 483 0.493
24 NYM 958 468 490 0.489
25 SEA 884 430 454 0.486
26 DET 786 374 412 0.476
27 BAL 858 402 453 0.470
28 CHC 964 450 513 0.467
29 PIT 900 414 485 0.461
30 KCR 874 378 496 0.432
Now, the quantile (bin) cut points for ten groups and the number of teams expected and observed in each (expect 3 in each, since 10 x 3 = 30 teams total).
Min 10% 20% 30% 40% 50% 60% 70% 80% 90% Max
0.405 0.478 0.486 0.491 0.496 0.500 0.504 0.509 0.514 0.522 0.581
Team Win% Obs. Exp.
NYY 0.557 1 1
SFG 0.528 1 1
CAL 0.526 1 1
OAK 0.518 2 2
LAD 0.516 2 2
ATL 0.516 2 2
CIN 0.516 2 3
MIN 0.515 2 3
FLA 0.51 3 3
CHW 0.51 3 4
BOS 0.51 3 4
SDP 0.507 4 4
PHI 0.502 5 5
HOU 0.502 5 5
MIL 0.502 5 5
CLE 0.501 5 6
TEX 0.501 5 6
STL 0.499 6 6
TBD 0.498 6 7
ARI 0.497 6 7
TOR 0.496 6 7
COL 0.496 6 8
MON 0.493 7 8
NYM 0.489 8 8
SEA 0.486 9 9
DET 0.476 10 9
BAL 0.47 10 9
CHC 0.467 10 10
PIT 0.461 10 10
KCR 0.432 10 10
And finally, wrapping up:
The overall distribution of teams into the ten bins is not bad overall–it follows expectation reasonably closely, although not perfectly (for example, an excessive number of teams in the quantiles 2, 6 and 10, and a deficiency in quantiles 4 and 7-9, such that the bottom end of the distribution in particular is skewed). But overall, not too bad at all, giving some definite evidence that many games are in fact won by chance, especially by “middle of the pack” teams.
However, if we look at the individual teams on the two ends of the distribution there are some definite improbabilities there. Starting with the top end, the Yankees W% far exceeds that of the other teams. The p value for the Yanks’ .557 W% is very low, .00047 to be precise. So the evidence is strong that the Yankees are indeed skilled at winning close games, to a degree unmatched by any other team. However, the Giants come in at p = .048 also, the only teams to meet the traditional p = .05 statistical significance level, although the Angels are just shy of it, at p =.0615.
On the other end, the bottom four teams (Royals, Pirates, Cubs, Orioles) all come in at p < .05. The Royals are by far the worst of these, with p = .00004. No way to explain their bad record by chance.
I’m sure a good chunk of the Yankees success in one run games can be attributed to one person in particular.
I’m pretty sure that’s a big ol’ can of worms Ed. You have to develop the lead first before Rivera can protect it.
One last result:
The top 12 teams on that list all fall within the upper third of the expected distribution (expect 10), while only the bottom 7 or 8 teams fall within the lower third (expect 10 again). So, the distribution overall is shifted somewhat in the “successful” direction. The Royals by themselves seem to compensate for a good chunk of this shift by their bad record.
Jim,
Once again you’ve been very clear, but I’m not sure your analysis fits the question – at least the question I’d thought had been posed – which was whether particular teams, like the 2012 O’s, may be more skilled in one-run games than in other games, in which case we might expect good outcomes to continue as the season went on, or whether it was just a matter of luck, in which case we could expect normal W-L results from now on.
For example, you show that the Yankees win so many one-run games that it’s reasonable to think more than luck was involved. I’d agree, but I’d ask whether they’re unusually skilled in winning 1-r games or simply all games. Over the period 1992-2011, just averaging W-L percentages (I’m lazy, and I know 1994-95 would alter things a bit), the Yankees have a record of about .590 (!), much higher than their 1-r record of .557. Obviously, teams that stayed close to the Yankees would have a better chance of beating them than teams that did not, so maybe that’s not surprising, but it does not appear to show any incremental skill on the Yankees’ part in one-run games (of the type that would trigger Ed’s tribute at #16 to some unnamed player whose identity one could hardly guess).
Of course, to assess this better, I suppose we’d need to get a sense of how the 1-r context bends W-L outcomes for all teams. The Royals, for example, are awful in 1-r games (.432), but not much different in all games (about .436). Perhaps they’re more awful than it appears, since bad teams should win a higher percentage of the games they stay close in as a norm (the Pirates, for example, were a .439 team, but .461 in 1-r games). I think that to assess the data you’ve gathered, we’d need to know something about normal curves that relate team W-L pct. to 1-r W-L pct. (I’m too lazy to run the figures for 30 teams, as you have.)
But it’s also true that for the question originally (I think) posed, a 20-year timeline wouldn’t mean much, because the issue concerns teams, not franchises, and teams generally hold together for 2-3 years at most.
Of course, I may have misconstrued the entire issue, and in addition, your explanations notwithstanding, I don’t have a clue about statistics.
epm,
Right this analysis does not address the question of whether the Orioles have been skilled vs lucky in going 19-6 in 1-r games to date or whether that rate will continue, at least not directly. The time period is short and the sample small. But I can do the same kind of analysis as the above, on this year’s data alone, to get some insight into it, and probably will.
The point of the posts here was only to correct earlier mistakes on how to approach the more general question, because conceptual approach is by far the most important part of getting correct insight into these kind of things. Granted that most people are not statisticians here, but at least a couple of people were paying close attention, and trying to learn the basic approach. Anyway, on general principle, I could not leave the issue uncorrected.
I agree that teams generally change over time, usually gradually but sometimes abruptly, and this complicates the use of teams’ records from year to year as evidence of luck vs skill. Even over the course of one season it can be complicated by injuries, roster changes and schedule for example.
I do agree that ability in games likely correlates, possibly highly, with ability in close games. It certainly does for the Yankees and Royals, and probably the few other teams mentioned falling at the ends of the distribution. But Ive never been arguing one way or another on that. I’m not sure who first introduced that issue into the discussion, but it’s a separate one.
And I think you meant to say “more skilled in one-run games than [other teams]” right?
I may have mistaken the question of the earlier thread by reading my own expectations into it, Jim. Perhaps the issue was not the one I was responding to.
I actually did mean more skilled in 1-r games than other games. The issue for me was whether there were reasons other than chance why a team like the O’s so outperformed their overall record in 1-r games. It’s actually an issue I’ve always found interesting, having to do with the dynamic of close games (of which there are, of course, many different types, including 1-r games that only become close at the close). I do wonder whether for certain teams, some combination of skills, energy, and tactics makes them more (or less) successful in close games.
I suppose answers would tend to be ungeneralizable: informed, perhaps, by statistical regularities and data, such as what you wrote about, but perhaps dominated by qualitative analysis and individual cases.
In any case, I do want to add that I thought it was neat that after writing and arguing at great length on the original thread, you sent your I-was-wrong-about-(almost)-everything message out and then did ten times as much work to correct yourself. I often feel that way after my own messages, but my reaction is just to hope that no one read them and to think about something else.
OK, thanks for clarifying epm, and for the good words as well.
I also like the idea of looking at success/failure rates in the different types of 1-r games you mention. Another helpful thing to look at would be trends in 1-r success between say, teams’ first and second halves of seasons, where the rosters/lineups are typically less variable than they are between years. I’d like to look at that in more depth when I get some time. They would get more to the original issue of the likelihood of the Orioles to +/- maintain their current pace.
The issue of how generalizable statistical analysis results are to different situations is a topic of great fascination for me.
In general it would be helpful to show the overall W/L% for all teams in that 1992-2011 timeframe next to the 1-run game percentages. I would imagine, for example, the Pirates’ .461 percentage is significantly higher than their overall winning percentage in that 20 year span. Or, even better, the teams’ % in just the non-one-run games, if that’s possible.
Not sure what we’re finding out here even with that data, but it’s interesting all the same.
“•Attendance stands at 31,381 per game, the highest since 2008 and a pinch higher than pre-strike level of 31,256 in 1994″…Does any reader know what is the percentage of total attendance capacity if the average is 31,381?
I can get at least some of that data for you. Coming up in an hour or so.
According to Wikipedia the average seating capacity of all ML parks is 44,234. The percentage of capacity for 2012 is 70.94%.
That’s not weighted quite correctly, I don’t think, but it’s close. I will have a post up in a minute about this.
I see from your new blog that the percentage is 72.9%. I calculated the average capacity (43,250) from that ESPN attendance spreadsheet and saw that the number I got from Wikipedia is incorrect. I also found that the 2012 average attendance is 31,532. By using the correct values I got the same percentage that you did.
Somebody (Lincoln?) said something like “Being right for the wrong reasons doesn’t count”. Very true.
But I still feel a little vindicated after running the analysis described above on the 2012 data alone. Here are the results, where column p refers to the probability of winning the observed number of games, for teams with winning records, and of losing the observed number of games, for teams with losing records:
Rk Tm W L Win% p
1 BAL 20 6 0.769 0.0028
2 CLE 15 6 0.714 0.0133
3 ATL 16 8 0.667 0.0320
4 PIT 23 16 0.59 0.1668
5 OAK 17 12 0.586 0.1813
6 SFG 22 16 0.579 0.1914
7 WSN 19 16 0.543 0.3144
8 CHW 14 12 0.538 0.3493
9 LAD 20 18 0.526 0.3983
10 DET 14 13 0.519 0.4094
11 CIN 17 16 0.515 0.4379
12 TBR 17 16 0.515 0.4379
13 MIA 19 18 0.514 0.4465
14 MIN 18 17 0.514 0.4465
15 SDP 15 15 0.5 0.4606
16 TEX 12 12 0.5 0.4606
17 COL 11 11 0.5 0.4606
18 KCR 14 15 0.483 0.4300
19 BOS 11 12 0.478 0.4052
20 NYY 13 15 0.464 0.3537
21 NYM 13 15 0.464 0.3537
22 PHI 15 19 0.441 0.2631
23 MIL 18 23 0.439 0.2600
24 SEA 14 18 0.438 0.2600
25 STL 12 18 0.4 0.1370
26 CHC 11 18 0.379 0.0975
27 LAA 7 12 0.368 0.0786
28 ARI 8 14 0.364 0.0724
29 HOU 10 20 0.333 0.0325
30 TOR 5 13 0.278 0.0112
The odds of skill versus chance in determining the outcomes for those five teams are:
Orioles 355:1
Indians 74:1
Braves 30:1
Astros 30:1
Jays 88:1
So the evidence still strongly indicates that the Orioles (especially) but also the Indians and Braves, are skilled at winning 1-r games, while the Astros and Blue Jays are lacking in same. [Incidentally, the O’s have only played one 1-r game since the post last week, which they won.]
“The last temptation is the greatest treason:
To do the right thing for the wrong reason.”
from Murder in the Cathedral, by T.S. Eliot
Oh I like that!
Would be really nice to have a way to properly format data tables…