tag:blogger.com,1999:blog-38600807.post8744919936930345325..comments2018-06-02T14:19:34.554-04:00Comments on Advanced Football Analytics (formerly Advanced NFL Stats): Correlations between Team Offense and DefenseUnknownnoreply@blogger.comBlogger29125tag:blogger.com,1999:blog-38600807.post-76488499533173469672011-01-25T12:08:17.240-05:002011-01-25T12:08:17.240-05:00How to calculate the # of offensive & defensiv...How to calculate the # of offensive & defensive plays per game?Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-39524524300764374382011-01-23T10:52:07.562-05:002011-01-23T10:52:07.562-05:00Have you run defense correlations against run offe...Have you run defense correlations against run offense and pass offense separately? It seems to me that a good run game helps a defense, but a good pass game doesn't necessarily.Adamhttps://www.blogger.com/profile/02724539019761144400noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-22877795002390644102011-01-20T13:09:20.762-05:002011-01-20T13:09:20.762-05:00What if the study were limited to the first 3 quar...What if the study were limited to the first 3 quarters - before prevent defense strategies kick in?<br /><br />beprepnAnonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-48339684755947911172011-01-20T11:12:17.440-05:002011-01-20T11:12:17.440-05:00I think it's more difficult to score against a...I think it's more difficult to score against a defense when the same team's offense is good than when the offense is bad : not on one particular drive of course, but over the course of many drives. Just because good offense create lesser field position. <br />The strenght of the defense remains the same so, as expected and because of salary cap limitation, SR correlation is negative. <br />WPA correlation is however positive because it is harder to had lots of WP when the opposing offense scores tons of point and when you start every drive deep in your territoryAnonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-84101545873744022602011-01-20T06:48:31.708-05:002011-01-20T06:48:31.708-05:00" It appears that the more that context is co..." It appears that the more that context is considered, the more a team's offensive and defensive performance correlate. "<br /><br />Could contex be coaching? Would it be possible tor replot the data with teams (head coaches) identified? Maybe a subset of the data is driving the result.JMMnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-37070744476027158572011-01-19T22:05:49.147-05:002011-01-19T22:05:49.147-05:00Using the same situations cited in the article (in...Using the same situations cited in the article (inside the 50, 4th and 1, 2 or 3, and outside 4 min to play), the stats are correct. 18 opportunities, 13 conversion attempts, and 11 successes.<br /><br />The successes (ignoring the failures) were worth a total of 0.90 WPA (0.9 wins). Including failures, their 4th down attempts were worth 0.7 WPA.<br /><br />But interestingly enough, the attempts net 29.2 EPA, close to the 30 points claimed in the article!<br /><br />By far, the biggest play was a conversion against GB in the regular season, worth 0.20 WPA.<br /><br />However, this does not set them apart from the rest of the league in terms of winning games. Many teams do go for it in opponent territory on 4th and short, especially in no-man's-land. The average NFL team nets 0.72 WPA and 7.4 EPA in similar situations. <br /><br />So two things: 1) Good for ATL!, but 2)It doesn't explain their good fortune. I'll probably make a full post on this.Brian Burkehttps://www.blogger.com/profile/12371470711365236987noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-33988435859055279552011-01-19T21:26:47.362-05:002011-01-19T21:26:47.362-05:00Buzz-Definitely possible. Let me run some numbers....Buzz-Definitely possible. Let me run some numbers. Good catch!Brian Burkehttps://www.blogger.com/profile/12371470711365236987noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-12587785333175294892011-01-19T21:18:00.393-05:002011-01-19T21:18:00.393-05:00Brian - In regards to your tweet about coaching be...Brian - In regards to your tweet about coaching becoming more agreesive they talk about the falcons adding about 30 points and possibly 3 wins. Do you agree with those numbers? It seems a little high but I haven't analyzed the 2010 PBP data yet. If so it could help explain part of their outlier season in regards to their efficiency stats.Buzznoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-62645344369880124432011-01-19T19:34:03.226-05:002011-01-19T19:34:03.226-05:00Danny-Cool. Thanks. That's higher than I would...Danny-Cool. Thanks. That's higher than I would have expected.Brian Burkehttps://www.blogger.com/profile/12371470711365236987noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-31195646809261283942011-01-19T18:47:51.252-05:002011-01-19T18:47:51.252-05:00Andy-
i wasn't impugning the character of the...Andy-<br /><br />i wasn't impugning the character of the formula for a sample correlation. of course sample size is in there for a reason. i was just making a methodological point that you're more likely to find statistical significance as sample size increases simply because of the mathematics involved; not because there's anything meaningful about the correlation in a practical sense. or, as mark m. put it above , "statistically significant does not imply meaningful."<br /><br />i apologize if anyone, especially the correlation formula, was offended by my remark.Florida Dannyhttp://www.ninersnation.com/section/statistical-preview-and-renoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-36278841635028434942011-01-19T17:57:50.887-05:002011-01-19T17:57:50.887-05:00Well, when you put it in the context of being surp...Well, when you put it in the context of being surprised that there'd be a correlation at all, i suppose we can attribute some meaning to that -.20. I'd agree that there's really no obvious reason why'd there be even that large of a correlation between offensive SR and defensive SR. so, in that context, -.20 is a decent whodunit to solve. my comment was more of an absolute one, considering .2 to be really small in comparison to stuff that really matters (e.g., your correlation between SR and winning, or the R^2's that are found in game prediction models). believe me, i think the same thing when i see FO writing up pieces representing .2 correlations as meaningful.<br /><br />as to your question, same data set (2002-2008, n = 224) shows correlation between pass offense DVOA and run offense DVOA is .542. correlation between pass defense DVOA and run defense DVOA is .477. looking back at the SR-analogous correlation you found (r = .41 for both offense and defense), these are just a tad higher. of course, i think in that original SR piece of yours, you didn't use opponent adjustments, so the correlations are probably higher for opponent-adjusted SR, to the point where they're as high as the ones for DVOA, which of course is opponent-adjusted.Florida Dannyhttp://www.ninersnation.com/section/statistical-preview-and-renoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-46234255575093662712011-01-19T17:49:39.354-05:002011-01-19T17:49:39.354-05:00"Did you adjust def EPA/WPA so that more is b..."Did you adjust def EPA/WPA so that more is better? Since good numbers are low, or large negative, it would make sense that the correlation would have the opposite sign since higher is better for Def SR."<br /><br />You made it nice and easy to copy and paste the data so I ran it and the correlations all go in the same direction because of what I mentioned above. It's probably easier to explain using regression than correlation.<br /><br />Regressing OWPA on DWPA gives a coefficient of 0.21 (note: t-value is 3.06 so it's quite significant). DWPA is opponents' WPA against your D so 0.21 means that your offense getting 1 more WPA will cause your D to give up 0.21 WPA more (as usual, all else equal). Increasing offensive output 10 WPA wins will lead to your defense giving up 2 more WPA wins.<br /><br />Regressing OSR on DSR gives a coefficient of -0.25. Note that DSR is one minus opponent success rate against your D. The -0.25 means that, all else equal, increasing your offense's success rate four percentage points will cost your D one percentage point in success rate. In other words, your opponents will be successful against you one more time out of 100 plays for every four more times you are successful against them.<br /><br />I'm not sure how clear that was, but in both cases teams with better offensive stats tend to have worse defensive stats and vice versa.JaredLhttps://www.blogger.com/profile/08374219319148164119noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-71732983385270712492011-01-19T16:57:03.594-05:002011-01-19T16:57:03.594-05:00About the significance value and those articles po...About the significance value and those articles posted by Adam H (very interesting); <br />statistical significance is what it is, just dont make it out to be more than it claims to be. p = .05 just means that 95% of random data sets will not be as extreme as this data set. Sometimes that has meaning, sometimes it doesn't.<br />Also about the large sample size increasing the statistical significance, the formula isn't broken! its because the more samples you have that continue to show the same correlation the less likely it is due to chance. You can't just add 300 to the significance calc without having 300 more data points! and those 300 data points can easily be random and heavily decrease the correlation if there is no real trend. My point is the formula is right, there is a reason that sample size is in there.Andyhttps://www.blogger.com/profile/13033784825091918451noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-90282914527270450482011-01-19T16:26:06.321-05:002011-01-19T16:26:06.321-05:00As to statistical significance tests, the great (a...As to statistical significance tests, the great (and unique) economist Deirdre McCloskey warns not to become slaves to them. Countless findings that rigorously pass 95% significance are bunkum (<em>New Scientist</em>: "<a href="http://www.newscientist.com/article/dn7915--most-scientific-papers-are-probably-wrong.html" rel="nofollow">Most scientific papers are probably wrong</a>"), while if one has good reason to believe something previously unsuspected is true, and a test supports the idea but at only 65%, that can have serious real-world meaning. Not to digress from football stats. Just saying, for context. Statistical findings should aid judgment, not replace it.Jim Glassnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-38607134711142199022011-01-19T16:01:04.137-05:002011-01-19T16:01:04.137-05:004% shared variance is only true in the purest, str...4% shared variance is only true in the purest, strictest sense. There is another 16% shared by the effects of the two variables interacting with other variables.<br /><br />I think an r of -0.20 is quite remarkable considering that the offense and defense of team are never on the field at the same time. Random variation is going to be a very large part of the overall R in a 16-game season, so that -0.20 is a fair share of everything else that isn't random.<br /><br />The fact that the r for EPA is near zero is part of the point of the piece. As you add situational factors to the mix, the r goes from significantly negative, through zero, to significantly positive. <br /><br />Regarding DVOA, from what I understand, it's a SR metric with bonuses for big plays, plus it over-weights red zone plays and discounts 'trash time.' In a sense it's a kluge of all three concepts: SR, EPA (big play bonuses), and WPA (game situation). I'd expect its correlation between off and def to fall somewhere between SR and EPA.<br /><br />Here's one question I have. Does the DVOA of team rushing correlate with team passing? For offense or for defense?Brian Burkehttps://www.blogger.com/profile/12371470711365236987noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-182427886182698632011-01-19T15:22:51.817-05:002011-01-19T15:22:51.817-05:00i'd second mark m's comment above.
first,...i'd second mark m's comment above.<br /><br />first, even if you focus on the largest of these correlations, SR's -.20, you're saying that offensive SR and defensive SR share 4% of their variance (i.e., -.20^2). in other words, 96% of offensive SR has nothing to do with defensive SR, and vice versa. if there's anything cool found out about that 4%, great. i'd be much more interested in the other 96%.<br /><br />second, and more generally, the formula for a correlation coefficient has sample size in the denominator, so correlations are by definition highly sensitive to sample size. to me, that .06 correlation for EPA is only even marginally statistically significant because n = 318. increase the sample size to 600, and watch how it magically becomes full statistically significant. lower the sample size to 100, and watch how it magically loses its marginal statistical significance. add to this that .06^2 = .0036, meaning 99.64% of offensive EPA has nothing to do with defensive EPA (and vice versa), the only real practical conclusion here is to totally ignore that correlation altogether.<br /><br />i'll acknowledge that it's difficult to find larger correlations in NFL stat analysis, but correlations in the vicinity of .1 and .2 are woefully low even by that standard.<br /><br />p.s. i ran the correlation between offensive DVOA and defensive DVOA, and it's -.048 (2002-2008; n = 224). given that DVOA's just a gussied up version of SR that uses yards instead of WPA as its measure of "success," i find it interesting that the SR correlation you found is almost 4 times larger. there's a clear methodological implication of that, i think, which i'd discuss here if you wanted to go down that road.Florida Dannyhttp://www.ninersnation.com/section/statistical-preview-and-renoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-13112868078403805252011-01-19T14:46:00.744-05:002011-01-19T14:46:00.744-05:00Good follow up.
Did you adjust def EPA/WPA so tha...Good follow up.<br /><br />Did you adjust def EPA/WPA so that more is better? Since good numbers are low, or large negative, it would make sense that the correlation would have the opposite sign since higher is better for Def SR.JaredLhttps://www.blogger.com/profile/08374219319148164119noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-6062141331941022212011-01-19T14:31:22.517-05:002011-01-19T14:31:22.517-05:00@BIGKIDD325
WPA, EPA and SR all account for field...@BIGKIDD325<br /><br />WPA, EPA and SR all account for field position.Samponoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-8298294520239000512011-01-19T14:29:17.287-05:002011-01-19T14:29:17.287-05:00My theory: many many factors create positive and m...My theory: many many factors create positive and many many factors create negative correlations between a teams offense and defense. The surprising result (to me) is that they very nearly cancel each other out.<br /><br />Re: Statistical Significance...<br />http://www.nytimes.com/2011/01/11/science/11esp.html?_r=2<br />http://wmbriggs.com/blog/?p=2730Adam Hnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-1957483214710659222011-01-19T14:02:02.114-05:002011-01-19T14:02:02.114-05:00How about if you have a dominating defense that fo...How about if you have a dominating defense that forces an incredible amount of turnovers, particularly in the opponent's territory, it will increase the offenses chances of scoring.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-10339216955151303922011-01-19T13:50:09.046-05:002011-01-19T13:50:09.046-05:00Could you look at a breakdown of offensive SR, EPA...Could you look at a breakdown of offensive SR, EPA, and WPA depending on game situation (score margin and time remaining)? For instance, I'd expect teams that are losing by more than a touchdown in the 4th quarter to have a high offensive SR, since the defense will be playing safe to avoid giving up the big play. But their EPA and WPA might not be any higher than normal - they might even be lower, since the offense will take chances and risk turnovers to attempt high variance plays.<br /><br />That could produce a negative correlation between a team's offensive SR and defensive SR, since teams with bad defenses will get more chances to pad their offensive SR in garbage time and teams with good offenses will hurt their defensive SR in garbage time. But you wouldn't get the same negative correlation for EPA and WPA.Vincenoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-77892835180291906252011-01-19T13:28:13.798-05:002011-01-19T13:28:13.798-05:00Other thing to remember, statistically significant...Other thing to remember, statistically significant does not imply meaningful.Mark Mhttps://www.blogger.com/profile/00437341898224311426noreply@blogger.comtag:blogger.com,1999:blog-38600807.post-40216409581900614872011-01-19T12:51:46.812-05:002011-01-19T12:51:46.812-05:00He qualifies some of the correlations in terms of ...He qualifies some of the correlations in terms of statistical significance. Assuming a p value of .05 (the most liberal definition of statistical significance in many cases), the correlations would be due simply to chance only 1 in 20 times. Stated another way, 19 out of 20 times, they are not within error.<br /><br />Correlations can be very weak, but still statistically significant...it just usually takes a larger sample size.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-71347900337528204242011-01-19T12:41:55.583-05:002011-01-19T12:41:55.583-05:00Last comment was me (if anybody cares), not an ano...Last comment was me (if anybody cares), not an anonymous criticism. Just clicked the wrong thing.Xnoreply@blogger.comtag:blogger.com,1999:blog-38600807.post-79173960789712280402011-01-19T12:39:51.636-05:002011-01-19T12:39:51.636-05:00Have you considered that all of these could be zer...Have you considered that all of these could be zero correlation within error? Telling just-so stories about the noise in your data is not good statistics.Anonymousnoreply@blogger.com