Ryan Tannehill

"Vertical Limit wrote:

They got an excellent package for Jalen Ramsey though. They can stack picks and trade up for anyone they want.

it wouldnt shock me at all if they end up with Tua even though I dont see Tua being a successful quarterback in the NFL.

as of right now theyre picking 6th.. they may take the Tua gamble. And it wont be a successful gamble, Mariota 2.0..

They can have tons of picks and still flop. The organization is horrible

"PhinFan1968 wrote:

Yep, and they're also talking about whether to re-sign Henry...I don't see why they wouldn't do whatever they can to bring both back. Could be a 4 or 5 year power combo.

They can free up another 20M next year with Mariota moving on.

Is that ONLY PPG based on TDs by the offense? Does it take into account FGs?

Edit: Reason I ask, is Tennessee's kicking situation is horrible. 5 of 15 FGs on the year from their first 2 kickers...not to mention how that would impact a coach's decision based on confidence.

PPG includes all points scored by the team, including special teams and pick 6's by the defense. So if let's say the kicking situation is horrible on a particular team we could repeat the same analysis I just did but for passer rating vs. passing TD's instead of passer rating vs. PPG. Of course what's weird here is that passer rating includes passing TD's! But so be it.

Results:
The slope of the best-fitting line between z-score passer ratings and z-score passing TD's is 0.7786 which is VERY close to the slope of 0.7857 found for PPG. That's a really important piece of information because it's actual evidence that "averaging out" is occurring with things like pick 6's, FG's and rushing TD's across teams and seasons. In particular, this means we can directly compare whatever result we get here for passing TD's with the PPG result.

So.. Tannehill's passing TD z-score is 1.5049 (using once again 9.4706 games started due to splitting passing attempts with Mariota in Tannehill's first game) which is higher than his PPG z-score of 1.3286, supporting what you're saying that Tannehill is sometimes being let down by the kicking unit or possibly the running game (w.r.t. scoring). What's the expected z-score for TD's? Tannehill's passer rating z-score is 2.3806 so his expected TD z-score would be 2.3806*0.7786 = 1.8535.

As was the case with PPG, his expected passing TD's is higher than actual TD's, but by less than for PPG: 1.8535 - 1.5049 = 0.3486 versus 0.5418 found for PPG earlier. So yes some of the difference between expected and actual PPG could be accounted for by scoring (or lack thereof) due to other sources including the kicking unit or running game. Just remember that this kind analysis makes the weird assumption that the QB isn't in any way responsible for scores other than passing TD's (so even if the QB helped drive the team to the 1 yard line, if a RB runs it in for a TD you're assuming none of that scoring was due to the QB).

"The Guy wrote:

I hate to ask you to do more work, but can you determine whether sack percentage is a significant covariate?

Here's the thing. The correlation between between passer rating and PPG across NFL history is 0.7780. You know what the partial correlation between the two are when controlling for number of sacks allowed? It's 0.7602, and it's 0.7604 if you control for sack%.

In other words, controlling for sacks doesn't change the relationship at all (basically). So if you want to incorporate sacks into the argument, you actually need to incorporate that in the passer rating formula itself, for example by replacing Y/A with NY/A = (passing yards - sack yards)/(passing attempts + sack attempts). I haven't found the optimal formula for doing that.

But if you did that, then you could compare the effect on "actual vs. expected" z-scores and that would give you an answer.

"The_Dark_Knight wrote:

Brad, call me a simpleton, but your charts that you post look nothing more than an impact zone after an artillery barrage. I have no flipping idea what all of those dots are supposed to represent.

lol.

Well I can attempt an explanation. No guarantees it will help though!

Each open circle in that graph (each impact of an artillery shell) represents one team in one season in NFL history. Draw an imaginary vertical line through any open circle and look at where that line intersects at the bottom of the graph (called the x-axis). That tells you in z-scores the team's passer rating in that year. Don't worry about what z-scores are for now.

Similarly, draw an imaginary horizontal line through any open circle and see where it intersects the left-most part of the graph (called the y-axis) and that tells you the z-score of that team's points scored per game in that year.

So what are z-scores? It's best to think of z-scores as a standardized unit of measurement that applies no matter what you are trying to measure. For example, you could measure how fast a person is in the 100m sprint relative to that person's competition, or you could measure how high a person can jump, or how many points a team scores, all relative to their competition, and their z-score puts ALL those different measurements on the same axis so you can compare across sports.

However, unlike physical measures like "meters" or "seconds", a z-score is a measure relative to the competition, so it only works if you have a larger number of competitors (which isn't needed for "meters" or "seconds").

That red line is just the best estimate of a trend line for all those data (all the open circles), and its slope (how high it rises for every unit increase on the x-axis) basically tells you that for every z-score unit improvement in passer rating you get 0.7857 z-score unit improvement in points scored per game. Hopefully that helps explain the graph. How I used that info is explained in the post you quoted.

"cbrad wrote:

lol.

Well I can attempt an explanation. No guarantees it will help though!

Each open circle in that graph (each impact of an artillery shell) represents one team in one season in NFL history. Draw an imaginary vertical line through any open circle and look at where that line intersects at the bottom of the graph (called the x-axis). That tells you in z-scores the team's passer rating in that year. Don't worry about what z-scores are for now.

Similarly, draw an imaginary horizontal line through any open circle and see where it intersects the left-most part of the graph (called the y-axis) and that tells you the z-score of that team's points scored per game in that year.

So what are z-scores? It's best to think of z-scores as a standardized unit of measurement that applies no matter what you are trying to measure. For example, you could measure how fast a person is in the 100m sprint relative to that person's competition, or you could measure how high a person can jump, or how many points a team scores, all relative to their competition, and their z-score puts ALL those different measurements on the same axis so you can compare across sports.

However, unlike physical measures like "meters" or "seconds", a z-score is a measure relative to the competition, so it only works if you have a larger number of competitors (which isn't needed for "meters" or "seconds").

That red line is just the best estimate of a trend line for all those data (all the open circles), and its slope (how high it rises for every unit increase on the x-axis) basically tells you that for every z-score unit improvement in passer rating you get 0.7857 z-score unit improvement in points scored per game. Hopefully that helps explain the graph. How I used that info is explained in the post you quoted.

You're right....no guarantee LMAO

"cbrad wrote:

PPG includes all points scored by the team, including special teams and pick 6's by the defense. So if let's say the kicking situation is horrible on a particular team we could repeat the same analysis I just did but for passer rating vs. passing TD's instead of passer rating vs. PPG. Of course what's weird here is that passer rating includes passing TD's! But so be it.

Results:
The slope of the best-fitting line between z-score passer ratings and z-score passing TD's is 0.7786 which is VERY close to the slope of 0.7857 found for PPG. That's a really important piece of information because it's actual evidence that "averaging out" is occurring with things like pick 6's, FG's and rushing TD's across teams and seasons. In particular, this means we can directly compare whatever result we get here for passing TD's with the PPG result.

So.. Tannehill's passing TD z-score is 1.5049 (using once again 9.4706 games started due to splitting passing attempts with Mariota in Tannehill's first game) which is higher than his PPG z-score of 1.3286, supporting what you're saying that Tannehill is sometimes being let down by the kicking unit or possibly the running game (w.r.t. scoring). What's the expected z-score for TD's? Tannehill's passer rating z-score is 2.3806 so his expected TD z-score would be 2.3806*0.7786 = 1.8535.

As was the case with PPG, his expected passing TD's is higher than actual TD's, but by less than for PPG: 1.8535 - 1.5049 = 0.3486 versus 0.5418 found for PPG earlier. So yes some of the difference between expected and actual PPG could be accounted for by scoring (or lack thereof) due to other sources including the kicking unit or running game. Just remember that this kind analysis makes the weird assumption that the QB isn't in any way responsible for scores other than passing TD's (so even if the QB helped drive the team to the 1 yard line, if a RB runs it in for a TD you're assuming none of that scoring was due to the QB).

So essentially its another piece in the puzzle, is what I'm taking from that. That a fair assumption?

"PhinFan1968 wrote:

So essentially its another piece in the puzzle, is what I'm taking from that. That a fair assumption?

Of course, and this piece of the puzzle can at least partially be quantified.

btw.. I think a better approach than looking at passing TD's would be to give the QB partial credit for ANY score by the offense based on the proportion of total yards in the drive that were passing yards. That way you don't have the weird interpretation of not giving the QB any credit for a rushing TD or a FG even though most of the drive involved passes by the QB, or (conversely) falsely giving the QB full credit for a passing TD even if much of the drive was due to rushing.

"cbrad wrote:

Of course, and this piece of the puzzle can at least partially be quantified.

btw.. I think a better approach than looking at passing TD's would be to give the QB partial credit for ANY score by the offense based on the proportion of total yards in the drive that were passing yards. That way you don't have the weird interpretation of not giving the QB any credit for a rushing TD or a FG even though most of the drive involved passes by the QB, or (conversely) falsely giving the QB full credit for a passing TD even if much of the drive was due to rushing.

In the rating??? That would pull rating even farther away from actually describing QB performance than it currently is. Lol

"resnor wrote:

In the rating??? That would pull rating even farther away from actually describing QB performance than it currently is. Lol

No not in the rating. The rating stays the same (the x-axis stays the same). What was changed was the y-axis (the effect, not the cause), from PPG to passing TD. I'm saying you'd probably remove more of the effects of the running game and/or special teams if you didn't use PPG or passing TD's but instead used a weighted PPG based on how much of each scoring drive was due to passing.

"cbrad wrote:

Here's the thing. The correlation between between passer rating and PPG across NFL history is 0.7780. You know what the partial correlation between the two are when controlling for number of sacks allowed? It's 0.7602, and it's 0.7604 if you control for sack%.

In other words, controlling for sacks doesn't change the relationship at all (basically). So if you want to incorporate sacks into the argument, you actually need to incorporate that in the passer rating formula itself, for example by replacing Y/A with NY/A = (passing yards - sack yards)/(passing attempts + sack attempts). I haven't found the optimal formula for doing that.

But if you did that, then you could compare the effect on "actual vs. expected" z-scores and that would give you an answer.

I'm pretty surprised at that when the page linked below indicates that the correlation between offensive sack rate and win percentage (which goes well beyond only points scored) is -0.28.

https://www.footballperspective.com/correlating-passing-stats-with-wins/

Here's something you might find interesting. The offensive pass efficiency variable in the regression model is NY/A.

http://archive.advancedfootballanalytics.com/2007/07/what-makes-teams-win-3.html

"The Guy wrote:

I'm pretty surprised at that when the page linked below indicates that the correlation between offensive sack rate and win percentage (which goes well beyond only points scored) is -0.28.

https://www.footballperspective.com/correlating-passing-stats-with-wins/

The correlation between the variable you are trying to control for (sacks or sack%) and either of the two variables you are looking at in the (partial) correlation doesn’t on its own tell you what to expect from the partial correlation. The relationship could be anything, including that the variable you’re controlling for doesn’t matter.

And personally I think the result is intuitive: sacks and sack percentage should only minimally affect the strength of the relationship between passer rating (which doesn’t include sacks) and either PPG or win%, even if the correlation between sacks and both passer rating and win% is not zero. I mean.. why should passer rating become a better or worse predictor if you conditioned only on cases with few sacks or only on cases with a lot of sacks? It should remain similarly strong in both cases (and it does).

btw.. that link purportedly gives you correlations from 1990-2011, but those correlations are generally too low in magnitude because the author of that post, Chase Stuart, seems to have just combined the data across years without any adjustments. For example, the correlation between passer rating and wins is over 0.6 and not his listed 0.51, but you do get into that 0.5 range if you don’t take into account passer rating inflation!

"The Guy wrote:

Here's something you might find interesting. The offensive pass efficiency variable in the regression model is NY/A.

http://archive.advancedfootballanalytics.com/2007/07/what-makes-teams-win-3.html

On my phone right now so I can’t easily search for everything I’d want to, but that link explicitly defines offensive pass efficiency as Y/A right after that Wins = 5.31 etc... equation. Is there someplace else where he says he’s using NY/A?

I do like that he’s doing a regression analysis using z-scores.

@7089 - do speak to the mods about changing your username to "z-brad" . :tongue2:

"cbrad wrote:

The correlation between the variable you are trying to control for (sacks or sack%) and either of the two variables you are looking at in the (partial) correlation doesn’t on its own tell you what to expect from the partial correlation. The relationship could be anything, including that the variable you’re controlling for doesn’t matter.

And personally I think the result is intuitive: sacks and sack percentage should only minimally affect the strength of the relationship between passer rating (which doesn’t include sacks) and either PPG or win%, even if the correlation between sacks and both passer rating and win% is not zero. I mean.. why should passer rating become a better or worse predictor if you conditioned only on cases with few sacks or only on cases with a lot of sacks? It should remain similarly strong in both cases (and it does).

btw.. that link purportedly gives you correlations from 1990-2011, but those correlations are generally too low in magnitude because the author of that post, Chase Stuart, seems to have just combined the data across years without any adjustments. For example, the correlation between passer rating and wins is over 0.6 and not his listed 0.51, but you do get into that 0.5 range if you don’t take into account passer rating inflation!

Yeah the correlation between season sack percentage and points scored between 2016 and 2018 (96 cases) is -0.46, so I guess it's sufficient to just stay with that and not involve passer rating.

On my phone right now so I can’t easily search for everything I’d want to, but that link explicitly defines offensive pass efficiency as Y/A right after that Wins = 5.31 etc... equation. Is there someplace else where he says he’s using NY/A?

I do like that he’s doing a regression analysis using z-scores.

Yeah it was a four-part series, and he says this in part one on a different page:

Because sacks are an important factor in the passing game, I include plays that result in sacks as pass attempts for the purpose of calculating efficiency. Likewise, I also subtract sack yards from total passing yards.

http://archive.advancedfootballanalytics.com/2007/07/what-makes-teams-win-part-1.html

"cbrad wrote:

The correlation between the variable you are trying to control for (sacks or sack%) and either of the two variables you are looking at in the (partial) correlation doesn’t on its own tell you what to expect from the partial correlation. The relationship could be anything, including that the variable you’re controlling for doesn’t matter.

And personally I think the result is intuitive: sacks and sack percentage should only minimally affect the strength of the relationship between passer rating (which doesn’t include sacks) and either PPG or win%, even if the correlation between sacks and both passer rating and win% is not zero. I mean.. why should passer rating become a better or worse predictor if you conditioned only on cases with few sacks or only on cases with a lot of sacks? It should remain similarly strong in both cases (and it does).

btw.. that link purportedly gives you correlations from 1990-2011, but those correlations are generally too low in magnitude because the author of that post, Chase Stuart, seems to have just combined the data across years without any adjustments. For example, the correlation between passer rating and wins is over 0.6 and not his listed 0.51, but you do get into that 0.5 range if you don’t take into account passer rating inflation!

On my phone right now so I can’t easily search for everything I’d want to, but that link explicitly defines offensive pass efficiency as Y/A right after that Wins = 5.31 etc... equation. Is there someplace else where he says he’s using NY/A?

I do like that he’s doing a regression analysis using z-scores.

cbrad - Thanks for being such a level headed stats guy. You bring really good stuff to the forum. Obvious excellent knowledge of statistics with game knowledge and a healthy perspective that you cannot look at numbers in a vacuum. Kudos.

Merry Christmas.

"cuchulainn wrote:

@7089 - do speak to the mods about changing your username to "z-brad" . :tongue2:

That or Zorro since he's always "scratching" Z's all over the boards lol

Some interesting info on sacks (copied from different parts of the article):

What happens when a quarterback changes teams? Which performance stats remain most consistent, suggesting they are more the responsibility of the quarterback himself, and which are least consistent, suggesting that outside forces (such as teammates, game situation, and random luck) play a larger role?

Sack percentage checks out on top, primarily because those quarterbacks who were good at avoiding sacks tended to remain good at avoiding sacks. Of the top 21 in sack rate in year N (the top 20 and ties), 17 of them were above average in sack rate the following year on a different team.

While sacks can be the fault of the offensive line or the accomplishment of the defensive player, the evidence is pretty clear that the quarterback is at least as responsible for his team's sack rate as other passing performance measures that we readily attribute primarily to the quarterback. It's time that the NFL passer rating reflect as much. I suppose the argument against change would be that the passer rating is only measuring quarterbacks as passers, and a sack isn't a pass. This, to me, makes about as much sense as having a hitter rating in baseball that ignores walks and called third strikes, because, well, the guy wasn't swinging at the ball.

I recognize that I'm not the first the make an observation that quarterbacks might be more responsible for sack rate than we believe. In an article written in 2003, Michael David Smith, formerly of Football Outsiders, observed that quarterbacks on the same team showed different sack rates, when looking at a three year period.

https://www.pro-football-reference.com/blog/index2ad6.html?p=4152

Finally a fair ranking.

[ATTACH:full]

"cbrad wrote:

Here's the thing. The correlation between between passer rating and PPG across NFL history is 0.7780. You know what the partial correlation between the two are when controlling for number of sacks allowed? It's 0.7602, and it's 0.7604 if you control for sack%.

In other words, controlling for sacks doesn't change the relationship at all (basically). So if you want to incorporate sacks into the argument, you actually need to incorporate that in the passer rating formula itself, for example by replacing Y/A with NY/A = (passing yards - sack yards)/(passing attempts + sack attempts). I haven't found the optimal formula for doing that.

But if you did that, then you could compare the effect on "actual vs. expected" z-scores and that would give you an answer.

I was fooling with this (the highlighted portion above) a bit earlier. If you replace Y/A with NY/A throughout the passer rating formula, you end up with something that correlates with points scored at 0.84 over the past three seasons of play (2016-2018). That's an increase from 0.80 for traditional passer rating. So you get roughly a 6% increase in variance in points scored associated with passer rating by modifying passer rating in that manner.

What's interesting as well is that when you apply that to individual quarterbacks, you sometimes get an increase in passer rating, and other times a decrease. For example, Tannehill's 2019 passer rating goes from 116.5 to 112.8 as a function of his gaudy 10.43% sack rate, whereas Patrick Mahomes's 2018 passer rating of 113.9 goes up to 117.9 as a function of his much more meager 4.26% sack rate that year. So in that sense you get a more complete measure of how the quarterback is affecting his team.

I don't know if this is the optimal formula for incorporating sack data, but I thought it was interesting nonetheless. Fun with numbers. I'd be interested to hear your take on it.

EDIT: What gave rise to this for me initially was the 10-point loss by the Titans to the Saints despite the 15-point passer rating differential in their favor, with only a -1 turnover margin. That was highly unexpected.

Well in light of the above, if Tannehill's 15.6% sack rate and Brees's 7.3% sack rate in that game are figured into the above formula, the passer rating differential goes from 133.6 (Tannehill) to 118.2 (Brees), to 121.1 (Tannehill) to 118.5 (Brees). Tannehill loses 12.5 passer rating points due to sacks, while Brees's passer rating remains virtually unchanged.

Those figures are far more consistent with a 10-point loss when Tennessee's -1 turnover margin is considered.

"FinFaninBuffalo wrote:

Finally a fair ranking.

[ATTACH:full]

Only one person saw this coming...I don't think so.

Then again, I wouldn't expect them to read Dolphin fan forums haha.

I’ll never listen to another evaluator who dismissed Tannehill as garbage. I’ll do my own evaluations I suppose. Thanks.

May have a tough time finishing the season on top though...Brees is just behind and he's playing Carolina for a potential bye Sunday.

"The Guy wrote:

I was fooling with this (the highlighted portion above) a bit earlier. If you replace Y/A with NY/A throughout the passer rating formula, you end up with something that correlates with points scored at 0.84 over the past three seasons of play (2016-2018). That's an increase from 0.80 for traditional passer rating. So you get roughly a 6% increase in variance in points scored associated with passer rating by modifying passer rating in that manner.

What's interesting as well is that when you apply that to individual quarterbacks, you sometimes get an increase in passer rating, and other times a decrease. For example, Tannehill's 2019 passer rating goes from 116.5 to 112.8 as a function of his gaudy 10.43% sack rate, whereas Patrick Mahomes's 2018 passer rating of 113.9 goes up to 117.9 as a function of his much more meager 4.26% sack rate that year. So in that sense you get a more complete measure of how the quarterback is affecting his team.

I don't know if this is the optimal formula for incorporating sack data, but I thought it was interesting nonetheless. Fun with numbers. I'd be interested to hear your take on it.

EDIT: What gave rise to this for me initially was the 10-point loss by the Titans to the Saints despite the 15-point passer rating differential in their favor, with only a -1 turnover margin. That was highly unexpected.

Well in light of the above, if Tannehill's 15.6% sack rate and Brees's 7.3% sack rate in that game are figured into the above formula, the passer rating differential goes from 133.6 (Tannehill) to 118.2 (Brees), to 121.1 (Tannehill) to 118.5 (Brees). Tannehill loses 12.5 passer rating points due to sacks, while Brees's passer rating remains virtually unchanged.

Those figures are far more consistent with a 10-point loss when Tennessee's -1 turnover margin is considered.

Nice work!

Regarding "optimizing" the formulas (so that the methodology for arriving at them is the same), the simplest way to do this is through multiple linear regression, separately for each formula. The components of (traditional) passer rating are: COMP%, Y/A, TD%, INT% and a constant, while the components of a new "NY/A-modified passer rating" will be the same except that Y/A is replaced by NY/A.

If the goal is to maximize their prediction of win% in each year, the average coefficients you get across NFL history are:

Passer rating:
0.0004*COMP%
0.0686*Y/A
0.0489*TD%
-0.0469*INT%
-0.0349 = constant

NY/A-modified passer rating:
-0.0017*COMP%
0.0966*NY/A
0.0371*TD%
-0.0454*INT%
0.0343 = constant

A few things to note:
1) The ratings you get in the two cases can't be directly compared to each other or to traditional passer rating (they're not on the same scale.. each is on its own scale.. of course you can re-scale the ratings you get with the coefficients above to any range you want).
2) It's weird but true that the average coefficient for COMP% in the NY/A-modified rating is negative (it's only an average)! More important is to note that the weight on COMP% is very close to zero. In other words, traditional passer rating overweights COMP%.
3) Average correlation to win% for "passer rating" with those coefficients is 0.6486 so slightly above the 0.633 for current passer rating.
4) Average correlation to win% for "NY/A-modified passer rating" with those coefficients is 0.6696.

Summary:
When you "optimize" the weights (coefficients) on both passer rating and NY/A-modified passer rating, the difference in variance in win% explained is 0.6696^2 - 0.6486^2 = 2.77%, so including sack% in passer rating explains about 2.77% extra variance in win%.

I would argue that because of making down and distance more difficult, sacks are already generally (not always) incorporated to the degree they impact a game into the passer rating already.

"cbrad wrote:

Nice work!

Regarding "optimizing" the formulas (so that the methodology for arriving at them is the same), the simplest way to do this is through multiple linear regression, separately for each formula. The components of (traditional) passer rating are: COMP%, Y/A, TD%, INT% and a constant, while the components of a new "NY/A-modified passer rating" will be the same except that Y/A is replaced by NY/A.

If the goal is to maximize their prediction of win% in each year, the average coefficients you get across NFL history are:

Passer rating:
0.0004*COMP%
0.0686*Y/A
0.0489*TD%
-0.0469*INT%
-0.0349 = constant

NY/A-modified passer rating:
-0.0017*COMP%
0.0966*NY/A
0.0371*TD%
-0.0454*INT%
0.0343 = constant

A few things to note:
1) The ratings you get in the two cases can't be directly compared to each other or to traditional passer rating (they're not on the same scale.. each is on its own scale.. of course you can re-scale the ratings you get with the coefficients above to any range you want).
2) It's weird but true that the average coefficient for COMP% in the NY/A-modified rating is negative (it's only an average)! More important is to note that the weight on COMP% is very close to zero. In other words, traditional passer rating overweights COMP%.
3) Average correlation to win% for "passer rating" with those coefficients is 0.6486 so slightly above the 0.633 for current passer rating.
4) Average correlation to win% for "NY/A-modified passer rating" with those coefficients is 0.6696.

Summary:
When you "optimize" the weights (coefficients) on both passer rating and NY/A-modified passer rating, the difference in variance in win% explained is 0.6696^2 - 0.6486^2 = 2.77%, so including sack% in passer rating explains about 2.77% extra variance in win%.

I think if you took your second equation and re-scaled it so that it ranged from 0 to 100, you'd end up with probably the best and most user-friendly measure of quarterback play there is at present.