Ryan Tannehill

"Pauly wrote:

Why start in 2004?
The only rationalization appears to be that it is a round number. The problem with this is that you are truncating the data, and whenever you truncate the data you run the risk of an anomaly changing the outcome.
Dates that make sense for a cutoff dat, depending on what you are looking at.
2011 - introduction of the rookie salary pool.
2002 - the realignment into 4 divisions of 4 teams.
1994 - introduction of the salary cap
1978 - introduction of new passing rules that dramatically affected passing stats.
1970 - AFL - NFL merger

I would defer to @7089 on whether it was associated with a significant change in the nature of the passing game in the NFL (I think he may have actually generated those data at some point previously), but the rationale at least is this:

The rule stipulates that a defender can have only incidental contact with a receiver once the receiver is more than five yards downfield. The feeling among some coaches is that game officials have, in recent seasons, permitted defensive players more leeway. One head coach noted Sunday that while he embraces "a kind of 'let 'em play' attitude, there seems to have been a swing toward the defense lately."

By making the rule a point of emphasis, the competition committee essentially is saying that it expects illegal contact to be called much closer in 2004. There are no changes to the rule but, instead, it will be enforced as written.

https://www.espn.com/nfl/columns/story?columnist=pasquarelli_len&id=1771047

"The Guy wrote:

I would defer to @7089 on whether it was associated with a significant change in the nature of the passing game in the NFL (I think he may have actually generated those data at some point previously), but the rationale at least is this:

https://www.espn.com/nfl/columns/story?columnist=pasquarelli_len&id=1771047

My understanding is that the rules change was evolutionary - a change in long term trends not revolutionary - a sudden change that clearly separates different periods.

"Pauly wrote:

My understanding is that the rules change was evolutionary - a change in long term trends not revolutionary - a sudden change that clearly separates different periods.

I did a t-test on average season passer rating in the league for the groups 1978 to 2003 (M = 75.6) and 2004 to 2019 (M = 85.7). That was significant -- p < 0.001.

An ANOVA with three groups 1) 1978 to 1993 (M = 74), 2) 1994 to 2003 (M = 78.2), and 3) 2004 to 2019 (M = 85.7) was significant -- p < 0.001. Tukey and Scheffe post-hoc tests were significant among all three groups (highest p-value = 0.005).

"Pauly wrote:

Why start in 2004?
The only rationalization appears to be that it is a round number. The problem with this is that you are truncating the data, and whenever you truncate the data you run the risk of an anomaly changing the outcome.
Dates that make sense for a cutoff dat, depending on what you are looking at.
2011 - introduction of the rookie salary pool.
2002 - the realignment into 4 divisions of 4 teams.
1994 - introduction of the salary cap
1978 - introduction of new passing rules that dramatically affected passing stats.
1970 - AFL - NFL merger

"Pauly wrote:

My understanding is that the rules change was evolutionary - a change in long term trends not revolutionary - a sudden change that clearly separates different periods.

"The Guy wrote:

I did a t-test on average season passer rating in the league for the groups 1978 to 2003 (M = 75.6) and 2004 to 2019 (M = 85.7). That was significant -- p < 0.001.

An ANOVA with three groups 1) 1978 to 1993 (M = 74), 2) 1994 to 2003 (M = 78.2), and 3) 2004 to 2019 (M = 85.7) was significant -- p < 0.001. Tukey and Scheffe post-hoc tests were significant among all three groups (highest p-value = 0.005).

The change in passer rating is evolutionary since 1978, not change in passer rating differential. And you don't want to do a statistical test on passer rating itself – obviously due to passer rating inflation different periods are different which is why you need to adjust ratings – you want to do statistical tests on passer rating differential (since that's the question here).

Most obvious things to look at to get some idea of how things have changed are the mean and standard deviations of the distributions across years. The mean is irrelevant here because mean passer rating differential is always around zero. Here's the standard deviation over time:

That graph suggests that you probably won't see a major difference in most stats based on passer rating differential for different periods within 1978-2018, but that 1966-1977 might be different. In any case, no need to speculate here, we can do the calculations.

Here the probabilities of making the playoffs for the specific periods you described:

You can do statistical significance tests on these (and for anyone that wants to do that note that you have to modify the threshold for statistical significance because there are 6 tests being done simultaneously, so the thresholds is not 5%, it's 5/6 = 0.83%), and the two periods before the 1978 rule change are statistically significant relative to the league average (black line) but none of the other post-1978 periods are, though the 1994-2001 period is right on the borderline for statistical significance.

Anyway, for future reference, the equation for that black line is:

Playoff probability = 1/(1+e^(0.8746*PRD - 0.1187))

So in summary, you can treat periods within 1978-2018 (I left out 2019 since we still have one game lol) as coming from the same distribution as long as you use passer rating differential. No need to adjust ratings as long as you take the differential.

"cbrad wrote:

The change in passer rating is evolutionary since 1978, not change in passer rating differential. And you don't want to do a statistical test on passer rating itself – obviously due to passer rating inflation different periods are different which is why you need to adjust ratings – you want to do statistical tests on passer rating differential (since that's the question here).

Most obvious things to look at to get some idea of how things have changed are the mean and standard deviations of the distributions across years. The mean is irrelevant here because mean passer rating differential is always around zero. Here's the standard deviation over time:

That graph suggests that you probably won't see a major difference in most stats based on passer rating differential for different periods within 1978-2018, but that 1966-1977 might be different. In any case, no need to speculate here, we can do the calculations.

Here the probabilities of making the playoffs for the specific periods you described:

You can do statistical significance tests on these (and for anyone that wants to do that note that you have to modify the threshold for statistical significance because there are 6 tests being done simultaneously, so the thresholds is not 5%, it's 5/6 = 0.83%), and the two periods before the 1978 rule change are statistically significant relative to the league average (black line) but none of the other post-1978 periods are, though the 1994-2001 period is right on the borderline for statistical significance.

Anyway, for future reference, the equation for that black line is:

Playoff probability = 1/(1+e^(0.8746*PRD - 0.1187))

So in summary, you can treat periods within 1978-2018 (I left out 2019 since we still have one game lol) as coming from the same distribution as long as you use passer rating differential. No need to adjust ratings as long as you take the differential.

Thanks so much for doing that cbrad.

I expected all the post 1978 data to fit similar patterns. It was a little interesting to see the 2002-2010 mini-era being close to significantly different to the other eras.

The basic take home message (as I understand it) is if we want to do analysis based on passer rating differential then there is no reason not to use the full data from 1978 onwards.

Props to the extremists on both sides of this debate.

I FULLy expected to log on and see “Ryan threw a pic, see he sucks” and a “see Ryan threw a TD, he’s the man!”

And I saw neither.

Keep it up guy’s, proud of ya’s.

"Pauly wrote:

Thanks so much for doing that cbrad.

I expected all the post 1978 data to fit similar patterns. It was a little interesting to see the 2002-2010 mini-era being close to significantly different to the other eras.

The basic take home message (as I understand it) is if we want to do analysis based on passer rating differential then there is no reason not to use the full data from 1978 onwards.

I would say it's just the opposite, actually -- that if we've achieved a significant result by using some subset of the passer rating differential data since 1978, there's no reason to use all the data, because the effect size is highly unlikely to change significantly.

If for example the correlation between win percentage and passer rating differential from 2004 on is 0.81 (which it is), then the correlation between the same two variables from 1978 on is highly unlikely to be significantly different.

"cbrad wrote:

The change in passer rating is evolutionary since 1978, not change in passer rating differential. And you don't want to do a statistical test on passer rating itself – obviously due to passer rating inflation different periods are different which is why you need to adjust ratings – you want to do statistical tests on passer rating differential (since that's the question here).

Most obvious things to look at to get some idea of how things have changed are the mean and standard deviations of the distributions across years. The mean is irrelevant here because mean passer rating differential is always around zero. Here's the standard deviation over time:

That graph suggests that you probably won't see a major difference in most stats based on passer rating differential for different periods within 1978-2018, but that 1966-1977 might be different. In any case, no need to speculate here, we can do the calculations.

Here the probabilities of making the playoffs for the specific periods you described:

You can do statistical significance tests on these (and for anyone that wants to do that note that you have to modify the threshold for statistical significance because there are 6 tests being done simultaneously, so the thresholds is not 5%, it's 5/6 = 0.83%), and the two periods before the 1978 rule change are statistically significant relative to the league average (black line) but none of the other post-1978 periods are, though the 1994-2001 period is right on the borderline for statistical significance.

Anyway, for future reference, the equation for that black line is:

Playoff probability = 1/(1+e^(0.8746*PRD - 0.1187))

So in summary, you can treat periods within 1978-2018 (I left out 2019 since we still have one game lol) as coming from the same distribution as long as you use passer rating differential. No need to adjust ratings as long as you take the differential.

Do you see a need to create an era cutoff based on a significant change in run-pass ratio in the league if the analysis is based on a passing game statistic, i.e., passer rating differential? We do know teams are passing the ball far more now than they were in the past.

"Pauly wrote:

It was a little interesting to see the 2002-2010 mini-era being close to significantly different to the other eras.

1994-2001 period you mean.. not sure what the explanation is, but the steeper that curve is the MORE passer rating differential matters, so the passing game was slightly (but not significantly) more important during 1994-2001 than any of the other periods you listed.

And of course those pre-1978 periods show how little the passing game used to matter. It's not just the passing game, but also defense that mattered more up to 1977. That rule change in 1978 changed so much. Up to 1977 defense was on average more important than offense, but after 1978 that flipped and all that coincided with the massive increase in the importance of the passing game.

"Pauly wrote:

The basic take home message (as I understand it) is if we want to do analysis based on passer rating differential then there is no reason not to use the full data from 1978 onwards.

Yeah, that's one take home message. The other is that IF you happen to not have all the data from 1978 then what you find from a decent sized subset of that period will most likely apply to the full post-1978 period. But yes, it's best to use the entire data.

"The Guy wrote:

Do you see a need to create an era cutoff based on a significant change in run-pass ratio in the league if the analysis is based on a passing game statistic, i.e., passer rating differential? We do know teams are passing the ball far more now than they were in the past.

Yeah it's worth pointing out that just because passer rating differential from 1978 onwards doesn't need to be adjusted by era does NOT mean any other stat you're comparing it to doesn't need to be adjusted by era. Win% never needs to be adjusted, but technically playoff probabilities do need to be adjusted since 10 out of 28 in 1978 is different from 12 out of 32 today. However, the adjustment is so small it's not affecting how passer rating differential affects playoff probabilities (it might matter with other stats though). Run/pass ratio? I think that depends on context, but for most contexts I don't think it needs adjustment since it's a percent of total.

As far as an era cutoff based on rushing percent, I wouldn't do it. I'd just use rushing percent as a predictor variable. Also, rush percent has gradually declined even since 1978 (green circles are playoff teams, blue = SB winner), so you'd probably not want a simple cut-off anyway:

"cbrad wrote:

Yeah, that's one take home message. The other is that IF you happen to not have all the data from 1978 then what you find from a decent sized subset of that period will most likely apply to the full post-1978 period. But yes, it's best to use the entire data.

Heck the correlation (0.82) between passer rating differential and win percentage in 2018 alone is statistically significant, p < 0.001.

Yeah it's worth pointing out that just because passer rating differential from 1978 onwards doesn't need to be adjusted by era does NOT mean any other stat you're comparing it to doesn't need to be adjusted by era. Win% never needs to be adjusted, but technically playoff probabilities do need to be adjusted since 10 out of 28 in 1978 is different from 12 out of 32 today. However, the adjustment is so small it's not affecting how passer rating differential affects playoff probabilities (it might matter with other stats though). Run/pass ratio? I think that depends on context, but for most contexts I don't think it needs adjustment since it's a percent of total.

As far as an era cutoff based on rushing percent, I wouldn't do it. I'd just use rushing percent as a predictor variable. Also, rush percent has gradually declined even since 1978 (green circles are playoff teams, blue = SB winner), so you'd probably not want a simple cut-off anyway:

OK appreciate that.

"The Guy wrote:

Heck the correlation (0.82) between passer rating differential and win percentage in 2018 alone is statistically significant, p < 0.001.

Let's interpret that correctly though. All that means is that the true correlation is almost certainly not zero (less than 0.1% probability). It says nothing about whether you could use what you found in 2018 and apply it to 2017 or any other year. For that you actually need to do a t-test (actually Welch's test for unequal variances) given two different means and their estimated standard errors.

In other words, something like what I did with those curves.

"cbrad wrote:

Let's interpret that correctly though. All that means is that the true correlation is almost certainly not zero (less than 0.1% probability). It says nothing about whether you could use what you found in 2018 and apply it to 2017 or any other year. For that you actually need to do a t-test (actually Welch's test for unequal variances) given two different means and their estimated standard errors.

In other words, something like what I did with those curves.

Right but if what we're interested in is how the league is functioning at the present time, then shouldn't we make something of the fact that the 2018 correlation is what it is and is highly significant? And shouldn't that be even more relevant than what's gone on between 1978 and 2017?

"The Guy wrote:

Right but if what we're interested in is how the league is functioning at the present time, then shouldn't we make something of the fact that the 2018 correlation is what it is and is highly significant? And shouldn't that be even more relevant than what's gone on between 1978 and 2017?

I had just posted the results of some more tests but then realized that the results already shown are sufficient. So deleting what I just posted lol.. let's just say that the curves shown earlier already show that the relation between passer rating differential and playoff probabilities (and also win% btw) is statistically the same in the 1978-2018 period, so picking one year from that is best thought of as "small sample size" as long as the relation is to win% or playoff probabilities. That's not necessarily true for any other relation until proven.

"cbrad wrote:

I had just posted the results of some more tests but then realized that the results already shown are sufficient. So deleting what I just posted lol.. let's just say that the curves shown earlier already show that the relation between passer rating differential and playoff probabilities (and also win% btw) is statistically the same in the 1978-2018 period, so picking one year from that is best thought of as "small sample size" as long as the relation is to win% or playoff probabilities. That's not necessarily true for any other relation until proven.

In terms of the correlation between win percentage and passer rating differential (again 0.81), the 95% confidence interval for the sample from 2004 to 2018 (consisting of 480 team seasons) is 0.773 to 0.836. I can't imagine that confidence interval is going to narrow significantly by using the data from 1978 on, and in fact it may widen as we move back in time toward an era in which the league featured the run game far more than it has more recently.

"The Guy wrote:

In terms of the correlation between win percentage and passer rating differential (again 0.81), the 95% confidence interval for the sample from 2004 to 2018 (consisting of 480 team seasons) is 0.773 to 0.836. I can't imagine that confidence interval is going to narrow significantly by using the data from 1978 on, and in fact it may widen as we move back in time toward an era in which the league featured the run game far more than it has more recently.

If you combine the data from 1978-2018 the correlation between win percentage and passer rating differential is 0.7939 with the 95% CI going from [0.7723, 0.8136] so it's slightly narrower than from 2004 to 2018. However, you do notice how large the confidence interval is for 2018 alone. It goes from [0.6635, 0.9099] which shows you why one year alone (e.g., 2018) isn't anywhere near as reliable.

This needs to now be moved to the Stats forum. If we had one.

Yeah. You guys have ruined this thread with stats.

"Mcduffie81 wrote:

Yeah. You guys have ruined this thread with stats.

NERDS!

"Finatik wrote:

This needs to now be moved to the Stats forum. If we had one.

Once the SB is over, I'll start a statistical methods in football thread (Irishman's idea) and we can at least move statistical methodology discussions there. That won't remove statistical analysis from other threads, but it can move the discussion about methodological issues to that thread (and many of the "stats" posts here are of that kind), leaving mostly the results of the statistical analysis in the original thread, thus reducing the overall number of stats posts in the original thread. So that might help a bit.

However, the real issue here is something else I've said previously: those posters wanting a different type of discussion need merely to post more themselves. I'm guessing each poster likes their own posts? lol.. But if you're not willing to do that, there's not much else others can or even should do other than to stop posting, which isn't going to happen. So post more yourself!

"Finatik wrote:

This needs to now be moved to the Stats forum. If we had one.

How do people analyze football/football players?

Some people will say its all about film study and looking at players. Others will say it’s all about scheme and Xs and Os. Some people look at the stats. There are even some people who go on about how the team/player make them feel and whether or not they approve based on their social/political beliefs.

At this point of the season we have run out of film and Xs and Os regarding Tannehill, well there was the pro-bowl performance but no one mentioned that. That leaves the only live subjects in the thread stats and how stats hurt your feelings (show me on this doll where the bad stat hurt you).

I will always listen to someone who is more knowledgeable than me on a subject they have expertise in and I am interested in the subject matter. If it’s not a subject I’m interested in I’ll leave it alone.

"Pauly wrote:

At this point of the season we have run out of film and Xs and Os regarding Tannehill,.

Well then that's probably a good sign to put this thread to bed. Almost 200 pages. All good. Been there. done that. Goodnight.

"Pauly wrote:

How do people analyze football/football players?

Some people will say its all about film study and looking at players. Others will say it’s all about scheme and Xs and Os. Some people look at the stats. There are even some people who go on about how the team/player make them feel and whether or not they approve based on their social/political beliefs.

At this point of the season we have run out of film and Xs and Os regarding Tannehill, well there was the pro-bowl performance but no one mentioned that. That leaves the only live subjects in the thread stats and how stats hurt your feelings (show me on this doll where the bad stat hurt you).

I will always listen to someone who is more knowledgeable than me on a subject they have expertise in and I am interested in the subject matter. If it’s not a subject I’m interested in I’ll leave it alone.

I'd also like to know what affect coaching has on the player's performance.

"cbrad wrote:

Once the SB is over, I'll start a statistical methods in football thread (Irishman's idea) and we can at least move statistical methodology discussions there. That won't remove statistical analysis from other threads, but it can move the discussion about methodological issues to that thread (and many of the "stats" posts here are of that kind), leaving mostly the results of the statistical analysis in the original thread, thus reducing the overall number of stats posts in the original thread. So that might help a bit.

However, the real issue here is something else I've said previously: those posters wanting a different type of discussion need merely to post more themselves. I'm guessing each poster likes their own posts? lol.. But if you're not willing to do that, there's not much else others can or even should do other than to stop posting, which isn't going to happen. So post more yourself!

What do the numbers say for the total in the SB Sunday? Asking for a friend..

"Fin-O wrote:

What do the numbers say for the total in the SB Sunday? Asking for a friend..

Yeah.. predictions based on stats for any single game are HIGHLY unreliable, keep that in mind. Stats are good for predicting trends, not individual games, but there are nevertheless a few trends that are interesting to know about w.r.t. total points scored in the SB.

We don't yet know who the winner and loser will be, but from 1970-2018 the correlation between the sum of regular season z-scores for both offense and defense (a measure of how much better both units together were than league average) and points scored for the SB loser is -0.0623 so basically zero. That means that regular season performance for whichever team is the loser doesn't help predict points scored by the loser and that the best estimate is the average points scored by the loser in the SB during that time, which is 16.7.

So let's say the stats predict the loser will score 17 points.

With the winner you do have some predictive power but not that much. The correlation between the sum of regular season z-scores for both offense and defense and points scored by the winner in the SB is much better at 0.3854 (still not that high though), and the best-fitting line to the data points is 3.8x + 21.35 where x is the sum of z-scores.

If KC is the winner that gives you an estimated 29.51 points scored, while for SF it's 30.9. So the prediction would either be 30-17 if KC wins or 31-17 if SF wins. Tons of variations there if you look at the stats, but you get the idea: if the O/U is 54.5 the stats say Under wins.

It's also interesting how similar z-score wise the two teams are. For KC the offense is at 1.2749 while the defense is at 0.8726 while for SF it's 1.69 for offense and 0.842 for defense. Very similar teams in terms of production.

Anyway.. don't take the predictions too seriously since it's a single game prediction, but with the types of regular season stats these teams have, you'd expect the Under to win.

"cbrad wrote:

Yeah.. predictions based on stats for any single game are HIGHLY unreliable, keep that in mind. Stats are good for predicting trends, not individual games, but there are nevertheless a few trends that are interesting to know about w.r.t. total points scored in the SB.

We don't yet know who the winner and loser will be, but from 1970-2018 the correlation between the sum of regular season z-scores for both offense and defense (a measure of how much better both units together were than league average) and points scored for the SB loser is -0.0623 so basically zero. That means that regular season performance for whichever team is the loser doesn't help predict points scored by the loser and that the best estimate is the average points scored by the loser in the SB during that time, which is 16.7.

So let's say the stats predict the loser will score 17 points.

With the winner you do have some predictive power but not that much. The correlation between the sum of regular season z-scores for both offense and defense and points scored by the winner in the SB is much better at 0.3854 (still not that high though), and the best-fitting line to the data points is 3.8x + 21.35 where x is the sum of z-scores.

If KC is the winner that gives you an estimated 29.51 points scored, while for SF it's 30.9. So the prediction would either be 30-17 if KC wins or 31-17 if SF wins. Tons of variations there if you look at the stats, but you get the idea: if the O/U is 54.5 the stats say Under wins.

It's also interesting how similar z-score wise the two teams are. For KC the offense is at 1.2749 while the defense is at 0.8726 while for SF it's 1.69 for offense and 0.842 for defense. Very similar teams in terms of production.

Anyway.. don't take the predictions too seriously since it's a single game prediction, but with the types of regular season stats these teams have, you'd expect the Under to win.

Interestingly enough, that is the top play of most "pro's". The Under...

I'm going 31-27 Chiefs personally.