A look at stat priorities for Defensive Linemen.

[7hawk77]

[div align=\\\"center\\\"]A look at stat priorities for Defensive Linemen[/div]

Hey everybody,

Today I'm going to be going over the defensive linemen stats and what to prioritize. I'll also be doing some minor guesswork and analytics on the data once I present it.

To do this, I'll be using data for the 14 game season. I'm only using players that played all 14 games. I think there is only one player that didn't play all games so I wanted it to be more reflective of players starting out and updating along the entire season to smooth out any statistical outliers.

Below is the data for the defensive linemen tackles/sacks etc.








Now lets also look at player stats.





Now that we have both of those, lets look at the correlation between them.

First lets look at tackling. Which stats correlated best to making tackles?

First let me talk about the unimportant stats. Correlation does not mean causation. Pass blocking run blocking, arm strength all have positive correlations? Why is that? My guess is that the players that have higher stats of these have purchased all of the equipment which means they probably have bumped up their TPE more than other players as well as invested more time advancing their character. Because of this they are outperforming their peers and yielding an irrelevant stat correlation.
Anyways, After disregarding irrelevant stats, we get the following stat priority to make tackles.

Strength > Speed > Tackling > Endurance > Intelligence > Hands > Agility

This seems about right to me but strength being so high is kind of surprising. I imagined that it would be something like Speed>Tackling>Strength>Endurance.
My biggest guess for this is pancakes. Offensive linemen pick up pancakes but defensive linemen don't get a negative stat saying that they were pancaked. My guess is that without sufficient strength, the player will get pancaked and be unable to really contribute at all to the play. Either that or you just need strength to bring the player down.

Next lets look at tackle for loss correlations.



So this stat priority is:

Speed > Agility > Tackling > Hands > Strength > Endurance > Intelligence

Again this is interesting. Strength is quite a bit lower and hands looks like a big outlier. I honestly don't know how hands beat out strength. Speed, agility and tackling make perfect sense on tackling for loss, but I don't have a great explanation on why hands would be more useful than strength. Maybe this is just an outlier. There is a drop off between tackling and hands and then strength and endurance. The trend looks like the top 3 stats on a priority list are very important and then things start dropping off.


Next lets look at Forced Fumbles.


I'm not that interested in fumble recoveries because I feel like it's more random than forcing fumbles. Also forcing a fumble seems to make the fumble recovery more likely.

Stat priority:
Strength > Tackling > Speed> Endurance > Intelligence > Agility > Hands

This makes sense. You get to them fast and hard. Strength tackling and speed are all very important which again shows that top 3 coefficient.


Finally, lets look at everyones favorite. Sacks.




Stat priorty

Strength > Speed > Tackling > Endurance > Intelligence >Hands > Agility

Now this is pretty surprising. Agility is very far down as far as sacks are concerned. Also Strength and Speed are so much more valuable than any other stat. I'm starting to think that tackling counters agility, and really there is only 1 mobile quarterback and who knows what will happen to him with the bridges he burned. Because of this, I think that tackling isn't as important for sacks until quarterbacks start investing in agility.


I already wrote a piece about sacks and more specifically team sacks so if you are looking for more information on that, go read about it here: http://nsfl.jcink.net/index.php?showtopic=1363

Anyways, If you want to model your defensive linemen to be really good at a certain aspect of the game, feel free to use these stat priorites at your own risk.

One final note that I'd like to add. I currently have the excel sheet set up where I can easily do the same analysis with update stats. In the future I'm going to see how the correlation coefficients shift and speculate on why that is happening.

If you have any questions about the above, feel free to ask.

Code:

727 words

[SimmerDownBruhh]

Very awesome read, bro! I love all of your analysis articles.

[7hawk77]

Very awesome read, bro! I love all of your analysis articles.

I appreciate the kind words. I honestly don't know that much about statistics, so I was a little nervous throwing this out there.

[SimmerDownBruhh]

I appreciate the kind words. I honestly don't know that much about statistics, so I was a little nervous throwing this out there.

Stepping out of your comfort zone is never a bad thing, man. It's always good to see things in other perspectives when it comes to stats. I say you should keep these going, it's definitely something that I look forward to seeing.

[timeconsumer]

I appreciate the kind words. I honestly don't know that much about statistics, so I was a little nervous throwing this out there.

You're off to a good start. If you want to go a little more in-depth I would examine your p-values to see if the relationship you are seeing with the correlation coefficient is considered significant. If you return a correlation with a p-value of something like 0.5 you're probably looking at a bunch of statistical noises. But if you see a p-value of 0.1 to 0.01 you could be on to something.

Also when it comes to correlation coefficients just because it's positive doesn't necessarily mean much. So when you found a coefficient of 0.16 for blocking it really didn't mean anything. 0 means absolutely no linear relationship at all and 1 means a completely linear relationship so if you look at the picture here:



You can see that 0.16 is basically a big blob. But finding something like a 0.5 or 0.6 (strength) is a good start if you have it associated with a significant p-value, and something you can definitely formulate a conclusion off of.

(disclaimer: not a statistician by any means, but I dabble)

[7hawk77]

You're off to a good start. If you want to go a little more in-depth I would examine your p-values to see if the relationship you are seeing with the correlation coefficient is considered significant. If you return a correlation with a p-value of something like 0.5 you're probably looking at a bunch of statistical noises. But if you see a p-value of 0.1 to 0.01 you could be on to something.

Also when it comes to correlation coefficients just because it's positive doesn't necessarily mean much. So when you found a coefficient of 0.16 for blocking it really didn't mean anything. 0 means absolutely no linear relationship at all and 1 means a completely linear relationship so if you look at the picture here:



You can see that 0.16 is basically a big blob. But finding something like a 0.5 or 0.6 (strength) is a good start if you have it associated with a significant p-value, and something you can definitely formulate a conclusion off of.

(disclaimer: not a statistician by any means, but I dabble)

I gotcha. So some of my conclusions might still be right, but instead of me guessing at the statistical significance, I can use P-Values to determine that. I'll try and figure out how to calculate P-values on the above data when I find some time. Also, is there a simple way to do that in excel? I ended up using the data analysis to do an F-Test on Two sample for Variances but I think I just confused myself a bit.

[Molarpistols]

I gotcha. So some of my conclusions might still be right, but instead of me guessing at the statistical significance, I can use P-Values to determine that. I'll try and figure out how to calculate P-values on the above data when I find some time. Also, is there a simple way to do that in excel? I ended up using the data analysis to do an F-Test on Two sample for Variances but I think I just confused myself a bit.

Under the data analysis tool you could run a regression that'll give you a separate p-value for each variable (agi/str/etc/etc). It'll also give you a p-value for the entire relationship, as well as F-values and r^2. It likely won't tell you much unless you have quite a large data set for each variable however.

If you can't find the regression in the data analysis tool, you may have to download it by going into Excel->Options->Add-ins

[timeconsumer]

I gotcha. So some of my conclusions might still be right, but instead of me guessing at the statistical significance, I can use P-Values to determine that. I'll try and figure out how to calculate P-values on the above data when I find some time. Also, is there a simple way to do that in excel? I ended up using the data analysis to do an F-Test on Two sample for Variances but I think I just confused myself a bit.

I'm a big fan of the regression analysis. The output will give you the r-squared value which tells you how much variance in the dependent variable can be explained by the independent variable(s).
Then you can check your ANOVA table underneath that to see the Significance F value to check the statistical significance of the whole test (and the associated r2).
Then finally your coefficients table gives you the p-value for each attribute while also showing you a coefficient attempting to explain what an increase of 1 in x does to y.

But here's the big issue with a pearson correlation and a regression (and most other tests we would use as amateurs) they rely on continuous data. And with my understanding of this sim attributes and their affect on the player are not continuous. This means that going from 70 to 80 strength does not give as much of a benefit as going from 90 to 100 strength. So instead we have to weight our attributes to account for this. I have a method of weighting that I use, and it's working okay for me, but unfortunately it's secret Otter property. Play around with weighting it, you might get a better system than me.

[7hawk77]

I'm a big fan of the regression analysis. The output will give you the r-squared value which tells you how much variance in the dependent variable can be explained by the independent variable(s).
Then you can check your ANOVA table underneath that to see the Significance F value to check the statistical significance of the whole test (and the associated r2).
Then finally your coefficients table gives you the p-value for each attribute while also showing you a coefficient attempting to explain what an increase of 1 in x does to y.

But here's the big issue with a pearson correlation and a regression (and most other tests we would use as amateurs) they rely on continuous data. And with my understanding of this sim attributes and their affect on the player are not continuous. This means that going from 70 to 80 strength does not give as much of a benefit as going from 90 to 100 strength. So instead we have to weight our attributes to account for this. I have a method of weighting that I use, and it's working okay for me, but unfortunately it's secret Otter property. Play around with weighting it, you might get a better system than me.

This is super helpful information. I'll probably do a follow up article that shows all of the summary outputs and what it means once I figure that out.

However, with how I did the first article, that's 7 regression tests for each stat just to see how it effects tackling or sacking or 1 other aspect. This is going to be a lot of stuff to go through but pretty cool to look at. I definitely get what you mean with the weighted stats, I'm looking forward to playing around with that.

Edit: Disregard some of my post, For some reason I didn't think I could do a multivariable regression and would have to look at them individually.

[7hawk77]

Under the data analysis tool you could run a regression that'll give you a separate p-value for each variable (agi/str/etc/etc).  It'll also give you a p-value for the entire relationship, as well as F-values and r^2.  It likely won't tell you much unless you have quite a large data set for each variable however.

If you can't find the regression in the data analysis tool, you may have to download it by going into Excel->Options->Add-ins

Gotcha, Yeah I've got it added in.

Thanks!