Which defensive line position is the most valuable?

SIS Football Analytics Challenge | July 2020

Sam Struthers & Adrian Cadena

Sports Info Solutions mentioned this project as one of the highlights of the competition. Source code here.

Challenge

Not all situations are created equal.

What is Talent?

What is positional value?

  • Position “a” (low variance): the difference in impact from a high-end to a low-end player is not very pronounced.
  • Position “b” (high variance): there is a massive difference in impact from a high-end to a low-end player.
  • Position “b” is more valuable than “a.”

Positions

  • Effective Positions: NT, DT, DE, and OLB
  • Broad Positions: EDGE (DE+OLB) and IDL(NT+DT)
  • IDLs will sometimes be referred to as interior defenders and EDGE as EDGE defenders.

Run Defense Analysis: Evaluation Method

Following the principle from [Yurko, Ventura, Horowitz 2019], we will be finding the individual defensive points added (iDPA) for each player:

“We refer to an intercept estimating a player’s average effect as their individual points/probability added (iPA), with points for modeling EPA and probability for modeling WPA. Similarly, an intercept estimating a team’s average effect is their team points/probability added (tPA)” [Yurko, Ventura, Horowitz 2019]

Our goal will be to take iDPA a step further by only looking at plays in which the player had a chance to be involved; we will call it: Credited Individual Defensive Points Added (cIDPA)

Mixed Effects Model — Run: Finding a player’s intercept

For a description of the variables, please refer to our submission.

  • Response variable: Resulting EPA on the play
Fixed Effects = mean_dist + min_dist + min_dist_2 + distance_from_gap* + num_DT + num_NT + num_DE + num_OLB + Yardline_100 + ScoreDiffRandom Effects = PlayerId:distance

Here is where IDPA turns into cIDPA. We are trying to credit EPA only when the defender is close enough to impact the play.

  • By making each player interact with distance, we can look at each player’s intercepts at each span.
  • We kept only each player’s intercepts when they are close or at mid-distance from the designed run.
  • We grouped by the player using their mean intercept (close and mid-distance). The mean is each player’s cIDPA

Mixed Effects Model — Run: Results

In run defenses, from high to low, the order of positional value is the following:

  1. IDL (variance: 4.67e-04)
  2. OLB (variance: 2.99e-04)
  3. DE (variance: 6.03e-05)

In a front 3 run defense, when comparing our two broad positions (IDL vs. EDGE), we find they have similar variance. However, IDL Defenders still carry a bit more variance.

In Front 4:

Our model found that DE is significantly more ‘valuable’ than IDL in a front 4 run defense.

Run Defense Analysis: Interpretation

Pass Defense Analysis: Evaluation Method

We are controlling for specific parameters in this analysis and only accounting when player i is rushing. We are not predicting whether the QB will be pressured during the play, but whether player i will pressure the QB

We call our final product: iLog-Odds of Pressure.

How impactful are QB pressures?

As our plot shows, QB pressures significantly reduces avg completion % regardless of Air Yards.

During weeks 7–19 of the 2019 season:

  • Avg EPA/Pass without pressure: 0.204
  • Avg EPA/Pass with pressure: -0.40

Mixed Effects Logistic Regression — Pass: Finding a player’s intercept

For a description of the variables, please refer to our submission.

Response variable = Player i pressuring the QB during play nFixed Effects = ToGo:Down + Yardline_100 + Down + ToGo + ScoreDiff +shotgun + blitz + (number of players from other positions pressuring)Random Effects = PlayerId:IsRushing

By making each player interact with the IsRushing variable, we can look at each player’s intercepts when rushing the pass. We kept only each player’s intercepts when they were rushing the passer: iLog-Odds of Pressure.

Mixed Effects Logistic Regression — Pass: Results

In Front 3 pass defenses, from high to low, the order of positional value is the following:

  1. OLB (variance: 0.130)
  2. DE (variance: 0.028)
  3. IDL (variance: 0.022)

Front 4:

We are analyzing broad positions as we did in the run analysis. IDL is slightly more valuable than DE in front 4 pass defenses

Front 3

When comparing broad positions, EDGE Defenders are more valuable than IDL in a front 3 pass defense.

Pass Defense Analysis: who helps who?

To address this, we looked at:

  • Avg. EPA for each player lined-up in position a when they sacked or pressured the QB during the play, and position b was pressuring the QB during the play
  • Same as previous, but when position b was not pressuring the QB during the play
  • Then, we ran simulations and compared the sample-mean distributions of both scenarios for both positions.

Generally, IDL players' average impact sharply increases when there is one or more EDGE pressuring the QB during the play. The difference between the mean of these two distributions, in these simulations, was 0.52 (pressure — no pressure from other position)

Generally, EDGE players' average impact increases when there is one or more IDL pressuring the QB during the play. The difference between the mean of these two distributions, in these simulations, was 0.19 (pressure — no pressure from other position)

After looking at both results, we determined that EDGEs help IDLs to a greater extent.

Pass Defense Analysis: Interpretation

Works Cited

2.Hermsmeyer, J. (2020) Exactly How Much Does A Great Pass Rush Hurt An Offense? https://fivethirtyeight.com/features/exactly-how-much-does-a-great-pass-rush-hurt-an-offense/

3.Riske, T. (2020) PFF Data Study: Debunking the myth of the “sack artist https://www.pff.com/news/nfl-pff-data-study-sack-artist-pass-rushers

4.Walder, S. SethWalder. (2019, Dec 19).Double team rate as an edge rusher (x) by pass-rush win rate as an edge rusher (y) https://twitter.com/SethWalder/status/1205222343120957449?ref_src=twsrc%5Etfw%7Ctwcamp%5Etweetembed%7Ctwterm%5E1205222343120957449%7Ctwgr%5E&ref_url=https%3A%2F%2Ftheramswire.usatoday.com%2F2019%2F12%2F13%2Fnfl-rams-aaron-donald-pass-rush-win-rate%2F

Hello, I’m a Sr. Data Analyst and aspiring Data Scientist. My interests include sports analytics, particularly NFL, economics, and finance.