NFL Plus/Minus: Myles Garrett is the Defensive Player of the Year

Using historical on-off the field splits and player clustering to estimate value of defenders

In this post, I’m going to detail the methodology and results applying my NFL Plus/Minus metric to defenders this season, which includes many of the top candidates for NFL honors, including Defense Player of the Year and All-Pro selections, plus some names you might not expect.

My NFL Plus/Minus methodology will evolve as time goes on, and the results will also change. Even so, the first runs of results over the last two years point to a few potential gaps in our current understanding of where defender value comes from, and gives us some ideas on how to apply historical on/off splits to make assumptions about the broader values of different position groups.

Yesterday I detailed the NFL Plus/Minus methodology and results for all non-quarterback offensive players (showing that Tyreek Hill is probably the most valuable non-quarterback, with Christian McCaffrey not far behind), and in this post we turn to defenders. It’s more difficult to find relevant, objective metrics to judge defenders than we have for offensive players, specifically offensive skill position players. Therefore, there’s a higher reliance on PFF grading in this analysis, especially for run defense, and coverage to a lesser degree.

METHODOLOGY

Using participation data and on- and off-the-field splits have a rich history in other sports for building player valuation metrics. No matter how much we try, the human eye and brain cannot notice the contribution to every player on every play while calculating the exact impact of those actions on the game in a comparable metric. By studying the differences in team success — in this case, using expected points added — when a player is on and off the field, we can capture the entire effect in useful numbers for comparison. 

The issue with using on-off splits in football is the time for all players on the field is limited in a 17-game season, and the substitution patterns for different positions don’t allow for much comparison. Without injury, cornerbacks can play every snap of the game, whereas certain interior defenders are rotated consistently. The solution I came up with was to look at the on-off splits of groups of similar players rather than individual players.

I used the nflreadR participation data going back to 2016 — along with other traditional and advanced stats — to cluster similar players for each position in each facet of the game and then compiled the on-off splits of those groups. In doing so, we increase the total samples of measurement and reduce the noise of single-player numbers. 

Before we get into the results of the analysis for every defensive position, let’s walk through a specific example of how clustering works, using the coverage facet for cornerbacks.

Without getting bogged down in the details, the graph above divides all cornerback seasons with at least 25 coverage snaps from 2016 to 2023 into eight different color-coded clusters. The metrics used to group the clusters include coverage grade, percentage of snaps targeted and percentage of snaps in the slot. This is a simplified version of the actual clusters used in the analysis, which are more numerous and contain fewer players per cluster.

I’ve highlighted a handful of cornerback seasons from 2023, showing that young studs like Sauce Gardner and Patrick Surtain are in the direction of higher grading, and the opposite direction of higher yard and touchdown rate (per coverage snaps) where Mike Ford resides. Michael Carter II is an example of a strong slot corner, which falls in the lower lefthand side. Alontae Taylor wasn’t as strong in cover, and spent significant time in the slot. DaRon Bland and Derek Stingley weren’t quite on the level of Gardner or Surtain in grading, but their interception rates (per coverage snap) were very strong, in particular Bland, who led the NFL with nine INTs. James Bradberry fell into the mediocre category for grading and receiving efficiency allowed.

Once we have players assigned in clusters, we total up the team-level expected points added with the clustered players on- and off-the-field then come up with a per-attempt number for the average value added by that cluster versus their teams giving coverage snaps to other cornerbacks.

Unsurprisingly, historical on-off splits show that teams miss cornerbacks like Carter, Gardner and Surtain when they’re off the field. There’s also a material positive effect for the Bland/Stingley cluster, though I’ve found that corners who rely more heavily on interceptions for value aren’t consistent producers going forward.

For the NFL Plus/Minus results, this exercise of clustering by role is applied to every position in all three facets of defensive play: pass rush, run defense and coverage. The results below detail the points prevented (negative is good) numbers for the top-10 defenders in every position group and then the top-12 overall. I’ll post more commentary after on the results and how I’m thinking about them.