At 2-14, the 2019 Cincinnati Bengals were the worst team in the NFL, but the franchise had a promising future ahead of them in their next two drafts. After landing the number one overall pick, the team unsurprisingly took quarterback Joe Burrow: arguably the best prospect at the position in years. Following a marginal improvement in a four-win season, the Bengals once again had a top five draft pick for 2021. This time, the Bengals found themselves in a new dilemma. They already had a long-term plan at quarterback; the question was whom they would pick as Burrow’s offensive teammate.
To paint with a very broad brush, the decision fell between two stars: wide receiver Jamar Chase and offensive tackle Penei Sewell. On one hand, the Bengals could potentially bring Burrow a homegrown receiver that would both draw fans and immediately be a big play threat every time he stepped out on a field; on the other hand, Sewell could possibly contribute every single down and protect Burrow from opposing pass rushes. Ultimately, the Bengals chose Chase, and Sewell fell down two spots later to the Detroit Lions.
Did Cincinnati make a bad choice here? Based on my research last week, as well as the underrated methodological strategy of “having eyes,” the answer is obviously no. Along with being one of the five best skill position players of this decade, Chase has helped bring the Bengals to relevancy; to a Super Bowl appearance where he had one of the most dominant playoff receiving performances of all-time. Still – it’s worth noting that Sewell in this same time period was fairly dominant within his own position, playing a large role in the revival of the Detroit Lions franchise. What would have happened if he and Chase swapped places?
In last week’s column, I expanded my personal Wins Above Replacement methodology to cover wide receivers, tight ends, and running backs. However, for today’s piece, I’ve chosen to try applying a slightly different methodology for WAR at the tackle, guard, and center position.
Challenges With Offensive Lineman and WAR
To avoid lengthy explanations, what you need to know about my approach to WAR is that I use Expected Points Added (EPA) as its primary baseline. This is a statistic that comes from nflFastR, which tracks all play-by-play data in the NFL from 1999 onward; for the purpose of my multi-week research project, I’ve limited my scope of analysis to 2021 to now. Using each player’s attributed EPA, the calculated EPA for their bottom 25 percentile of their respective position group (quarterback and skill position), and projected EPA-per-actual win, I am able to calculate WAR in the following formula: (EPA – ReplacementEPA) /(EPA per Win). percentile of players at their position group (quarterback or skill position), and EPA.
EDITOR’S NOTE: In the case of quarterbacks, I used a balanced version of EPA that factored in predicted EPA by Completion Percentage Over Expected and Success Rate as an equalizer to the otherwise unsteady and volatile raw EPA; this was in order to prevent values from inflating.
Offensive lineman, however, throw a wrench into my current approach for EPA. Unlike quarterbacks, running backs, tight ends, and wide receivers, linemen are not actively assigned EPA because their play-by-play data doesn’t exist in any public capacity. The forefathers of my work (Ronald Yurko, Samuel Ventura, and Maksim Horowitz) acknowledge this in their nflWAR paper, though they lay out a conceptual groundwork that would treat ‘unit-wide blocking’ as a more pertinent point for emphasis than individual linemen tracking.
This was reasonable enough, but I’m a stickler for consistency. For the purpose of an exploratory graduate project on assessing positional value, this just did not seem adequate, but as a writer, it also felt narrative unsatisfying to me. Therefore, I had to come up with a looser approach.
All Paths Lead to PFF
At the risk of sounding like a shill for Pro Football Focus – who has not hired me to write this – I thankfully had a premium membership that allowed me to access season-long blocking data for offensive linemen (tackles, guards, and centers). This included prominent categories for offensive lineman such as snap count, number of pass blocks, number of rush blocks, penalties, pressures allowed (including sacks, hurries, and quarterback hits as separate categories too), and, controversially, PFF grades for each of run blocking, pass blocking, and overall offense.
It was here, when I had to make a judgment call: spend time synthesizing pressures, sacks, and other blocking outcomes into my own composite metric or just use PFF grades. Ultimately, I landed on sticking with PFF grades, which largely already factor these statistics into account for context, as evaluated by a team of scouts. In the future, I may eliminate PFF grades entirely and try to find a better way of creating a composite stat for individual pass blocking and applying a flat portion of rushing EPA to run blocking, but for now, this was the most convenient solution. Because the data for these players is so limited otherwise, I had no other convenient choice.
With these steps completed, I then compiled the average offensive EPA for each team in a given season for my data set (2021-2024). Following that, I ran team-aggregated combinations for PFF grades and, out of curiosity, individual statistics, making sure to use snap-weighted averages and sums when necessary. I also aggregated positional information related to these statistics so that I could measure any noticeable differences in guards, centers, and tackles before merging my two tables together of aggregated individual statistics with team attributions and team-wide EPA statistics.
Going through each of these categories and more, I then built multiple regression models, the most important of which was my model measuring total offensive grades to total offensive EPA. This had an R-squared of 0.264, so while it wasn’t entirely explanatory, it did explain a moderate amount of variation, with minimal improvements when accounting for other pass blocking statistics. Ultimately, what I found was that roughly a one point increase in team pass block grade predicted 5.30 more offensive EPA over a season, while a one point increase in team run block grade predicted 5.09 more offensive EPA over a season.
Calculating WAR
Following team-wide calculated attributions of EPA to the offensive line, I then proceeded as usual with the rest of my WAR process, first establishing a replacement level (50 minimum snaps and bottom 25 percentile of graded players) for each individual position (T, G, and C) of the offensive line group. For each lineman, I calculated their grade minus the replacement-level grade for their position, making sure to do this for pass blocking and run blocking alike. To guarantee that players with minimal snaps weren’t inflated, I weighted their contributions by playing time.
Next, I used previously calculated regression coefficients for pass blocking (5.30) and run blocking (5.09) alike to articulate team-level relationships between lineman grades and overall line performance. I then applied a divisor of 2.5 – derived from manual trial-and-error between divisors from 1 to 5 and measured against team wins correlation – to ‘counter’ potential grade inflation and account for offensive lineman’s impact on each other’s performance.
Using Grade Above Replacement and my chosen divisor, I then multiplied the following result by my coefficients and snap shares for each of pass blocking and run blocking before running the results to have my final EPA numbers attributed to individual offensive linemen. Lastly, with EPA numbers, I followed the same process as in my previous columns, taking Individual EPA above Replacement divided by the EPA per win attributed to an offensive line times the offensive line’s projected share of wins. The final set for calculating WAR had approximately 1153 seasons.
Initial Insights
I ended up with a total of 476 tackle seasons, 465 guard seasons, and 212 center seasons (for timeliness, I did not differentiate between right or left variants of guards or tackles). Across these specific positions, the average WAR was the highest for tackles (0.157), but guards (0.152), and centers (0.147) weren’t too far behind. Interestingly, however, the biggest outliers were found in the center and guard positions. Each of the highest guard (1.033) and center (0.966) seasons exceeded the maximum WAR for a tackle (0.861) season.
| Top 10 Cumulative WAR for Offensive Lineman (2021-2024) | ||
| Player | Position | Total WAR |
| Creed Humphrey | C | 3.30 |
| Chris Lindstrom | G | 3.25 |
| Jordan Mailata | T | 2.83 |
| Penei Sewell | T | 2.79 |
| Trent Williams | T | 2.77 |
| Joe Thuney | G | 2.72 |
| Joel Bitonio | G | 2.58 |
| Lane Johnson | T | 2.39 |
| Kevin Zeitler | G | 2.36 |
| Kolton Miller | T | 2.34 |
If you’ll remember from last week – Jamar Chase’s four-season WAR sat at 2.78. So to bring this back to the question I posed at the start of this column, the answer between Chase and Sewell was practically a tie (even if Sewell has a 0.01 lead in two more games).
Validation with AV
Like what I’ve done with quarterbacks and skill position players, I gathered together a sheet of offensive lineman from 2021 to 2024 organized by Approximate Value in order to measure my WAR’s correlation with it. I then used a bootstrapped set of 240 players as my point of comparison for the two statistics (Cumulative AV and WAR for offensive linemen).
Among matching players, the model had an R-Squared value 0.591, meaning that it explained well over half of the results. Conversely, the Pearson correlation (0.769) was fairly high across position groups, with each of tackles (0.798), guards (0.772), and centers like (0.707) having a high number. This demonstrates that both metrics (AV and my variant of WAR) largely measure the same thing in terms of contributed volume from an offensive lineman across a whole time span. The difference, of course, is that my methodology implements more subjective positional grades for individuals and EPA estimates derived from them, while AV applies a more static distributional share of value based on the performance of other positions.
It’s here where my multi-week project takes its first L. When I ran a correlation test between the two metrics, AV (0.564) AV slightly outperforms WAR (0.512). Although the difference is tiny, it is noticeable. Then again, I do think that AV probably benefits from its innate distributional value model; by having a direct, even if static, tie-in with how other positions perform, it has a noticeable advantage in terms of its correlation to wins versus WAR, which is more individual-specific.
Offensive WAR Conclusion
With the third week of this project out of the way, I have now calculated WAR for the following position groups: quarterbacks, running backs, wide receivers, tight ends, tackles, guards, and centers. For each of these position groups, I’ve taken a slightly different approach to WAR, but the individual EPA-centric foundation has largely stayed the same.
Before heading onto defense for next week, I was curious to check two more points of references – the first one being Pro Football Focus’s old attempt at WAR in 2018, built from far more seasons worth of data, with a lower baseline for replacement level, more complex methodology (Massey matrices implementation), and higher snap thresholds. How did the offensive values for WAR differ from each other?
| Position Group | Sample Size (Me) | Sample Size (PFF) | Average Total WAR (Me) | Average Total WAR (PFF) | Net Diff |
| QB | 209 | 994 | 1.05 | 1.63 | -0.58 |
| HB (including FB) | 400 | 2373 | 0.06 | 0.10 | -0.04 |
| TE | 167 | 1621 | 0.15 | 0.18 | -0.03 |
| WR | 431 | 2864 | 0.30 | 0.28 | +0.02 |
| T | 476 | 1543 | 0.16 | 0.09 | -0.07 |
| G | 465 | 1604 | 0.15 | 0.10 | -0.05 |
| C | 212 | 708 | 0.15 | 0.10 | -0.05 |
| Offense | 2360 | 11,707 | 2.02 | 2.47 | -0.32 |
The first step here was to run a quick correlation test for my version of WAR against PFF’s version of WAR in its paper. Ultimately, my raw correlation (0.992) was nearly identical.
With that said, I was not satisfied. To quote Deep Throat – or at least the film adaptation of him from All The President’s Men – the key here was to follow the money. Although NFL teams are not necessarily always making the correct financial decisions, my next step was to compare what I and PFF had as separate inputs and compare it to market rates, as measured by Over The Cap in 2021. I chose these specific numbers both out of convenience (this came up as one of the first hits in my research) and out of confidence that the general numbers wouldn’t have changed too much in four years.
Both in terms of raw numbers (0.949), my WAR methodology largely reflected my stand-in market numbers in statistically significant fashion. PFF WAR was similarly strong in terms of numbers (0.930). Based on the mean numbers though (and assuming most of the 2021 trends stay true today), the following can be inferred about the tracked position groups.
- PFF emphasizes the importance of quarterbacks; I do it to a lesser degree, and the market has not caught up.
- All three of me, PFF, and the market are aligned on wide receiver and tight end value.
- The market values tackles far more than me and especially PFF.
- The market values centers and guards slightly more than me and far more than PFF.
- PFF and I agree that the market overpays running backs.
- For non-quarterbacks, the general rule of thumb is that the further you are from touching the ball/on the field, the more valuable you are. However, if you play significantly more snaps than another position, you can make up the difference.
Determining who has it right on a positional basis is an entire endeavor for another column. I will say though: one factor both my model and PFF undervalue is the importance of insurance. In other words, teams may value offensive linemen far more because of their stability and ‘protection’ of other offensive assets, like a star quarterback’s ability to stay on the field. This is largely a limitation of a pure WAR or EPA methodology though, and not something within the scope of what I’m evaluating.
Still – for a graduate student that’s spent most of this semester researching football, I think I’ve done a fine enough job. In next week’s piece, I’ll take a look at defensive players as a whole.
Appendix