Quarterback spending has the clearest payoff, +0.31 points per game for each point of cap share, almost all of it through the passing game. Receivers, linemen, edge rushers and cornerbacks pay off on the side of the ball where they play. Tight end looks valuable too, but it fails a basic placebo check, so I treat it as noise. All of it together explains about 10% of how teams perform in these seasons, and closer to 4–8% when predicting seasons the model has not seen. Within a single team, tight end spending shows no effect at all.
Salary-cap spending by position against team results, 2013–2025: 416 team-seasons, 32 teams. Starters are separated from bench depth using snap counts. QB returns taper above roughly 14% of the cap (about $30M+ today), though the data cannot say whether they turn negative.
What one more point of cap share (about $1.6M) buys when moved from bench depth into a position’s starters. The last row, dead money and unspent cap, is cap not tied to any tracked player.
Click a row to open it in the explorer below; the shaded row is the one open there now. TE is tagged because it fails the offense-versus-defense placebo check. Terms are explained under How to read this.
How “spending predicts” is calculated. It is the fitted value from the same model as the effects chart: that outcome regressed on every position’s starter cap share plus dead money, with a separate baseline for each season. So the prediction is that season’s league-average result, adjusted up or down for how this team split its cap. The gap is actual minus predicted. The model was fit on these same team-seasons, so the gaps are slightly smaller than they would be for a truly new season.
Spending explains only about 16% of results, so most of these gaps are not about money: schedule, coaching, depth and luck all sit here, and so do injuries. The injury figures are context (the planned starting QB is the highest-cap QB on the roster; injured-reserve data starts in 2016), not an adjustment.
From the notebook (sections 9 to 13). Bench spending looked wasteful at first; most of it is injury cost.
| Starting QB misses 8 weeks | −5 pts/game | About 0.64 points per game for each missed week. |
|---|---|---|
| A replacement QB one standard deviation better | ~1/5 back | Close to what the mechanics predict. The data cannot tell it from zero, and cannot rule out a benefit up to about two-fifths. |
| Last year’s injured-reserve money | ~0 carry-over | It predicts nothing about this year, so it is an injury effect, not a badly built roster. |
Data. Per-player cap hits (Over the Cap), snap counts, results and EPA/play from nflverse, 2013–2025. The cap data covers part of the official cap, so everything is expressed as shares.
Starter means the top-N players at a position by snaps (QB1, RB1, WR3, TE1, OL5, DL2, EDGE2, LB2, CB3, S2). Everyone else is bench.
Caution. These are associations, not causal effects. Eleven rows across several outcomes means a few significant results are expected by chance. The full write-up, with all checks, is in the project notebook.