How to Adjust NFL Stats for Opponent Strength

Start with one phase, play type, and game state. For example, passing efficiency can be measured as EPA per dropback, including sacks but excluding spikes and kneel-downs. Do not blend it with rushing EPA, special teams, or total scoring; each responds differently to opponent quality.

Prefer rates whose denominators represent genuine opportunities:

• Pressure rate: pressures per opponent dropback
• Run-defense success rate: unsuccessful opponent runs per qualifying carry
• Points per drive: points per non-kneel possession

The denominator must remain consistent across teams and opponents. Raw sack totals or season points reward extra snaps and possessions, obscuring efficiency. When using EPA in matchup analysis, document play exclusions, garbage-time filters, and whether turnovers retain their full value.

Success rate shows how often an offense stays on schedule; explosive-play rate captures large gains. Comparing success rate with explosive-play efficiency separates consistency from volatility: equal EPA can reflect steady gains or a few breakaways.

Match the unit to the question. Compare passing offense with pass defense per dropback, rushing per designed carry, and special teams per kick, punt, or return. Never pool offense with defense or mix plays, drives, games, and per-attempt rates. Incompatible denominators create precise-looking but meaningless adjustments.

Opponent adjustment cannot repair mismatched inputs. Freeze season and week boundaries, EPA or success-rate definitions, scoring baselines, and overtime treatment across all teams. Classify sacks, scrambles, penalties, spikes, kneels, two-point tries, and special-teams plays explicitly; exclusions must be symmetric for offense and defense.

Choose the observation level

• Play-level models preserve situation and opponent detail, but require clustered errors because plays within games are dependent.
• Drive-level models capture possession outcomes, provided starting field position is controlled.
• Game-level models are simpler but discard context and produce noisier estimates.

Tempo must match the target. Per-game production reflects pace and efficiency; per-play rates isolate efficiency. For volume forecasts, model pace separately rather than treating extra snaps as superior execution.

Filter garbage time with a reproducible rule based on score, clock, and pre-snap win probability—not hindsight. Report results with and without the filter when thresholds materially change rankings. Recency weighting is appropriate for forecasting amid injuries, quarterback changes, or tactical shifts; full-season evaluation generally favors equal weighting, with postseason weeks handled consistently.

Baselines must use the same unit, filters, and weighting as the target statistic. To adjust offensive early-down EPA per play, calculate defensive early-down EPA allowed per play—not points allowed, total EPA, or an all-down measure. The league baseline should apply identical season, play-type, garbage-time, and recency rules.

For target offense T, estimate each opponent’s defensive effect after removing every play against T:

Opponent effect = EPA allowed excluding T − league EPA baseline

This leave-one-team-out calculation prevents a circular result. If an offense posts 0.30 EPA per play against a defense, including that game makes the defense appear easier and discounts part of the performance being evaluated.

Average the opponent effects across the target’s schedule, weighted by relevant opportunities. Then calculate:

Adjusted offensive EPA = observed EPA − weighted opponent effect

An opponent effect of +0.04 indicates easier-than-average defenses, so 0.12 observed EPA becomes 0.08 adjusted EPA. Recompute ratings for every target; one universal opponent table retains contamination.

The simplest correction is additive because EPA already has a meaningful linear scale:

Adjusted EPA/play = observed EPA/play − opponent effect

Suppose an offense produced +0.08 EPA/play. Its opposing defense typically allows offenses to perform +0.03 EPA/play above neutral. The schedule-neutral estimate is therefore:

+0.08 − (+0.03) = +0.05 EPA/play

The sign matters. A permissive defense has a positive opponent effect, which lowers the adjusted result. A defense that suppresses EPA by 0.03 has an effect of −0.03; subtracting it raises the offense’s estimate.

Additive treatment works best when equal absolute differences carry similar meaning across the scale. It is less suitable when effects are naturally proportional. Completion percentage, pressure rate, and other bounded percentages may be modeled through odds ratios or logits, avoiding impossible adjusted values below 0% or above 100%.

For rare events such as interceptions or blocked punts, a multiplicative rate ratio or log-rate model often fits better: doubling a 1% rate is more meaningful than adding one percentage point. Zero counts require smoothing before taking logs.

For observation (g), model performance as

[ yg = \mu + O{i(g)} – D_{j(g)} + X_g\beta + \varepsilon_g ]

Here, (\mu) is the league mean, (O_i) and (D_j) are offensive and defensive strengths, and (X_g) contains controls such as home field, rest, weather, quarterback availability, and garbage-time indicators. The sign convention must remain consistent: a larger (D_j) above means a stronger defense suppresses the outcome.

Because adding a constant to every team effect can leave fitted values unchanged, impose an identification constraint such as sum-to-zero offense and defense ratings. Observation weights can reflect plays, drives, recency, or estimated measurement precision.

Three estimation approaches are common:

• Iterative SRS-style updates alternate between recalculating team ratings and opponent adjustments until convergence. They are transparent, but controls, uncertainty, and unequal weights become cumbersome.
• Least squares estimates all effects simultaneously, accommodates contextual variables naturally, and exposes residual variance and standard errors.
• Ridge regression adds penalties to team effects, shrinking weakly supported ratings toward league average.

Ridge is usually preferable early in the season, when sparse or nearly disconnected schedules make unpenalized estimates unstable. The league mean and well-established contextual effects are typically left unpenalized. Select the penalty with time-aware validation, then inspect residuals for heteroskedasticity and within-team correlation; clustered or robust errors are often necessary.

Opponent strength can absorb unrelated advantages unless the model separates major contextual effects. Add only controls with a credible mechanism, reliable data, and enough observations to estimate them.

• Treat home field, rest differential, and season or era as explicit covariates; league scoring environments and home advantage change over time.
• Segment quarterback samples when a starter is absent rather than assigning every snap to one offensive rating.
• Model weather selectively—wind matters more for deep passing and kicking than for interior rushing.
• Exclude garbage time or control for score, time remaining, and win probability so game script does not masquerade as opponent quality.
• Represent injuries at the position-group level. Detailed personnel evidence from PFF Premium matchup research can support a specific adjustment, but subjective injury grades should not dominate the estimate.

Prefer a transparent core model, then run sensitivity checks: remove backup-quarterback games, split home and away samples, vary garbage-time thresholds, and compare recent seasons with era-normalized history. If opponent ratings move sharply, report a range or flag instability instead of adding more parameters. The goal is not maximum fit; it is a schedule adjustment that remains interpretable and robust.