The 49ers’ injury problem is real.
Explaining it is harder.

Adjusted Games Lost (AGL)

Football Outsiders / FTN’s estimate of injury-related availability lost by expected starters and important contributors during the regular season. Missed games add to the total; playing while listed as injured can add a fraction. It is not a count of diagnoses, and a star is not weighted more than another starter.

Injury index

A comparison scale: 100 is that season’s NFL average. A team at 140 has 40% more AGL than the league average; a team at 50 has half as much. Higher means more injury burden. It is not an individual player’s chance of getting hurt.

Playing time in earlier games

The model uses a player’s total snaps and offensive/defensive snap share in the last earlier game, plus snaps and games in the previous 28 days. All workload comes strictly before the game leading to the next report. Example: predicting a report before Week 8, with Week 7 as the current game, uses snap history before Week 7. Week 7 could already have been shortened by an injury. This does not measure first-half versus second-half workload, practices, collisions or prior-year fatigue.

Earlier injury reports

How often a player appeared on public physical-injury or condition reports in the preceding 28 and 365 days, plus time since the last report. Several weekly listings may all refer to one lingering injury; this is report history, not a tally of separate injuries.

Player-snaps

Participation summed across players. An ordinary offensive play with 11 players contributes about 11 offensive player-snaps. The team totals here add offensive and defensive participation. They do not measure distance run, collision force, practice work or actual fatigue.

The uncertainty range

A 95% confidence interval describes uncertainty in an estimated average or association under the model’s assumptions. It is not the range containing 95% of individual teams. When an estimated effect’s interval includes zero, these data do not clearly distinguish it from no association; a wide interval can still allow meaningful harm.

Team profiles (clusters)

Groups of team-seasons with similar age, continuity, playing-time concentration, volume, playoff history and rushing share. Injury outcomes were not used to form them. Their boundaries overlap, so they are descriptive summaries rather than distinct medical risk categories.

Age weighted by playing time

The average age of the players who actually played, with more influence given to those who took more offensive or defensive snaps. Age is measured on September 1 of the earlier season. It does not describe every player on the next season’s opening roster.

Historical roster continuity

The share of an earlier season’s snaps taken by players who also played for that club the year before. For the 2024 profile, SF’s 82.5% means 2023 snaps by players who also played for SF in 2022. It does not measure how many of those players returned for 2024.

Concentration of playing time

The share of a team’s offensive and defensive player-snaps taken by its 22 most-used players. A higher number means the core handled more of the work. It is not an overuse threshold; good health itself can allow a stable core to play more.

Rushing share

Carries divided by carries plus pass attempts plus sacks. It includes quarterback scrambles and kneel-downs, and reflects score and game situation. It is not a pure measure of designed runs or coaching philosophy.

Predictive importance

We shuffle one group of inputs and see how much the model’s predictions worsen on data it did not train on. A larger drop means that information helped this model more. It does not say what caused an injury, or what percentage of injuries a factor explains.

Average precision (AP)

A 0-to-1 score for how well a model brings positive cases toward the top of its predictions. Larger is better. Chart values show the change multiplied by 100: a loss of 2.8 AP percentage points means the AP score fell by 0.028. This is not a change in injury probability.

Share of injury reports

The proportion of public physical-injury/condition listings assigned to a body region. If a listing names two different regions, each gets half a unit. One persistent injury can appear in many weeks. Players sent directly to injured reserve may be missing entirely.

A new reporting episode

A player’s first observed listing for a body region, or a listing at least 28 days after the previous report in that region. This helps reduce repeated weekly listings, but it does not establish a new medical injury or confirmed recovery.

Timed injury announcements

The first play-by-play announcement per player per game explicitly stating that he was injured during the play, assigned to the quarter of that play. Return-to-game updates are excluded. This captures visible game announcements, not every injury or the full medical record.

Injury index = 100 × a team’s AGL divided by that season’s 32-team mean. Era rankings equally weight nine annual indexes. Franchise relocations are unified. AGL covers the regular season, not postseason losses, and reflects the publisher’s judgments about contributor eligibility; it is not talent-weighted clinical incidence.

We compiled successive FO/FTN public releases, using revised comparison columns where available. The 2017 table is an older vintage that cannot be fully restated from recovered sources. These are not all historical seasons rerun under one latest algorithm. Year normalization reduces scale differences but cannot eliminate methodology changes.

Direct COVID absences are excluded in 2020–2021. Pandemic changes to preparation, scheduling and reporting can still matter. The 2022 AGL source counts the cancelled Buffalo–Cincinnati game from report/IR status; workload counts only played games. Excluding pandemic years and excluding 2017 leave SF near the top. The conventional outlier check uses the 32 franchise means and the upper quartile + 1.5 interquartile ranges.

We classify all four primary/secondary injury and practice-report fields into 11 body-region/condition categories. Each physical player-game listing contributes one unit divided equally across distinct categories; repeated weekly listings remain separate observations. Illness, rest and administrative-only listings are excluded. Concussion-related includes protocol/evaluation language and is not necessarily a confirmed concussion.

The sample includes 41,531 physical listings, including 1,354 for SF, across 2017–2025. The other-31 comparison is standardized to the selected team’s season-by-broad-position distribution among its physical reports. This is composition conditional on being listed, not risk per player or per snap. Report coverage is present for 4,759 of 4,766 completed team-games, including every SF game.

Full-era SF uncertainty uses 4,000 bootstrap replicates of SF player trajectories and other-team franchise histories, recomputing standardization each time. Displayed difference intervals cover the 11-category family approximately simultaneously. None excludes zero; the exploratory global maximum-statistic p-value is .0895. This is conditional on a team and window selected after observing its history, and shared calendar shocks and player movement are not fully captured. Other team/year filters show descriptive estimates without intervals.

Checks add age bands, exclude 2020–2021, restrict to 2022–2025 or Out/Doubtful listings, and reduce repeats with first-region/player-season or a 28-day reporting-episode proxyⓘ. Direct-to-IR cases can be absent. Report location does not identify tissue damage, mechanism or severity; calf/Achilles is not a count of Achilles ruptures. The classification audit and sensitivity tables are in the downloads.

Exposure seasons are 2017–2024; outcome seasons are 2018–2025. Primary regression: following-season injury index on prior playoff games, prior injury index, prior regular-season win percentage, prior regular-season snap-weighted age, and outcome-year fixed effects. Super Bowl appearance and total player-snaps replace playoff games in separate models. Snaps sum offense plus defense and include the postseason; special teams and practice load are omitted.

Reported intervals use franchise-clustered standard errors, finite-sample corrections and Student-t critical values with 31 degrees of freedom. Restricted wild-cluster bootstrap checks use 9,999 draws. For prior playoff games, p=.315 (wild-bootstrap p=.324); for total volume, p=.940; for a Super Bowl appearance, p=.367. These p-values do not establish absence of harm.

We checked raw AGL, added team effects, removed SF, excluded pandemic outcomes or transitions, and split the playoff-expansion eras. Main full-sample conclusions are inconclusive. The small pre-expansion subset alone is borderline positive (p=.047), whereas the expanded-era subset is not (p=.877). Those exploratory, uncorrected checks show instability rather than a selected positive conclusion. Team effects with a lagged outcome in a short panel can introduce bias; that version is a sensitivity.

Prior health and wins help explain playoff selection but cannot randomize it. Team totals do not follow the same returning players, measure meaningful-game intensity, or observe actual offseason recovery. The 16 Super Bowl follow-ups involve seven franchises. Three-year playoff measures use 2015–2016 schedules for initial lookback windows.

All 256 profiles use completed pre-outcome information. Six features are standardized across the pooled sample. K-means uses 100 starts; k=2 through k=6 are compared using silhouette without inspecting the injury outcome. The best score is 0.181 at k=2. Median franchise-bootstrap adjusted Rand agreement is 0.597; Ward clustering agreement is only 0.217. Algorithm and feature choices materially affect membership.

Age is weighted by regular-season offensive/defensive snaps and measured September 1 of the feature year. Historical continuity is the share of those snaps from players who also played for the same club in the preceding year: SF’s 82.5% entering-2024 value means 2023 snaps by players also on the field for SF in 2022. It does not measure 2023-to-2024 retention. Top-22 concentration is partly influenced by previous injuries.

Rushing share is carries / (carries + attempts + sacks), including scrambles and kneels. Volume is scrimmage player-snaps per regular-season game. Playoff history covers three prior seasons. The two-dimensional PCA chart preserves 53.4% of feature variation. Nearest peers are selected using all six standardized features, not their apparent distance on the map.

Outcome intervals resample franchises 5,000 times and condition on the displayed cluster labels. They do not include uncertainty from choosing and fitting the clusters. SF entering 2024 retains its broad assignment in all 500 matched-label resamples; that stability is not a prediction of its injuries. Descriptive grouping labels do not apply identically to every member.

The unit is a participating player’s transition to the next scheduled team regular-season game. Eligibility requires positive index-game snaps, no current physical injury/condition listing, and public-report coverage for both team-games. The target is a new physical injury/condition report before the next game. Illness, rest and explicit non-injury reasons are excluded. A new listing is not necessarily new tissue damage.

Two fixed candidates—logistic regression and histogram gradient boosting—train on 2017–2022. 2023 average precision selects gradient boosting; it is refit through 2023 and tested once on 2024–2025. Held-out n=41,858 with 3,795 positive appearances. AUC=.663, average precision=.151, Brier=.0802 versus .0824 for a constant-probability baseline.

Earlier-workload predictors exclude the index game itself, which could be shortened by injury. Prior report weeks can represent continuing conditions. Grouped permutation importance measures test average-precision loss after scrambling related columns. Bands are approximate game-cluster bootstrap intervals conditional on the fitted model and a fixed permutation; they omit training uncertainty and some repeated-player/team dependence. Correlated groups can substitute for each other or create unrealistic shuffled combinations. These scores are not additive causal shares.

Direct-to-IR cases, nonparticipants, camp injuries and off-roster outcomes can be missed. The outcome differs from AGL and cannot establish which feature explains SF’s total excess. Surface and schedule associations cannot reliably identify the time or place an injury occurred. Do not interpret low importance as refuting diagnosis-specific medical studies.

We parse 2017–2025 regular-season nflverse play-by-play descriptions for the explicit named-player phrase “was injured during the play.” Team attribution comes from that player’s team token, not possession. Multiple injured players on a play are retained, duplicate player/play mentions are removed, and the primary endpoint takes the first announcement per player-game. Quarter belongs to the play; return updates and nonspecific injury timeouts are excluded.

The extraction finds 7,454 first player-game announcements across 2,383 games. Regulation includes 286 SF and 7,116 other-team announcements; 52 overtime announcements (one SF) are excluded from half shares. First half is quarters 1–2; second half is quarters 3–4. The SF-minus-other second-half difference is −3.20 percentage points, with a 95% game-bootstrap interval of −8.69 to +2.20. The ratio of second/first-half announcement rates, SF relative to others, is .876 (95% interval .704–1.093).

Rates use 1,000 recorded team game-play opportunities, counting offense, defense and special-teams plays once per team in the game; eligible penalty/no-play records are included. They are not individual athlete-exposure rates. The bootstrap resamples whole games within season and SF-involvement strata 10,000 times, keeping SF opponents together. It does not fully model repeated players or changing recording practices. Sensitivities retain later announcements, restrict to scrimmage plays, remove each half’s final two minutes and split the era; none produces a larger SF second-half skew.

All 7,662 standardized source phrases matched the parser. An independent nearest-team/jersey check found no attribution mismatches; 72 seeded examples, spaced names and ambiguous nonstandard entries were audited. Missingness is demonstrable: in 2024, 119 of 609 player-games with return updates lacked any initial announcement. Public timestamps are not a complete injury-onset record, and this cannot decompose AGL or estimate injury severity. Selected source columns, extraction audits, hashes and reproduction scripts are included.

The downloads include row-level source URLs, input hashes, definitions, model coefficients, uncertainty and analysis scripts. Headline regressions, clustering summaries and held-out performance were independently recalculated before publication.

1. FTN: 2025 AGL and revised 2024 comparison
2. FTN: 2024 AGL and revised 2023 comparison
3. FTN: 2023 AGL and revised 2022 comparison
4. FO: COVID-excluded 2021 comparison
5. FO: 2020, without COVID
6. FO: 2019 and revised 2018
7. FO: original 2017 AGL
8. nflverse / PFR snap counts
9. nflverse schedules
10. nflverse injury reports
11. nflverse play-by-play descriptions and clocks
12. nflverse birth-date records
13. nflverse team statistics
14. Team marks: ESPN’s team-logo assets, identified through nflverse team metadata. Current franchise marks are used across historical seasons for consistent identification.