Projects Cumulative Head Trauma in MMA
Cumulative Head Trauma in MMA
Does absorbing head strikes wear a fighter down, separately from ageing, selection, and the sport itself changing? A ten-study programme on 17,434 fighter-fights and 2,708 fighters, 1994–2026. It grew out of my 2023 masters project, which proposed the question as future work.
- Links
- Read the studies
- Stack
- Python · pandas · statsmodels · lifelines · SQLite · React
![Forest plot of four ratio-scale effects on a log axis, each a point estimate with its 95% interval. Within-fight KO hazard at 1.38 [1.32, 1.44], damage carried into the UFC at 1.22 [1.15, 1.29] and the recurrent-KO clock at 1.11 [1.04, 1.19] all clear the null of 1, in blue; head KO versus ground TKO sits on it at 0.98 [0.82, 1.18], in grey.](/_astro/head-trauma-mma-card.DjyZJsJF_Z25doWr.webp)
The short version
Every fighter in the UFC absorbs head strikes. The question is whether that dose wears them down over a career, and whether you can separate that from three things that move with it: fighters get older, the ones still fighting are the durable ones, and the sport in 2024 is not the sport of 1998.
I built a panel of 17,434 fighter-fights across 2,708 fighters, with every covariate computed only from fights before the one being predicted. KO losses are placed at their exact fight. Head-strike dose is normalised by era. Active careers are right-censored. Every estimate has a 95% interval, bootstrapped by resampling fighters rather than fights.
Some of what it shows:
- Each extra 10 head strikes absorbed in earlier rounds raises the odds of being knocked out in the current round by about 38% (OR 1.38, CI 1.32–1.44, across 38,690 rounds).
- After a first KO loss, a fighter’s win rate falls 6.4 percentage points more than after a first decision loss, in a matched difference-in-differences design (CI −10.2 to −2.8).
- After a career peak, rating declines about 2.4 Elo per year faster for every 100 head strikes absorbed before the peak, once age at peak, peak rating, fights at peak and division are controlled (CI −4.8 to −0.1).
- Judges’ margins appear to erode with dose. Control for age and that effect disappears. I report it as confounded by ageing, not as a finding.
What this is not. There are no concussion diagnoses, no medical suspensions, no imaging. Every outcome is a performance proxy. The findings are consistent with cumulative trauma. They don’t prove it.
My 2023 masters project, Mixed Martial Arts Victor Analysis, listed “tracking the total number of head blows absorbed by a fighter over time” as future work. This is that work.
How it works
All eleven headline estimates sit in one plot, on two scales because they are not all ratios.
![The full forest plot from the ten-study programme: eleven estimates, each a point with its 95% interval, split into a ratio-scale panel with the null at 1 on a log axis and a difference-scale panel with the null at 0. Intervals that exclude the null are drawn in blue, those that cross it in grey. On the ratio scale, the within-fight KO hazard is 1.38 [1.32, 1.44] per 10 head strikes absorbed in prior rounds over 38,690 rounds; damage carried into the UFC is 1.22 [1.15, 1.29] per pre-UFC KO/TKO loss on the record over 16,519 fights; the recurrent-KO clock is 1.11 [1.04, 1.19] per additional prior KO loss over 1,969 gaps; and head-KO versus ground-strikes TKO sits on the null at 0.98 [0.82, 1.18] over 4,275 fights. On the difference scale, career dose to KO-loss risk is +1.07 [−0.02, +2.16] percentage points and crosses the null; pre-peak dose to post-peak slide is −2.45 [−4.79, −0.11] Elo points per year over 506 retired fighters; a first KO loss moves win rate −6.4 [−10.2, −2.8] points and the chance of being knocked down +6.3 [+2.5, +10.5] points, both over 755 matched pairs; short rest to re-KO on return is −3.6 [−12.4, +7.6] and crosses the null; and the judge-margin dose gradient is −0.23 [−0.27, −0.18] judge points with no age control, becoming +0.02 [−0.08, +0.13] once age, opponent Elo and year are added. The footer reads 17,434 fighter-fights, 2,708 fighters, 2,823 KO/TKO losses, 38,690 fighter-rounds, 1994–2026, data to 2026-06-27.](/_astro/head-trauma-mma.72Ur7jUr_1Gag7I.webp)
The unit of analysis is one fighter in one fight. There are 17,434 of those rows across 2,708 fighters, and every covariate on a row is built only from fights that happened earlier. Nothing about a fight informs its own row. The within-fight study narrows the exposure further. It counts only head strikes absorbed in prior rounds, which keeps the finishing sequence out of its own dose.
Dose is era-adjusted. I express it as cumulative head strikes absorbed per 100, normalised so a strike landed in 1998 and one landed in 2024 are comparable units. The round-level study carries an era bucket as a control on top of that: 2005–09, 2010–14, 2015–19, 2020–26. Without the adjustment, a lot of what I would be measuring is the sport getting busier.
Careers that are still running are censored rather than guessed at. The post-peak study sets aside 707 of the 2,702 fighters with an identified peak because they are still active. The recurrent-KO study treats 1,577 unclosed gaps as censored against 1,246 closed ones. The event study censors first losses whose 24-month return window runs past the data horizon of 2026-06-27. Never returning is counted as an outcome of its own, not as missing data.
Uncertainty is resampled on fighters, not on rows. One fighter contributes many correlated observations, and a row-level bootstrap would report intervals narrower than the data supports. The bootstrap clusters on fighter id. Model-based estimates use clustered standard errors for the same reason. That means fight-clustered errors in the round study, fighter-clustered in the judging study, and HC3 with one observation per fighter after the peak.
The first-KO study needs a control group that is not simply “fighters who were never knocked out”, because those are the durable ones. Its control is each fighter’s first career loss by decision. Both arms start from the same zero-prior-losses baseline by construction, and no fighter appears in both. I match one to one, nearest neighbour without replacement, caliper 2.0 SD, on pre-fight age, Elo, UFC fight number and calendar year. That gives 755 matched pairs at a mean match distance of 0.26, with balance reported as standardised mean differences.
What the design cannot reach. There are no concussion diagnoses here, no medical suspensions, no imaging. Every outcome is a performance proxy. Dose is UFC-recorded landed head strikes and nothing else. A grazing jab and a flush hook count the same, with no force, location or timing attached. Sparring, amateur bouts and fights in other promotions are invisible. Damage from before the UFC enters only as bout counts and KO/TKO losses on the record.
The harder limit is that dose, age and experience accumulate together within a career. Their within-fighter correlation is 0.92. None of these designs fully separates them. The judging study is the worked example. Judges’ margins erode clearly with dose on its own, and the gradient disappears once age is in the model. I report that as confounded by ageing rather than attributing it to trauma.
The findings are consistent with cumulative trauma. They do not prove it.
The long version
This section is being written.