Who is who: injury rates between players and within each player in professional team sports
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Who is who".
Jane: Here's the gist of the paper. Football injury incidence is conventionally reported per 1000 hours of exposure, and where recorded minutes track the workload under study, that division removes part of the association it should scale The denominator gradient exceeds the negligible threshold of 0.05 in all 15 leagues measured here Quantifying Denominator Defects The study aimed to quantify the two defects contaminating a recorded-minute denominator, outcome truncation and exposure-dependent playing time The first defect is that a same-day event may curtail the appearance it occurs in, which contaminates the denominator with the outcome The second defect is that a player with more recent minutes is likelier to start, which contaminates it with the exposure Assessing Outcome Truncation Outcome truncation was isolated by replacing each event row’s minutes with the non-event mean for its lineup role and refitting This defect was visible as a shifted distribution rather than a shifted mean, with the mean difference being −21.0 minutes (95% CI −23.1 to −19.0) across all appearances The study found that changing only the offset moved the seven-day estimate from a rate ratio of 1.27 to 1.09, an attenuation of 0.151 that divided away 14.0% of the rate ratio Measuring Exposure Dependence Exposure dependence was isolated by refitting all denominators in every lineup stratum with 1000 player-resampled intervals and by adjusting for lineup role rather than stratifying by it Within starters, the attenuation all but vanished, while among substitutes and among rows of unknown lineup status it was larger than either The gradient tracks how much room the offset has to move, with the measured sweep running from 0.94 where gamma is 0.001 to 2.01 where gamma is 0.303 League-Specific Recommendations The gradient exceeded the negligible threshold of 0.431 to 0.655 in men’s competitions and 0.444 to 0.647 in women’s Restricting to starters returned it below that threshold in eight of eight men’s leagues but five of seven women’s, with the bound reaching 0.058 in Sweden The recommendation is therefore to restrict to starters in 13 of the 15 leagues and report per appearance in the remaining two Recorded playing time is not a neutral denominator, and the gradient should be reported before any per-minute rate The quantity that decides whether this matters is the denominator gradient, which requires no injury data and exceeded the negligible threshold in all 15 leagues measured here Where the pooled gradient’s lower bound exceeds the threshold, a per-minute rate is not a neutral summary The remedy that works is restriction to starters, and it works because starters have almost no minutes to vary over — an interquartile range of two minutes — not because a role has been held fixed Where the within-starter gradient still reaches past the threshold, as in Spain and Sweden, the safer report is risk per appearance The distinction between a rate and a per-exposure risk is long established The gap this paper fills is still live: a 2025 treatment of injury-rate estimation models the count carefully and treats exposure purely as a standardising divisor, without asking whether it moves with the factors in the model Limitations and Context Public records do not meet surveillance consensus standards, and same-day records may reflect delayed coding, training events or unrelated absence The women’s panel is one season per league against a multi-season men’s window and comes from a different provider, so the populations are compared on mechanism rather than magnitude No analysis was pre-registered and every specification was chosen after inspecting the data, so intervals cover sampling uncertainty and not unmeasured misclassification or selection The context is not hypothetical: FIFPRO monitors load, travel and recovery for 300 women’s players across more than 30 leagues The gradient is a league-level property of squad rotation: it needs every appearance in a competition rather than a cohort of consenting clubs The context is not hypothetical: FIFPRO monitors load, travel and recovery for 300 women’s players across more than 30 domestic leagues The designs are complementary, and their investigators could run this check on their own records The context is not hypothetical: FIFPRO monitors load, travel and recovery for 300 women’s players across more than 30 domestic leagues Practical Applications
Paper summary: Tom: To wrap up "Who is who: injury rates between players and within each player in professional team sports," the paper really hammers home that recorded playing time isn't a neutral denominator when you're trying to estimate injury rates. It forces you to consider two specific problems they quantified: outcome truncation and exposure-dependent playing time >
Jane: And the key finding, which I think is what matters most for us, is measuring that denominator gradient across men's and women's leagues. They found that this gradient consistently exceeded the negligible threshold in all fifteen leagues they measured >
Lu: The conclusion is that a per-minute rate isn't a neutral summary unless you account for this gradient first because it requires no injury data to check, and it exceeded the negligible threshold in every league studied >
Meng: So for someone building an AI model, the implication is that using recorded playing time as your primary divisor without checking this gradient means your results are potentially biased by how players are selected or scheduled >
Jane: The remedy they propose is simple: restrict to starters. They found this works because starters have almost no minutes to vary over—an interquartile range of two minutes—which fixes the issue for many leagues >
Tom: But they also admit that even with that restriction, the gradient still goes past the threshold in places like Spain and Sweden, so the safer report there is risk per appearance >
Lu: That distinction between a rate and a per-exposure risk is what this work solidifies; it shows that we need to treat exposure differently depending on whether we're looking at a fixed role or just raw minutes >
Meng: For us shipping models, it means the choice of denominator isn't just about mathematical convenience; it’s about which underlying assumption you are making about player selection and workload variability >
Jane: So the main thing to remember is that if you're using recorded playing time, you need to know the denominator gradient first because it determines whether that summary statistic is actually neutral or not >
Conclusion: Tom: So we’ve gone through how they checked for those weird biases in the data, and now we’re looking at what this paper is actually claiming about injury rates between players and within each player.
Jane: Basically, they're saying that just using total minutes played isn't enough to get an accurate picture of how often people get hurt when you look at the numbers.
Tom: Right, the title itself says "Who is who," and it points to this problem with how we usually measure injury risk based on time spent on the field.
Lu: The core idea they’re pushing is that recorded playing time isn't a neutral way to divide by when you calculate an injury rate. They found two main ways the data got messed up: outcome truncation and exposure dependence.
Meng: Outcome truncation means sometimes an event happens, and it cuts off the record of how long the player *would* have played in that situation.
Jane: And exposure dependence is about players who play more recently being more likely to start, which messes with how much time they’re actually exposed to injury risk.
Tom: The big takeaway here is that you can't just use a simple per-minute rate unless you first check this denominator gradient.
Lu: Exactly. And they showed that this gradient is high in almost every league they studied, meaning the way you count minutes matters a lot.
Jane: So for someone who just listens to the show, it means that if you’re building an injury model using playing time, you need to be really careful about how you divide by exposure.
Tom: And what this changes for us is that we have to go beyond simple per-minute rates and look at whether the denominator itself is stable or not.
Jane: That opens up a whole new way of thinking about risk estimation in sports data. It makes you realize that the measurement method is part of the finding itself.
School of Computer Science and Statistics, Trinity College Dublin · Discipline of Anatomy, School of Medicine, Trinity College Dublin
stat.AP
Submitted: 2026-07-29
Updated: 2026-11-15
Comments: 14 pages, 2 figures, 1 table. Supplementary material included as an ancillary file. Observational analysis; not peer reviewed. Version 4 replaces the earlier denominator-gradient paper with a new paper on injury rates between players and within each player. Data and code: https://doi.org/10.5281/zenodo.24243290
Code: https://github.com/Gustolandia/football-denominator-gradient
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 87/100
The gist: Football injury incidence is conventionally reported per 1000 hours of exposure, and where recorded minutes track the workload under study…
Key concepts
- Outcome Truncation
- This defect occurs when an event happens during a recorded period, but only part of it is counted in the data. For example, if an injury occurs mid-session, the recorded minutes might not fully capture the risk associated with that specific time slot.
- Exposure Dependence
- This flaw means players who have more recent playing minutes are statistically more likely to be selected for future play. This biases the denominator by making it dependent on how recently a player has played, rather than just their role in the team.
- Denominator Gradient
- This measures how much the calculated rate changes based on the denominator used (e.g., minutes vs. appearances). If this gradient is high, using minutes as a denominator leads to inaccurate results because it is not a neutral measure of risk.
Terminology
Summary
Football injury incidence is conventionally reported per 1000 hours of exposure, and where recorded minutes track the workload under study, that division removes part of the association it should scale The denominator gradient exceeds the negligible threshold of 0. p. 005 in all 15 leagues measured here
Quantifying Denominator Defects
The study aimed to quantify the two defects contaminating a recorded-minute denominator, outcome truncation and exposure-dependent playing time The first defect is that a same-day event may curtail the appearance it occurs in, which contaminates the denominator with the outcome The second defect is that a player with more recent minutes is likelier to start, which contaminates it with the exposure
Assessing Outcome Truncation
Outcome truncation was isolated by replacing each event row’s minutes with the non-event mean for its lineup role and refitting This defect was visible as a shifted distribution rather than a shifted mean, with the mean difference being −21.0 minutes (95% CI −23. p. 01 to −19.0) across all appearances The study found that changing only the offset moved the seven-day estimate from a rate ratio of 1. p. 027 to 1.09, an attenuation of 0. p. 0151 that divided away 14. p. 00% of the rate ratio
Measuring Exposure Dependence
Exposure dependence was isolated by refitting all denominators in every lineup stratum with 1000 player-resampled intervals and by adjusting for lineup role rather than stratifying by it Within starters, the attenuation all but vanished, while among substitutes and among rows of unknown lineup status it was larger than either The gradient tracks how much room the offset has to move, with the measured sweep running from 0.94 where gamma is 0.001 to 2.01 where gamma is 0.303
League-Specific Recommendations
The gradient exceeded the negligible threshold of 0.431 to 0.655 in men’s competitions and 0. p. 0444 to 0.647 in women’s Restricting to starters returned it below that threshold in eight of eight men’s leagues but five of seven women’s, with the bound reaching 0. p. 0058 in Sweden The recommendation is therefore to restrict to starters in 13 of the 15 leagues and report per appearance in the remaining two
Conclusion on Denominator Use
Recorded playing time is not a neutral denominator, and the gradient should be reported before any per-minute rate The quantity that decides whether this matters is the denominator gradient, which requires no injury data and exceeded the negligible threshold in all 15 leagues measured here Where the pooled gradient’s lower bound exceeds the threshold, a per-minute rate is not a neutral summary The remedy that works is restriction to starters, and it works because starters have almost no minutes to vary over — an interquartile range of two minutes — not because a role has been held fixed Where the within-starter gradient still reaches past the threshold, as in Spain and Sweden, the safer report is risk per appearance The distinction between a rate and a per-exposure risk is long established The gap this paper fills is still live: a 2025 treatment of injury-rate estimation models the count carefully and treats exposure purely as a standardising divisor, without asking whether it moves with the factors in the model
Limitations and Context
Public records do not meet surveillance consensus standards, and same-day records may reflect delayed coding, training events or unrelated absence The women’s panel is one season per league against a multi-season men’s window and comes from a different provider, so the populations are compared on mechanism rather than magnitude No analysis was pre-registered and every specification was chosen after inspecting the data, so intervals cover sampling uncertainty and not unmeasured misclassification or selection The context is not hypothetical: FIFPRO monitors load, travel and recovery for 300 women’s players across more than 30 leagues The gradient is a league-level property of squad rotation: it needs every appearance in a competition rather than a cohort of consenting clubs The context is not hypothetical: FIFPRO monitors load, travel and recovery for 300 women’s players across more than 30 domestic leagues The designs are complementary, and their investigators could run this check on their own records The context is not hypothetical: FIFPRO monitors load, travel and recovery for 300 women’s players across more than 30 domestic leagues
Practical Applications
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
1
No longer assume that dividing football injury counts by recorded playing time is a neutral rescaling, as it is not; instead, AI systems should first fit the denominator gradient using only appearance records to determine if the per-minute rate is attenuated.
2
AI systems can perform a check before any outcome model by fitting the gradient: log recorded minutes on the workload under study, from appearance records alone — no injury data required.
If this lower bound exceeds 0.05, the system should report that the per-minute rate is attenuated.
3
When a gradient is significant, AI systems should restrict analysis to starters: Where it does, restrict the analysis to regular starters and confirm that the within-starter gradient falls below the threshold in your own data.
If restriction cannot be made due to missing lineup data, the system should report risk per appearance instead.
Sources
Related papers
- Unraveling time-varying causal effects of multiple exposures: integrating Functional Data Analysis with Multivariable Mendelian Randomization
- Strategic Causal Policy Learning with Manipulable Eligibility
- BayesJudge: Uncertainty-Aware Bayesian Meta-Evaluation of Human and LLM Judgments
- Automating structural reliability analysis with a multi-agent large language model framework
- Improving Minority Population Sampling with BISG Probabilities: Evidence from a Survey of Jewish Americans
- When Do Sports Leagues Stop Being Competitive? Bayesian Detection of Competitive-Balance Eras from Match-Level Data