Playing Through Pain — NBA Career Longevity

Abstract

NBA careers are shaped by talent, physical workload, and injury history, but it is not obvious how much each of these actually matters for how long a player lasts. This project looks at the predictors of NBA career length using three publicly available Kaggle datasets compiled from Basketball Reference and official NBA records: per-season player statistics covering 4,486 players from 1947–2026, a game-level box score dataset with about 1.67 million individual player records, and an injury log of roughly 37,600 injury list transactions with free-text injury descriptions spanning 1951–2023. We restrict the analysis to the modern NBA era (1980–2022) so that the data is reasonably complete.

We apply four modelling approaches. An Ordinary Least Squares regression models total seasons played as a function of early-career workload, injury history, entry age, and position. A logistic regression classifier predicts whether a player will reach a "long career" of ten or more seasons using only their first three years of data, and reaches a ROC-AUC of 0.84. Kaplan-Meier survival curves and a Cox Proportional Hazards model examine how injury type and entry age affect the hazard of a career ending. Finally, K-Means clustering groups players into distinct workload profiles and compares career longevity across clusters.

The main takeaway is that talent selection and survivorship bias drive most of what we observe. High early-career minute load is positively associated with career length, since high-usage players tend to be more talented and hold onto roster spots longer. Players who suffer severe injuries (ACL, Achilles) actually average longer careers for the same reason: only established players spend enough time on the court to pick up these injuries in the first place. Players who enter the league at 18–19, almost all of whom are elite lottery picks, average over three more seasons than players entering at 22–23. Together, these results show how hard it is to separate a causal workload effect from talent.

Introduction

Background and Motivation

An NBA career is one of the most physically demanding careers in professional sports. The average career lasts around four to five seasons, and only a handful of players manage productive careers past fifteen years. Figuring out what separates long-career players from short-career ones has real implications for team management (roster construction, load management decisions), player development (draft evaluation, minutes allocation for young players), and sports medicine (identifying which injury types carry the highest long-term risk).

There are two competing narratives in sports media. The first argues that heavy early-career workloads, especially for players who enter the league as teenagers, wear the body down and shorten careers. The second argues that player quality is the dominant factor: elite players both receive heavy minutes AND sustain long careers, so the workload-longevity relationship looks positive in the raw data even if the true causal effect is neutral or negative. Telling these two stories apart requires careful statistical modelling that controls for talent proxies.

In terms of motivation, we are both basketball fanatics, and love to play, watch, and talk basketball. When brainstorming project ideas we knew immediately that we wanted to revolve it around the sport, but were unsure which direction to take it in. We are both Clippers fans, and anyone who knows anything about basketball knows that Kawhi Leonard is one of the most injury-prone players in the NBA today. Because of this, we decided to do this project about injuries in basketball, and how they affect long-term careers. We were very pleased with our results, and learned a great deal throughout this process.

Research Questions

This project addresses four specific questions motivated by the proposal:

Does high early-career minute load (before age 25) shorten career length?
Which injury types (ACL, Achilles, back) carry the highest risk of career termination or lasting decline?
Do players entering the league at 18–19 have shorter careers than those entering at 21–23?
Can we predict long-career players (10+ seasons) from their first three years of data?

Methodology

3.1 Ordinary Least Squares Linear Regression

We model total career seasons Y as a linear function of a feature matrix X:

Y = Xβ + ε, ε ~ Normal(0, σ²I)

The OLS estimator minimises the residual sum of squares, yielding the closed-form solution:

β̂ = (XᵀX)⁻¹XᵀY

Model fit is assessed via the coefficient of determination R² = 1 − SS_res / SS_tot and root mean squared error RMSE = √( (1/n) Σ(yᵢ − ŷᵢ)² ). Residual plots are used to check the homoscedasticity and normality assumptions. The F-statistic tests the joint significance of all predictors. Inference uses a 80/20 train–test split.

3.2 Logistic Regression (Classification)

To predict a binary outcome (long career, Y=1, ≥10 seasons, vs short career, Y=0) we fit a logistic regression model:

P(Y=1 | X) = σ(Xβ) = 1 / (1 + e^−Xβ)

Parameters β are estimated by maximising the log-likelihood via gradient descent. Features are standardised before fitting so the coefficient magnitudes are comparable. Model evaluation uses:

Confusion matrix: counts of true/false positives and negatives.
ROC curve and AUC: the area under the receiver operating characteristic curve; AUC = 1 is perfect, AUC = 0.5 is no better than random.

3.3 Survival Analysis

Kaplan-Meier Estimator. The non-parametric survival function is estimated as:

Ŝ(t) = ∏_tᵢ≤t ( 1 − dᵢ/nᵢ )

where dᵢ is the number of career endings (events) at time tᵢ and nᵢ is the number of players still active just before tᵢ. Players still active at the end of the observation window (2022) are treated as right-censored. Group differences are tested with the log-rank test, whose statistic under H₀ follows a χ²₁ distribution.

Cox Proportional Hazards Model. The semi-parametric Cox model specifies:

h(t | X) = h₀(t) · exp(Xβ)

where h₀(t) is an unspecified baseline hazard. The hazard ratio exp(β_j) gives the multiplicative change in retirement hazard for a one-unit increase in predictor j. Parameters are estimated via partial likelihood, which eliminates the nuisance baseline hazard. The penalizer λ = 0.1 provides mild L₂ regularisation to improve numerical stability.

3.4 K-Means Clustering

K-Means partitions n players into K clusters by minimising the within-cluster sum of squared Euclidean distances:

min_{Cₖ} Σₖ Σ_x∈Cₖ ‖x − μₖ‖²

where μₖ is the centroid of cluster Cₖ. Lloyd's algorithm alternates between assigning each point to its nearest centroid and recomputing centroids until convergence. Features are standardised to unit variance before clustering. The optimal K is chosen with the elbow method: we plot within-cluster inertia against K and pick the value where the marginal gains start to level off. Career longevity is then compared across clusters to see whether workload archetype is associated with career length.

Data

4.1 Data Sources and Sample Description

We use three publicly available datasets from Kaggle, all compiled from Basketball Reference and official NBA records:

Data Sources
Dataset	Rows	Key Variables	Source
NBA Player Per-Season Statistics	~33,000	Games, minutes, points, position, age, shooting splits	Basketball Reference via Kaggle
NBA Player Injury Log	~37,600	Date, team, player, free-text injury description	Basketball Reference via Kaggle
Advanced Per-Season Statistics	~33,000	BPM, VORP, usage%, win shares	Basketball Reference via Kaggle
Player Career Info	~5,400	Birthdate, debut, career span (from–to)	Basketball Reference via Kaggle

The unit of observation in the per-season files is a player-season pair. The injury log records each individual injury list (IL) placement transaction. We restrict the analysis to the modern NBA era (1980–2022) to avoid the sparse pre-IL period identified in the proposal. A season only counts toward career length if the player appeared in at least 20 games, which filters out brief call-ups.

The resulting sample is a convenience sample of all NBA players active between 1980 and 2022. Since we observe the entire population of active NBA players rather than a random draw from it, formal frequentist inference should be read as describing effect sizes within this population rather than generalising to some broader population.

4.2 Cleaning & Feature Engineering

Data cleaning, missingness handling, injury-type parsing from free text, and the full feature-engineering pipeline (early-career workload, injury aggregation, career-level rollups) are covered step by step in the notebook. — see full notebook →

4.3 Exploratory Data Analysis

The visualizations below look at the distribution of career lengths, differences across positions and eras, injury patterns, and how early-career workload relates to longevity.

Career length distribution and era comparison

fig. 1 Career length is heavily right-skewed — most players last only a few seasons — and median career length has fallen sharply since the 2000s.

Career length by position and entry age

fig. 2 Centers hold a slight longevity edge over forwards and guards; entry age shows a clear downward trend in career length.

Injury frequency by type and career impact

fig. 3 Ankle and knee injuries are the most common; players who sustain ACL and Achilles injuries average longer careers — a survivorship effect.

Early-career minute load vs career length

fig. 4 Players with heavier early-career minute loads tend toward longer careers, consistent with talent selection rather than physical wear.

Results

We apply the four modelling approaches in order. Each subsection presents the model fitting results, goes over performance metrics, and interprets the findings in the context of our four research questions.

5.1 Linear Regression: Predicting Total Career Seasons

We use OLS regression to model total career seasons as a linear function of early-career workload, injury history, entry age, and position. The statsmodels summary below provides coefficient estimates, standard errors, p-values, and the F-statistic. We do not standardise features here so that the coefficients stay interpretable in natural units (seasons per MPG, etc.).

                            OLS Regression Results
==============================================================================
Dep. Variable:         career_seasons   R-squared:                       0.447
Model:                            OLS   Adj. R-squared:                  0.445
Method:                 Least Squares   F-statistic:                     196.2
Prob (F-statistic):          2.41e-273   Log-Likelihood:                -5757.0
No. Observations:                2192   AIC:                         1.153e+04
Df Residuals:                    2182   BIC:                         1.159e+04
Df Model:                           9
Covariance Type:            nonrobust
=======================================================================================
                          coef    std err          t      P>|t|      [0.025      0.975]
---------------------------------------------------------------------------------------
const                  -0.1843      1.061     -0.174      0.862      -2.265       1.897
early_mpg               0.1075      0.013      8.166      0.000       0.082       0.133
early_games_per_szn     0.0959      0.006     15.781      0.000       0.084       0.108
entry_age              -0.0723      0.041     -1.745      0.081      -0.154       0.009
total_injuries          0.1231      0.017      7.205      0.000       0.090       0.157
severe_injury           0.6460      0.284      2.272      0.023       0.088       1.204
had_back_injury         0.5863      0.208      2.813      0.005       0.178       0.995
had_knee_injury         1.1116      0.182      6.123      0.000       0.756       1.468
pos_F                  -0.8296      0.246     -3.370      0.001      -1.312      -0.347
pos_G                  -1.0574      0.244     -4.338      0.000      -1.535      -0.579
========================================================================================

fig. 5 OLS coefficients (effect on career seasons): knee injuries, severe injury, and back injuries carry the largest positive association — position (guard/forward vs. center) the largest negative.

fig. 6 Residual plot: no strong pattern in residuals vs fitted values, though variance narrows for very short predicted careers.

Linear Regression Test-Set Performance
Metric	Value
Test RMSE (seasons)	3.076
Test R²	0.556

5.2 Classification: Predicting Long vs Short Career

We train a logistic regression classifier on first-three-year features to predict whether a player will reach a long career (≥10 seasons). Features are standardised before fitting, and the model is evaluated on the held-out 20% test set.

              precision    recall  f1-score   support

Short career       0.83      0.95      0.88       425
 Long career       0.64      0.33      0.44       124

    accuracy                           0.81       549
   macro avg       0.73      0.64      0.66       549
weighted avg       0.79      0.81      0.78       549

ROC-AUC: 0.841

fig. 7 Confusion matrix, ROC curve (AUC = 0.841), and standardised logistic regression coefficients.

5.3 Survival Analysis: Career Duration and Attrition Risk

Kaplan-Meier curves estimate the survival function (the probability of still being active after t seasons) for subgroups defined by injury type and entry age. Log-rank tests check whether the group differences are statistically significant. The Cox PH model then jointly estimates the effect of all predictors on the retirement hazard while controlling for confounders.

Players with to >= 2022 are treated as right-censored (still active at the end of the observation window). The event indicator career_ended = 1 marks confirmed retirements.

Kaplan-Meier: severe injury vs no severe injury

fig. 8 Players with a severe injury (ACL/Achilles) show no worse survival than uninjured players — consistent with survivorship bias.

Kaplan-Meier: entry age group

fig. 9 Players entering the league at 18–19 have markedly better survival than those entering at 22+.

Cox Proportional Hazards model

Cox Proportional Hazards — Coefficients
Feature	coef	exp(coef) [HR]	z	p
entry_age	0.12	1.13	10.67	<0.005
early_mpg	−0.05	0.95	−16.21	<0.005
early_bpm	−0.11	0.90	−14.32	<0.005
log_injuries	−0.38	0.68	−13.29	<0.005
severe_injury	0.00	1.00	0.06	0.95
had_back_injury	−0.16	0.86	−2.63	0.01
had_knee_injury	−0.12	0.89	−2.15	0.03

Concordance = 0.82 · 2,741 observations, 2,200 events observed · log-likelihood ratio test = 1,624.31 on 7 df

fig. 10 Cox PH coefficients with 95% CI. Bright markers cross zero on neither side (statistically significant); entry age is the only predictor associated with higher retirement hazard.

5.4 Unsupervised Learning: K-Means Workload Clustering

We apply K-Means to cluster players by early-career workload profile (MPG, games per season, PPG, entry age), using the elbow method to guide the choice of K. We then compare career longevity across the cluster profiles to see whether workload archetype is associated with career length.

fig. 11 Elbow plot of within-cluster inertia vs K, guiding the choice of K = 4.

Cluster Profiles — Mean Feature Values
	early_mpg	games/szn	early_ppg	entry_age	career_seasons
Cluster 0	13.19	48.31	4.39	26.83	3.08
Cluster 1	31.16	72.35	15.27	22.41	9.54
Cluster 2	10.44	38.25	3.50	23.09	3.01
Cluster 3	19.48	65.32	7.21	22.89	6.69

fig. 12 Mean career length by cluster. Cluster 1 (“heavy-usage stars”) averages the longest careers; Clusters 0 and 2 (low usage) are shortest-lived.

fig. 13 Full per-cluster breakdown: career-length distribution (left) and early-workload vs. career-length scatter (right) by cluster.

Conclusion

6.1 Summary of Findings

This project applied four complementary statistical methods to investigate the predictors of NBA career longevity using modern-era data (1980–2022). Here is what we found for each research question:

Q1: Does high early-career minute load shorten career length?

The OLS regression finds that early-career MPG is positively associated with career length (β ≈ +0.11, p < 0.001). This fits a talent-selection story: players who get heavy minutes early are disproportionately talented and hold onto roster spots longer. Nothing in the observed data suggests that heavy early usage mechanically shortens careers on average.

Q2: Which injury types carry the highest risk of career termination?

Counterintuitively, players who sustain ACL, Achilles, and back injuries average longer careers than uninjured players. This is a survivorship bias effect, since only high-usage star players are on the court long enough to pick up these injuries. The Cox PH model confirms that severe injuries carry no statistically significant additional retirement hazard once talent proxies (BPM, MPG) are controlled for. That does not mean these injuries are harmless; it means their career-ending effect is masked by the positive selection of who gets injured in the first place.

Q3: Do 18–19 year-old entrants have shorter careers than 21–23 entrants?

No. Players entering the NBA at 18–19 average roughly 3.3 more seasons than those entering at 22–23. This is driven by selection: players who enter the league as teenagers are almost exclusively lottery picks of exceptional talent. Players entering at 24+ have the shortest careers, which reflects their status as undrafted or late-round players with narrow roster windows.

Q4: Can we predict long-career players from first-three-year data?

Yes. The logistic regression reaches a ROC-AUC of 0.84, which indicates strong predictive power. Games played per season is the single most predictive feature, followed by BPM and injury history. In other words, early availability (staying healthy and on the floor) is a more reliable signal of long-term potential than raw statistical output.

6.2 Limitations

Survivorship bias: Players who exit the league early have fewer records, which biases injury and workload associations upward for longer-career players.
Talent confounding: Both playing time and career length are driven by talent, so observational regression alone cannot isolate a causal workload effect.
Injury data completeness: Historical injury list reporting was inconsistent before the formal IL system was fully established, so injury counts for pre-1990 players are likely underestimates.
Right-censoring: Players still active in 2022 are right-censored, meaning their eventual career lengths are unobserved.

6.3 Proposed Continuations

Instrumental variable regression using draft position as an instrument for playing time, to better separate the causal effect of early-career workload from talent confounding.
Random survival forests (a non-parametric survival method) would relax the proportional hazards assumption of the Cox model and could capture non-linear interactions between predictors.
Natural language processing on the injury notes to extract richer injury severity signals (for example, distinguishing a torn ACL from a partial tear, or one-off injuries from chronic conditions).
Panel data / mixed-effects models to account for within-player correlation across seasons and to model the trajectory of performance decline rather than just career-end events.
Comparison across leagues (European basketball, G-League) to test whether the findings hold beyond the NBA's particular selection mechanism.

References

Basketball Reference (2024). NBA & ABA Player Statistics. Sports Reference LLC. Retrieved from basketball-reference.com
Sumit Rodatta (2023). NBA Player Stats since 1950 [Dataset]. Kaggle.
P. Loria (2023). NBA Player Injury Stats (1951–2023) [Dataset]. Kaggle.
Kaplan, E.L. and Meier, P. (1958). Nonparametric estimation from incomplete observations. Journal of the American Statistical Association, 53(282), 457–481.
Cox, D.R. (1972). Regression models and life-tables (with discussion). Journal of the Royal Statistical Society: Series B, 34(2), 187–220.
Lloyd, S.P. (1982). Least squares quantization in PCM. IEEE Transactions on Information Theory, 28(2), 129–137.
Davidson-Pilon, C. (2019). lifelines: survival analysis in Python. Journal of Open Source Software, 4(40), 1317.
Pedregosa, F. et al. (2011). Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research, 12, 2825–2830.
Sekhon, J.S. (2008). The Neyman causal model for average treatment effects and what it implies about the regression model. Political Analysis, 16(4), 408–427.
Coupat, M. and Pilet, B. (2022). Survivorship bias in sports analytics: a framework for observational career-length studies. Journal of Sports Analytics, 8(3), 201–218.