Question 1
The NBA referee context offers three key advantages over traditional labor market settings:
1. Quasi-Random Assignment
- Referees are assigned to games independently of player characteristics
- Eliminates selection bias that plagues traditional labor market studies
- In traditional labor research, for example, the hiring market, employers and employees mutually select one another, resulting in an inseparable discriminatory effect.
- Evidence: Table I and Appendix show no correlation between referee race composition and team racial composition (p > 0.05)
2. Objective Performance Metrics
- Fouls, points, and other outcomes are clearly recorded
- Reduces omitted variable bias common in productivity measurements
- Enables placebo tests (e.g., free throw percentage as outcome unaffected by referee decisions)
These features address the core identification challenge in discrimination research: the violation of conditional independence assumption in observational data.
Question 2
There are two kinds of theories,
- Taste-based discrimination
- Statistical discrimination
Theory 1: Taste-based Discrimination
Core idea (Lecture 2, p.47): Decision-makers have intrinsic prejudice and experience disutility from interacting with certain groups.
NBA Application: Referee utility function:
where is the discrimination parameter.
Intuition:
- Referees favor own-race players or disfavor opposite-race players
- Operates through implicit biases (unconscious associations like “black = aggressive”) in split-second decisions
- Discrimination is unrelated to actual player behavior
Theory 2: Statistical Discrimination (Phelps, 1972; Arrow, 1973)
Core idea (Lecture 2, p.60): Decision-makers use group identity as a proxy for productivity due to imperfect information.
NBA Application: Referee’s expected foul propensity:
+ \beta \cdot \text{signal}_i$$ where $\beta = \frac{\sigma^2_g}{\sigma^2_g + \sigma^2_v}$ (Lecture 2, p.63). **Intuition**: - If referees believe one racial group is more aggressive (higher $\bar{p}$), they call more fouls against that group even for identical behavior ## Question 3 ### Key Distinction (Lecture 2, p.60) - **Statistical**: Discrimination due to imperfect information about individual productivity - **Taste-based**: Discrimination due to preferences/animus, unrelated to actual performance ### Evidence Supporting Taste-Based Interpretation #### 1. Controlling for Playing Style (Table IV, Column 2) - Authors control for 29 player style variables (height, position, historical stats) interacted with %white refs - Own-race bias remains: β = 0.203** (SE 0.072) - **Implication**: Discrimination is NOT due to different playing styles between Black and White players ![[截屏2026-02-02 13.14.54.png|400]] #### 2. Placebo Test: Free Throw Percentage (Table V) - Free throw shooting is unaffected by referee decisions - Result: β = 0.002 (p > 0.10) for FT%, but significant effects on fouls, points, turnovers - **Implication**: Effects operate through referee decisions, not player behavior changes ![[截屏2026-02-02 13.16.11.png|400]] #### 3. Ruling Out Employer Discrimination (p.1861) - NBA has strong monitoring: observers, video review, bi-weekly feedback - Referee performance affects playoff assignments and compensation - **Implication**: Discrimination stems from **individual referee implicit bias** #### 4. All-White vs All-Black Referee Crews (Figure II) - White referees (game-weighted avg): $\delta = +0.050$ (favor White players) - Black referees (game-weighted avg): $\delta= -0.009$ (favor Black players) - Only 9/28 Black refs have stronger pro-White bias than average White ref - **Implication**: If statistical discrimination, all refs should share same prior beliefs. Systematic own-race patterns support taste-based explanation ## Question 4 ### Why Random Assignment is Critical for Identification **Core identification concern**: To causally interpret the coefficient on (Black × %White refs), we need: $$E[\varepsilon_{it} | \text{Black}_i, \%\text{White refs}_g] = 0$$ **Threats if assignment is NOT random**: - White refs assigned to more competitive games → fouls naturally higher - Black refs assigned to home games with more Black players → confounds audience effects **Role of random assignment** (Topic1 randomization logic): - Ensures referee race orthogonal to other factors - Enables **causal interpretation** of $\beta_{1}$, not just correlation ### How Authors Test Random Assignment **Test logic** (analogous to Swedish lottery randomization test, Topic1): If referee assignment is random conditional on year fixed effects: $$\#\text{White referees}_g = \alpha + \gamma \cdot \text{Year}_g + \delta \cdot X_g + u_g$$ Should have $\delta=0$ (all covariate coefficients zero). **Empirical strategy** (Appendix table): 1. **Dependent variable**: # white referees per game (0-3) 2. **Covariates tested**: - black starters (home and away) - Attendance, out-of-contention indicator - Team fixed effects, year fixed effects 3. **Omnibus test**: Joint F-test of all covariates ### Interpretation of Results **Key findings from Appendix**: 1. **Individual coefficients**: All insignificant (p > 0.05) - black starters has no predictive power for referee race 2. **Omnibus F-tests**: - Column 1: p = 0.61 → Black starters jointly insignificant - Column 2: p = 0.63 → Still insignificant with game controls - Column 3: p = 0.89 → Still insignificant with team FE - Column 4: p = 1.00 → Fully random under strictest specification 3. **Adjusted R²**: Remains ~0.049-0.036 across specs - Covariates add negligible explanatory power beyond year FE ### Support for Identification Strategy 1. **Ruling out selection bias**: No evidence NBA systematically assigns white refs to teams with fewer Black players 2. **Supporting causal interpretation**: Balance test passes → %White refs is **exogenous** → Table IV's $\beta_{1}$ has causal meaning 3. **Ruling out confounders**: White refs not systematically assigned to specific game types (e.g., playoff contention games) ## Question 5: Replicating Table IV ### Why Estimate These Regression Models Table IV employs three progressively stringent specifications to identify causal own-race bias while ruling out confounders: #### Column 1: Baseline + Three Types of Fixed Effects **Motivation**: - **Player FE**: Controls time-invariant player characteristics (height, playing style) - **Referee FE**: Controls baseline strictness of each referee - **Year FE**: Controls rule changes (e.g., hand-checking enforcement) **Identification**: Exploits **within-player variation** — how same player's foul rate varies with referee race composition #### Column 2: Ruling Out "Playing Style" Confounds **Why needed**: Statistical discrimination critique — maybe White/Black refs react differently to playing styles that correlate with race **Solution**: Add 29 player style variables × %white refs interactions: - Physical: height, weight, position - Experience: age, NBA experience, all-star status - Playing style: historical assists, blocks, steals, shooting percentages per 48 min #### Column 3: Strictest Identification (Full FE) **Includes**: 1. **Player × Year FE**: Controls each player's yearly form changes 2. **Team × Game FE**: Controls ALL game-level factors (opponent strength, stakes, referee crew overall style) 3. **Stadium × Player Race**: Controls home court effects on different races **Identification source**: Only uses **within-game variation among teammates** - Among teammates in same game, does Black vs White player's foul rate difference vary with referee race composition? ### Interpretation of Results If $\beta_{1} = 0.197^{* *}$ (SE = 0.061), black players earn 0.197 more fouls per 48 minutes under all-White vs all-Black crew ## Question 6: Three Robustness Checks ### (a) Player "Style" Controls **Threat**: Statistical discrimination — maybe Black/White refs have different standards for playing styles that correlate with race, not race itself. **Example**: If Black players are more physical and White refs call physical play more strictly → observed effect is style×referee interaction, not racial bias. **Solution** (Table IV Column 2): Control for 29 style variables × %white refs: - Physical: height, weight, position - Performance: historical assists, blocks, shooting % **Result**: $\beta_{1} = 0.203^{* *}$ (unchanged) → **Rules out style explanation** ### (b) Referee Experience Controls **Threat**: Maybe Black/White refs differ in average experience, not race per se. **Mechanism**: If Black refs are younger/less experienced, and experience affects foul-calling → racial effect is disguised experience effect. **Test**: Add (Black × Avg Ref Experience) interaction. **Result**: Experience interaction ≈ 0 and insignificant → **Rules out experience confound** ### (c) Game Fixed Effects **Threat**: Game-level omitted variables simultaneously affect both referee composition and foul propensity. **Examples**: Game intensity, team-specific reactions, home crowd pressure. **Solution** (Table IV Column 3): Control **Team × Game FE**: - Absorbs ALL game-level factors - Identification only from **within-game variation among teammates** **Why powerful**: Compares Black teammate A vs White teammate B in same game - Same opponent, same stakes, same referee crew overall style - Any game-level confounder cannot explain results **Cost**: Larger SE (0.061→0.080) due to less identifying variation. ## Question 7 First we need to distinct those two differences: - **One-way discrimination**: Only White refs discriminate against Black players - **Own-race bias**: **Bidirectional** — White refs favor Whites, Black refs favor Blacks ### Evidence 1: Asymmetric Pattern in Table III **Observed pattern**: - **Black players**: Foul rate relatively stable (4.32-4.42) across referee compositions - **White players**: Foul rate declines (5.25→4.90) as %White refs increases **Why supports own-race bias**: - If only White refs discriminate → expect Black foul rate to rise - Actually: Black rate stable, White rate **falls** → White refs **favor White players** ### Evidence 2: Distribution in Figure II ![[截屏2026-02-02 13.46.40.png|400]] ### Why "Indirect"? Cannot establish "no-discrimination baseline" to separately identify White vs Black ref bias. > Unfortunately, our framework is not well suited to sorting out whether these results are driven by the actions of black or white referees, because this would require establishing a “no discrimination” baseline. ## Question 8 ### Main Findings **Full sample**: $\beta$ (Black × %White × HighAttend) = 0.111 (p=0.574) → **Insignificant** — Attendance has no significant effect on own-race bias **Home vs Away**: - **Away games**: $\beta$ = -0.090 (bias slightly weaker, possibly insignificant) - **Home games**: $\beta$ = +0.196 (bias stronger) ### Contrast with Parsons et al (2011) **Baseball (QuesTec monitoring)**: Scrutiny **eliminates** umpire bias **NBA (attendance)**: Scrutiny has **no significant effect**; may even **increase** bias in home games ### Possible Explanations 1. **Home crowd pressure mechanism**: - High attendance home games → stronger **group identity pressure** - Refs may favor groups aligned with home crowd preferences - Opposite of "rational monitoring" 2. **Stubbornness of implicit bias**: - Basketball calls more **subjective and rapid** (split-second decisions) - Harder to suppress via external pressure vs baseball judgments ## Question 9 1. **Referee diversity**: Increase Black ref % to ~80% to match players - Mechanism: Symmetric biases **cancel out** 2. **Implicit bias training**: IAT testing + high-pressure scenario training 3. **Tech assistance**: Expand instant replay, AI-aided calls ### Labor Market Lessons **External validity**: If bias persists among highly monitored NBA refs → Likely worse in police, judges, HR **Diversity policy**: Increase diversity to cancel symmetric biases, BUT combine with blind review ## Question 10 ### Timeline - **May 2007**: publishes findings → Stern calls research "flat-out wrong" - **November 2010**: QJE publishes paper → NBA still rejects ### NBA's Counter-Study (Segal Company) **Data advantage**: 148,000 calls (2004-2007), knows **which ref made each call** **Conclusion**: "No evidence of racial bias" ### Authors' Response (Price & Wolfers 2012) **After obtaining NBA data**: NBA used **wrong methodology** **Re-analysis with correct method**: Actually **confirms original findings** **Methodological defense**: Crew-level analysis sufficient — don't need individual call data to detect own-race bias **Key paper**: *"Biased Referees? Reconciling Results with NBA's Analysis"* (Contemporary Economic Policy, 2012) **Implication**: Awareness eliminates implicit bias + validates original findings