How Interview Order Shapes Graduate Exam Scores: A Simple Mathematical Model

Problem Context

Graduate‑school interview panels in China typically consist of five faculty members who score each candidate independently. Scores are averaged (or extreme values are discarded). A single interview session lasts from morning to afternoon and includes 10–20 candidates, each receiving 10–20 minutes of questioning.

Mathematical Model of the Observed Score

The observed score S_i for the i ‑th candidate is modeled as an additive combination of four components: S_i = T_i + B_i + F_i + \varepsilon_i T_i : the candidate’s true ability score, determined by knowledge and performance; independent of interview order.

B_i : a reference‑bias term that captures anchoring and contrast effects caused by judges’ memory of earlier candidates.

F_i : a fatigue term that reflects the change in judges’ discriminative power over the course of the day.

\varepsilon_i : random noise (e.g., variation in question difficulty, individual judge style).

Reference‑Bias Component

Judges form an implicit reference point from the average true ability of all previously interviewed candidates: \mu_{<em>prev</em>}(i) = \frac{1}{i-1}\sum_{j=1}^{i-1} T_j The bias applied to candidate i is proportional to the difference between the candidate’s true ability and this reference: B_i = \alpha\,(T_i - \mu_{<em>prev</em>}(i)) where \alpha (0 ≤ α ≤ 1) is the reference‑strength coefficient. The first candidate has no reference bias ( B_1 = 0). If T_i > \mu_{prev} (the candidate is stronger than earlier ones), the bias slightly lowers the score; if T_i < \mu_{prev}, the bias slightly raises it. The magnitude of the effect grows with larger \alpha and with a larger gap between the candidate and the prior average.

Fatigue Component

Judges’ attention level A(t) is modeled as a piecewise linear function of the normalized time t (0 = start of the session, 1 = end). The function includes a decay coefficient for the morning ( \beta_m), a decay coefficient for the afternoon ( \beta_a), and a short recovery boost \rho at the lunch break (approximately at t = 0.5).

A(t) = \begin{cases} 1 - \beta_m\,t, & 0 \le t < t_{\text{lunch}} \\ 1 - \beta_m\,t_{\text{lunch}} + \rho - \beta_a\,(t - t_{\text{lunch}}), & t_{\text{lunch}} \le t \le 1 \end{cases}

The fatigue term that modifies the discriminative power of the judges is then F_i = -\gamma\,(1 - A(t_i))\,(T_i - \bar{T}) where \gamma scales the impact of fatigue, t_i is the interview time of candidate i , and \bar{T} is the overall mean true ability of all candidates. When judges are fresh ( A≈1), the term is near zero and scores reflect true differences. As fatigue increases ( A declines), the term compresses scores toward the mean: strong candidates are slightly penalized, weak candidates are slightly boosted.

Combined Effects on Different Candidate Levels

Integrating the reference‑bias and fatigue terms yields distinct patterns for strong, medium, and weak candidates:

Strong candidates : Early slots give them the anchoring advantage (no reference pressure) and benefit from high judge energy, preserving their lead. In later slots their scores may be modestly compressed unless earlier candidates were unusually weak.

Medium candidates : Early slots expose them to stricter judges and more probing questions, increasing risk of score reduction. Later slots benefit from judge fatigue, which reduces scrutiny and can stabilize scores, though a strong early anchor may still depress their relative standing.

Weak candidates : Early slots are evaluated by fully attentive judges, making deficiencies obvious and scores low. Later slots may receive a small lift because fatigue reduces discriminability, but the overall score remains low.

Key Conclusions

The model predicts that order effects are real but modest compared with the intrinsic ability differences among candidates. Institutional reforms—such as mandatory video recording, sealed question pools, independent scoring, and uniform interview durations—have already reduced the magnitude of both reference bias and fatigue.

In marginal cases where two candidates have nearly identical true abilities, the combined bias can tip the ranking, but such situations are rare.

Practical Implications for Candidates

Early slot : Judges are fresh; present material clearly and confidently without over‑performing. You set the initial reference point.

Late slot : Judges may be fatigued; avoid mimicking earlier candidates and instead highlight personal insights to re‑engage attention.

Any slot : Focus on answering questions accurately; the reference bias is beyond your control.

Broader Relevance

The same instability of reference frames in small‑sample sequential evaluations appears in job interviews, performance judging, and academic defenses, where human evaluators experience fatigue and rely on relative comparisons. Mitigation strategies at the system level include multiple independent judging panels, blind evaluation, and dynamic score adjustments.