The Effect of Simulated College Admissions with AI Advice on Understanding and Strategic Behavior in China's Parallel Admission Mechanism: A Randomized Controlled Trial

2026-06-28

2026-07-04

Measured on each student's actual college application and admission (not the in-experiment game):

Admitted-college quality — the selectivity rank of the college the student is actually admitted to, measured by its 2025 admission cutoff-score provincial rank (primary), with average-score provincial rank as the pre-specified alternative. Lower rank = more selective. (Source: official admission records; available for all randomized students.)
First-choice aggressiveness — the selectivity rank of the college the student lists as their first choice, by the same 2025 cutoff-score (and average-score) ranking. Lower rank = a more ambitious/aggressive first choice. (Source: endline self-report.)
Admission choice-order (被第几志愿录取) — the position, within the student's submitted ROL, of the college they are admitted to. (Source: endline self-report / records.)
A pre-specified standardized index of the cutoff-rank measures is the lead summary endpoint for each construct; the average-rank versions are confirmatory robustness.

College-selectivity rankings. Each college in a student's application or admission record is assigned a selectivity rank by merging it to province-specific reference tables of 2025 admission scores. The primary measure is the college's 2025 cutoff-score provincial rank; the alternative is its 2025 average-score provincial rank. Rankings are computed separately for Shaanxi and Shandong (distinct exam/score distributions and ROL systems) and oriented so that a lower value denotes a more selective college. "Admitted-college quality" applies this to the college a student is admitted to; "first-choice aggressiveness" applies it to the college listed first on the student's ROL.
Admission choice-order (被第几志愿录取). The integer position, within the student's submitted rank-ordered list, of the college to which they are admitted (1 = first listed choice). Reassignment/non-admission cases are coded per a pre-specified rule documented in the analysis plan.
Match-efficiency / "高分低就" gap. If used, this is constructed by placing the student's own exam rank and their admitted college's 2025 cutoff rank on a common within-province percentile scale and taking the difference (own-rank percentile − admitted-college percentile); larger positive values indicate greater under-placement relative to ability.

We conduct a three-arm randomized controlled trial with high-school students who have completed the 2026 Gaokao and received their scores but have not yet submitted their college rank-ordered list (志愿). Participants are recruited from three high schools (two in Shaanxi, one in Shandong) during the window between score release and the application deadline.

Through an online application, all participants receive standardized information about how China's "parallel volunteer" (平行志愿) admission mechanism works and complete a short incentivized comprehension quiz. Participants are then randomly assigned, at the individual level and with equal probability, to one of three groups: (1) a control group that receives only the information and quiz; (2) a simulation group that additionally plays a six-round incentivized game simulating the parallel-volunteer admission process, with feedback after each round; and (3) an AI-advice group that plays the same simulation while also receiving AI-generated suggestions on each round. The study tests whether learning the mechanism experientially (simulation) and with AI assistance improves students' real college applications and admission outcomes, relative to receiving information alone. Outcomes are measured from a follow-up survey and from official college-admission records after the application window closes. Participants receive a participation payment plus performance-based earnings.

Not available

randomization through oTree code

individual

No

900 students

900 students

300 control
300 simulation
300 AI

Effect sizes and outcome standard deviations are anchored to last year's two-arm pilot (n = 109; Simulation vs. Information). In that pilot the Simulation-vs-Control effect was ~0.20 SD on admitted-college quality (the binding, smallest effect) and ~0.45–0.49 SD on first-choice aggressiveness. We size the study off the smaller welfare effect. At the target N = 300 per arm, a pairwise comparison detects effects of ~0.23 SD (two-sided) or ~0.20 SD (one-sided). Three features of the pre-registered analysis improve effective power on the welfare endpoint beyond this pairwise benchmark: (i) the directional AI > simulation > control hypothesis licenses one-sided tests; (ii) the primary test of the ordering is a linear-trend contrast (control = 0, simulation = 1, AI = 2), roughly twice as efficient as a pairwise contrast — minimum detectable per-step effect ≈ 0.12 SD; and (iii) regression adjustment for exam score/ability (Lin-type, with interactions) removes a large share of admission-outcome variance. The behavioral endpoints (pilot effect ≈ 0.45 SD) are powered above 99% at 300/arm and retain ≥90% power even at a 50% endline response rate (≈150/arm). For reference, detecting the conservative welfare effect of 0.20 SD at 80% power in a single pairwise two-sided test would require ~393/arm; this is the case for relying on the trend contrast and one-sided directional tests as the primary inferential approach for admitted-college quality. (Pilot n = 109 implies wide confidence bands on these anchors; we adopt the conservative 0.20–0.25 SD range deliberately.)