The paper proposes the Maximum Rule Concentration Index (MRCI) to measure how parsimoniously menu-dependent decision rules explain repeated binary lottery choices, provides an MIQP and a scalable greedy heuristic to compute it, and applies a permutation test to show that average subject-level MRCI ≈ 0.545 (Neff ≈ 1.85) with 64.1% of subjects rejecting a menu‑independent random benchmark at 1% — modal-payoff focusing (MAP) and salience (SAL) account for most concentration gains .
速读内容
Framework & Identification
The paper models binary risky choice as menu-specific transformations drawn from a finite rule library of 10 parameter-free rules (ID, MMn, MMx, MAP, PW, SAL, REG, DIS, A1, A2) and imposes Local Discriminability: at each menu the assigned rule must make the chosen alternative a strict First-Order Stochastic Dominance improvement in the perceived menu .
Key definition: MRCI = maximum Herfindahl–Hirschman index over all admissible menu-by-menu rule assignments; effective number of rules Neff = 1 / MRCI .
Computational approach & inference
Exact computation via MIQP is provided but scales poorly; the authors develop a greedy heuristic (precompute admissibility, greedy assignment by rule-order, random restarts) that matches MIQP closely in benchmarks (mean absolute error < 0.001; >96% within 1% tolerance) and runs in negligible time for full datasets .
For inference they construct a finite-sample permutation test under an exchangeability null (permute left/right choices conditional on observed left-count Z), which is finite-sample valid and consistent against Random Rule Models with excess latent concentration .
Empirical results — summary statistics
Statistic
Value
Subjects (N)
686
MRCI (mean)
0.545
MRCI (sd)
0.05
Neff (mean)
1.85
Share reject random (p ≤ 0.01)
0.641
Computed subject-by-subject on CPC18 risk problems (500–700 decisions per subject); average MRCI ≈ 0.545 implies Neff ≈ 1.85, i.e., most subjects behave as if using fewer than two effective rules .
Rule-importance diagnostics (Concentration Gain φ & Stability κ)
Concentration Gain (φ): MAP (modal-payoff focus) is the largest contributor (≈17% of attainable concentration on average), SAL ≈7%, REG & extremum heuristics (MMx/MMn) ~4–5% each; PW, ID negligible; DIS effectively non-operative in CPC18 under this operationalization .
Stability Score (κ): MAP is most often indispensable (~40% stability), REG and SAL also show substantial stability (≈25% and ≈20%), indicating substitution among active rules even when MAP drives average gains .
Heterogeneity: small but significant gender difference — women slightly lower MRCI (≈0.01) and use a more diverse rule mix; men slightly more concentrated on MAP/SAL .
Practical takeaways
MRCI provides an interpretable scalar of behavioral complexity (high MRCI → few unifying rules); pairing it with φ and κ identifies which rules unify choices vs. act as substitutes .
The heuristic + permutation test make the framework feasible for standard large repeated-choice datasets and allow testing whether observed concentration exceeds what menu‑independent randomness can produce .
思维导图
深度阅读
下面是对 “Revealed Decision Rules in Choices Under Risk” (Avner Seror,January 2026)一文的详尽、逐节解构性分析。我的目标是解释每一项重要论点、数据点、假设与结论,逐图逐表地剖析其内容、意义与局限,并在涉及原文结论或推断处给出页码溯源标识(格式为 [page::页码])。文中所有主要图表也会以文中给出的相对路径嵌入以便直观核对。摘要性要点先行,然后按报告结构逐章细读并给出批判性评注与综合结论。
一、元数据与总体概览(引言与报告概览)
报告标题为 “Revealed Decision Rules in Choices Under Risk”,作者 Avner Seror,发布日期 January 2026,主题是关于二元风险选择中“菜单依赖感知”的识别与量化问题,方法学与实证均针对风险下决策规则库与浓缩度度量展开。
Concentration Gain φ(f):删除 f 后 MRCI 相对下降的比例,衡量规则作为“统一器”的边际贡献。
计算策略:MIQP 与启发式(Section 5)
问题为 NP-hard:最大化凸的 HHI over 离散 assignment;作者给出 MIQP 精确式(附录 D),但该方法对 T 增长敏感,计算不可扩展。故提出启发式:先计算可行矩阵 a_{t,f}(规则在菜单 t 是否严格可辨别),然后对规则按覆盖数(coverage)排序并进行贪心分配,最后多次随机重启并保留最大 HHI 的解(popularity-first + random restarts)。 ,
局限:表中“Share Single Rule (Neff 1) = 0” 需谨慎解读——并不能排除存在规则仅在某些子集主导,但在样本总体上没有一人完全由单一非注意规则覆盖。表中“Share Non-Random”使用了 B=500 排列近似,Monte Carlo 的数值稳定性取决于 B 与重启次数的选取。