AI 中文总结
研究诚实置信区间在DiD、IV、RD和因子模型设置中的局部渐近功效，其由偏差界与抽样率速率决定有三种情况，通过极小极大论证表明损失是诚实固有，标准偏差感知区间速率最优，模拟和实证应用说明情况并给出实际建议。
AI 中文摘要
在双重差分法（DiD）、工具变量法（IV）、断点回归法（RD）和因子模型设置中，针对不可检验偏差具有保守性的置信区间，即偏差感知或诚实置信区间，现已成为标准。本文刻画了它们所诱导检验的局部功效。功效由偏差界相对于抽样率的速率决定，分为三种情况：当界比标准误差消失得更快时，保守性渐近自由；当二者同阶时，会有一个有界的明确代价；当界占主导时，即在参数速率的典型情况下，诚实检验的局部功效为零，无法拒绝局部备择假设的概率趋近于1。一个极小极大论证表明这种损失是诚实本身固有的，而非任何特定构造的属性。没有诚实程序能恢复它，标准偏差感知区间在速率上是最优的。一般来说，任何宽度收缩不够快的置信区间在其描绘集合内部没有局部功效，在边界处至多有单侧功效。部分识别是该论证的极限情况。模拟和两个实证应用说明了这三种情况。实际建议是在报告偏差感知区间的同时，报告功效‘死区’的半宽度。
英文摘要
Confidence intervals that are conservative against an untestable bias, called bias-aware or honest, are now standard in DiD, IV, RD, and factor-model settings. This paper characterises the local power of the tests they induce. Power is governed by the rate of the bias bound relative to the sampling rate, giving three regimes: when the bound vanishes faster than the standard error, conservatism is asymptotically free; when the two are of the same order it costs a bounded, explicit amount; and when the bound dominates, the typical case at the parametric rate, the honest test has zero local power: it covers every local alternative with probability approaching one. A minimax argument shows this loss is intrinsic to honesty itself, not a property of any particular construction. No honest procedure recovers it, and the standard bias-aware interval is rate-optimal. Broadly, any confidence interval whose width fails to shrink fast enough has no local power in the interior of the set it traces out, and at best one-sided power at the boundary. Partial identification is the limiting case of this argument. Simulations and two empirical applications illustrate the regimes in which the loss binds. The practical recommendation is to report the power "dead zone" half-width alongside bias-aware intervals.
