arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2404.00319stat.MEstat.AP
收藏

方向偏好置信区间

Direction Preferring Confidence Intervals

Tel Aviv University(特拉维夫大学);Brown University School of Public Health(布朗大学公共卫生学院)

Tzviel Frostig, Orestis A. Panagiotou, Yoav Benjamini, Ruth Heller

AI总结:

本文提出方向偏好置信区间,通过扩展Modified Pratt CIs在无参数选择和有参数选择两种场景下提升推断灵活性,并证明控制FCR的CIs在符号判定上功效更高且条件CIs能控制mFCR。

摘要:

置信区间(CIs)在统计分析中起着重要作用,为参数提供区间估计。在现代统计学中,选择性推断很常见,即仅突出显示某些参数。然而,这种选择性方法可能会使推断产生偏差,导致一些人主张使用CIs而非p值。为了增加置信区间的灵活性,我们引入了方向偏好CIs,使分析师能够专注于呈特定方向趋势的参数。我们在两种情况下提出了这类CIs:第一,当没有参数选择时;第二,对于涉及参数选择的情况,我们提供了方向偏好CIs的条件版本。这两种方法都建立在Modified Pratt CIs的基础上,后者依赖非等变接受域,以通过延长区间换取更好的符号排除。我们表明,对于从m > 1个初始关注参数中选出的参数,旨在控制错误覆盖率(FCR)的CIs与条件CIs相比,在确定符号方面具有更高的功效。我们还表明,条件置信区间在任何依赖关系下都能控制边际错误覆盖率(mFCR)。

英文摘要:

Confidence intervals (CIs) are fundamental in statistical analysis, providing interval estimates for unknown parameters. We introduce direction-preferring CIs that prioritize inference for parameters with a specific sign (e.g., positive sign). Unlike previous methods of constructing CIs, our intervals are based on the inversion of acceptance regions that differ for positive and negative parameters, to enhance the probability of determining the sign in the prespecified preferred direction while maintaining standard coverage. We apply our suggestion in the typical context that multiple parameters are considered, but only some of the constructed CIs are of interest, e.g., when CIs are reported only for the subset of parameters that are found to differ significantly from zero. This leads us to suggest direction-preferring CIs with false coverage rate (FCR) control, as well as conditional direction-preferring CIs that adjust for selection by conditioning on the selection event. We compare the two selection adjustment methods in theory and using numerical experiments. The procedures we develop are useful for inference on secondary outcomes in medical applications, a setting we highlight through two motivating examples.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑