文献速递流行病学与生物统计组学、生物样本库与精准医学

对「使用Cosinor模型检测组学生物节律的实验设计与功效计算」评论的回复

Response to the Commentary on "Experimental Design and Power Calculation in Omics Circadian Rhythmicity Detection Using the Cosinor Model".
PubMed 文献 · Stat Med · 2025-01-01

作者认可Westermark博士对功效计算中非中心参数(ncp)的更正,承认推导中存在数学错误。在均匀采样设计下(如每2小时采样一次、共24小时、每个时间点4个生物学重复,12×4=48个样本),更正后与原文的非中心参数及统计功效差异为0;在均匀或接近均匀分布的设计下差异为0或接近0,而在中度至极度不均匀分布设计下差异变得越来越大,原错误ncp倾向于高估统计功效。作者已更新GitHub上的CircaPower R包并引用该评论文章。

要点

  • 作者承认原功效计算中非中心参数(ncp)的推导存在数学错误。
  • 均匀采样设计下更正前后非中心参数及统计功效差异为0。
  • 不均匀分布设计下原错误ncp倾向于高估统计功效。

核心结论

  • 作者认可Westermark博士对非中心参数(ncp)的更正,并承认原推导存在数学错误。
  • 在均匀采样设计下更正前后非中心参数及统计功效差异为0,在不均匀分布设计下原错误ncp倾向于高估统计功效。

研究设计与样本

  • 通过模拟研究考察不均匀采样设计下统计功效误算的程度,考虑24小时内的两种采样设计:均匀分布(Uniform)和单峰高斯分布(Unimodal),后者δ=2、4、6、8。
  • 模拟参数包括振幅噪声比r=A/σ=0.4、0.6、0.8和相位φ=0、3、6、9,并指出因对称性φ=12、15、18、21的结果与φ=0、3、6、9相同。

主要发现

  • 在均匀或接近均匀分布的设计(Uniform、Unimodal-sd6、Unimodal-sd8)下,更正后与原文的功效计算差异为0或接近0。
  • 在中度至极度不均匀分布的设计(Unimodal-sd2和Unimodal-sd4)下,差异变得越来越大,原错误ncp倾向于高估统计功效。
  • 在均匀采样设计示例(每2小时采样一次、共24小时、每个时间点4个生物学重复,12×4=48个样本)下,更正后与原文的非中心参数及统计功效差异为0。

关键数值与推荐

  • 均匀采样设计示例:每2小时采样一次、共24小时、每个时间点4个生物学重复,12×4=48个样本。
  • 模拟设置:振幅噪声比r=A/σ=0.4、0.6、0.8;相位φ=0、3、6、9;单峰高斯分布δ=2、4、6、8。

临床怎么用

  • 作者已更新GitHub上的CircaPower R包(https://github.com/circaPower/circaPower),突出更正内容并引用评论文章。

局限性

  • 原文未报告局限性。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

We agree and highly appreciate the correction of the non-central parameter (ncp) in the power calculation by Dr. Westermark. We acknowledge the mathematical mistake in our derivation. As indicated in the commentary, under the "evenly spaced sampling design" (e.g., sampling every 2 h for 24 h with four biological replicates at each time point; 12 × 4 = 48 $$ 12\times 4=48 $$ samples), discrepancy of the non-central parameter and the resulting statistical power between the corrected calculation and ours is 0. Below, we conduct further simulations to investigate the degree of statistical power miscalculation in "unevenly-spaced sampling design" in case potential users may have applied our formula in their experimental planning. We consider two types of sampling designs throughout the 24 h of a day: (i) Uniformly distributed (uniform): t i ∼ iid UNIF ( 0 , 24 ) $$ {t}_i\overset{iid}{\sim}\mathrm{UNIF}\left(0,24\right) $$ ; (ii) Unimodal Gaussian distributed (unimodal): t i ∼ iid N 0 , δ 2 $$ {t}_i\overset{iid}{\sim }N\left(0,{\delta}^2\right) $$ with δ = 2 , 4 , 6 , 8 $$ \delta =2,4,6,8 $$ . The following figure shows results for amplitude-to-noise ratio r = A / σ = 0.4 , 0.6 , 0.8 $$ r=A/\sigma =0.4,0.6,0.8 $$ and phase ϕ = 0 , 3 , 6 , 9 $$ \phi =0,3,6,9 $$ . Note that results of ϕ = 12 , 15 , 18 , 21 $$ \phi =12,15,18,21 $$ are identical to ϕ = 0 , 3 , 6 , 9 $$ \phi =0,3,6,9 $$ due to symmetry. As expected, for uniformly or almost uniformly distributed designs (i.e., Uniform, Unimodal-sd6, and Unimodal-sd8), the difference between the corrected and previous power calculations is 0 or close to 0. For moderate to extremely uneven distributed design (Unimodal-sd2 and Unimodal-sd4), the discrepancies become more and more substantial. The previous incorrect npc tends to inflate the estimated statistical power. We have updated the CircaPower R package on the Github ( https://github.com/circaPower/circaPower), highlighted the correction, and cited the commentary paper.

来源:Stat Med,查看原文。版权归原作者所有。

引用

  1. Yin R, Tseng GC. Response to the Commentary on "Experimental Design and Power Calculation in Omics Circadian Rhythmicity Detection Using the Cosinor Model". Stat Med. 2025. PMID: 40397012. doi:10.1002/sim.70079

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。