科研 Skill

聚类置换统计检验

neuro-cluster-stats-interpreter

Explains how to set up and correctly interpret cluster-based permutation tests and TFCE for ERP or time-frequency data. Use to avoid onset or localization overclaims.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

Cluster-based permutation tests: run and read them right

A cluster test controls the family-wise error for the hypothesis "the conditions do not differ anywhere". A significant cluster rejects that null. It does not say the effect starts at the cluster's first sample, is confined to its channels, or is absent elsewhere (Maris & Oostenveld 2007; Sassenhagen & Draschkow 2019).

1. Set up

  • Define the data dimensions (time, frequency, channels or vertices) and the adjacency (channel neighbours from the montage).
  • Choose the test statistic for the design: independent or paired t, F for several conditions; within-subject designs permute condition labels within subject (sign-flip for paired differences).
  • Cluster-forming threshold: commonly the t value for p < 0.05 two-sided. It is arbitrary and changes cluster shape; decide it before looking, or use TFCE (Smith & Nichols 2009; in MNE pass threshold=dict(start=0, step=0.2)), which removes the threshold choice.
  • Permutations: at least 1000 for a reported p; more for small p.
  • Restrict the analysis window a priori if you have a hypothesis; a smaller search space gains power honestly.

2. Report what the test supports

Say: "the conditions differed (cluster-based permutation test, p = X); the difference was most pronounced over channels ... between about T1 and T2". Do not say: "the effect started at T1", "the effect was located at channels ...", or "no effect at other channels".

For onset or latency claims, use a method built for it: jackknife-based onset estimation on averaged waveforms, fractional area latency, or bootstrap confidence intervals of onset.

3. Common mistakes

  • Running separate tests per channel or time window, then using clusters.
  • Changing the threshold or window after seeing results.
  • Reporting several clusters from one test as independent findings with their own exact extents.
  • Using cluster p-values to compare effect sizes between conditions.

Output

Design, statistic, adjacency, threshold or TFCE, permutations, window, p-values per cluster, and wording that stays within what the test shows.

示例提示词

帮我对这组ERP数据做聚类置换检验并正确解读显著性结果。

所属插件

神经科学面向神经科学研究:EEG/MEG、fMRI、BIDS、spike sorting、连接组、脑机接口解码评估、神经刺激安全与公开神经数据检索。

安装该插件会一并获得同领域的 45 个 Skill / MCP。

同插件的其他科研 Skill

常见问题

聚类置换统计检验能做什么?

Explains how to set up and correctly interpret cluster-based permutation tests and TFCE for ERP or time-frequency data. Use to avoid onset or localization overclaims。

如何安装聚类置换统计检验?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「聚类置换统计检验」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。

聚类置换统计检验需要哪些权限?

该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。

聚类置换统计检验经过官方审核吗?

是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。