科研技能库/统计分析与质量控制
数据分析
未发现用户侧风险

统计分析与质量控制

统计分析与质量控制

文件预览

1 个文件
SKILL.md
2.2 KB · 可预览
---
name: statistics
description: # Statistical Analysis & Quality Control
---

# Statistical Analysis & Quality Control

## Overview
Comprehensive statistical methodology for scientific research. Covers test selection, assumption verification, power analysis, effect size reporting, and reporting standards.

## Test Selection Guide

| Data Type | Groups | Paired? | Normal? | Recommended Test |
|-----------|--------|---------|---------|-----------------|
| Continuous | 2 | No | Yes | Independent t-test |
| Continuous | 2 | No | No | Mann-Whitney U |
| Continuous | 2 | Yes | Yes | Paired t-test |
| Continuous | 2 | Yes | No | Wilcoxon signed-rank |
| Continuous | 3+ | No | Yes | One-way ANOVA + post hoc |
| Continuous | 3+ | No | No | Kruskal-Wallis + Dunn |
| Continuous | 3+ | Yes | Yes | Repeated measures ANOVA |
| Categorical | 2x2 | — | — | Chi-square / Fisher's exact |
| Time-to-event | 2+ | — | — | Log-rank + KM curves |
| Time-to-event | Adjusted | — | — | Cox proportional hazards |
| Continuous | Prediction | — | — | Linear/logistic regression |

## Assumption Checks
- **Normality**: Shapiro-Wilk (n < 50), Kolmogorov-Smirnov (n > 50), Q-Q plot visual
- **Homoscedasticity**: Levene's test, Bartlett's test
- **Independence**: study design review (not a statistical test)
- **Proportional hazards**: Schoenfeld residuals, log-log plot

## Multiple Comparison Correction
- **Bonferroni**: conservative, for few comparisons
- **Holm-Bonferroni**: step-down, less conservative than Bonferroni
- **FDR (Benjamini-Hochberg)**: for many comparisons (e.g., genomics)
- **Tukey HSD**: for all pairwise comparisons after ANOVA

## Effect Size Guidelines
| Measure | Small | Medium | Large |
|---------|-------|--------|-------|
| Cohen's d | 0.2 | 0.5 | 0.8 |
| Pearson r | 0.1 | 0.3 | 0.5 |
| Odds Ratio | 1.5 | 2.5 | 4.3 |
| R-squared | 0.02 | 0.13 | 0.26 |

## Reporting Standards
- Always report: test statistic, degrees of freedom, exact p-value, effect size, 95% CI
- Follow STROBE (observational), CONSORT (RCTs), PRISMA (reviews), ARRIVE (animal)
- Never say "trend toward significance" for p > 0.05
- Report non-significant results honestly

SKILL.md

元数据
namestatistics
description# 统计分析与质量控制

统计分析与质量控制

概述

科学研究的综合统计方法。涵盖检验选择、假设验证、功效分析、效应量报告和报告标准。

检验选择指南

数据类型组数是否配对?是否正态?推荐检验
连续2否是独立样本 t 检验
连续2否否Mann-Whitney U 检验
连续2是是配对 t 检验
连续2是否Wilcoxon 符号秩检验
连续3+否是单因素方差分析 + 事后检验
连续3+否否Kruskal-Wallis 检验 + Dunn 事后检验
连续3+是是重复测量方差分析
分类2x2——卡方检验 / Fisher 精确检验
时间-事件2+——对数秩检验 + Kaplan-Meier 曲线
时间-事件调整——Cox 比例风险回归
连续预测——线性/逻辑回归

假设检验

  • 正态性:Shapiro-Wilk 检验(n < 50),Kolmogorov-Smirnov 检验(n > 50),Q-Q 图可视化
  • 方差齐性:Levene 检验,Bartlett 检验
  • 独立性:研究设计审查(非统计检验)
  • 比例风险假设:Schoenfeld 残差,log-log 图

多重比较校正

  • Bonferroni:保守,适用于少量比较
  • Holm-Bonferroni:逐步降低,比 Bonferroni 保守性低
  • FDR (Benjamini-Hochberg):适用于大量比较(如基因组学)
  • Tukey HSD:用于方差分析后的所有配对比较

效应量指南

测量指标小中大
Cohen's d0.20.50.8
Pearson r0.10.30.5
比值比1.52.54.3
R-squared0.020.130.26

报告标准

  • 始终报告:检验统计量、自由度、精确 p 值、效应量、95% 置信区间
  • 遵循 STROBE(观察性研究)、CONSORT(随机对照试验)、PRISMA(系统综述)、ARRIVE(动物实验)指南
  • 切勿对 p > 0.05 使用“趋向显著”的说法
  • 如实报告不显著的结果