为临床试验从业者提供的易用样本量与检验效能计算工具。后台依托 R + rpact/gsDesign/TrialSize/PowerTOST 等 20+ 专业 R 包,自然语言驱动,支持 37 种检验。可应要求提供可复现 R 代码;默认英文输出,中文操作系统自动切换中文。 / Easy-to-use sample size and power calculation tool for clinical trial practitioners. Backed by R + 20+ professional R
临床试验样本大小和功率 / 临床试验样本量与检验效能计算专家是一项面向实际任务的技能,主要用于⚠️ 完成计算与解释;⚠️ → preview preview → → → → → → → → → → → → → → → → → → → → → → → → → → → 完成计算与解释;
实际使用前应先确认任务范围、数据来源、运行环境、必要权限和关键参数,再依据技能说明逐步执行;若输入条件不完整,应先补齐信息或采用保守配置,避免因错误假设导致结果偏离需求。执行过程中需要关注工具调用是否成功、接口或依赖是否可用、输出格式是否符合预期,并对异常提示、缺失字段和边界情况进行处理;
涉及批量任务时,还应保存进度,避免中断后重复操作。该技能适合用于一次性任务,也可以接入自动化工作流,与其他技能或上层代理配合完成更完整的业务链路;在组合使用时,应明确每一步的输入输出关系,并避免不同步骤之间出现参数冲突。
Auto detect → recommend → calculate & explain / 自动检测环境 → 推荐最优工具 → 完成计算与解释
⚠️ Safe by default — preview, not execute / 默认安全:仅预览、不执行: By default the skill generates R code and shows it in a SAFE PREVIEW (no execution). Execution requires an explicit opt-in: CLI
--yes/-y, or the agent adds--yeswhen it needs the actual number.--show-codeshows code;--dry-runis the default preview mode. / 技能默认生成 R 代码并以安全预览模式展示(不执行)。执行需显式确认:命令行用--yes/-y;Agent 在需要真实数值时显式加上--yes。--show-code展示代码;--dry-run即默认预览模式。Output language / 输出语言: Default = English for all user-facing content; auto-switch to Chinese when the OS environment is Chinese (locale contains
zh/CN, e.g.LANG=zh_CN.UTF-8or Windows UI languagezh-CN). Common modules must prepare BOTH English & Chinese prompt content; complex/rarely-used modules may temporarily be English-only. This setting does not affect code output. / 默认全部英文,但当操作系统为中文环境时自动切换为中文(locale 含zh/CN,如LANG=zh_CN.UTF-8或 Windows UI 语言zh-CN)。常用模块须同时准备英文与中文两套提示内容;复杂/少用模块可暂只提供英文。此设置不影响代码输出。
Detailed bilingual rules (detection, module list) in
references/language_policy.md. Key: English by default; auto-switch to Chinese when OS locale is zh/CN; common modules prepare EN+ZH, complex/rare EN-only. Code output unaffected (always English). / 详细双语细则(检测方式、模块清单)见references/language_policy.md。要点:默认英文;检测到 OS 为中文环境(locale 含zh/CN)时自动切换中文;常用模块备英/中两套,复杂/少用模块暂只英文。代码输出不受影响(始终英文)。
~/.workbuddy/MEMORY.md):属统计分析类且已发布 ClawHub,故需双语。ttest_*, anova, proportion_*, odds_ratio, risk_ratio, roc, poisson, non_inferiority, superiority_margin, be_tost, equivalence, survival, ni_survival, cluster, dunnett + report template / quick menu / flag reference. / 常用模块(英/中双套):上述常用检验类型 + 报告模板/快速菜单/参数速查。group_sequential, adaptive, adaptive_simulate, mixed_model, bayesian, win_ratio, historical_controls, assurance, conditional_power, dose_escalation, vaccine_efficacy, mams, survival_exact, mediation etc. / 复杂/少用模块(暂只英文):上述复杂/少用检验类型。This skill provides clinical trial researchers with an easy-to-use, comprehensive sample size & power calculation tool. Powered by R and 20+ professional R packages (rpact, gsDesign, TrialSize, PowerTOST, etc.), users can perform 37 complex calculations through natural language prompts — English by default, auto-Chinese on Chinese-OS (locale zh/CN). Reproducible R code is available on request. / 本技能为临床试验从业人员提供一整套简单易用的样本量与检验效能计算工具。后台以 R 软件及 rpact/gsDesign/TrialSize/PowerTOST 等 20+ R 工具包为依托,用户只需使用自然语言对话方式的提示词,即可完成(默认英文输出,OS 中文环境时自动切换中文)37 种复杂专业的样本量与检验效能计算工作。可复现代码按需提供(无 R 时附 Python 代码)。
Beyond basic "solve n given target power", the skill offers three advanced capabilities covering the most common analysis needs in the design phase:
除基础的"给定目标效能 → 求解样本量"外,本技能还提供三项进阶能力,覆盖方案设计阶段最常见的分析需求:
| Capability 能力 | Description 说明 | Typical Scenario 典型场景 |
|---|---|---|
| ① Sample size ⇄ Power (bidirectional) / ① 样本量 ⇄ 检验效能(双向计算) | Solve n given target power, AND solve achievable power given fixed n. --power (forward) and --nobs (reverse) are mutually exclusive; covers all 37 types. / 除正向求解样本量外,支持给定固定样本量反解可达检验效能(power)。--power(正向)与 --nobs(反向)互斥,覆盖全部 37 种检验类型。 |
Sample size fixed, evaluate if power meets target / 样本量已定,评估把握度是否达标 |
| ② Power curve / ② 样本量变化曲线图 | Given a sample-size sequence, batch-compute and plot the Power curve (x=sample size, y=power), with a target-power reference line. / 给定一组样本量序列,批量计算并绘制 Power 曲线(横轴=样本量,纵轴=检验效能),直观呈现"样本量越大、效能越高"的走势,并叠加目标效能参考线。 | Sample-size sensitivity analysis, protocol reporting / 样本量敏感性分析、方案汇报 |
| ③ Sample-size curve / ③ 检验效能变化曲线图 | Given a power-target sequence, batch-compute and plot the sample-size curve (x=target power, y=required n). / 给定一组效能目标序列,批量计算并绘制 样本量曲线(横轴=目标效能,纵轴=所需样本量),一眼看出"要达到某把握度需多少样本"。 | Resource planning, feasibility assessment / 资源规划、可行性评估 |
"20,40,200" or auto-generated "20:20:200" (start:step:stop); multiple effect-size curves can be overlaid for sensitivity analysis, default PNG output with a data table. / ②③ 曲线模式支持两种序列输入:显式列表 "20,40,200" 或自动生成 "20:20:200"(起:步:止),并可叠加多条效应量曲线做敏感性分析,默认输出 PNG 并附数据表。references/cli_examples.md. / 详细命令参数、37 种检验示例与曲线支持的检验类型清单,分别见下方 Implementation / 实施 与 references/cli_examples.md。Full test-type cross-reference table and all CLI examples:
references/cli_examples.md(37 tests, bidirectional solve, curve mode, R package install). / 完整检验类型对照表与全部命令行示例见references/cli_examples.md(含 37 种检验、双向求解、曲线模式、R 包安装)。
How to choose / 如何选择: 1️⃣ endpoint type 2️⃣ hypothesis direction 3️⃣ design complexity; if unsure, look up the table in that reference file. / 1️⃣ 终点类型 2️⃣ 假设方向 3️⃣ 设计复杂度;不确定时从该参考文件查表。
Beyond the 37 analytic tests, --test adaptive_simulate runs a Monte-Carlo simulator to validate adaptive / group-sequential designs empirically: power, type I error, expected sample size, early-stop probabilities. Primary engine: a standalone, source()-able pure base-R function library (scripts/adaptive_sim.R) — source() it in R to call run_adaptive_sim() (one-shot with report) or the individual simulate_group_sequential() / simulate_adaptive_reestimate() / simulate_drop_the_loser() / optimize_power() functions directly. The CLI shows this source(...) + run_adaptive_sim(...) code in SAFE PREVIEW (like every other test) and executes it with --yes; it needs only base R (no extra packages). Fallback: when R is not installed, the skill automatically runs the equivalent pure-Python module scripts/adaptive_simulator.py so the user still gets results. Designs: group_sequential, adaptive_reestimate (promising-zone SSR with Cui-Hung-Wang statistic), drop_the_loser (multi-arm). Spending: obrien_fleming / pocock / power_family. Also supports --futility, --optimize (min-N search) and --visualize. Full guide: index.md / references/adaptive_simulator.md. / 除 37 种解析检验外,--test adaptive_simulate 运行蒙特卡洛仿真器经验验证自适应/成组序贯设计的功效、I 类错误、期望样本量与早停概率。主引擎:独立、可直接 source() 的纯 base-R 函数库(scripts/adaptive_sim.R)——在 R 中 source() 后可直接调用 run_adaptive_sim()(一键出报告)或各底层函数 simulate_group_sequential()/simulate_adaptive_reestimate()/simulate_drop_the_loser()/optimize_power()。CLI 默认以安全预览展示该 source(...) + run_adaptive_sim(...) 代码、用 --yes 执行,仅需 base R(无需额外 R 包)。备用:当本机未安装 R 时,技能自动改跑等价的纯 Python 模块 scripts/adaptive_simulator.py,保证仍能出结果。设计含 group_sequential、adaptive_reestimate(promising-zone + CHW 统计量)、drop_the_loser(多臂);消耗函数 obrien_fleming/pocock/power_family;支持 --futility、--optimize、--visualize。完整说明见 index.md / references/adaptive_simulator.md。
| Requirement 要求 | Details 详情 |
|---|---|
| R | ≥ 4.1.0 (install on demand, see references/r_packages_zh.md) / ≥ 4.1.0(按需安装,详见 references/r_packages_zh.md) |
| Python | ≥ 3.8 + statsmodels≥0.14.2, numpy≥1.24.3, scipy≥1.11.4. Used by the adaptive_simulate no-R fallback and --visualize (matplotlib≥3.4). R is the primary engine for adaptive_simulate. / 用于 adaptive_simulate 的无 R 备用路径与 --visualize(matplotlib≥3.4)。adaptive_simulate 的主引擎是 R。 |
--yes/-y to explicitly execute and compute / R 代码默认不执行,运行于安全预览模式(展示代码、不运行);使用 --yes/-y 才显式执行并计算This skill's behavior is explicitly disclosed here for auditing:
本技能的行为已在此显式披露,便于审计:
| Behavior 行为 | Description 说明 |
|---|---|
| Local process call / 本地进程调用 | Run DYNAMICALLY GENERATED (never raw user input) R code locally via subprocess.run([Rscript, '--vanilla', tmp]), timeout 300s, NO shell. Execution is OPT-IN: by default the code is only previewed; it runs only when --yes is passed. Every user string that reaches the R code is validated against a strict allowlist first, so it cannot break out of an R string (no RCE). / 在本地运行动态生成(绝非原始用户输入)的 R 代码,通过 subprocess.run([Rscript, '--vanilla', tmp]),超时 300s,无 shell。执行是显式确认的:默认仅预览代码,仅当传入 --yes 才运行。所有进入 R 代码的用户字符串都会先经严格白名单校验,无法逃逸出 R 字符串(杜绝 RCE)。 |
| R code source / R 代码来源 | All generated by the skill's built-in templates (scripts/r_templates/), no static .R files, no remote script download. By default the skill runs in SAFE PREVIEW — generated R code is shown but NOT executed; --yes executes, --show-code reveals the code. / 全部由本技能内置模板参数化生成(scripts/r_templates/),无静态 .R 文件、不下载远程脚本。默认运行于安全预览模式——生成的 R 代码会展示但不执行;--yes 才执行,--show-code 展示代码。 |
| Output sanitization / 输出脱敏 | R stdout/stderr passes through sanitize_output() to strip local absolute paths and truncate over-long content before display, avoiding environment leakage or content injection. / R 的 stdout/stderr 经 sanitize_output() 过滤本地绝对路径并截断超长内容后再展示,避免泄露环境信息或注入过长内容。 |
| Network access / 网络访问 | Default analysis is fully offline, zero network. The only network touchpoint is the optional R-package install: by default it only prints install.packages() commands, not executed; you must explicitly add --run-install to download & install from CRAN (supply-chain risk triggered only after user is informed). The permission manifest declares network: "optional" (used only by --run-install). The full install R code is printed before execution. / 默认计算全程离线、零联网。唯一联网点是可选的 R 包安装:默认仅打印 install.packages() 命令、不执行;须显式追加 --run-install 才会从 CRAN 下载安装(供应链风险由用户知情后触发)。权限清单已据此声明 network: "optional"(仅 --run-install 使用)。执行前会完整打印将要运行的安装 R 代码。 |
| Filesystem / 文件系统 | Writes only a temp R script in the skill dir / system temp, discarded after use; never reads/writes user data files. / 仅在技能目录/系统临时目录写入临时 R 脚本,用后即弃;不读写用户数据文件。 |
Default = SAFE PREVIEW (dry-run): generate & show R code, do NOT execute. Add
--yes/-yto execute and compute the result;--show-codealso shows the code during execution. / 默认安全预览(dry-run):生成并展示 R 代码,但不执行。追加--yes/-y才执行并计算;--show-code在执行时也会展示代码。 Data format guide / 数据格式指南:references/data_format_guide.md
All 37 test CLI examples / 全部 37 种检验的命令行示例 see
references/cli_examples.md→ Implementation Examples.
Default --power (or omitted) is forward: given target power → solve required sample size n. Pass --nobs N to switch to reverse: given sample size → solve achievable power. --power and --nobs are mutually exclusive; when both are passed, --nobs wins.
默认 --power(或省略)为正向:给定目标效能 → 求解所需样本量 n。传入 --nobs N 切换为反向:给定样本量 → 求解可达检验效能(power)。--power 与 --nobs 互斥,同时传入时以 --nobs 为准。
Bidirectional solving: omit --nobs (or set --power) to solve for n given target power; pass --nobs N to solve for achieved power given a fixed sample size. The two flags are mutually exclusive.
反向求解命令行示例见 references/cli_examples.md → Reverse Examples.
Covers all 37 test types / 覆盖全部 37 种检验类型. Reverse-solve strategy:
pwr.* (pwr.t.test(n=) auto-reverses), PowerTOST::power.TOST(n=), rpact::getPowerMeans/getPowerSurvival(n=) — exact. / pwr.*(pwr.t.test(n=) 等自动反解)、PowerTOST::power.TOST(n=)、rpact::getPowerMeans/getPowerSurvival(n=),精确反解。z_b via non-centrality, then power = pnorm(z_b). / 自写检验(ROC、Poisson、疫苗效力、多终点、贝叶斯、Win Ratio、MAMS 等)通过非中心参数逆推 z_b,再 power = pnorm(z_b)。bland_altman returns achievable CI half-width (precision, not power); dose_escalation is heuristic design (power N/A); conditional_power/assurance --nobs maps directly to planned/assurance sample size. / bland_altman 回报可达 CI 半宽(精度而非效能);dose_escalation 为启发式设计,效能不适用;conditional_power/assurance 的 --nobs 直接映射至计划样本量/保证样本量。Note: some types (e.g. roc needs --auc1/--effect, mixed_model needs --effect_name) require mandatory params in both directions; missing params returns a validation error (expected). / 注:部分类型(如 roc 需 --auc1/--effect、mixed_model 需 --effect_name)在双向均要求提供必需参数,缺失时返回参数校验错误(符合预期)。
Implementation architecture / 实现架构: All test algorithms are provided as pre-written R functions (
ss_*) inscripts/r_templates/; the main dispatcher only injects params and calls.format()— no scattered R code. Types that use R packages auto-fall back to analytic closed-form approximations (Schoenfeld / O'Brien-Fleming etc.) on package failure — stable, reproducible, zero crash. / 所有检验算法以预编写的 R 函数(ss_*)形式提供,存放于scripts/r_templates/;主分发脚本仅做参数注入与模板.format()调用,无散落 R 代码。含 R 包的类型在包调用失败时自动回退至解析闭式近似(Schoenfeld / O'Brien-Fleming 等),结果稳定可复现、零崩溃。Security hardening & version history / 安全加固与版本历史 see
CHANGELOG.md(v3.3.5 fix of ClawHub 9 findings, v3.3.8 permission-declaration fix). / 见CHANGELOG.md(含 v3.3.5 对 ClawHub 审计 9 findings 的整改、v3.3.8 权限声明修复)。
--side one|two (default two): test direction. one = one-sided (alternative="greater", expecting treatment better); two = two-sided. Affects t-test, proportion tests (proportion_one/proportion_two/proportion_paired/odds_ratio/risk_ratio) and the significance level / required n in curve mode. Proportion tests since v3.3.2 uniformly use R functions (ss_prop_one/ss_prop_two/ss_prop_paired/ss_or_rr) wired to alternative. / 检验方向。one = 单侧(alternative="greater",预期处理组更优);two = 双侧。影响 t 检验、比例类检验(proportion_one / proportion_two / proportion_paired / odds_ratio / risk_ratio)及曲线模式的显著性水平与所需样本量。比例类自 v3.3.2 起统一以 R 函数(ss_prop_one / ss_prop_two / ss_prop_paired / ss_or_rr)形式提供,并接 alternative。--sd FLOAT (optional): when provided, --effect is treated as the raw mean difference Δ (e.g. 0.5 mmol/L) and the script auto-computes Cohen's d = Δ / sd; when omitted, --effect is directly Cohen's d. Avoids manual conversion, fits the medical literature "difference + SD" reporting. / 提供时,--effect 视为原始均差 Δ(如 0.5 mmol/L),脚本自动折算 Cohen's d = Δ / sd;不提供时 --effect 直接作为 Cohen's d。避免手动换算,更贴合医学文献"差值 + 标准差"的报法。Example (two independent groups, one-sided, Δ=0.5, sd=0.8 → d=0.625) / 示例(两独立组、单侧、Δ=0.5、sd=0.8 → d=0.625):
python scripts/samplesize_power.py --test ttest_ind --side one --sd 0.8 --effect 0.5 --power 0.9 # → n = 45 / group (one-sided); two-sided → n = 53 / group
On top of bidirectional solving, supports batch curve plotting to visualize the sample-size ⇄ power relationship.
在双向求解基础上支持批量绘制曲线,直观展示样本量 ⇄ 检验效能关系。
--n_seq "20,40,200" or "20:20:200" (start:step:stop) → Power curve (x=sample size, y=power) / --n_seq "20,40,200" 或 "20:20:200"(起:步:止)→ Power 曲线(x=样本量, y=效能)--power_seq "0.8,0.9" → sample-size curve (x=power, y=sample size) / --power_seq "0.8,0.9" → 样本量曲线(x=效能, y=样本量)--plot_effects "0.3,0.5,0.8": overlay multiple effect-size curves (sensitivity analysis, some types); --out path.png sets output / 多效应量叠曲线(敏感性分析,部分类型支持);--out path.png 指定输出png()+plot()), no ggplot2 needed / 绘图用基础 R 图形(png()+plot()),无需 ggplot2references/cli_examples.md → Curve Mode Examples); other types print a clear "not yet covered" notice at runtime / 支持 22 种检验类型(详见 references/cli_examples.md → Curve Mode Examples);其余类型运行时会提示暂未覆盖For supply-chain safety, the skill does not auto-install R packages over the network: --install-all-packages only prints install.packages() commands (not executed); appending --run-install actually downloads & installs the 10 packages from CRAN. Full list: references/cli_examples.md → R 包安装 and references/r_packages_zh.md.
出于供应链安全,本技能不自动联网安装 R 包:--install-all-packages 仅打印 install.packages() 命令(不执行);追加 --run-install 才从 CRAN 下载安装 10 个包。完整清单见 references/cli_examples.md → R 包安装 及 references/r_packages_zh.md。
Formulas / 公式: references/formulas_zh.md | Full functions / 完整函数: references/extended_functions.md
| Scenario 场景 | Formula 公式 |
|---|---|
| Independent t | $n_1 = 2(frac{Z_{1-alpha/2} + Z_{1-beta}}{d})^2$ |
| Proportion (arcsin) | $n = 2(frac{Z_{1-alpha/2} + Z_{1-beta}}{h})^2$ |
| Survival (Schoenfeld) | $d = frac{(Z_{1-alpha/2} + Z_{1-beta})^2}{(log HR)^2}$ |
| Cluster DEFF | $DEFF = 1 + (m - 1) times ICC$ |
| Error 错误 | Fix 修复 |
|---|---|
| "Rscript not found" | Install R or specify path / 安装 R 或指定路径 |
| "package not found" | install.packages("xxx") / 安装缺失包 |
| ImportError: statsmodels | pip install statsmodels / 安装 statsmodels |
| simr timeout | Reduce --nsim / 减小 --nsim |
Version: v3.4.4 | Updated: 2026-07-18 | License: MIT
相关专题
热门下载
相关下载
精品课程
共6课时 | 54.7万人学习
共89课时 | 134.1万人学习
共49课时 | 82.6万人学习