科研 Skill

随机对照试验设计

rct-core-design

Designs randomized controlled trials: treatment assignment, control groups, and experimental structure for field experiments. Use when designing an RCT for causal impact evaluation.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

<skill_content>

RCTs answer "does X cause Y?" with minimal assumptions.

<mandatory_requirements>

</mandatory_requirements>

<thinking_process> When designing an RCT:

  1. Articulate research question and theory of change
  2. Identify outcomes and measurement strategy (primary vs. secondary)
  3. Conduct power analysis for required sample size
  4. Choose randomization level (individual vs. cluster) based on spillovers
  5. Design ethical treatment assignment (phase-in, lottery, encouragement)
  6. Prepare IRB protocol and obtain approval
  7. Register trial and file pre-analysis plan
  8. Implement with fidelity monitoring
  9. Plan for compliance, attrition, and threats to validity </thinking_process>

<implementation_pattern>

<code_template>

@app.cell
def power_analysis_rct():
    # Calculate required sample size for RCT design
    # Based on Duflo et al. (2007) power calculation framework

    import numpy as np
    from scipy.stats import norm

    # Parameters
    alpha = 0.05  # Significance level
    power = 0.80  # Statistical power
    mde = 0.25  # Minimum detectable effect (SD units)
    p = 0.50  # Proportion assigned to treatment

    # Calculate critical values
    z_alpha = norm.ppf(1 - alpha/2)  # Two-tailed test
    z_power = norm.ppf(power)

    # Basic sample size
    n_base = ((z_alpha + z_power)**2) / (p * (1-p) * mde**2)

    # Adjust for clustering if applicable
    icc = 0.05  # Intra-cluster correlation
    cluster_size = 30  # Average cluster size
    design_effect = 1 + (cluster_size - 1) * icc
    n_clusters = n_base * design_effect / cluster_size

    # Adjust for attrition
    attrition_rate = 0.20
    n_final = n_base * design_effect / (1 - attrition_rate)

    print(f"POWER ANALYSIS FOR RCT")
    print(f"Parameters: α={alpha}, power={power}, MDE={mde} SD")
    print(f"\nBase sample size: {int(np.ceil(n_base))}")
    print(f"Design effect: {design_effect:.2f}")
    print(f"Clusters needed: {int(np.ceil(n_clusters))}")
    print(f"Final N (with attrition): {int(np.ceil(n_final))}")

    return int(np.ceil(n_final)),

</code_template>

</implementation_pattern>

import pandas as pd
import numpy as np

# Context: Limited program slots, everyone gets treatment eventually
# Solution: Randomize TIMING rather than ACCESS

n_units = 300
n_cohorts = 3  # Roll out in 3 waves

# Assign units to cohorts randomly
cohorts = np.random.choice(range(n_cohorts), size=n_units)

# Cohort 1: Immediate (Year 1)
# Cohort 2: Delayed (Year 2)
# Cohort 3: Delayed (Year 3)

df = pd.DataFrame({
    'unit_id': range(n_units),
    'cohort': cohorts,
    'year_treated': cohorts + 1  # Everyone gets treatment
})

print("PHASE-IN / STEPPED WEDGE DESIGN")
print("=" * 50)
print(df.groupby('cohort').size())

print("\nEthical advantages:")
print("- No one permanently denied treatment")
print("- Addresses capacity constraints")
print("- Still enables causal inference")

print("\nAnalysis approach:")
print("- Compare early vs. late cohorts")
print("- Use DID or event study framework")
print("- Longer follow-up for early cohorts")

return df,
</code>
<lesson>
When direct denial is unethical: (1) Phase-in/stepped wedge (randomize timing), (2) Encouragement design (randomize encouragement, not access), (3) Lottery (when slots limited), (4) Oversubscription (randomize among eligible). Never deny established entitlements.
</lesson>
</example>

</examples>

<common_mistakes>

<mistake severity="critical">
  <what>Not conducting power analysis before starting</what>
  <consequence>Underpowered study wastes resources, fails to detect real effects, misleads policy</consequence>
  <prevention>Always run power calculations with realistic effect sizes and account for clustering/attrition</prevention>
</mistake>

<mistake severity="critical">
  <what>Starting data collection before IRB approval and registration</what>
  <consequence>Research misconduct, inability to publish, ethical violations</consequence>
  <prevention>IRB approval and trial registration are prerequisites, not afterthoughts</prevention>
</mistake>

<mistake severity="high">
  <what>Individual randomization when spillovers likely</what>
  <consequence>SUTVA violation, biased estimates, contaminated control group</consequence>
  <prevention>Use cluster randomization or buffer zones when spillovers expected</prevention>
</mistake>

<mistake severity="high">
  <what>Measuring outcomes before treatment fully delivered</what>
  <consequence>Premature measurement finds null effects when treatment hasn't had time to work</consequence>
  <prevention>Theory of change specifies timing - measure outcomes after sufficient exposure period</prevention>
</mistake>

<mistake severity="medium">
  <what>Not planning for non-compliance</what>
  <consequence>Surprised by low take-up, insufficient power for actual treatment received</consequence>
  <prevention>Pilot to estimate take-up, power for ITT given expected compliance, design encouragement</prevention>
</mistake>

</common_mistakes>

<interpretation_guide>

<design_checklist>
Before launching RCT:
- [ ] Research question clearly specified
- [ ] Theory of change documented
- [ ] Power analysis conducted (accounts for clustering, attrition)
- [ ] Outcomes and measurement plan specified
- [ ] Randomization method chosen (individual vs. cluster)
- [ ] Ethical design confirmed (no unjustified denial)
- [ ] IRB protocol submitted and approved
- [ ] Trial registered (AEA RCT Registry or equivalent)
- [ ] Pre-analysis plan filed
- [ ] Implementation fidelity plan created
- [ ] Compliance monitoring plan ready
- [ ] Attrition tracking procedures established
</design_checklist>

<when_not_to_use_rct>
Avoid RCTs when:
- Randomization is unethical (denying life-saving treatment)
- Spillovers are unavoidable and large (market equilibrium effects)
- Sample size insufficient for adequate power
- Treatment is national-level policy (no counterfactual)
- Cost vastly exceeds value of information gained
- Results won't inform decisions (no policy window)
</when_not_to_use_rct>

</interpretation_guide>

<references>
<paper>Duflo, E., Glennerster, R., & Kremer, M. (2007). Using randomization in development economics research: A toolkit. Handbook of Development Economics, 4, 3895-3962.</paper>
<paper>Glennerster, R., & Takavarasha, K. (2013). Running Randomized Evaluations: A Practical Guide. Princeton University Press.</paper>
<paper>Christensen, G., & Miguel, E. (2018). Transparency, reproducibility, and the credibility of economics research. Journal of Economic Literature, 56(3), 920-980.</paper>
<paper>Button, K.S., et al. (2013). Power failure: why small sample size undermines the reliability of neuroscience. Nature Reviews Neuroscience, 14(5), 365-376.</paper>
<resource>J-PAL Research Resources: https://www.povertyactionlab.org/research-resources</resource>
</references>

</skill_content>

示例提示词

我要设计一个田野实验,帮我确定处理分配和对照组设置方案。

所属插件

社会学、政治学与城市研究面向社会科学研究:问卷与访谈设计、质性编码、文本即数据、社会网络、ABM、比较政治与选举数据、RCT 与 DID、城市与区域分析。

安装该插件会一并获得同领域的 58 个 Skill / MCP。

同插件的其他科研 Skill

常见问题

随机对照试验设计能做什么?

Designs randomized controlled trials: treatment assignment, control groups, and experimental structure for field experiments. Use when designing an RCT for causal impact evaluation。

如何安装随机对照试验设计?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「随机对照试验设计」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。

随机对照试验设计需要哪些权限?

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。

随机对照试验设计经过官方审核吗?

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。