科研 Skill

抽样方法与样本量把关

alterlab-ssci-sampling-gate

Checks that the sampling frame, method, and size match the inference paradigm before data collection. Use when planning recruitment or judging sample generalization.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

Sampling Gate — The Sample Decides Who the Answer Is About

Skill type: DISCIPLINE-ENFORCING. Before a single case is collected, this gate fixes three things — the frame, the method, and the size logic — and ties the generalization claim to them. It does not run the study or pick the test; it refuses to let sizing and recruitment happen by habit.

The Core Rule

WHO YOU SAMPLE AND HOW MANY MUST FOLLOW THE INFERENCE YOU WANT —
POWER FOR TESTS, PRECISION FOR ESTIMATES, SATURATION FOR THEORY.
A BIGGER N NEVER FIXES A BROKEN FRAME.

Sample size is not one calculation; it is whichever logic matches the inference. A hypothesis test needs an a-priori power analysis (effect size + alpha + power ⇒ N). An estimation goal needs a precision target (a margin of error at a confidence level). A qualitative study is governed by saturation / information power, not a formula. And none of these matters if the frame — the list you actually draw from — omits or over-represents part of the target population. Coverage error and self-selection are not cured by collecting more.

When to Use This Skill

  • "How many participants / interviews / respondents do I need?"
  • "Is my sample big enough to detect the effect?" (← a-priori power)
  • "I surveyed 300 students; can I generalize to all adults?" (← frame + method)
  • "My result was not significant — should I just collect more until it is?" (← no; pre-specify)
  • "What sampling method should I use — stratified, cluster, quota…?"

Does NOT Trigger

The request is really about…Route toWhy not this skill
Writing / wording the questionnaire itemsalterlab-survey-designInstrument construction, not sampling.
Which statistical test to run once data are inalterlab-test-selection-guardTest choice, downstream of sampling.
Executing the power analysis / analysis in softwarealterlab-statistical-analysisComputation; this gate sets the logic, hands execution off.
Deep grounded-theory / coding mechanicsalterlab-qualitative-methodsThis gate sets saturation logic; that skill does the qual depth.
Choosing the design & identifying assumptionalterlab-ssci-design-gateDesign routing, upstream of sampling.

Hand-off rule. When the request matches a row above, open your answer by naming that skill by its exact id (for example alterlab-survey-design) and saying in one line why it fits better, then stop applying this skill's procedure. Name it even when you also need to ask for missing input. If only part of the request is out of scope, name the skill for that part and do only the in-scope part. A brief general pointer is fine; the named skill does the work.

Three sizing logics — pick by the inference, not by habit

Inference goalSizing logicInputsCommon failure
Test a hypothesisa-priori power analysisexpected effect size, α, target power (usually .80/.90), testpowering off an inflated pilot effect; post-hoc "observed power"
Estimate a quantityprecision / margin of errordesired half-width, confidence level, expected variance/proportionreporting N with no CI target
Build/refine theory (qual)saturation / information powerscope, sample specificity, dialogue quality, analysis strategyquoting a fixed "N=12" as a rule instead of arguing information power

A stdlib calculator for the first two (two-group mean or two-proportion designs): scripts/sample_size.py. Full logic, stratified/cluster design effects, and finite-population correction: references/sampling_and_power.md.

Frame and method before size

  1. Frame vs target population. Name the target population, then the frame you can actually draw from. The gap between them is coverage error — state it. A sample of your university's students is not a sample of adults.
  2. Probability vs non-probability. Only a probability sample (each unit a known non-zero selection probability: SRS, stratified, cluster, systematic) supports statistical generalization with a sampling-error quantification. Non-probability samples (convenience, quota, snowball, purposive) support only analytical/theoretical generalization — say so and do not attach a margin of error as if it were random.
  3. Non-response & attrition. A high non-response rate reintroduces selection bias even from a good frame. Plan for it; report it; consider weighting.

Excuse vs Reality

ExcuseReality
"I'll collect as many as I can get."Convenience size is not power. Name the target effect and compute N — or state you are estimating, not testing.
"It wasn't significant, so I'll add data until it is."Optional stopping inflates the false-positive rate. Pre-register N (or a sequential design with corrected boundaries).
"300 is a big sample, so it generalizes."Generalization comes from the frame and method, not the count. 300 self-selected students still only speak for self-selected students.
"Observed (post-hoc) power shows the test was fine."Post-hoc power is a deterministic function of the p-value; it carries no new information. Use a-priori power.
"For qualitative work I'll do N=12 because that's standard."Saturation/information power is argued from scope and data quality, not a fixed number copied from another study.

The Design Passport (hand-off)

Append: target_population, sampling_frame (+ named coverage gap), sampling_method (probability/non-probability + specific design), size_logic (power | precision | saturation) with its inputs and resulting N (or saturation argument), expected_nonresponse, and generalization (statistical | analytical). alterlab-ssci-inference-gate later audits the final generalization claim against sampling_method + generalization.

Self-Check Before Advancing

  • Is the sizing logic the one that matches the inference (power / precision / saturation)?
  • For a test, is N from an a-priori power analysis on a defensible effect size — not a pilot's inflated one?
  • Is the frame stated, with its coverage gap from the target population named?
  • Is the sample probability or non-probability, and is the generalization claim scoped to match?
  • Is non-response/attrition planned for, not ignored?

References

  • references/sampling_and_power.md — sizing logics, design effects, finite-population correction, reporting templates.
  • scripts/sample_size.py — stdlib a-priori N for two-group mean (mean --d), two-proportion (prop --p1 --p2) and proportion-precision (precision --margin --p) designs. Inputs must be finite; alpha, power, p, p1, p2 strictly in (0, 1) with power > alpha; d ≠ 0 and p1 ≠ p2 (a zero effect has no finite N); margin in (0, 1) on the proportion scale (0.03 = ±3 points). Invalid input exits 2.

Part of the AlterLab Academic Skills suite.

示例提示词

帮我检查这项调查的抽样框和样本量是否撑得起要做的推论。

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

抽样方法与样本量把关能做什么?

Checks that the sampling frame, method, and size match the inference paradigm before data collection. Use when planning recruitment or judging sample generalization。

如何安装抽样方法与样本量把关?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「抽样方法与样本量把关」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。

抽样方法与样本量把关需要哪些权限?

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。

抽样方法与样本量把关经过官方审核吗?

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。