科研 Skill

社科研究设计把关

alterlab-ssci-design-gate

Routes a study to its research design (experiment, quasi-experiment, observational, qualitative) and pins the identifying assumption first. Use when choosing a study design.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

Design Gate — Pin the Identifying Assumption Before Anything Else

Skill type: DISCIPLINE-ENFORCING. This is the entry point of the social-science methods spine, not an analysis engine. It routes a study to the right design family and forces the one decision every downstream gate depends on: what identifying assumption licenses the causal claim? It does not run models — for execution it hands off to the analysis skills.

The Core Rule

A CAUSAL CLAIM IS ONLY AS GOOD AS ITS IDENTIFYING ASSUMPTION — NAME IT FIRST.

Design is chosen by the question and the data-generating process — who was selected, who was assigned, what varies and when — not by which method is fashionable or convenient. A quasi-experiment is an observational design that earns a causal interpretation only by committing, up front, to an assumption that makes the effect identified. State that assumption before touching an estimator; if you cannot defend one, the claim is associational, not causal.

When to Use This Skill

Trigger the design gate whenever a design is being chosen, defended, or implied by a claim:

  • "What research design should I use for this question?"
  • "I have observational survey data and want to claim X improves Y." (← pin the assumption)
  • "Should this be an experiment or a quasi-experiment?"
  • "Can a difference-in-differences / IV / regression-discontinuity design answer this?"
  • "Is my before/after comparison enough to claim the program worked?"

Does NOT Trigger

Route these adjacent requests to the real sibling skill. This gate picks the design and pins the assumption; it does not execute, choose the test, or write.

The request is really about…Route toWhy not this skill
Running the analysis / a specific model once the design is fixedalterlab-statistical-analysis / alterlab-statsmodelsExecution, not design choice.
Which statistical test to use (t-test vs Mann-Whitney, etc.)alterlab-test-selection-guardTest choice is downstream of design.
Deep qualitative design (grounded theory, phenomenology, coding)alterlab-qualitative-methodsThis gate only routes to qual; that skill does it.
Integrating qual + quant strands, joint displaysalterlab-mixed-methodsMixed-methods design mechanics.
Grading evidence quality, confounders, bias upstreamalterlab-scientific-thinkingStudy-validity judgment, not design routing.
Designing the instrument/questionnaire itselfalterlab-survey-designMeasurement instrument, not design family.

Hand-off rule. When the request matches a row above, open your answer by naming that skill by its exact id (for example alterlab-statistical-analysis) 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.

In scope for this gate means choosing, defending or questioning a design or the causal claim it licenses; a design the user states as already fixed is context, so a hand-off adds no design type, identifying assumption or claim type.

The Design Decision Tree

Walk top-down. Each branch is decided by the data-generating process, not the desired claim.

1. Is the CAUSE randomly ASSIGNED by the researcher?
   ├─ YES → TRUE EXPERIMENT (RCT / lab / field experiment)
   │         identifying assumption: randomization → ignorability (assignment ⊥ potential outcomes)
   └─ NO  → 2. Is there exogenous variation you can exploit?
            ├─ a policy/treatment turned on for some units at some time → DIFFERENCE-IN-DIFFERENCES
            │     assumption: PARALLEL TRENDS (treated & control would have moved together absent treatment)
            ├─ an "as-good-as-random" nudge affecting treatment but not the outcome directly → INSTRUMENTAL VARIABLES
            │     assumption: EXCLUSION RESTRICTION + relevance (instrument affects Y only through the treatment)
            ├─ treatment assigned by a threshold on a running variable → REGRESSION DISCONTINUITY
            │     assumption: CONTINUITY of potential outcomes at the cutoff (no manipulation of the score)
            ├─ one unit observed before/after an intervention over many periods → INTERRUPTED TIME SERIES
            │     assumption: the pre-trend + modeled counterfactual would have continued absent the intervention
            ├─ repeated observations, unit/time confounders → FIXED EFFECTS / panel
            │     assumption: no time-varying confounders (selection-on-observables within unit)
            └─ none of the above, adjust for measured confounders only → OBSERVATIONAL / SELECTION-ON-OBSERVABLES
                  assumption: CONDITIONAL IGNORABILITY (no unmeasured confounding) — the weakest, name it as such

3. Is the aim interpretive / theory-building rather than effect estimation?
   ├─ YES → QUALITATIVE design (route to alterlab-qualitative-methods)
   │         "power" is SATURATION / information power, not a sample-size formula (see alterlab-ssci-sampling-gate)
   └─ combine strands → MIXED METHODS (route to alterlab-mixed-methods)

Full branch logic, the five quasi-experimental designs and their threats, and worked routing examples: references/design_decision_tree.md. A stdlib router that prints the design family and its required assumption from your answers: scripts/design_router.py.

"You Buy Credibility by Burning Information"

Every step from a true experiment toward pure observation trades statistical control for an assumption you must defend. Name the trade honestly:

ExcuseReality
"It's basically an experiment — people just chose their own group."Self-selection is exactly what randomization removes. This is observational; name the confounding you are assuming away.
"I have before-and-after data, so it's causal."A single pre/post has no counterfactual. You need parallel trends (DiD) or a modeled counterfactual (ITS) — state which.
"I controlled for everything relevant."You controlled for what you measured. Conditional ignorability assumes no unmeasured confounders — the strongest, least testable assumption.
"Instrumental variables fix endogeneity."Only if the exclusion restriction holds. An instrument that affects the outcome through any other path is invalid — defend exclusion, don't assert it.
"Regression discontinuity is quasi-random at the cutoff."Only if units cannot manipulate the running variable and outcomes are continuous there. Check for sorting/bunching.

The Design Passport (hand-off)

On exit, emit or update the pipeline's YAML Design Passport with: research_question, design_type, identifying_assumption (the named assumption + why it is defensible), and claim_type (causal | associational | descriptive). Downstream, alterlab-ssci-measurement-gate and alterlab-ssci-sampling-gate append to it, and alterlab-ssci-inference-gate audits final claims against design_type + identifying_assumption.

Self-Check Before Advancing

  • Is the design chosen by the data-generating process, not the desired conclusion?
  • Is exactly one identifying assumption named, with a sentence on why it is defensible here?
  • If observational, is the claim explicitly downgraded to associational unless a QED assumption holds?
  • For qualitative aims, is "how many" deferred to saturation logic (not a power formula)?
  • Is the Design Passport populated for the next gate?

References

  • references/design_decision_tree.md — full branch logic, the five QEDs, threats, examples.
  • scripts/design_router.py — stdlib router printing the design family + required assumption.

Part of the AlterLab Academic Skills suite.

示例提示词

帮我判断这个课题该用准实验还是观察性设计,并锁定识别假设。

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

社科研究设计把关能做什么?

Routes a study to its research design (experiment, quasi-experiment, observational, qualitative) and pins the identifying assumption first. Use when choosing a study design。

如何安装社科研究设计把关?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「社科研究设计把关」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。

社科研究设计把关需要哪些权限?

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。

社科研究设计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。