科研 Skill

案例过程追踪

process-tracing

Guides within-case causal process tracing for comparative politics and IR: evidence typing, causal-mechanism mapping, and rival-explanation testing. Use for qualitative case-study causal inference.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

Process Tracing for Case-Study Causal Inference

Purpose

Process tracing is the qualitative method for establishing that a hypothesized causal mechanism actually produced an outcome in a specific case, rather than merely co-occurring with it. It is the within-case complement to the cross-case, large-N causal designs covered elsewhere in this skill family (difference-in-differences, instrumental variables, regression discontinuity). Use this skill when the question is "did X cause Y in this case, through what mechanism, and how confident can we be?" — typically in comparative politics, international relations, historical sociology, or any single- or small-N case study where random assignment is impossible and the evidence is documentary, interview-based, or archival rather than a large structured dataset.

This skill does not run statistical models. It gives you the evidentiary discipline (the four Van Evera/Bennett tests), a structured workflow for building and defending a causal-mechanism narrative, and templates for the tables a reviewer will expect to see.

When to Use

Use when the request mentions:

  • process tracing, causal mechanism, causal-process observation (CPO)
  • case study causal inference, within-case analysis
  • congruence method, structured focused comparison
  • "how do we know X caused Y" in a single case or small set of cases
  • critical juncture, path dependence argument needing mechanism evidence

Do not use for:

  • cross-case statistical causal identification (see difference-in-differences, alterlab-causal-inference, rdd-design in the economics-management plugin)
  • descriptive case narration without a causal claim (plain case study writing)
  • large-N qualitative comparative analysis of set relations (see alterlab-qca)

Core Workflow

Step 1 — State the Causal Claim and Candidate Mechanism(s)

Write the claim as: "Cause C, operating through mechanism M (steps M1 → M2 → M3), produced outcome O in case X." Decompose the mechanism into discrete, observable steps — a mechanism that cannot be broken into steps that leave traces is not testable by process tracing.

List every plausible rival mechanism or rival cause at this stage, before collecting evidence. A process-tracing account that only tests the favored hypothesis and never seriously tries to confirm or eliminate rivals is the single most common reviewer objection to this method.

Step 2 — Build the Evidence Inventory

For each step in the mechanism (and each rival), list:

  • what observable evidence (CPOs — causal-process observations) would exist if that step actually occurred
  • where that evidence would be found (archives, interview transcripts, press, meeting minutes, correlation of timing)
  • what evidence would exist instead if a rival mechanism were operating

A CPO is evidence about how something happened (a smoking memo, a timed sequence, an actor's stated reasoning), distinct from a data-set observation (a measured value used for cross-case comparison). Process tracing lives or dies on CPOs.

Step 3 — Classify Each Piece of Evidence with the Four Tests

Use the Van Evera / Bennett typology to grade the inferential power of each piece of evidence found:

TestPassing itFailing itInferential power
Straw-in-the-windRaises plausibilitySlightly weakensWeak either way
HoopNecessary but not sufficientEliminates the hypothesisStrong at elimination
Smoking-gunSufficient but not necessaryWeakens but doesn't eliminateStrong at confirmation
Doubly decisiveNecessary and sufficientEliminatesStrong both ways

Write each item into a table: Evidence | Test type | Hypothesis tested | Pass/Fail | Confidence update. This table is the backbone of the final write-up and the first thing a methods-savvy reviewer will look for.

Step 4 — Test Rivals, Not Just the Favored Hypothesis

For each rival explanation from Step 1, run it through the same evidence. A rival that survives every hoop test applied to the favored hypothesis and is not eliminated by any smoking-gun evidence against it has not been ruled out — say so honestly rather than quietly dropping it.

Step 5 — Assess Overall Confidence and Scope

State explicitly:

  • which hypotheses were eliminated (failed a hoop test) and on what evidence
  • which hypothesis survived best and what doubly-decisive or smoking-gun evidence, if any, supports it
  • the scope of the claim: single-case mechanism evidence supports "this mechanism operated in this case," not automatically "this mechanism operates in general" — extending to other cases requires either a most/ least-likely case selection argument or additional cases (structured focused comparison).

Step 6 — Write Up with Source Traceability

Every claimed piece of evidence must cite its source (archive box/folder, interview ID and date, document and page). A process-tracing narrative with unsourced "it is known that..." claims is not evidence-based and will be flagged in review.

Case Selection Note

If you are choosing which case(s) to trace (rather than being handed one), state the selection logic: most-likely case (a hard test if the mechanism fails here), least-likely case (a strong test if it succeeds here), or a deviant case (an outcome the general pattern does not predict). Random case selection is usually wrong for process tracing — it is a small-N method whose leverage comes from deliberate case choice, not representativeness.

Prerequisites

  • Qualitative source material: archival documents, interview transcripts, press records, meeting minutes, or secondary historical accounts.
  • A specific, falsifiable causal claim and mechanism — process tracing cannot be run on a vague claim like "institutions matter."
  • Familiarity with the empirical context well enough to identify what evidence would distinguish rival mechanisms (domain knowledge, not just method knowledge, drives the inferential leverage here).

Limitations

  • Process tracing establishes mechanism plausibility and elimination within a case; it is not a substitute for cross-case causal identification when a population-level effect size is the actual research question.
  • The four-test typology is a heuristic for grading evidence strength, not a formal probability calculus; treat the confidence language ("hoop," "smoking gun") as qualitative, not as point estimates.
  • Interview- or archive-based evidence carries the reliability limitations of its source (memory, selective record-keeping, translation); note these explicitly rather than treating all CPOs as equally trustworthy.
  • For building and defending qualitative trustworthiness more generally (positionality, triangulation, audit trail), pair this skill with alterlab-ssci-reflexivity-gate.

Quality Checklist

  • Causal claim and mechanism steps stated before evidence collection
  • At least one rival hypothesis listed and seriously tested, not just the favored one
  • Every piece of evidence classified by test type (straw-in-the-wind / hoop / smoking-gun / doubly decisive)
  • Every evidence item traceable to a specific source
  • Case-selection logic (most-likely / least-likely / deviant) stated if the case was chosen rather than given
  • Final claim's scope (this case vs. general mechanism) stated honestly

示例提示词

我想用过程追踪法检验案例中因果机制是否成立,并排除竞争性解释。

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

案例过程追踪能做什么?

Guides within-case causal process tracing for comparative politics and IR: evidence typing, causal-mechanism mapping, and rival-explanation testing. Use for qualitative case-study causal inference。

如何安装案例过程追踪?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「案例过程追踪」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。

案例过程追踪需要哪些权限?

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。

案例过程追踪经过官方审核吗?

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。