Education News Reporting 媒體技能：Education News Reporting

Distilled from education-journalism curricula at Spencer Foundation, Education Writers Association (EWA), Columbia Journalism School (Education Track), and Taiwan educational institutions (師大新聞系 education track, NTU journalism education reporting). Covers four main education-news sub-types: policy reform / research findings / campus events / student data.

IRON LAW: Effect Size + Population, Not Just "Research Shows"

Education research is widely sensationalized into "study finds X improves Y by Z%".
The LLM tendency is to lead with the headline effect and skip the methodology footer.
Instead: always report (a) effect size (Cohen's d, NNT, % point change), (b) sample size
and demographic (N=500 Taiwan Grade 4, etc.), (c) replication status (single study vs
meta-analysis vs unpublished), (d) source funding (ministry, private foundation, etc.).

This is not optional. A study with d=0.08 is "statistically significant" but educationally
meaningless; a study of 35 suburban Grade 5 students cannot generalize to national policy.
Readers must have this context to judge whether the news is real improvement or noise.

Default LLM failure mode: "A new study shows bilingual education boosts test scores by 12%"
(leading effect, no Cohen's d, no sample demographic, no replication context).

Correct: "A 2024 study of 240 Grade 4 students in Taipei bilingual programs found a 0.6
standard-deviation improvement in reading (Cohen's d=0.6), sustained in a follow-up cohort
but not replicated in rural schools. The National Taiwan University research was funded by
the Language Ministry. Previous international meta-analyses show effect sizes ranging d=0.2
to d=0.5 depending on classroom intensity."

Why this is non-obvious: the headline % is true, the study is real, the writing flows naturally — but the reader cannot judge whether the news is a meaningful education breakthrough or a statistically-significant artifact of a small, unrepresentative sample. This is how education policy gets made on bad evidence.

Rationalization Table — these justifications DO NOT override the Iron Law:

Trigger conditions:

• User supplies education material — 教育部新聞、校園事件、教育研究論文摘要、課程改革公告、升學統計、教師訪談、學生表現數據 — and asks for a news piece.
• User asks for "教育新聞" / "校園報導" / "教育研究新聞" / "education policy story" / "campus event coverage" / "student achievement news".
• User paraphrases: "寫一篇教育政策新聞", "整理校園事件成報導", "幫我解讀這份 PISA 排名", "draft an article on this research finding".

Input signals:

• Named schools, students (or deidentified cohorts), educators, education institutions, student data, curriculum changes, test scores, research findings.
• Education-specific terminology: 教育部、國教院、頂大、全教總、會考、學測、108 課綱、升學率、GPA、PISA、effect size.

When NOT to use:

• School / education institution press release in the institution's own voice → use pr-press-release.
• Teacher training / pedagogy guidance / school-internal communication → use tech-teaching or domain-specific skill.
• Student recruitment / marketing ("discover our innovative bilingual program") → use mkt-*.
• Pure curriculum design or lesson-planning → out of scope for journalism.

Step 0: Defer general workflow to med-news-reporter

Read or have already loaded med-news-reporter for: material audit, fact-checking, source-strength tagging, balance principle, media-ethics check, media-literacy self-check. Do not re-implement those steps here. This file specializes Steps 1–3, adds education-specific Step 3.5 (Research Evidence Audit), and modifies Step 4 (ethics) to include education-specific red lines.

Step 1: Classify the education-story sub-type

If ambiguous, ask the user — do not guess.

Step 2: Source vetting & demographic tagging

Every education claim involving data or outcomes must carry demographic context at first mention:

• ❌ Vague: 「研究顯示,雙語教育提升閱讀成績。」
• ✅ Specific: 「一項由台灣大學 2024 年進行的研究,針對 240 位台北市雙語班四年級學生,發現閱讀成績提升 0.6 個標準差(Cohen's d=0.6)。」
• ✅ With limitation: 「該研究樣本侷限於市區雙語班;在農村地區試行時效果未再現。」

Education source tier tagging (extends med-news-reporter):

Step 3: Education-specific risk check

Beyond med-news-reporter's general ethics check, add:

1. 兒少法 §69 — do NOT disclose name + school + identifying details of minors:

   • ❌ "12 歲的王小華就讀範例國中"
   • ✅ "一位 12 歲女學生" / "示範國中一名四年級男童"
   • Even with parental consent, err toward deidentification.

2. 升學率 / 錄取率口徑混淆:

   • "升學率"(進入高等教育比例) ≠ "錄取率"(申請者中被錄取比例) ≠ "申請成功率"(報考人數 ÷ 錄取人數)
   • Always cite official source (聯招中心、教育部) and definition. If ambiguous in source, flag as [待查證: 升學率定義].

3. 研究結果過度延伸(Goodhart's Law):

   • "One school improved writing with Method X" does not support "Method X should be national policy".
   • "PISA ranking rose 1 position" does not support "reform worked" (within error margin; trends matter more than rank).
   • Mark speculative cause-effect as conditional: "如果..." / "可能" / "若要推廣至全國,須進一步驗證".

4. 教師受訪許可:

   • Teachers employed by schools usually need school approval to speak publicly.
   • 「受訪時曾言及」need confirmation: did the school grant permission? Mark risk if unclear.

5. 校園隱私與實驗倫理:

   • Classroom data (test scores, attendance, behavioral notes) may contain PII.
   • Educational research on students requires IRB approval or school consent + parental consent.
   • If data source unclear, ask before citing.

Step 3.5: Research Evidence Audit (education-specific addition)

For every research-based claim, extract and verify:

1. Effect size: Convert headline claim to Cohen's d, odds ratio, percentage point change, or NNT (number needed to treat).

   • If not in abstract, read Methods + Results. If absent, flag as [待查證: 效果量統計值].
   • Benchmark: education interventions with d < 0.2 are weak; d > 0.5 are strong.

2. Sample size and demographic:

   • Who? (Grade level, school type, region, SES if known, language background for language research).
   • How many? (N = total sample size).
   • If N < 50 or sample is non-representative (private urban school, bilingual cohort only), note as "preliminary evidence" or "context-specific".

3. Replication and consistency:

   • Is this a single published study, a pre-print, a meta-analysis, or a series of replications?
   • If single study: "a study found" not "studies show".
   • If meta-analysis: report heterogeneity (I²), range of effect sizes across studies, and which studies drove the mean.

4. Funding and conflict of interest:

   • Who paid for the research? (Ministry, private foundation, school, textbook company?).
   • Relevant conflict (e.g., a bilingual-program-company-funded study of bilingual-program effectiveness has incentive skew).
   • Disclose in piece: "The research was funded by [source]" or include in demographic tag.

5. Publication status:

   • Peer-reviewed journal > preprint > press release > blog.
   • If data is unpublished or under review, note explicitly: "preliminary findings" / "not yet peer-reviewed".

Use the med-news-reporter base format, with these education additions to the meta footer:

[Headline / sub-headline / body paragraphs per med-news-reporter]

---

**稿件類型**: 教育政策新聞 / 研究新聞 / 校園事件 / 升學新聞
**字數**: approx. XXX
**消息來源層級**: 教育部公開資料 N / 具名教育者 N / 研究論文 N / 學校 N / 利益相關團體 N / 學生/家長 N
**教育專業檢核**:
- 人口統計完整性: ✅ / ⚠️ (列出缺項: 樣本數 / 地區 / 年級 / 家庭背景)
- 效果量稽核: ✅ / N/A / ⚠️ (報告 Cohen's d / 百分點 / 其他指標 + 樣本)
- 研究複製狀態: ✅ / ⚠️ (單一研究 vs 後續複製 vs 後設分析)
- 兒少法 §69 保護: ✅ / ⚠️ (無名字 + 學校組合 / 數據去識別)
- 升學率定義澄清: ✅ / N/A / ⚠️ (列出採用之定義與來源)
**經費與利益揭露**: 〔研究經費來源、利益關係人〕
**待查證事項**: ...
**倫理 / 識讀檢核摘要**: 〔交給 med-news-reporter 的 Step 4-5 footer〕

Good Example

Scenario: User supplies (a) 國家教育研究院 2024 年一份教科書閱讀理解研究摘要(樣本 640 名中部六年級學生),報告採用新編版與舊版教科書的效果比較,Cohen's d=0.45,95% CI [0.28, 0.62];(b) 教育部新聞稿回應;(c) 親子天下與報導者過往類似研究的對比。要求寫 900 字教育新聞。

Analysis:

1. Step 1: classified as research findings (研究新聞),焦點是教科書介入的效果證據。
2. Step 2: source tier — 國教院論文屬 Tier 1(academic); 教育部新聞稿屬 Tier 1(institutional); 先前研究為背景參考。
3. Step 3: risk check — 樣本為國小六年級,中部地區,無兒少法問題(因為報告已去識別)。
4. Step 3.5 (Research Evidence Audit):
   • 效果量: Cohen's d=0.45 —— 屬中等效果,教育上有意義。
   • 樣本: N=640, 台灣中部六年級 —— 樣本大,但侷限於一地區、一年級,全國推廣需後續驗證。
   • 複製狀態: 此為首次研究;未來應進行rural區校驗。
   • 經費: 國教院主導,中立來源。
5. Output footer 標示效果量、樣本demographic、複製狀態,以及後續驗證需求。

Result: 讀者清楚知道:改革有evidence support(d=0.45),但證據來自特定地區特定年級,推廣需謹慎與後續評估。

Bad Example

Scenario: Same input. Writer produces piece that (a) leads with "教科書改革提升學生閱讀成績達 12%" without effect size or sample context, (b) omits sample demographic ("中部六年級" → 改為泛稱「台灣學生」), (c) cites 親子天下 過往發現 as "一致證據" without reporting that past study had N=85 and d=0.2 (much weaker), (d) removes methodological footer because "it looks clean".

What went wrong:

• (a) The "12%" headline is 真的,但12 % point ≠ 0.45 Cohen's d。讀者無法判斷是meaningful reform還是統計雜訊。
• (b) Demographic omission makes readers think evidence applies to "all Taiwan students", not "600+ students in Central region". Single-region, single-grade studies do not automatically generalize.
• (c) Treating d=0.45 and d=0.2 as equally "consistent evidence" misleads. The new study is stronger; the historical comparison is less so. Reader needs that hierarchy.
• (d) Removing methodology footer hides the real limitation: this study alone cannot support national policy.

Net:每個句子都technically true,但讀者會高估evidence strength,導致政策決定可能過度樂觀。

Related skills:

• med-news-reporter — general news workflow (this skill specializes it)
• med-political — for education-policy stories with strong political dimension
• stat-hypothesis-testing — for deep methodological critique of research
• stat-eda — exploratory data analysis on education datasets
• grad-survey-design — for evaluating educational surveys and sampling
• hum-source-criticism — source vetting frameworks