---
id: wiki-2026-0508-secondary-research
title: Secondary Research
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Desk Research, Literature Review, Existing-Data Analysis]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [research, methodology, literature-review, knowledge-synthesis]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: agnostic
  framework: research-methods
---

# Secondary Research

## 매 한 줄
> **"매 secondary research = 매 existing 의 published / collected data 의 매 analysis"**. 매 primary research (raw 새 data 수집) 의 반대. 매 lit review, 매 meta-analysis, 매 industry report 분석, 매 dataset reuse 다 포함. 매 2026 년 LLM-assisted secondary research 가 매 dominant — 매 single researcher 의 매 weeks → 매 hours.

## 매 핵심

### 매 vs primary research
- **Primary**: 매 직접 collect — survey, interview, experiment, observation. 매 control 큼, 매 cost 큼.
- **Secondary**: 매 already-published 의 reuse — books, papers, gov stats, industry reports, internal docs. 매 cheap, 매 fast, 매 control 작음.

### 매 source taxonomy
- **Academic**: peer-reviewed papers (PubMed, arXiv, Google Scholar, Semantic Scholar, OpenAlex).
- **Government**: census, BLS, OECD, World Bank, KOSIS.
- **Industry**: Gartner, Forrester, IDC, McKinsey, CB Insights, Statista.
- **Internal**: company analytics, post-mortems, design docs.
- **Community**: HN, Reddit, GitHub, blog posts (lower trust, higher recency).

### 매 응용
1. **Lit review**: 매 새 paper 의 매 background section.
2. **Market analysis**: 매 startup 의 매 TAM/SAM/SOM 추정.
3. **Competitor research**: 매 product strategy 의 매 input.
4. **Meta-analysis**: 매 multiple studies 의 매 effect size 통합.
5. **Due diligence**: 매 investment / 매 acquisition 의 매 background.

## 💻 패턴

### Pattern 1: LLM-assisted lit review
```python
import anthropic, asyncio

client = anthropic.AsyncAnthropic()

async def summarize_paper(abstract: str, question: str):
    msg = await client.messages.create(
        model="claude-opus-4-7",
        max_tokens=512,
        system="You are a careful research assistant. Cite verbatim.",
        messages=[{
            "role": "user",
            "content": f"Question: {question}\n\nAbstract:\n{abstract}\n\nIs this relevant? If yes, extract key findings + methodology in 3 bullets.",
        }],
    )
    return msg.content[0].text

async def lit_review(question: str, abstracts: list[str]):
    results = await asyncio.gather(*[summarize_paper(a, question) for a in abstracts])
    return [r for r in results if "not relevant" not in r.lower()]
```

### Pattern 2: arXiv / Semantic Scholar fetch
```python
import requests

def search_semantic_scholar(query: str, limit=20):
    r = requests.get(
        "https://api.semanticscholar.org/graph/v1/paper/search",
        params={
            "query": query,
            "limit": limit,
            "fields": "title,abstract,year,authors,citationCount,openAccessPdf",
        },
    )
    return r.json()["data"]
```

### Pattern 3: Citation graph traversal
```python
def expand_citations(seed_papers, depth=2):
    frontier = list(seed_papers)
    seen = set(p["paperId"] for p in seed_papers)
    for _ in range(depth):
        next_frontier = []
        for paper in frontier:
            r = requests.get(
                f"https://api.semanticscholar.org/graph/v1/paper/{paper['paperId']}/references",
                params={"fields": "title,abstract,year,citationCount"},
            )
            for ref in r.json().get("data", []):
                pid = ref["citedPaper"]["paperId"]
                if pid and pid not in seen:
                    seen.add(pid)
                    next_frontier.append(ref["citedPaper"])
        frontier = next_frontier
    return list(seen)
```

### Pattern 4: Source-trust scoring
```python
def trust_score(source: dict) -> float:
    base = {
        "peer-reviewed": 0.9,
        "preprint": 0.7,
        "government": 0.85,
        "industry-paid": 0.6,
        "blog": 0.4,
        "social": 0.2,
    }.get(source["type"], 0.3)
    age_yrs = 2026 - source["year"]
    decay = max(0.5, 1 - 0.05 * age_yrs)
    citations = min(1.0, source.get("citations", 0) / 100)
    return base * decay * (0.6 + 0.4 * citations)
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| 매 새 topic 빠른 overview | LLM survey + 매 5-10 review papers |
| 매 medical / safety claim | Cochrane / systematic review only |
| 매 market size estimation | Triangulate 3+ sources (Gartner + government + internal) |
| 매 historical trend | Government/longitudinal data |
| 매 cutting-edge tech | arXiv (acknowledge non-peer-reviewed) |

**기본값**: 매 source diversification — 매 single source 의 매 trust X. 매 triangulate ≥3.

## 🔗 Graph
- 부모: [[Research Methodology]]
- 응용: [[Literature Review]]
- Adjacent: [[Knowledge Synthesis]]

## 🤖 LLM 활용
**언제**: 매 abstract 의 매 relevance filter, 매 cross-paper synthesis, 매 lit review draft.
**언제 X**: 매 LLM 의 매 hallucinated citations — 매 always 매 source verify.

## ❌ 안티패턴
- **Single-source bias**: 매 매 1 paper / 매 1 industry report 만 의 매 conclusion.
- **Citation laundering**: 매 LLM 생성 citation 의 매 unverified copy-paste.
- **Stale data**: 매 fast-moving field (LLM, crypto) 의 매 2-yr-old report 의 매 current 처럼 사용.

## 🧪 검증 / 중복
- Verified (Cooper *Research Synthesis and Meta-Analysis* 5th ed; PRISMA 2020 guidelines).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — Secondary Research 의 vs primary, source taxonomy, LLM lit-review pipeline, citation graph, trust scoring 정리 |