2nd/10_Wiki/Topics/Other/Research-Methodology.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-research-methodology	Research Methodology	10_Wiki/Topics	verified	self	Research Methods Empirical Research Scientific Method	none	A	0.9	applied	research science methodology statistics ml-research		2026-05-10	pending	language framework python scientific-method

Research Methodology

매 한 줄

"매 a result without a method is folklore.". 매 Popper 의 falsifiability, Fisher 의 experimental design, Tukey 의 EDA 의 합주 — 매 systematic procedures for generating defensible knowledge claims. 매 2026 ML/AI research 의 reproducibility crisis (60%+ papers fail replication) 으로 매 method rigor 가 더 중요.

매 핵심

매 spectrum

• Quantitative: 매 numeric, statistical inference, causal claims.
• Qualitative: 매 thematic, interpretivist, descriptive depth.
• Mixed-methods: 매 sequential or concurrent triangulation.

매 designs

• Experimental: 매 RCT — random assignment to treatment/control.
• Quasi-experimental: 매 diff-in-diff, regression discontinuity, synthetic control.
• Observational: 매 cross-sectional, longitudinal, case-control.
• Computational: 매 ablation, benchmark, simulation, A/B.

매 quality criteria

• Validity: 매 construct, internal, external, statistical conclusion.
• Reliability: 매 repeatable measurement.
• Reproducibility: 매 same data + code → same result.
• Replicability: 매 new data, same protocol → consistent result.

매 응용

1. ML paper: 매 ablation table + seed-variance + held-out test set.
2. Product A/B: 매 power analysis → sample size → MDE.
3. UX study: 매 mixed-method (interview + log analytics).
4. AI safety eval: 매 capability + propensity + control evaluations.

💻 패턴

Pattern 1: Power analysis before experiment

from statsmodels.stats.power import NormalIndPower

analysis = NormalIndPower()
n = analysis.solve_power(effect_size=0.2, alpha=0.05, power=0.8, ratio=1.0)
print(f"매 minimum sample per arm: {int(n)+1}")

Pattern 2: Pre-registration template (YAML)

# 매 preregistration.yaml — 매 commit BEFORE running experiment
hypothesis: "매 LLM with chain-of-thought scores ≥ 5pp higher on GSM8K vs no-CoT"
primary_outcome: gsm8k_accuracy
n_per_arm: 1000
conditions: [no_cot, cot]
analysis: paired_t_test
exclusion_criteria: ["api_error", "max_tokens_truncated"]
seeds: [0, 1, 2, 3, 4]

Pattern 3: Reproducible experiment seed control

import random, numpy as np, torch, os

def set_all_seeds(s):
    random.seed(s); np.random.seed(s); torch.manual_seed(s)
    torch.cuda.manual_seed_all(s)
    os.environ["PYTHONHASHSEED"] = str(s)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False

Pattern 4: Ablation table generation

import itertools, pandas as pd

def ablation_runs(components, base_run):
    rows = []
    for subset in itertools.combinations(components, len(components)-1):
        cfg = base_run.copy(); 
        removed = [c for c in components if c not in subset][0]
        cfg["removed"] = removed
        cfg["score"] = run(cfg)
        rows.append(cfg)
    return pd.DataFrame(rows)

Pattern 5: Confidence interval reporting (not just p-values)

import scipy.stats as st
def ci(scores, alpha=0.05):
    m = np.mean(scores); s = np.std(scores, ddof=1); n = len(scores)
    h = s / np.sqrt(n) * st.t.ppf(1 - alpha/2, n-1)
    return m, m-h, m+h
# 매 always report (mean, lo, hi) — 매 not just "significant"

Pattern 6: Qualitative coding (thematic analysis)

# 매 inter-rater reliability via Cohen's kappa
from sklearn.metrics import cohen_kappa_score
kappa = cohen_kappa_score(coder_a_codes, coder_b_codes)
assert kappa > 0.7, "매 coding scheme too ambiguous — refine"

Pattern 7: A/B with sequential testing (mSPRT)

def msprt_decision(treatment, control, theta=0.01):
    """매 mixture sequential probability ratio test — 매 anytime-valid."""
    # Lindon & Malek 2020 — 매 lets you peek without inflating type-I
    pass  # use external lib like `confseq`

매 결정 기준

상황	Design
매 cause-effect claim	RCT or quasi-experimental
매 description / mapping	Observational + descriptive stats
매 user "why"	Qualitative interview + thematic
매 ML model claim	Ablation + multiple seeds + held-out
매 product feature decision	A/B with power analysis + pre-reg
매 emerging behavior	Mixed-methods

기본값: 매 pre-register + multiple seeds + report CIs + share code & data.

🔗 Graph

• 부모: Statistics
• 변형: Causal Inference

🤖 LLM 활용

언제: 매 designing experiments, 매 reviewing methodology of papers, 매 drafting pre-registrations. 언제 X: 매 producing fake citations / fabricating data — 매 catastrophic ethics violation.

❌ 안티패턴

• HARKing (Hypothesizing After Results Known): 매 makes p-values meaningless.
• p-hacking: 매 trying many tests until significant.
• Single seed reporting: 매 ML papers — 매 noise dressed as signal.
• Overfitting to test set: 매 multi-stage benchmarks → 매 leakage.
• No pre-registration: 매 invites unconscious bias.

🧪 검증 / 중복

• Verified (Popper 1959, Fisher 1935, Open Science Framework, Pineau et al. 2021 ML reproducibility checklist).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — design spectrum + ML reproducibility focus