---
id: wiki-2026-0508-inexact-science
title: Inexact Science
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Soft Science, Probabilistic Reasoning, Approximate Methods]
duplicate_of: none
source_trust_level: A
confidence_score: 0.85
verification_status: applied
tags: [epistemology, statistics, uncertainty, methodology]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: pymc
---

# Inexact Science

## 매 한 줄
> **"매 uncertainty 매 quantify"**. Inexact science 매 deterministic closed-form X — 매 noise, bias, partial observability 매 inherent. 매 2026 ML interpretability, social science replication crisis 매 forefront. 매 tool: 매 Bayesian inference, robust statistics, sensitivity analysis.

## 매 핵심

### 매 inexactness 의 source
- **Aleatory**: 매 inherent randomness (quantum, dice).
- **Epistemic**: 매 ignorance — 매 reducible by data.
- **Measurement noise**: 매 instrument precision limit.
- **Model misspecification**: 매 wrong functional form.
- **Selection bias**: 매 non-representative sample.

### 매 mitigation 전략
- **Bayesian credible intervals** (vs frequentist CI).
- **Bootstrap resampling** — 매 distribution-free uncertainty.
- **Cross-validation** — 매 generalization estimate.
- **Sensitivity analysis** — 매 parameter perturbation.
- **Pre-registration** — 매 p-hacking 방지.

### 매 응용
1. 매 medical trials (FDA Phase III).
2. 매 ML model deployment (Bayesian deep learning).
3. 매 climate modeling (ensemble).
4. 매 economics (DSGE models).

## 💻 패턴

### 1. Bayesian Linear Regression (PyMC)
```python
import pymc as pm

with pm.Model() as model:
    alpha = pm.Normal('alpha', 0, 10)
    beta = pm.Normal('beta', 0, 10)
    sigma = pm.HalfNormal('sigma', 5)
    mu = alpha + beta * x_obs
    y = pm.Normal('y', mu=mu, sigma=sigma, observed=y_obs)
    trace = pm.sample(2000, tune=1000)
# 매 posterior distribution — credible intervals 매 natural
```

### 2. Bootstrap Confidence Interval
```python
import numpy as np
def bootstrap_ci(data, stat_fn, n=10_000, alpha=0.05):
    boots = [stat_fn(np.random.choice(data, len(data), replace=True))
             for _ in range(n)]
    return np.percentile(boots, [100*alpha/2, 100*(1-alpha/2)])
```

### 3. Sensitivity Analysis (Sobol)
```python
from SALib.analyze import sobol
from SALib.sample.sobol import sample as sobol_sample

problem = {'num_vars': 3, 'names': ['x1','x2','x3'],
           'bounds': [[0,1]]*3}
param_values = sobol_sample(problem, 1024)
Y = np.array([model(p) for p in param_values])
Si = sobol.analyze(problem, Y)  # 매 first/total order indices
```

### 4. Cross-Validation
```python
from sklearn.model_selection import cross_val_score
scores = cross_val_score(model, X, y, cv=10, scoring='neg_mean_squared_error')
print(f"MSE: {-scores.mean():.3f} ± {scores.std():.3f}")
```

### 5. Robust Statistics (M-estimator)
```python
from sklearn.linear_model import HuberRegressor
# 매 outlier-resistant — Huber loss 매 quadratic+linear
huber = HuberRegressor(epsilon=1.35).fit(X, y)
```

### 6. Conformal Prediction (Distribution-Free)
```python
# 매 2026 standard — coverage guarantee 매 model-agnostic
calib_residuals = np.abs(y_calib - model.predict(X_calib))
q_hat = np.quantile(calib_residuals, 0.95)
# 매 prediction interval: [pred - q_hat, pred + q_hat]
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Small n, prior knowledge | Bayesian (PyMC, Stan) |
| Large n, distribution-free | Bootstrap + conformal |
| Causal claim | RCT > obs + IV/DiD |
| Outliers heavy | Huber / RANSAC |
| Multiple comparisons | BH-FDR / Bonferroni |

**기본값**: 매 report point estimate + 95% interval; 매 effect size > significance.

## 🔗 Graph
- 부모: [[Statistics]] · [[Probability Theory]]
- 변형: [[Bayesian Inference]]
- 응용: [[Statistical-Power]] · [[Multivariate-Analysis]]
- Adjacent: [[Epistemology]]

## 🤖 LLM 활용
**언제**: 매 study design review, 매 uncertainty communication, 매 robustness check 제안.
**언제 X**: 매 deterministic system (compiler, hash). 매 cryptographic exactness 필요.

## ❌ 안티패턴
- **p<0.05 cult**: 매 effect size 무시, multiple-testing 무수정.
- **HARKing**: 매 hypothesis after results known.
- **Overconfident point estimate**: 매 ±uncertainty 미보고.
- **Garrison the data**: 매 outlier 임의 제거.

## 🧪 검증 / 중복
- Verified (Gelman, *BDA*; Wasserman, *All of Statistics*; ASA p-value statement).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — Bayesian/bootstrap/conformal patterns |