拡大推計の使い所とアンチパターン：サンプルから全体を語るときに何を間違えるか

import pandas as pd
# サンプル50店舗の月中売上sample = pd.DataFrame({    "store_id": range(1, 51),    "sales": [120, 95, 230, 88, ...],  # 50件})
TOTAL_STORES = 1000inflation_factor = TOTAL_STORES / len(sample)
estimated_total = sample["sales"].sum() * inflation_factorprint(f"全店推定売上: {estimated_total:,.0f} 千円")

# サンプル50店舗 + 全店舗の補助情報（店舗面積 m²）sample = pd.DataFrame({    "store_id": range(1, 51),    "sales": [120, 95, 230, 88, ...],    "area": [50, 40, 100, 35, ...],  # 店舗面積})
TOTAL_AREA = 65000  # 全1000店舗の合計面積（事前に分かってる）
# 単位面積あたり売上 × 全店面積sales_per_area = sample["sales"].sum() / sample["area"].sum()estimated_total = sales_per_area * TOTAL_AREAprint(f"比推定: {estimated_total:,.0f} 千円")

# 全店舗の層別店舗数（事前に把握）total_by_stratum = {    "都心_大型": 200, "都心_小型": 300,    "郊外_大型": 150, "郊外_小型": 350,}
# サンプル中の層別データsample = pd.DataFrame({    "store_id": range(1, 51),    "stratum": ["都心_大型"] * 15 + ["都心_小型"] * 15 +               ["郊外_大型"] * 10 + ["郊外_小型"] * 10,    "sales": [...],  # 50件})
# 層別に単純拡大 → 合算total = 0for stratum, n_total in total_by_stratum.items():    s = sample[sample["stratum"] == stratum]    if len(s) == 0:        continue    avg = s["sales"].mean()    total += avg * n_total    print(f"  {stratum}: 平均 {avg:.0f} × {n_total}店 = {avg*n_total:,.0f}")
print(f"層化推定 合計: {total:,.0f} 千円")

import numpy as npimport pandas as pd
# 全体の周辺分布（既知）target_sex = {"M": 0.49, "F": 0.51}target_age = {"20s": 0.20, "30s": 0.25, "40s": 0.25, "50s+": 0.30}
# サンプル（500人）np.random.seed(42)sample = pd.DataFrame({    "sex": np.random.choice(["M", "F"], 500, p=[0.6, 0.4]),  # 偏り: 男性多め    "age": np.random.choice(["20s", "30s", "40s", "50s+"], 500, p=[0.4, 0.3, 0.2, 0.1]),    "score": np.random.normal(70, 10, 500),})sample["weight"] = 1.0
# 反復比例フィッティングfor _ in range(20):    # 性別軸の補正    for s, t in target_sex.items():        cur = sample[sample["sex"] == s]["weight"].sum() / sample["weight"].sum()        if cur > 0:            sample.loc[sample["sex"] == s, "weight"] *= t / cur    # 年代軸の補正    for a, t in target_age.items():        cur = sample[sample["age"] == a]["weight"].sum() / sample["weight"].sum()        if cur > 0:            sample.loc[sample["age"] == a, "weight"] *= t / cur
# 補正前後の平均スコアprint(f"単純平均: {sample['score'].mean():.2f}")print(f"Raking後: {(sample['score'] * sample['weight']).sum() / sample['weight'].sum():.2f}")

import numpy as np
def bootstrap_total(sample_sales, n_total, B=1000):    estimates = []    for _ in range(B):        resampled = np.random.choice(sample_sales, len(sample_sales), replace=True)        estimates.append(resampled.sum() * n_total / len(sample_sales))    return np.percentile(estimates, [2.5, 50, 97.5])
low, med, high = bootstrap_total(sample["sales"].values, n_total=1000)print(f"全店推計: {med:,.0f}（95% CI: {low:,.0f} - {high:,.0f}）")