正态分布特征描述

结论与要点

• 次日留存率在当前样本下可视为近似正态（正态性检验p=0.20），可安全使用基于正态分布的区间估计与显著性检验。
• 总体估计：均值42%，标准差3.5%，n=800。标准误约0.124个百分点，95%置信区间为[41.76%，42.24%]，估计精度高。
• A/B 对比：A=41.8%，B=43.1%，差异1.3个百分点；若两组样本量相等且合并标准差为3.6%，差异的标准误约0.255个百分点，Z≈5.11，p<10^-6，差异在统计上显著；差异的95%置信区间约为[0.8，1.8]个百分点。标准化效应量d≈0.36（小到中等）。
• 样本量与检验力（基于σ=3.6%，差异δ=1.3个百分点）：达到80%/90%检验力所需的每组样本量约为120/162；若当前为均分800（每组≈400），检验力>99%。请确认各组实际样本量以最终定论。

正态分布的关键特征（结合本指标解读）

• 由两个参数完全刻画：均值μ与标准差σ。对称、钟形，均值=中位数=众数。对次日留存这种比例型指标，当样本量足够大且p不极端（此处p≈0.42），样本均值在中心极限定理下近似正态。
• 标准化与区间：任意观测可用Z=(X−μ)/σ转为标准正态。对样本均值，需使用标准误SE=σ/√n。当前SE≈0.124个百分点，意味着均值估计的随机波动很小。
• 经验法则（68-95-99.7）：X~N(μ,σ^2)时，约68%/95%/99.7%的值落在μ±1σ/±2σ/±3σ内。以μ=42%，σ=3.5%计，μ±3σ≈[31.5%，52.5%]，仍在0–100%边界内，正态近似合理。
• 线性与可加性：独立正态的线性组合仍为正态。对比A/B的均值差（在大样本下）也近似正态，便于用Z检验/置信区间评估差异。
• 推断友好性：正态假设允许使用z/t检验、置信区间、效应量与功效分析，便于运营评估中的快速决策。

A/B 评估要点与计算

• 假设与方法：检验H0: μA=μB。已知各组方差近似相等（合并σ≈3.6%），可用两独立样本z/t检验。差异的标准误SE(Δ)=σp√(1/nA+1/nB)。
• 在均分nA=nB=400的示例下：SE(Δ)≈0.036×√(1/400+1/400)≈0.255个百分点；Z=1.3/0.255≈5.11，p<10^-6；95% CI≈[0.8，1.8]个百分点；Cohen’s d=0.013/0.036≈0.36。
• 样本量指引（两侧α=0.05）：为检测δ=1.3个百分点，σ=3.6%，
  • 80%检验力：每组约120
  • 90%检验力：每组约162
• 注：如nA≠nB或方差不齐，应改用Welch检验；如存在多轮查看或多指标并行，对显著性需做校正或使用序贯方法。

注意事项

• 指标边界与分布：留存率∈[0,1]。在当前p与n下，用正态近似评估均值与均值差是恰当的；若p接近0或1或样本很小，应改用二项/比例检验或广义线性模型（如logit）。
• 稳健性：关注异常值与方差同质性；必要时进行Levene/Brown-Forsythe检验。
• 行动建议：确认A/B各组样本量并复算SE与p值；报告差异的置信区间与效应量；结合业务阈值评估1.3个百分点的实际价值，并规划后续发布或扩量验证。

Objective Describe the characteristics of a normal distribution and apply them to your marketing metrics: conversion rate (CR) and average order value (AOV), with validity checks and actionable implications.

Key characteristics of a normal distribution

• Shape and symmetry: Bell-shaped, perfectly symmetric around the mean; mean = median = mode.
• Defined by two parameters: Mean (μ) sets the center; standard deviation (σ) sets the spread. Variance is σ².
• Empirical rule (68–95–99.7):
  • About 68% of values lie within ±1σ of μ.
  • About 95% within ±2σ.
  • About 99.7% within ±3σ.
• Z-scores: Standardizing (z = (x − μ) / σ) allows probability calculations and anomaly detection.
• Additivity/approximation: Sums/averages of many independent effects tend toward normality (central limit theorem), making normal a practical model for aggregated marketing metrics.
• Tails: The tails are thin; extreme values are rare but possible and quantifiable.

Application to your metrics Normality checks

• CR normality test p = 0.09; AOV p = 0.15. Both > 0.05 → fail to reject normality; “approximately normal” is reasonable for these data.
• Note: The provided CR variance (0.0016) is inconsistent with σ = 0.4% if σ is in proportion units (0.004), which implies variance 0.000016. I’ll use σ for interval and probability estimates.

Conversion rate (μ = 3.2%, σ = 0.4%, n = 3000)

• One-sigma range (≈68% of observations): 2.8% to 3.6%.
• Two-sigma range (≈95%): 2.4% to 4.0%.
• Three-sigma range (≈99.7%): 2.0% to 4.4%.
• Example probabilities:
  • P(CR > 3.9%): z = (3.9 − 3.2) / 0.4 = 1.75 → ≈ 4.0%.
  • P(CR < 2.5%): same z magnitude → ≈ 4.0%.
• Precision of the mean (95% CI for μ): SE = 0.4% / √3000 ≈ 0.0073%; CI ≈ 3.2% ± 0.014% → [3.186%, 3.214%]. The mean estimate is very stable.

Average order value (μ = 86, σ = 12, n = 3000)

• One-sigma range (≈68%): 74 to 98.
• Two-sigma range (≈95%): 62 to 110.
• Three-sigma range (≈99.7%): 50 to 122.
• Precision of the mean (95% CI for μ): SE = 12 / √3000 ≈ 0.219; CI ≈ 86 ± 0.429 → [85.571, 86.429].

Anomaly assessment: Thursday AOV = 125

• z = (125 − 86) / 12 = 3.25.
• One-sided tail probability ≈ 0.0006; two-sided ≈ 0.0012. This exceeds the 3σ upper bound (~122), making it a statistically rare event under normality.
• Business implications: Treat as an outlier likely driven by a discrete factor (promotion, campaign targeting, mix shift, data error). Investigate and document; consider robust handling (e.g., winsorize or analyze with/without the point).

Practical notes for marketing analysis

• CR is bounded between 0 and 1 and often arises from binomial processes; at μ = 3.2% with modest σ and large n, the normal approximation is acceptable, but beta/binomial modeling can be considered for finer inference.
• Normal modeling supports:
  • Setting alert thresholds (e.g., 2σ bands for monitoring).
  • Quantifying the rarity of spikes/drops.
  • Building confidence intervals for KPIs to separate signal from noise.

Conclusions

• Both CR and AOV are approximately normal in your sample, enabling reliable use of z-scores, sigma bands, and confidence intervals.
• Expected operational ranges:
  • CR: 2.4%–4.0% (95% band); mean tightly estimated at ~3.2%.
  • AOV: 62–110 (95% band).
• The Thursday AOV = 125 is a statistically rare outlier (>3σ); investigate cause and treat carefully in reporting and optimization.