📊 Chapter 4.1: Decoding Cancellation Rates — How Age, Gender & Location Influence Customer Behavior

Welcome back! In the previous chapter, we laid the foundation for robust data analysis by cleaning and preparing the TheLook eCommerce dataset, transforming raw data into a reliable asset for actionable insights. Now, we dive into Chapter 4.1: Decoding Cancellation Rates — Age, Gender & Location, where we apply inferential statistics to uncover patterns behind customer behavior. The analysis demonstrates a robust statistical methodology using z-tests and GLM modeling to evaluate cancellation rates across demographics. However, since the data is synthetic, observed patterns—such as higher cancellation rates in older age groups or regional differences—should not be interpreted as real-world behavioral insights. Instead, this post highlights how to apply inferential statistics to customer data, emphasizing hypothesis testing, confidence intervals, and model interpretation in a controlled environment. The goal is to showcase analytical rigor, not to draw definitive business conclusions.

🔍 Why Use a Z-Test for Cancellation Rates?

In large eCommerce datasets, proportions like cancellation rates can be compared using a two-sample z-test, which is ideal when:

• Sample sizes are large (n > 30 per group)
• Data follows a binomial distribution (canceled vs. not canceled)
• We want to test if two population proportions differ significantly

The z-test leverages the Central Limit Theorem, ensuring that the sampling distribution of the proportion is approximately normal, even if the underlying data isn’t.

      from statsmodels.stats.proportion import proportions_ztest

# Example: Compare male vs. female cancellation rates
counts = [cancelations_male, cancelations_female]
nobs = [orders_male, orders_female]

z_stat, p_val = proportions_ztest(counts, nobs)
print(f"Z-statistic: {z_stat:.3f}, P-value: {p_val:.3f}")
    

🧪 Hypothesis Testing Framework

• Null Hypothesis (H₀): Cancellation rates for males and females are equal.
• Alternative Hypothesis (H₁): Cancellation rates differ between genders.

With a p-value of 0.284, we fail to reject H₀ — no significant difference exists in overall cancellation behavior by gender.

However, two months showed significance:

• September 2023 (p = 0.034)
• May 2024 (p = 0.043)

This suggests short-term behavioral shifts possibly due to promotions, stock issues, or UX changes.

📈 Cancellation Trends by Demographics

Age Group Impact

-💡 Insight: Older customers show higher cancellation rates — possibly due to stricter quality expectations or fit concerns.

Country Comparison (Top 3)

China’s higher rate may reflect competitive market dynamics or fulfillment delays.

🌍 Business Implications

🧮 Confidence Interval Example (Python)

      from statsmodels.stats.proportion import proportion_confint

# 95% CI for female cancellation rate
ci_lower, ci_upper = proportion_confint(cancelations_female, orders_female, alpha=0.05)
print(f"95% CI: [{ci_lower:.3f}, {ci_upper:.3f}]")
    

💬 SQL: Extract Cancellation Rates by Demographics

      SELECT
  CASE 
    WHEN u.age < 20 THEN 'Youth'
    WHEN u.age < 40 THEN 'Early_Adulthood'
    WHEN u.age < 60 THEN 'Middle_Adulthood'
    WHEN u.age < 75 THEN 'Young_Old'
    ELSE 'Old_Old'
  END AS age_group,
  u.country,
  u.gender,
  COUNT(CASE WHEN o.shipped_at IS NULL THEN 1 END) * 1.0 / COUNT(*) AS cancellation_rate
FROM `bigquery-public-data.thelook_ecommerce.users` u
JOIN `bigquery-public-data.thelook_ecommerce.orders` o ON u.id = o.user_id
GROUP BY 1, 2, 3
ORDER BY cancellation_rate DESC;
    

⚠️ Model Limitations

While a logistic regression (via GLM) is appropriate for modeling cancellation probability, the current results are unreliable due to data distortions — particularly extreme coefficient estimates like Austria’s (–19.9), which suggest quasi-complete separation from sparse or zero-event data. These issues invalidate standard errors and p-values, making inference misleading. This highlights the need for diagnostic checks or Consider collapsing categories (e.g., group low-sample countries) or using regularization methods before deploying such models in production.

✅ Key Takeaways

• Statistical testing moves beyond descriptive stats — it validates whether observed differences are real or random.
• Age has more influence than gender on cancellation behavior.
• Regional trends require localized strategies, especially in high-cancel regions like China.

As we move forward, the next chapter — Chapter 4.2: Customer Segmentation in Practice — will build on this foundation by applying clustering techniques to group users based on behavior, enabling personalized marketing strategies grounded in data.

• 📣 Explore the full dashboard
• 🔗 View code on Colab