CODE:

# Install libraries if not already installed

install.packages(c("tidyverse", "skimr", "ggplot2", "janitor", "broom", "dplyr", "ggcorrplot", "caret", "pROC", "ROCR"))

# Load libraries

library(tidyverse)

library(skimr)

library(ggplot2)

library(janitor)

library(broom)

library(ggcorrplot)

library(pROC)

library(ROCR)

# Read in the data

df <- read_csv("OnlineBookClub.csv")

# Clean column names

df <- clean_names(df)

# Preview data

skim(df)

To better understand the customer base and prepare the dataset for modeling, a full data scan was conducted using summary statistics from skim(). The dataset is clean, with no missing values across any variables. Numeric fields show a wide range of behaviors, including varied spending levels, purchase frequencies, and time since first or last purchase—suggesting strong potential for segmentation and predictive modeling. Categorical fields such as gender and subscription status are well-balanced and contain no unexpected or malformed values. Overall, the data is complete, well-structured, and ready for advanced analysis.

CODE:

# Basic stats

summary_stats <- df %>%

  summarise(

    avg_total_spent = mean(total, na.rm = TRUE),

    sd_total_spent = sd(total, na.rm = TRUE),

    avg_books_purchased = mean(purch, na.rm = TRUE),

    sd_books_purchased = sd(purch, na.rm = TRUE),

    avg_recency = mean(last, na.rm = TRUE),

    sd_recency = sd(last, na.rm = TRUE)

  )

print(summary_stats)

# Gender & subscription stats

gender_subs <- df %>%

  group_by(gender, subscribe) %>%

  summarise(count = n()) %>%

  pivot_wider(names_from = subscribe, values_from = count, values_fill = 0) %>%

  mutate(

    total = yes + no,

    sub_rate = yes / total

  )

print(gender_subs)

Subscription Rate by Gender

The subscription rate among M/NB customers (12.77%) is nearly double that of female customers (7.17%). While there are more total female customers, they are proportionally less likely to subscribe to the online book club. This suggests gender-based differences in engagement that may require targeted marketing strategies or a review of how the offering is positioned to different segments.

Descriptive Statistics: Spend, Purchases, and Recency

On average, customers have spent $208 in offline stores, purchased about 4 books, and have not made a purchase in approximately 12 months. The high standard deviation for both spend and recency reflects a wide range of engagement levels, with some customers being highly active and others potentially lapsed. This variability reinforces the need for personalized outreach strategies based on recency and spending behavior.

CODE:

lm_model <- lm(total ~ is_female + first + child + youth + cook + do_it + refernce + art + geog, data = df)

# Tidy output

tidy(lm_model)

Linear Regression Output: Predicting Total Spend

This model estimates how customer characteristics influence total offline spend. Nearly every book category, child, youth, cook, do_it, reference, art, and geography, has a positive and statistically significant relationship with spending, with p-values near zero, confirming their strong predictive value. Each estimate reflects the dollar increase in total spend associated with an additional unit of that category (e.g., one more cookbook purchase adds about $15.66 to total spend on average).

Variables like gender (is_female) and tenure (first) are not statistically significant (p > 0.05), suggesting that demographics or how long someone has been a customer do not meaningfully predict spend once purchase behavior is accounted for.

CODE:

# Create binary variable for female

df <- df %>%

  mutate(

    is_female = ifelse(gender == "F", 1, 0),

    subscribe_bin = ifelse(subscribe == "yes", 1, 0)

  )

# Correlation among numeric variables

corr_data <- df %>%

  select(is_female, first, last, total, purch, child, youth, cook, do_it, refernce, art, geog)

corr_matrix <- cor(corr_data)

ggcorrplot(corr_matrix, lab = TRUE, type = "lower", title = "Correlation Matrix")

CODE:

logit_model <- glm(subscribe_bin ~ last + total + is_female + child + youth + cook + do_it + refernce + art + geog,

                   data = df, family = binomial)

summary(logit_model)

# Tidy view

tidy(logit_model)

The regression model uses behavioral and demographic variables to predict the probability of subscription. Significant negative predictors include:

• Recency (last): The longer it has been since a customer’s last purchase, the less likely they are to subscribe.

• Is_female: Female customers are significantly less likely to subscribe, even after accounting for behavior.

• Book categories like child, youth, cook, and do_it: These are all negatively associated with subscription, indicating these buyers may be more traditional, less inclined toward digital formats, or more value-oriented.

Positive predictors include:

• Total spend: Slight but significant positive effect, showing that more engaged spenders are more likely to subscribe.

• Reference and art books: Strongest positive relationships. Customers interested in these categories may value curated or intellectual reading experiences, aligning well with a subscription model.

CODE:

# Predict subscription probabilities

df$predicted_prob <- predict(logit_model, type = "response")

# Decile assignment

df <- df %>%

  mutate(decile = ntile(predicted_prob, 10))

# Profile top and bottom decile

top_decile <- df %>% filter(decile == 10) %>% summarise_all(mean, na.rm = TRUE)

bottom_decile <- df %>% filter(decile == 1) %>% summarise_all(mean, na.rm = TRUE)

print(top_decile)

print(bottom_decile)

Predict Subscription Probabilities | Top Decile

Predict Subscription Probabilities | Bottom Decile

The predicted probabilities were divided into deciles to compare those most and least likely to subscribe.

Top Decile:

• Average predicted probability of subscription is approximately 39 percent.

• Customers are more recently active (recency 7 months), have higher spending totals ($257), and purchase more books on average (6.5).

• Higher engagement in niche categories like art, reference, and geography.

• Lower proportion of females compared to the full sample, aligning with regression findings.

Bottom Decile:

• Average predicted probability is extremely low, around 0.7 percent.

• Recency is much higher (26 months), total spend is lower ($204), and customers have fewer overall purchases (~4.2).

• Behavior is concentrated in more traditional or low-engagement categories like cookbooks and non-book items.

• Much higher proportion of female customers, reinforcing the gender disparity in likelihood to subscribe.

To drive subscriptions to the online book club, an email campaign should target customers who recently made a purchase and have shown interest in art, reference, or geography books. These individuals ranked highest in predicted subscription likelihood based on behavioral data. The campaign should prioritize customers in the top subscription probability decile and exclude those who are inactive or heavily oriented toward categories like cookbooks and DIY, which were negatively associated with subscription.

Timing:

Since calendar-based timing is not available, the message should be triggered based on relationship recency. Emails should be sent within 30 to 60 days after a customer’s last offline purchase. This timing ensures the brand is still top of mind and leverages recent engagement to prompt an online conversion.

Metrics to track:

• Open rate

• Click-through rate (CTR)

• Conversion rate (email-to-subscription)

• Unsubscribe rate

• Bounce rate

A/B test design:

To test effectiveness, run an A/B test on the subject line:

• Version A: “Your Next Favorite Read Is Waiting – Let Us Pick It For You”

• Version B: “Two Free E-Books. Every Month. Yours to Explore.”

The test will help determine whether emotional curiosity or clear value-based messaging drives more engagement. Performance will be measured by open rate and subsequent click-through rate to refine future campaign strategies.