### Project 2: A/B Test ###

"""
A/B Testing (Imre Deak)

Data: https://www.kaggle.com/datasets/faviovaz/marketing-ab-testing
Target: To measure success of an ad campaign - did it successfully alter sales? 

Structure: 
    > importing packages
    > loading/cleaning data, basic EDA
    > Univariate analysis (visualising variables, independent of each other)
    > Bivariate analysis (boxplots)
    > Statistical testing:
        > Chi-squared test (variable dependence test)
        > Picking suitable test based on assumption checks:
            > T-test - if data is normalised & variance is even
            > Mann-Whitney U test (nonparametric) - if any of the conditions above aren't met
                > Neither met: Mann-Whitney U test

"""

### importing packages ----

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns

### loading, cleaning data & visualisation palette ----

df = pd.read_csv("H:/my_directory/Website_Github/Portfolio/Project2\marketing_AB.csv")
df.head()
df.info()

df.duplicated(subset = 'user id').sum() # sum of total of non-distinct data points
df.drop(['Unnamed: 0', 'user id'], axis = 1, inplace = True) # removing data index & id
df.columns

df_cat = df[['test group', 'converted', 'most ads day', 'most ads hour']] # subset of categorical variables
df_cat.nunique() # equivalent of 'levels' in R.

# listing 'levels'
for i in df_cat.columns:
    print(i.upper(), ":", df_cat[i].unique())
    
# colour palettes    
palette = sns.color_palette("RdBu")
pie_colors = sns.color_palette("RdBu")
    
### Univariate analysis ----

    # test group 

variable = 'test group'

plt.figure(figsize = (6,4))
# Count plot
plt.subplot(1,2,1)
sns.countplot(x=variable, data=df_cat, palette = palette)
plt.title(f'Count plot - {variable}')

# Pie chart
plt.subplot(1,2,2)
counts = df_cat[variable].value_counts()
plt.pie(counts, labels=counts.index, autopct='%0.2f%%', colors=pie_colors)
plt.title(f'Pie chart - {variable}')

# Show plot
plt.show()

    # Conversion rates 

variable = 'converted'

plt.figure(figsize = (6,4))
# Count plot
plt.subplot(1,2,1)
sns.countplot(x=variable, data=df_cat, palette = palette)
plt.title(f'Count plot - {variable}')

# Pie chart
plt.subplot(1,2,2)
counts = df_cat[variable].value_counts()
plt.pie(counts, labels=counts.index, autopct='%0.2f%%', colors=pie_colors)
plt.title(f'Pie chart - {variable}')

# Show plot
plt.show()

    # Most ads received a day 

variable = 'most ads day'

plt.figure(figsize = (6,4))
# Count plot
plt.subplot(1,2,1)
sns.countplot(x=variable, data=df_cat, palette = palette, order = df_cat['most ads day'].value_counts().index) # value_counts() does descending order arrangement.
plt.title(f'Count plot - {variable}')
plt.xticks(rotation=90)

# Pie chart
plt.subplot(1,2,2)
counts = df_cat[variable].value_counts()
plt.pie(counts, labels=counts.index, autopct='%0.2f%%', colors=pie_colors)
plt.title(f'Pie chart - {variable}')

# Show plot
plt.tight_layout()
plt.show()

    # most ads hour 
    
variable = 'most ads hour'

plt.figure(figsize = (8,4))
# Count plot
plt.subplot(1,2,1)
sns.countplot(x=variable, data=df_cat, order = df_cat['most ads hour'].value_counts().index)
plt.title(f'Count plot - {variable}')
plt.xticks(rotation=90)

# Pie chart
plt.subplot(1,2,2)
counts = df_cat[variable].value_counts()
plt.pie(counts, labels=counts.index, autopct='%0.2f%%')
plt.title(f'Pie chart - {variable}')

# Show plot
plt.tight_layout() # makes spacing nicer
plt.show()

    # total ads (non-categorical numeric) 
    
variable = 'total ads'

plt.figure(figsize=(6,4))
# Histogram
plt.subplot(1,2,1)
sns.histplot(x=variable, data=df, palette=palette)
plt.title(f'Histogram - {variable}')

# Pie chart
plt.subplot(1,2,2)
sns.boxplot(y = variable, data = df, color=pie_colors)
plt.title(f'Boxplot - {variable}')

# Layout/show plot
plt.tight_layout()
plt.show() # <-- large data N, hence many outliers. Poor visualisation w/out filter
 
df['total ads'].describe() # descriptive stats

    # Repeating previous plot but w/ a filter (trial and error estimation of upper boundary)

variable = 'total ads'

plt.figure(figsize=(6,4))
# Histogram
plt.subplot(1,2,1)
sns.histplot(x=variable, data=df[df['total ads'] < 50] )
plt.title(f'Histogram - {variable}')

# Pie chart
plt.subplot(1,2,2)
sns.boxplot(y = variable, data = df[df['total ads'] < 50] )
plt.title(f'Boxplot - {variable}')

# Layout/show plot
plt.tight_layout()
plt.show() # <-- boxplot spread is easily visible, covers all quartiles.


### Bivariate analysis ----


# Crosstab = creates a 2x2 table for two variables
ct_conversion_test_group = pd.crosstab(df['test group'], df['converted'], normalize = 'index') 
ct_conversion_test_group
ct_conversion_test_group.plot.bar(stacked = True) # visualising crosstab

    # repeating w/ other variables against 'converted'

ct_conversion_day = pd.crosstab(df['most ads day'], df['converted'], normalize = 'index')
print(ct_conversion_day.sort_values(by = True, ascending = False))
ct_conversion_day.plot.bar(stacked = True)

ct_conversion_day = pd.crosstab(df['most ads hour'], df['converted'], normalize = 'index')
print(ct_conversion_day.sort_values(by = True, ascending = False))
ct_conversion_day.plot.bar(stacked = True)

    # Raw & filtered boxplot comparison, converted = x-axis category
sns.boxplot(x = 'converted', y = 'total ads', palette=palette, data = df)
sns.boxplot(x = 'converted', y = 'total ads', palette=palette, data = df[df['total ads'] < 50])

'''
Plot above shows that the median number of times people who purchased the product saw the ad was 25,
as opposed to ~10 for those who have not purchased the product (ie. converted)
    -> repeated targeting seems to yield a higher sales rate

'''

### Statistical testing ----

    # Chi-squared test (Automated @ 5% significance level)

from scipy.stats import chi2_contingency

alpha = 0.05 

for variable in df_cat.columns:
    if variable != 'converted':
        # Create crosstab
        contingency_table = pd.crosstab(df_cat[variable], df_cat['converted'])
        
        # Chi-squared test
        chi2, p, _, _ = chi2_contingency(contingency_table)
        
        # Display results
        print(f"\nChi-squared test for {variable} vs. converted:")
        print(f"Chi-squared test value: {chi2}")
        print(f"p-value: {p}")
        
        # Check for significance
        if p < alpha:
            print(f"The difference in conversion rates across {variable} is statistically significant.")
        else:
            print(f"There is no significant difference in conversion rates across {variable}.")
   
    # Checking assumptions for final test (t-test vs nonparametric alt.)
    
from scipy.stats import shapiro, levene, ttest_ind, mannwhitneyu

# Normality assumption
shapiro_stat_true, shapiro_p_value_true = shapiro(df[df['converted'] == True]['total ads'])
shapiro_stat_false, shapiro_p_value_false = shapiro(df[df['converted'] == False]['total ads'])

print(f"Shapiro-Wilk test for normality (True group): p-value - {shapiro_p_value_true}")
print(f"Shapiro-Wilk test for normality (False group): p-value - {shapiro_p_value_false}")

# Equality of variances assumption
levene_stat, levene_p_value = levene(df[df['converted']]['total ads'], df[~df['converted']]['total ads'])
print(f"Levene's test for equality of variances: p-value = {levene_p_value}")

    # Automation of correct test based on assumption test results

alpha = 0.05

if shapiro_p_value_true > alpha and shapiro_p_value_false > alpha and levene_p_value > alpha:
        # if assumptions are met - deploys t-test (based on mean)
        t_stat, t_p_value = ttest_ind(df[df['converted']]['total ads'], df[~df['converted']]['total ads'])
        print(f"Independent two-sample t-test: p-value = {t_p_value}")
else:
        # otherwise - use Mann-Whitney U test (based on median)
        u_stat, u_p_value = mannwhitneyu(df[df['converted']]['total ads'], df[~df['converted']]['total ads'])
        print(f"Mann-Whitney U test: p-value = {u_p_value}")
        
'''
Non-parametric Mann-Whitney U test, with p-value of 0.0
    -> There is sufficient evidence to reject null hypothesis
        -> Median amount of total ads seen makes a statistically significant difference on aggregate purchasing decision

'''