Analyzing Player Retention and Engagement Using Data Analysis

The PowerWash simulator dataset was created to support efforts to understand the effect of video games on people's well-being. While many studies exist, the researchers aimed to investigate this effect more extensively, gathering data from 11,080 volunteers, from 39 countries, covering 222 days [2].

Since the dataset [1] is vast, starting with an understanding of player engagement and retention through demographic data would be a good starting point. To this end, we delve into the following topics:

1. Data and Data Processing

- Missing Values

- Outliers

2. Exploratory Data Analysis

3. Retention Analysis — Creating New Features

4. Retention Analysis by Demographics

5. Hypothesis Testing

- ANOVA

- Assumption Checks

- Kruskal-Wallis Test

- Post-Hoc Tests — Tukey’s HSD vs Dunn’s Test

6. Conclusions And Recommendations

A Quick Overview of What’s to Come

After a preprocessing step to clean and organize the data, we first investigated the data to get some initial observations using the frequency distribution of player logins, player responses, player age, and country.

Then we created two more variables to investigate player retention: retention by days using login dates, and retention score using login days and number of logins. Using the distribution of these retention features, we understand the general tendency.

Afterward, we investigated the effect of demography on player retention. Using player retention by country and age, we understand that, while the number of players is mostly from the USA (Image 13), the most retention comes from Israel (Image 11,12).

Also, while most of the players are between 18–30 (Image 8), they don’t play for longer times compared to older players (Image 14), but they log in more often during this period (Image 15).

As we see a difference, to understand if age has a statistically significant difference based on customer retention, we conducted hypothesis tests. As we understand it has an effect, we conducted a posthoc test to see exactly which age groups have a remarkable difference.

The results should clearly guide future marketing strategies. If you want to start reading the analysis, let's begin.

1. DATA AND DATA PROCESSING

Let's start by importing the necessary packages;

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

I use Google Collab for data analysis. If you don't prefer to do so, ensure that the dataset_path is correct.

from google.colab import drive
drive.mount('/content/drive')
dataset_path = '/content/drive/My Drive/Powerwash Simulator Dataset/data/demographics.csv'
df = pd.read_csv(dataset_path)

After extracting unnecessary columns, we can print the information about the DataFrame.

df.drop(columns=['n_demographics'], inplace=True)
df.info()

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 11080 entries, 0 to 11079
Data columns (total 8 columns):
 #   Column       Non-Null Count  Dtype  
---  ------       --------------  -----  
 0   pid          11080 non-null  object 
 1   responses    10042 non-null  float64
 2   logins       11080 non-null  int64  
 3   first_login  11080 non-null  object 
 4   last_login   8248 non-null   object 
 5   country      10670 non-null  object 
 6   age          10964 non-null  float64
 7   gender       11018 non-null  object 
dtypes: float64(2), int64(1), object(5)
memory usage: 692.6+ KB

In this dataset, pid indicates a unique identifier for each player, and responses indicate responses to 6 psychological survey questionnaires via in-game pop-ups, explained in reference 2. Other column names are self-explanatory.

From the output, we can observe the missing values, 1,038 for responses, 2,832 for last_login, 410 for country, 116 for age, and 62 for gender.

To quickly see if the DataFrame object has the right data types, and have an overview, we run

df.head()

Image 1: Quick overview of the Data Frame

To obtain a summary of descriptive statistics to understand the central tendency and shape of the distribution, we run

df.describe()

Image 2: Descriptive Statistics

The max value for responses (3140) and logins (274) are significantly higher than the median, suggesting a skewed distribution and indicating potential outliers. Especially from logins, the difference between the median (5) and the maximum value (274), we can understand that a group of players is highly engaged with the game, suggesting further analysis. Also, The median age is 27, and the age value of 118 is quite absurd, especially in this context. It may need further analysis.

Handling Missing Values

Age and responses are continuous variables with skewed distribution. Replacing missing values with the median sounds reasonable. In this case, it minimizes the effect of extreme values, and it is less sensitive to skewed distributions compared to using the mean.

df['age'].fillna(df['age'].median(), inplace=True)
df['responses'].fillna(df['responses'].median(), inplace=True)

Since gender and country are categorical variables, filling missing values with the mode sounds reasonable, as it maintains the most common characteristics in the data.

df['gender'].fillna(df['gender'].mode()[0], inplace=True)
df['country'].fillna(df['country'].mode()[0], inplace=True)

Filling the missing values for last_login is tricky. See the head of the dataset in image 1, there are some cases in which the player logs in only once, and last_login is missing. For those cases, we assign the first_login value as the last_login value. For the remaining, we group the players by login value. if calculate the median for each group and add it to the first_login value.

df['first_login'] = pd.to_datetime(df['first_login'], errors='coerce')
df['last_login'] = pd.to_datetime(df['last_login'], errors='coerce')

median_deltas = df.dropna(subset=['last_login']).groupby('logins').apply(lambda x: (x['last_login'] - x['first_login']).median())

def impute_last_login(row):
    if pd.isna(row['last_login']):
        if row['logins'] == 1:
            return row['first_login']
        median_delta = median_deltas.get(row['logins'], pd.Timedelta(days=0))
        return row['first_login'] + median_delta
    return row['last_login']

df['last_login'] = df.apply(impute_last_login, axis=1)

Handling Outliers

From descriptive statistics, we know that responses, age, and login have skewed distributions, and contain potential outliers. One of the best ways to analyze it is to draw a box plot.

plt.figure(figsize=(10, 6))
sns.boxplot(x=df['responses'])
plt.title('Box Plot of Player Responses')
plt.xlabel('Number of Responses')
plt.show()

Image 3: Box plot to see the outliers

Most players seem to have a relatively low number of responses, there are many outliers on the right side of the maximum line, with some players having responses exceeding 500, 1000, and even up to 3000. The majority of the data is clustered on the left, indicating that the distribution is heavily right-skewed. These extreme values are significantly beyond the typical range of responses, which suggests they are statistical outliers. Let’s cap them with the 95th percentile

percentile_95 = df['responses'].quantile(0.95)
df['responses'] = np.where(df['responses'] > percentile_95, percentile_95, df['responses'])

As for player logins,

plt.figure(figsize=(10, 6))
sns.boxplot(x=df['logins'])
plt.title('Box Plot of Player Logins')
plt.xlabel('Number of Logins')
plt.show()

Image 4: Box plot to see the outliers

From the box plot of Player logins, we observe similar characteristics as with the responses data, so I want to deal with similarly:

percentile_95_logins = df['logins'].quantile(0.95)
df['logins'] = np.where(df['logins'] > percentile_95_logins, percentile_95_logins, df['logins'])

Finally, for age

plt.figure(figsize=(10, 6))
sns.boxplot(x=df['age'])
plt.title('Box Plot of Player Age')
plt.xlabel('Age')
plt.show()

Image 5: Box plot to see the outliers

While most players seem to fall between 18 and 40 years old, a very far outlier of 118 exists as an age value. This shall be a misspelling of 18 to 118, or maybe a troll. For other cases, such as 60–100, I believe it is also absurd.

Capping them with the 95th percentile would range the values up to 40,45 which sounds reasonable for the context.

percentile_95_age = df['age'].quantile(0.95)
df['age'] = np.where(df['age'] > percentile_95_age, percentile_95_age, df['age'])

2. EXPLORATORY DATA ANALYSIS

Plot distributions for key metrics

Before we delve deeper into extracting retention from this data, we should dig some more into the data.

plt.figure(figsize=(10, 6))
sns.histplot(df['logins'], kde=True, bins=30)
plt.title('Distribution of Player Logins')
plt.xlabel('Number of Logins')
plt.ylabel('Frequency')
plt.show()

This yields the distribution of player logins which is one of the key variables to understand retention in this data.

Image 6: The distribution of player logins

The distribution is heavily right-skewed, indicating that most players log in only a few times, while fewer players log in many times. Actually, many of the players play the game only once.

There is a long tail extending towards the right, with logins going as high as 35. These high login counts, though rare, indicate a small group of highly engaged players who log in frequently.

Here we can put a clear threshold for the difference in engagement between frequent players and a small number of more engaged players. This could be useful for segmenting players based on their engagement to define a high-risk churn player vs. an engaged player. Estimating high-risk churn players is another important research topic, this shall be one of our next topics.

plt.figure(figsize=(10, 6))
sns.histplot(df['responses'], kde=True, bins=30)
plt.title('Distribution of Player Responses')
plt.xlabel('Number of Responses')
plt.ylabel('Frequency')
plt.show()

This yields the distribution of player responses which is one of the key variables to understand player engagement in this data.

Image 7: The distribution of player responses

The distribution of player responses is heavily right-skewed, meaning that most players have relatively low numbers of responses, while a smaller number have very high response counts.

The long tail on the right-hand side of the distribution suggests that some players are extremely engaged, contributing far more responses compared to the average player. The peak in 250 is most likely due to the capping of outliers with the 95th percentile.

Age is one of the most important variables to investigate since it is commonly used in ad campaigns. Understanding the most engaged players helps to target the correct audience.

plt.figure(figsize=(10, 6))
sns.histplot(df['age'], kde=True, bins=30)
plt.title('Distribution of Player Age')
plt.xlabel('Age')
plt.ylabel('Frequency')
plt.show()

Image 8: The distribution of player ages

The distribution shows that the majority of players fall between 18 and 30 years old. The peak occurs around age 20, indicating that younger players form the largest group. The frequency decreases steadily after age 20, which should suggest that the game is most popular among young adults.

There is a slight uptick at age 45, which could be due to the capping of outliers to the 95th percentile. Based on this data, we shall suggest focusing marketing efforts on the 18–30 age group to attract even more players in this demographic.

3. Retention Analysis

Based on the given data, we could calculate two types of retention

1. Calculate the retention period in days for each player using the first login and the last login
2. Create a retention score using the first login, the last login, and the number of times the player logs in.

# 1st approach: Calculate retention period in days for each player
df['retention_days'] = (df['last_login'] - df['first_login']).dt.days

# 2nd approach: Retention score is calculated using both retention days and number of logins
df['retention_score'] = df['logins'] / (df['retention_days'] + 1)

The retention score measure shows how often a player logs in to the game relative to their retention period.

A higher retention score indicates that a player logged in frequently during their retention period, suggesting higher engagement.

A lower retention score means that even though the player may have been around for a long time, they did not log in frequently.

print(df['retention_days'].describe())

We obtain the following descriptive statistics;

count    11080.000000
mean        52.482130
std         69.661722
min          0.000000
25%          0.000000
50%         11.000000
75%        109.000000
max        224.000000
Name: retention_days, dtype: float64

The mean retention period is 52.48 days, while the median is 11 days. This significant difference between the mean and median suggests that some players have high retention periods.

The median of 11 days indicates that half of the players were engaged for less than two weeks, suggesting that most players may lose interest relatively quickly.

The minimum retention period is 0 days, which means some players logged in only once and never returned. The maximum is 224 days, indicating that some players remained active over a much longer period — suggesting a small group of highly committed players.

The standard deviation is 69.66 days, which is quite large, indicating a lot of variability in player retention. Some players might have extremely long retention periods, while we know that some players have 0 retention periods.

print(df['retention_score'].describe())

count    11080.000000
mean         0.768154
std          0.791513
min          0.009050
25%          0.146530
50%          0.666667
75%          1.000000
max         12.000000
Name: retention_score, dtype: float64

From the previous descriptive statistics, we saw a player with 224 days of retention. To get a better understanding, let's see his retention score

player_with_max_retention = df[df['retention_days'] == 224]
print(player_with_max_retention[['pid', 'retention_score']])

pid  retention_score
1402  p11290         0.111111

We can conclude from retention days that the player with 224 days has a high engagement, but its retention score is 0.1111, which has a maximum of 12. I believe combining retention scores and retention days gives a better idea of how players are engaged with the game.

plt.figure(figsize=(10, 6))
sns.histplot(df['retention_days'], kde=True, bins=30)
plt.title('Distribution of Player Retention Period (in Days)')
plt.xlabel('Retention Period (days)')
plt.ylabel('Frequency')
plt.show()

Image 9: The distribution of player periods

The distribution is heavily right-skewed, with a large number of players having a retention period of 0 days. This suggests that many players only logged in once or played for a very short time before stopping entirely.

The long tail extends up to 200+ days, indicating a small number of players are highly engaged and play for extended periods. There is also a minor uptick around 150 days, which could indicate some players coming back at specific times. It could be due to events or content updates.

plt.figure(figsize=(10, 6))
sns.histplot(df['retention_score'], kde=True, bins=30)
plt.title('Distribution of Player Retention Score')
plt.xlabel('Retention Score')
plt.ylabel('Frequency')
plt.show()

Image 10: The distribution of retention score

The retention score distribution is highly right-skewed, with the majority of players having low retention scores. This indicates that most players either did not log in frequently or logged in over a short period.

The peak occurs near a retention score of 0, indicating that many players have a very low frequency of logins.

There appear to be multiple small peaks in the distribution. This could suggest different player segments: players with occasional engagement versus those with moderate engagement.

Retention Analysis By Demographics

While demographics are one of the most important factors in ad campaigns or retention strategies, analyzing retention by demographics is important.

Retention By Country

retention_by_country = df.groupby('country')[['retention_days', 'retention_score']].mean().sort_values(by='retention_days', ascending=False).head(10)

plt.figure(figsize=(12, 8))
retention_by_country['retention_days'].plot(kind='bar', color='seagreen', alpha=0.7)
plt.title('Average Retention Period by Country (Top 10)')
plt.xlabel('Country')
plt.ylabel('Average Retention Period (days)')
plt.show()

Image 11: Average retention period by country.

As we have two metrics for retention, we shall plot the retention score, as well.

plt.figure(figsize=(12, 8))
retention_by_country['retention_score'].plot(kind='bar', color='seagreen', alpha=0.7)
plt.title('Average Retention Score by Country (Top 10)')
plt.xlabel('Country')
plt.ylabel('Average Retention Score')
plt.show()

Image 12: Average retention score by country.

Player Count By Country

player_count_by_country = df['country'].value_counts().head(10)

plt.figure(figsize=(12, 8))
player_count_by_country.plot(kind='bar', color='cornflowerblue', alpha=0.7)
plt.title('Top 10 Countries by Player Count')
plt.xlabel('Country')
plt.ylabel('Number of Players')
plt.show()

Image 13: Player count by country.

From Image 13, clearly, the USA has the highest number of players, followed by the UK, and Canada. But from Image 11, 12, we see that Israel has the highest retention period and retention score. This would mean the campaign used in the USA has been successful, or the CPI videos work the best in the USA while players from Israel, Austria, and Hungary are more engaged with the game.

Retention By Age

age_groups = pd.cut(df['age'], bins=[18, 25, 30, 35, 40, 45], labels=['18-25', '25-30', '30-35', '35-40', '40-45'])
df['age_group'] = age_groups

retention_by_age = df.groupby('age_group')[['retention_days', 'retention_score']].mean()
print(retention_by_age)

plt.figure(figsize=(10, 6))
retention_by_age['retention_days'].plot(kind='bar', color='skyblue', alpha=0.7)
plt.title('Average Retention Period by Age Group')
plt.xlabel('Age Group')
plt.ylabel('Average Retention Period (days)')
plt.show()

We obtain the following retention results for each age group;

age_group                                 
18-25           47.513201         0.784340
25-30           52.401076         0.768064
30-35           60.168682         0.723733
35-40           57.668588         0.767846
40-45           66.588115         0.730752

Image 14: The average retention period across different age groups

As we have two metrics for retention, we shall plot the retention score, as well.

plt.figure(figsize=(10, 6))
retention_by_age['retention_score'].plot(kind='bar', color='skyblue', alpha=0.7)
plt.title('Average Retention Score by Age Group')
plt.xlabel('Age Group')
plt.ylabel('Average Retention Score')
plt.show()

Image 15: The average retention score across different age groups

Solely based on the data from Image 8, we suggested that the distribution shows that the majority of players fall between 18 and 30 years old, i.e. the game is most popular among young adults.

Considering Images 14 and 15, we conclude that the retention period tends to increase as players grow older, peaking in the 40–45 age group. The 18–25 age group shows the lowest retention period but the highest retention score, which might indicate younger players are less committed or may have a higher tendency to switch to other games or activities after binge playing.

The game might be more appealing to older age groups (30 and above) as they stay longer. To make them visit more, notifications should be used. To boost engagement among younger players (18–25), and keep them in the game longer, specific marketing efforts such as events or in-game features that appeal more to this demographic could be implemented.

While there is a difference in retention period across different age groups, from Image 14, we understand that some age groups are more engaged than others in terms of retention period. To see if this is a statistically significant difference or if it is due to a random variation, let's go for a statistical test.

Hypothesis Tests: ANOVA (Analysis of Variance) vs. Kruskal-Wallis Test

Before we go for a hypothesis test, let us first put this into perspective:

Null hypothesis (H₀): There is no significant difference in retention periods between different age groups, meaning that any observed difference could just be noise.

Alternative hypothesis (H₁): There is a significant difference in retention periods between different age groups.

If the p-value is less than 0.05, we reject the null hypothesis, if it is greater than or equal to 0.05, we fail to reject the null hypothesis, meaning that the observed differences are meaningless.

Since we compare multiple groups, the first test that comes to mind is ANOVA which allows us to compare the means of multiple groups. To apply ANOVA, we first need to check if it satisfies some assumptions;

Assumption Check

1. Independence: Each player’s data (first_login, last_login, logins) represents a separate, individual event, and these players are not interacting in a way that influences each other’s retention. In other words, the data we are working with is player-specific. We gladly agree that independence is satisfied.
2. Normality: Since we have a relatively large sample size, taking power from the Central Limit Theorem we would suggest that each group is approximately normally distributed. (Wait for it!!)
3. Homogeneity of Variances (Almost! Equal Variances): Visual examination of the box plots suggested that variances across groups were not drastically different, allowing us to assume this requirement was met. (Wait for it!!)

For now, let’s assume that assumptions are satisfied and go for the test. First, prepare the data for ANOVA

import scipy.stats as stats

age_group_data = [
    df[df['age_group'] == '18-25']['retention_days'].dropna(),
    df[df['age_group'] == '25-30']['retention_days'].dropna(),
    df[df['age_group'] == '30-35']['retention_days'].dropna(),
    df[df['age_group'] == '35-40']['retention_days'].dropna(),
    df[df['age_group'] == '40-45']['retention_days'].dropna()
]

and conduct the test.

f_stat, p_value = stats.f_oneway(*age_group_data)
print("ANOVA F-Statistic:", f_stat)
print("ANOVA P-Value:", p_value)

It yields,

ANOVA F-Statistic: 20.681847739459926
ANOVA P-Value: 5.4070924691962256e-17

Therefore we reject the null hypothesis. Concluded that there is a significant difference in retention periods between age groups.

Genuinely Check Assumptions

13 years ago, I was warned by my lecturer how people tend to underestimate assumptions, easily suggesting that the assumptions are met. So let's genuinely check if the assumptions are met.

Normality Check With Shapiro-Wilk Test

from scipy.stats import shapiro

for age_group in age_groups:
    retention_data = df[df['age_group'] == age_group]['retention_days'].dropna()

    if len(retention_data) >= 3:  
        stat, p_value = shapiro(retention_data)
        print(f"Age Group: {age_group}, Shapiro-Wilk Test P-Value: {p_value}")
        if p_value < 0.05:
            print("  -> Data does not appear to be normally distributed.")
        else:
            print("  -> Data appears to be normally distributed.")
    else:
        print(f"Age Group: {age_group}, Insufficient data for Shapiro-Wilk test (n < 3).")

It yields,

Age Group: 18-25, Shapiro-Wilk Test P-Value: 6.740440315755131e-64
  -> Data does not appear to be normally distributed.
Age Group: 25-30, Shapiro-Wilk Test P-Value: 2.6269219378838177e-52
  -> Data does not appear to be normally distributed.
Age Group: nan, Insufficient data for Shapiro-Wilk test (n < 3).
Age Group: 40-45, Shapiro-Wilk Test P-Value: 3.574905529805361e-33
  -> Data does not appear to be normally distributed.
Age Group: 30-35, Shapiro-Wilk Test P-Value: 2.805396752554377e-43
  -> Data does not appear to be normally distributed.
Age Group: 35-40, Shapiro-Wilk Test P-Value: 1.936810790834937e-36
  -> Data does not appear to be normally distributed.

So even though a general explanation convinces you that the data is normally distributed, this is not the case. The assumption is clearly violated.

Homogeneity of Variances (Almost! Equal Variances) with Levene’s test

Let’s genuinely check if the data satisfies the homogeneity of variances

from scipy.stats import levene

age_group_18_25 = df[df['age_group'] == '18-25']['retention_days'].dropna()
age_group_25_30 = df[df['age_group'] == '25-30']['retention_days'].dropna()
age_group_30_35 = df[df['age_group'] == '30-35']['retention_days'].dropna()
age_group_35_40 = df[df['age_group'] == '35-40']['retention_days'].dropna()
age_group_40_45 = df[df['age_group'] == '40-45']['retention_days'].dropna()

stat, p_value = levene(age_group_18_25, age_group_25_30, age_group_30_35, age_group_35_40, age_group_40_45)

print("Levene's Test Statistic:", stat)
print("Levene's Test P-Value:", p_value)

if p_value < 0.05:
    print("Reject the null hypothesis")
else:
    print("Fail to reject the null hypothesis")

It yields,

Levene's Test Statistic: 15.370718890023284
Levene's Test P-Value: 1.544767618857263e-12
Reject the null hypothesis

The homogeneity of variance assumption is also clearly violated. Since two of the assumptions are clearly violated, we cannot expect ANOVA to perform well. While ANOVA compares the mean across different groups, the diverse variation in each group does not allow us to compare the means.

But don’t worry, there is a non-parametric way that we can conduct it.

Kruskal-Wallis Test

from scipy.stats import kruskal

h_stat, p_value = kruskal(*age_group_data)
print("Kruskal-Wallis H-Statistic:", h_stat)
print("Kruskal-Wallis P-Value:", p_value)

if p_value < 0.05:
    print("Reject the null hypothesis")
else:
    print("Fail to reject the null hypothesis")

It yields,

Kruskal-Wallis H-Statistic: 131.38133953702751
Kruskal-Wallis P-Value: 1.9722850648238523e-27
Reject the null hypothesis

So we reject the null hypothesis i.e. there is a significant difference in retention periods between age groups.

Post-Hoc Test

With ANOVA or the Kruskal-Wallis test, based on whether your data satisfies assumptions or not, we understand that there is a significant difference between the age groups. The next step is to understand for which age groups this difference occurs. Post-hoc tests are used for this purpose. If your assumptions were satisfied to conduct ANOVA, you should go for Tukey’s HSD. In our case, we conducted a non-parametric test, so as a post-hoc test, we picked Dunn’s test.

Dunn’s Test

import scikit_posthocs as sp

dunn_results = sp.posthoc_dunn(age_group_data, p_adjust='bonferroni')

print("Dunn's Test Results:")
print(dunn_results)

Dunn's Test Results:
              1             2             3             4             5
1  1.000000e+00  1.540693e-03  2.331977e-13  2.766074e-08  1.335662e-20
2  1.540693e-03  1.000000e+00  8.247766e-04  3.182397e-02  1.381457e-09
3  2.331977e-13  8.247766e-04  1.000000e+00  1.000000e+00  3.526140e-02
4  2.766074e-08  3.182397e-02  1.000000e+00  1.000000e+00  2.622075e-02
5  1.335662e-20  1.381457e-09  3.526140e-02  2.622075e-02  1.000000e+00

Interpreting the results from Dunn’s test, for the p-values less than 0.05, we conclude that variables for corresponding rows/columns have a significant difference.

Group 1 (Ages 18–25) has a significant difference between all the groups, especially 3,4 and 5.

Groups 3 and 4 have no significant difference, meaning they have similar retention in terms of days.

Group 2 and Group 5 have a significant difference.

We plotted the age groups by retention period and observed a difference. To see if this difference is statistically significant, we performed a hypothesis test. Eventually, we understand that this difference is statistically different. Using Post-Hoc test we understood for which age groups it holds a difference.

plt.figure(figsize=(12, 6))
sns.boxplot(x='age_group', y='retention_days', data=df)
plt.title('Retention Days by Age Group')
plt.xlabel('Age Group')
plt.ylabel('Retention Days')
plt.xticks(rotation=45)
plt.show()

Let's see the difference with a boxplot;

Image 16: Box plot is more descriptive to investigate the difference across age groups.

As we see there is clearly a difference between retention periods of age groups, let's further investigate it visually. From Image 16, we conclude that the 40–45 age group shows the highest median retention and also the greatest maximum retention value. Statistically, they stay in the game for longer periods.

6. Conclusions and Recommendations:

• Efforts should be focused on understanding why retention drops so drastically after the initial session (see Image 6) and developing strategies to engage new players longer. The minimum retention period is 0 days, which means some players logged in only once and never returned . This may suggest the game has appealing UI and graphics but may have issues with its mechanics that need urgent investigation.
• Segmenting players into high-risk churn and engaged groups based on their login patterns makes sense (see Image 6). Further investigation could make marketing efforts more effective.
• The game is most popular among the 18–30 age group (Image 8), but their retention period is relatively low for the 18–25 age group as they tend to churn more (see Image 14). New in-game events or strategies are needed to engage younger players.
• Older players have a higher retention period (Image 14), but most players are aged 18–30. Efforts should focus on attracting more older players, as they stay longer.
• Since older players tend to stay longer, this could mean that targeted marketing for these age groups might lead to better In-App Purchase revenue as they have money to spend.
• Since many players are from the USA but have low retention, organizing USA-based events seems like a reasonable approach.
• Countries like Israel, Sweden, and Hungary have high retention rates but a smaller player base. Comparing CPI videos and campaign variables with the USA could help identify and address potential issues.

Thanks for reading. Please leave feedback, and write me for suggestions and corrections.

Bests.

References:

1. Data: https://osf.io/wpeh6/
2. https://www.nature.com/articles/s41597-023-02530-3
3. Seçer, İ. (2013). SPSS ve LISREL ile pratik veri analizi: Analiz ve raporlaştırma. Anı Yayıncılık.