RPubs link information

• Rpubs link comes here: https://rpubs.com/beetcaps2/1194264

Introduction

Social media is crucial for modern marketing, with platforms like Facebook playing a key role in brand building and customer engagement. This analysis uses a dataset from Moro et al. (2016), which investigates the performance metrics of Facebook posts by a renowned cosmetics brand in 2014. The dataset includes 500 posts with 19 attributes, focusing on pre-publication features and post-engagement metrics. The goal is to identify key predictors of post performance through statistical analysis, helping to optimise social media strategies for enhanced reach and interaction.

Introduction Cont.

• Social media metrics are essential for understanding and optimising the impact of marketing efforts.
• This analysis aims to leverage statistical methods to uncover patterns and relationships within the data, focusing on metrics such as total reach, engagement, and interaction.
• By doing so, we can provide actionable insights for enhancing social media strategies, ultimately contributing to more effective brand building and customer engagement.
• This study will utilise descriptive statistics, hypothesis testing, and regression analysis to thoroughly investigate the data and draw meaningful conclusions.

Problem Statement

• Overall Problem/Question: How can the performance of Facebook posts, as measured by metrics such as total reach, engagement, and interaction, be optimised using statistical analysis?
• Statistical Approach:
  • Descriptive Statistics: Summarise the data to understand the central tendencies and dispersion of key metrics.
  • Hypothesis Testing: Determine if there are statistically significant differences or associations between different types of posts and their performance metrics.
  • Regression Analysis: Model the relationship between post characteristics (such as type, time of posting, and category) and performance metrics to identify key predictors of success.

Data

• The dataset was collected and curated by Moro et al. (2016) and is publicly available for research purposes. The data pertains to Facebook posts published throughout 2014 on the page of a well-known cosmetics brand.
• The dataset includes 500 instances and 19 attributes, divided into features known prior to post publication and features for evaluating post impact.
• No missing attribute values are reported, ensuring a complete dataset for analysis.
• The dataset is a subset of a larger collection originally consisting of 790 rows. The selection of these 500 rows was likely random to maintain confidentiality while providing a representative sample for analysis.
• The full dataset and its analysis are detailed in the paper by Moro, S., Rita, P., & Vala, B. (2016). Predicting social media performance metrics and evaluation of the impact on brand building: A data mining approach. Journal of Business Research. Available at: http://dx.doi.org/10.1016/j.jbusres.2016.02.010.

Data Cont.

• Important Variables:

  • Post_Type: Indicates the type of post (Photo, Status, Link, Video). Each type can impact engagement differently.
  • Paid: Binary variable indicating if the post was promoted (1) or not (0).
  • Total_Reach: The number of unique users who saw the post, an essential metric for post visibility.
  • Engaged_Users: The number of unique users who interacted with the post, reflecting engagement levels.
  • Total_Interactions: Sum of all interactions (comments, likes, shares), a key metric for overall engagement.

• Factors and Levels:

  • Post_Type: Levels include Photo, Status, Link, and Video.
  • Category: Represents different post categories such as Event or Product promotion.

• Scale of Numeric Variables:

  • Total_Likes, Total_Reach, Engaged_Users, Total_Interactions: These variables are count data, typically taking on non-negative integer values.
  • Paid: A binary variable (0 or 1).

Descriptive Statistics and Visualisation

• To understand the key metrics of the dataset, we perform descriptive statistical analysis on the numerical variables. The important variables in our dataset include:

  • Total_Reach: Important for understanding the breadth of exposure a post gets. Higher reach indicates better visibility of the content to the audience.
  • Total_Impressions: Crucial for measuring how often a post is seen. High impressions mean the post is frequently appearing on users’ feeds, increasing brand presence.
  • Engaged_Users: Reflects user interaction with the post. Engagement metrics are vital for assessing the effectiveness of content in capturing audience interest.
  • Post_Consumers: Indicates the number of users interacting with the post content. Higher numbers suggest more in-depth user interest and interaction.
  • Post_Consumptions: Shows the extent of user interaction with post content. More consumptions mean more clicks and interactions, highlighting content relevance.
  • Impressions_Liked_Page: Measures the impressions from users who liked the page, helping gauge loyalty and continuous engagement from fans.
  • Reach_Liked_Page: Represents the reach among users who liked the page. It’s important for understanding the core audience’s exposure to the content.
  • Engaged_Liked_Page: Counts engaged users who liked the page. This metric is significant for understanding how well content resonates with loyal followers.
  • Comment: Number of comments indicates how much the content prompts discussion and feedback, important for community engagement.
  • Like: The number of likes shows the approval and popularity of the post. It’s a straightforward measure of content attractiveness.
  • Share: Shares are crucial for measuring content virality. More shares mean the content is compelling enough for users to share with their networks.
  • Total_Interactions: Sum of likes, comments, and shares. This comprehensive metric shows overall engagement and interaction levels with the post, reflecting its success in engaging the audience.

These variables are essential for optimising social media strategies, as they provide insights into content performance, user engagement, and overall effectiveness in reaching and interacting with the target audience.

• The key numerical variables in the dataset show significant variability:

  • Total Reach and Total Impressions: Both have large ranges with Total Reach peaking at 7.29 and Total Impressions at 14.01.
  • Engagement Metrics: Engaged Users, Post Consumers, and Post Consumptions also display wide ranges with maxima of 10.66, 11.89, and 9.14, respectively, indicating high user interaction for some posts.
  • Liked Page Metrics: Impressions and Reach Liked Page reach up to 18.16 and 5.82, while Engaged Liked Page has a maximum of 6.12.
  • Interaction Metrics: Comments, likes, and shares peak at 17.13, 15.39, and 17.88, respectively, with Total Interactions reaching 16.03. These metrics highlight the potential for high engagement and are crucial for optimising social media strategies.

# Summary statistics for numerical variables
summary_stats <- facebook_data_scaled %>%
  select(Total_Reach, Total_Impressions, Engaged_Users, Post_Consumers, Post_Consumptions, 
         Impressions_Liked_Page, Reach_Liked_Page, Engaged_Liked_Page, Comment, Like, Share, Total_Interactions) %>%
  summary()

# Display summary statistics
knitr::kable(summary_stats, caption = "Summary Statistics for Numerical Variables")

Descriptive Statistics and Visualisation

• The box plot shows the distribution of Total Reach for different Post Types on Facebook.

  • Photos: Photos have the highest variability in Total Reach and the most outliers, indicating a wide range of audience reach.
  • Videos: Videos show a significant reach with moderate variability, suggesting they can reach large audiences consistently.
  • Links: Links have a lower median reach and less variability, showing consistent but lower audience engagement compared to photos and videos.
  • Status Updates: Status updates have the lowest median reach and variability, indicating they consistently reach fewer people.

# Boxplot for Total Reach by Post Type
p <- ggplot(facebook_data_scaled, aes(x = Post_Type, y = Total_Reach)) +
  geom_boxplot() +
  theme_minimal() +
  labs(title = "Total Reach by Post Type", x = "Post Type", y = "Total Reach")

# Convert ggplot object to plotly object
interactive_boxplot <- ggplotly(p)
interactive_boxplot

Descriptive Statistics and Visualisation

The graph indicates:

• Skewed Distribution: The histogram of Total Reach is highly skewed to the right, indicating that most posts have a relatively low reach.

• Majority Low Reach: A significant majority of the posts have a Total Reach close to zero, with the highest frequency in the first bin.

• Sparse High Reach: There are only a few posts with a Total Reach greater than 2, indicating that very few posts achieve high reach.

# Histogram for Total Reach
p_hist <- ggplot(facebook_data_scaled, aes(x = Total_Reach)) +
  geom_histogram(binwidth = 0.5, fill = "skyblue", color = "black") +
  theme_minimal() +
  labs(title = "Distribution of Total Reach", x = "Total Reach", y = "Frequency")

# Convert ggplot object to plotly object
interactive_histogram <- ggplotly(p_hist)
interactive_histogram

Descriptive Statistics Cont.

• Highest Mean Reach: Video posts have the highest mean Total Reach (295.86), followed by Status posts (217.04), indicating they generally perform better in terms of reach compared to other post types.

• Wide Range in Photo Posts: Photo posts have a broad range of Total Reach, from 0 to 6334, and a high standard deviation (407.57), suggesting that the performance of photo posts varies widely.

• Consistent Performance for Links: Link posts have the lowest median Total Reach (52.5) and the smallest standard deviation (95.72), indicating more consistent but generally lower performance in terms of reach.

facebook_data %>% group_by(Post_Type) %>% summarise(Min = min(Total_Interactions, na.rm = TRUE),Q1 = quantile(Total_Interactions, probs = .25, na.rm = TRUE), Median = median(Total_Interactions, na.rm = TRUE), Q3 = quantile(Total_Interactions, probs = .75, na.rm = TRUE), Max = max(Total_Interactions, na.rm = TRUE), Mean = mean(Total_Interactions, na.rm = TRUE),SD = sd(Total_Interactions, na.rm = TRUE), n = n(),
Missing = sum(is.na(Total_Interactions))) -> table1
knitr::kable(table1)

Hypothesis Testing

• Purpose: The chi-square test was used to assess whether the observed frequencies of different Facebook post types differ significantly from expected frequencies under the assumption that all post types are equally likely.
• Appropriateness: Since we are dealing with categorical data (Post_Type) and want to test the independence of categories, the chi-square test is suitable. It allows us to determine if the distribution of post types is uniform or if certain types are more prevalent.
• Assumptions: The chi-square test assumes that the expected frequency count for each category should be at least 5. We ensured this assumption was met by checking the frequency counts before conducting the test. \[H_0: \text{Observed frequencies match expected frequencies}\]

\[H_A: \text{Observed frequencies do not match expected frequencies}\]

The chi-squared test statistic is calculated as:

\[\chi^2 = \sum \frac{(O_i - E_i)^2}{E_i}\]

Where: - \(O_i\) is the observed frequency - \(E_i\) is the expected frequency

The chi-squared test resulted in a chi-squared statistic of 957.92 and a p-value of less than 2.2e-16, indicating that we reject the null hypothesis and conclude that the distribution of post types is not equal.

• Test Calculation:
  • The test compares the observed frequencies to the expected frequencies.
  • The Chi-Square statistic (X-squared) is calculated as 957.92 with 3 degrees of freedom (categories - 1).
• P-value Interpretation:
  • Since the p-value is extremely low, we reject the null hypothesis that all post types are equally distributed.
• Conclusion: The significant Chi-Square test result indicates that the distribution of post types on Facebook is not equal. Some post types (like Photos and Videos) are more frequent than others (like Links and Statuses).

# Observed frequencies
observed <- table(facebook_data$Post_Type)
expected <- c(0.25, 0.25, 0.25, 0.25) * sum(observed)
chi_square_test <- chisq.test(observed, p = expected / sum(expected))

chi_square_test

## 
##  Chi-squared test for given probabilities
## 
## data:  observed
## X-squared = 957.92, df = 3, p-value < 2.2e-16

Hypothesis Testing Cont.

For the confidence intervals, we calculated the 95% confidence intervals for the mean reach of each post type. The general formula for the confidence interval is:

\[CI = \bar{x} \pm t_{\alpha/2, n-1} \frac{s}{\sqrt{n}}\]

Where: \(\bar{x}\) is the sample mean, \(t_{\alpha/2, n-1}\) is the t-value for a 95% confidence interval, \(s\) is the sample standard deviation, \(n\) is the sample size

• Video Posts Have the Highest Reach: Among all post types, video posts have the highest mean reach by a significant margin. This suggests that video content is highly effective in engaging a larger audience on Facebook. Brands and social media managers should consider prioritising video content to maximise their reach.

• Photo Posts Perform Better Than Link and Status Posts: Photo posts have a higher mean reach compared to link and status posts, although not as high as video posts. This indicates that visual content, in general, tends to perform better in terms of reach. Incorporating more photo content into the posting strategy could improve engagement.

• Variability in Reach: The confidence intervals for each post type show the range within which the true mean reach is expected to lie with 95% confidence. Video posts have a very wide confidence interval, indicating high variability in their performance. This could be due to a few viral videos skewing the average. On the other hand, link, photo, and status posts have narrower confidence intervals, suggesting more consistent performance.

#Calculate confidence intervals for the mean reach of each post type
conf_intervals <- facebook_data %>%
  group_by(Post_Type) %>%
  summarise(
    Mean_Reach = mean(Total_Reach, na.rm = TRUE),
    CI_Lower = Mean_Reach - qt(0.975, n() - 1) * sd(Total_Reach, na.rm = TRUE) / sqrt(n()),
    CI_Upper = Mean_Reach + qt(0.975, n() - 1) * sd(Total_Reach, na.rm = TRUE) / sqrt(n())
  )
knitr::kable(conf_intervals, caption = "95% Confidence Intervals for Mean Reach by Post Type")

Regression Analysis

• Purpose: The chi-square test was used to assess whether the observed frequencies of different Facebook post types differ significantly from expected frequencies under the assumption that all post types are equally likely.
• Appropriateness: Since we are dealing with categorical data (Post_Type) and want to test the independence of categories, the chi-square test is suitable. It allows us to determine if the distribution of post types is uniform or if certain types are more prevalent.
• Assumptions: The chi-square test assumes that the expected frequency count for each category should be at least 5. We ensured this assumption was met by checking the frequency counts before conducting the test.

Results

• The R-squared value of 0.04044 indicates that approximately 4.04% of the variability in total reach is explained by the type of post. This is a relatively low percentage, suggesting that post type alone does not explain much of the variability in total reach.
• The Adjusted R-squared value of 0.03457, which accounts for the number of predictors in the model, is slightly lower than the R-squared.
• The F-statistic of 6.897 with a p-value of 0.000148 suggests that the overall model is statistically significant, meaning that at least one of the coefficients is significantly different from zero.

# Fit linear regression model
lm_model <- lm(Total_Reach ~ Post_Type, data = facebook_data)

# Summarise the model
summary(lm_model)

## 
## Call:
## lm(formula = Total_Reach ~ Post_Type, data = facebook_data)
## 
## Residuals:
##    Min     1Q Median     3Q    Max 
## -37662 -10154  -8189   -579 167205 
## 
## Coefficients:
##                 Estimate Std. Error t value Pr(>|t|)    
## (Intercept)        18545       4781   3.879 0.000119 ***
## Post_TypePhoto     -5269       4904  -1.074 0.283134    
## Post_TypeStatus    -5466       5833  -0.937 0.349229    
## Post_TypeVideo     32661       9730   3.357 0.000850 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 22420 on 491 degrees of freedom
## Multiple R-squared:  0.04044,    Adjusted R-squared:  0.03457 
## F-statistic: 6.897 on 3 and 491 DF,  p-value: 0.000148

Regression Analysis

• Residuals vs Fitted: Checks the linearity assumption. The red line being relatively flat indicates that the linearity assumption is reasonably met.
• Normal Q-Q: Checks the normality of residuals. Deviations from the line suggest deviations from normality.
• Scale-Location: Checks the homoscedasticity (constant variance) assumption. A flat red line suggests constant variance.
• Residuals vs Leverage: Identifies influential cases. Points outside Cook’s distance lines may be influential.

# Plot the residuals to check assumptions
par(mfrow = c(2, 2))
plot(lm_model)

Regression Analysis Cont

• The coefficients for other post types (Photo, Status, Video) represent the difference in the average total reach compared to the baseline category.
  • Photo: The coefficient is -5269.20, indicating that, on average, posts of type Photo have a total reach that is 5269.20 less than the baseline (Link), but this is not statistically significant (p = 0.283).
  • Status: The coefficient is -5465.70, indicating that, on average, posts of type Status have a total reach that is 5465.70 less than the baseline (Link), but this is not statistically significant (p = 0.349).
  • Video: The coefficient is 32661.12, indicating that, on average, posts of type Video have a total reach that is 32661.12 more than the baseline (Link), and this is statistically significant (p = 0.00085).

# Extract model coefficients and confidence intervals
tidy(lm_model, conf.int = TRUE)

Regression Analysis

• Post Reach Distribution:
  • Video posts exhibit the highest total reach, both in terms of median and variability, with the highest mean reach as indicated by the red dots.
  • Variability in Reach:
  • Photo posts have significant variability, including numerous outliers, suggesting they can achieve both very high and very low reach.
• Comparative Performance:
  • Link and Status posts show lower and more consistent reach, with fewer extreme outliers compared to Photo and Video posts.

# Visualise the relationship
ggplot(facebook_data, aes(x = Post_Type, y = Total_Reach)) +
  geom_boxplot() +
  geom_jitter(width = 0.2, alpha = 0.3) +
  stat_summary(fun = mean, geom = "point", shape = 20, size = 4, color = "red") +
  labs(title = "Total Reach by Post Type", x = "Post Type", y = "Total Reach")

Re-modelling Regression

• By applying a logarithmic transformation to Total_Reach and re-running the regression analysis.
• Observing the results the following is observed:

  • Residual Standard Error: The residual standard error has decreased from 22420 to 1.097, indicating that the transformed model fits the data better by reducing the error.

  • R-squared and Adjusted R-squared: The R-squared value has increased from 0.04044 to 0.0671, and the adjusted R-squared value has also increased from 0.03457 to 0.0614. Although these values are still relatively low, the increase suggests an improved fit.

  • Overall Model Significance: The F-statistic has improved from 6.897 to 11.77, with a corresponding p-value of 1.854×10^-7, indicating that the overall model is statistically significant.

# Apply logarithmic transformation
facebook_data$log_Total_Reach <- log(facebook_data$Total_Reach + 1)

# Fit the regression model with transformed data
regression_model_transformed <- lm(log_Total_Reach ~ Post_Type, data = facebook_data)

# Display the summary of the transformed regression model
summary(regression_model_transformed)

## 
## Call:
## lm(formula = log_Total_Reach ~ Post_Type, data = facebook_data)
## 
## Residuals:
##     Min      1Q  Median      3Q     Max 
## -3.2381 -0.6289 -0.2157  0.4788  3.3888 
## 
## Coefficients:
##                 Estimate Std. Error t value Pr(>|t|)    
## (Intercept)       9.1815     0.2339  39.259  < 2e-16 ***
## Post_TypePhoto   -0.4670     0.2399  -1.947  0.05216 .  
## Post_TypeStatus   0.2256     0.2854   0.791  0.42955    
## Post_TypeVideo    1.3045     0.4760   2.740  0.00636 ** 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 1.097 on 491 degrees of freedom
## Multiple R-squared:  0.0671, Adjusted R-squared:  0.0614 
## F-statistic: 11.77 on 3 and 491 DF,  p-value: 1.854e-07

# Plot diagnostic plots for the transformed regression model
par(mfrow = c(2, 2))
plot(regression_model_transformed)

Discussion

• Major Findings

  • Post Type Performance: Video posts achieved the highest average total reach, making them the most effective for audience engagement on Facebook. Photo posts also performed well but showed more variability.
  • Engagement Metrics: Metrics like Total Reach, Total Impressions, Engaged Users, and Total Interactions varied by post type, with videos and photos generally attracting more engagement than links and status updates.
  • Chi-Square Test Results: The test showed significant differences between the observed and expected frequencies of post types, indicating varying success rates.

• Strengths:

  • Comprehensive Dataset: The complete dataset with no missing values provided robust data for analysis.
  • Statistical Rigor: The use of descriptive statistics, hypothesis testing, and regression analysis gave meaningful insights.

• Limitations:

  • Single Brand Focus: The analysis was limited to one cosmetics brand, affecting generalisability.
  • Temporal Scope: Data from 2014 may not reflect current social media trends.
  • Simplified Variables: The analysis did not include factors like content quality and user sentiment.

Future Improvements & Conclusion

• Future Improvements
  • Broader scope: Include multiple brands and industries to improve generalisability.
  • Longitudinal analysis: Study changes in post effectiveness over time.
  • Advanced metrics: Consider additional variables like content quality and user sentiment.
  • Platform Comparison: Compare performance across different social media platforms.
• Final Conclusion

The analysis addressed the problem statement by identifying key predictors of post performance, specifically highlighting the effectiveness of video content. The hypothesis testing confirmed that different post types have significantly different engagement levels. Investing in high-quality video content can significantly enhance a brand’s reach and engagement on Facebook, making it a key component of effective social media strategies. The main takeaway is clear: Prioritise video content to maximise engagement and visibility on Facebook

References

• Class Modules 1-9

• Understanding How Your Videos Perform on Facebook Www.facebook.com. Retrieved June 13, 2024, from https://www.facebook.com/formedia/blog/understanding-how-your-videos-perform-on-facebook#:~:text=Metrics%20are%20available%20at%20the

• Teves, C. (2022, April 21). Facebook Analytics: how to analyze your data. Sprout Social. https://sproutsocial.com/insights/facebook-analytics/

• ChatGPT https://chatgpt.com/

• DataFlair Team. (2017, June 30). Introduction to Hypothesis Testing in R - Learn every concept from Scratch! - DataFlair. DataFlair. https://data-flair.training/blogs/hypothesis-testing-in-r/

• Calculating Confidence Intervals — R Tutorial. (n.d.). Www.cyclismo.org. https://www.cyclismo.org/tutorial/R/confidence.html

• Frost, J. (2018). How To Interpret R-squared in Regression Analysis. Statistics by Jim; Statistics By Jim. https://statisticsbyjim.com/regression/interpret-r-squared-regression/

• Keyhole. (2023, November 24). Facebook Analytics Guide 2024: How To Analyze & Use Your FB Data. Keyhole. https://keyhole.co/blog/facebook-analytics/

• Dopson, E. (2021, March 13). Videos vs. Images: Which Drives More Engagement in Facebook Ads? | Databox Blog. Databox. https://databox.com/videos-vs-images-in-facebook-ads