by Dorian’s Gs
superbowl is created by FiveThirtyEight, originally
collected through superbowl-ads.com. FiveThirtyEight is an American
Website that reports how statistics are relevant in current events.
There are 247 different Super Bowl commercials with 25 different
variables, but the FiveThirtyEight team observed only 10 different
brands. The dataset includes defining characteristics of Super Bowl ads
from these 10 popular brands, two of which are Toyota and Bud Light.
There are different types of variables included in
Variables like views, number of likes, and number of dislikes are
numeric values based off of YouTube statistics. Variables like brand,
title, and description act as identifiers for individual ads. There were
also 7 different ad categories, and a corresponding variable saying
whether the ad does or does not belong to the category. These logical
variables were determined by the FiveThirtyEight team as they watched
and categorized all of the advertisements. 3 of these 7 variables
include whether the ad is funny, patriotic, and/or includes animals. An
ad can belong to more than one of these 7 different categories.
Using the dataset
superbowl, we answer the following questions:
1. What factors contribute to a greater number of ad views and has the relationship between those factors and the views changed over time?
The relevant variables include:
animals: whether or not the ad contains animals
show_product_quickly: whether or not the ad show a product quickly
funny: whether or not the ad contains humor
year: the year the ad was released
brand: the brand that ran the ad
More specifically, we investigate how companies choose to include animals, humor, and show their product quickly. Even though the dataset provides a lot more variables related to ad characteristics, we chose to focus on these three because they occurred more than any of the other ad types within the dataset, leading us to believe that companies cared the most about these three characteristics. For the second part of the question, we primarily investigate how the trends and relationships we explored in the first part have changed over the years.
We were interested in this question because we have observed how influential ads can be if they include characteristics that appeal to our emotions, such as our adoration for a cute puppy, longing to be like a popstar, or simply our sexual appetite. However, we were curious to see if companies, the ones who provided the ads, thought the same as us. Emotional appeal would foster more views and we think it is important for advertisement agencies to know. Additionally, we observed that certain ad appeals, especially use of sexuality, have grown in popularity over the years, and would like to observe such trends with our data analysis.
Our initial plan was to create visualizations centered around the
relationship between the ad type (i.e., what characteristics an ad
possesses) and view count. After transforming and cleaning the data, we
ran into the issue of visualizations being overplotted, no distinct
relationships appearing, and outliers contributing to messy
visualizations (e.g., a Doritos ad from 2012 had greater than six times
the number of views than any other ad). To address these issues, we
shifted our focus to looking at commercials from a brand standpoint, as
well as decided to quantify ad usage by frequency (number of ads) and
not view count. To accomplish the latter, we added a new variable
to the data frame that was the number of occurrences of each ad type in
a year. To show the relationship between brand and ad type, we displayed
brand and ad type in a segmented bar chart. The segments showed how each
category was used by the ten brands in the dataset. To answer the part
of the question related to how the ad types have fared from 2000 to
2020, we looked at the frequency of the ad types over time, showing this
relationship in a line graph with the year on the x-axis and ad
occurrence on the y-axis.
We chose a bar chart as they are effective at displaying two variables, one continuous and the other discrete. It is easy to compare between bars, based on relative height of the bars and/or relative bar segments. For the other visualization for this question, we opted for a line chart as we were mapping a discrete variable over time. We eventually settled on using estimates of the means for each ad type as there was overplotting and using a line chart led to messy visualizations that were difficult to interpret.
superbowl_viz <- superbowl %>% select(year, brand, animals, celebrity, use_sex, funny, show_product_quickly, patriotic, danger) %>% pivot_longer(cols = c(-year, -brand), names_to = 'ad_type', values_to = 'have') %>% filter(have == TRUE) %>% count(brand, year, ad_type, name = "ads") %>% #group_by(brand, year, ad_type) %>% #summarise(ads = n()) %>% mutate(brand = if_else(brand == 'Hynudai', 'Hyundai', as.character(brand))) %>% mutate(category = case_when(brand == 'Bud Light' ~ 'Alcohol', brand == 'Budweiser' ~ 'Alcohol', brand == 'Doritos' ~ 'Food & Beverage', brand == 'Pepsi' ~ 'Food & Beverage', brand == 'Coca-Cola' ~ 'Food & Beverage', brand == 'Hyundai' ~ 'Automobile', brand == 'Toyota' ~ 'Automobile', brand == 'Kia' ~ 'Automobile', brand == 'E-Trade' ~ 'Financial Services', brand == 'NFL' ~ 'Sports')) %>% relocate(year, .before = brand) %>% relocate(category, .before = brand) #Brand Total Labels brand_totals <- superbowl_viz %>% filter(!(ad_type %in% c("celebrity", "use_sex", "danger", "patriotic"))) %>% count(brand, name = "brand_totals") #group_by(brand) %>% #summarise(brand_totals = n()) brand_viz <- superbowl_viz %>% filter(!(ad_type %in% c("celebrity", "use_sex", "danger", "patriotic"))) %>% count(brand, ad_type, name = "ads") %>% #group_by(brand, ad_type) %>% #summarise (ads = n()) %>% ggplot(aes(factor(brand, levels = c("Bud Light", "Budweiser", "Doritos", "Pepsi", "Coca-Cola", "Hyundai", "E-Trade", "Toyota", "Kia", "NFL")), ads, fill = ad_type)) + geom_bar(stat = "identity", position = position_stack(0.1)) + geom_text(data = brand_totals, aes(x = brand, y = brand_totals, label = brand_totals, fontface = 3, fill = NULL), size = 5, nudge_y = 2) + scale_y_continuous(breaks = seq(0, 60, 10)) + scale_fill_manual(values = c("#56B4E9", "#E69F00", "#009E73"), labels = c("Shows Product Quickly", "Funny", "Animals")) + labs(title = "Ad Occurrences by Brand", y = "Ad Occurrences", x = "Brand", fill = "Ad Type", caption = "Source: FiveThirtyEight") + theme_minimal() + theme(plot.title = element_text(face = "bold", size = 18, hjust = 0.5, vjust = 1), plot.caption = element_text(size = 12, hjust = 1, vjust = 6), legend.position = c(0.85, 0.7), axis.text.x = element_blank(), strip.text = element_blank(), axis.title.y = element_text(vjust = 1, size = 16), axis.text.y = element_text(hjust = 1.75), axis.title.x = element_text(vjust = 0, size = 16), axis.text = element_text(size = 13, lineheight = 1), legend.title = element_text(size = 16), legend.text = element_text(size = 13)) images <- axis_canvas(brand_viz, axis = 'x') + draw_image("https://1000logos.net/wp-content/uploads/2021/04/Bud-Light-logo-768x432.png", x = 0.5, scale = 0.75) + draw_image("https://logos-world.net/wp-content/uploads/2020/12/Budweiser-Logo-2016-present.jpg", x = 1.5, scale = 1.1) + draw_image("https://1000logos.net/wp-content/uploads/2020/07/Doritos-Logo-768x480.png", x = 2.5, scale = 1.1) + draw_image("https://1000logos.net/wp-content/uploads/2017/05/Pepsi-Logo-768x631.png", x = 3.5, scale = 0.9) + draw_image("https://1000logos.net/wp-content/uploads/2021/05/Coca-Cola-logo-768x432.png", x = 4.5, scale = 1) + draw_image("https://1000logos.net/wp-content/uploads/2018/04/Hyundai-logo-768x432.png", x = 5.5, scale = 1.1) + draw_image("https://res.cloudinary.com/crunchbase-production/image/upload/c_lpad,h_256,w_256,f_auto,q_auto:eco,dpr_1/v1476500866/ukg833gylj7adlryeuig.png", x = 6.5, scale = 1) + draw_image("https://1000logos.net/wp-content/uploads/2021/04/Toyota-logo-768x432.png", x = 7.5, scale = 0.8) + draw_image("https://1000logos.net/wp-content/uploads/2020/02/Kia-Logo-2012-1024x635.jpg", x = 8.5, scale = 0.9) + draw_image("https://1000logos.net/wp-content/uploads/2017/05/NFL-logo-768x518.png", x = 9.5, scale = 1.1) ggdraw(insert_xaxis_grob(brand_viz, images, position = "bottom"))
ad_labels <- superbowl %>% pivot_longer(cols = c(animals, celebrity, use_sex, funny, show_product_quickly, patriotic, danger), names_to = 'ad_type', values_to = 'have') %>% filter(have == TRUE, !(ad_type %in% c("celebrity", "use_sex", "danger", "patriotic"))) %>% count(ad_type, year, name = "ads") %>% #group_by(ad_type, year) %>% #summarise(ads = n()) %>% filter(year == 2005) %>% mutate(ads = case_when(ads == 6 ~ 4.1, ads == 11 ~ 9.9, ads == 8 ~ 8.5))
# message related to geom_smooth() method and formula, # which are both what we want superbowl %>% pivot_longer(cols = c(animals, celebrity, use_sex, funny, show_product_quickly, patriotic, danger), names_to = 'ad_type', values_to = 'have') %>% filter(have == TRUE, !(ad_type %in% c("celebrity", "use_sex", "danger", "patriotic"))) %>% count(year, ad_type, name = "ads") %>% #group_by(year, ad_type) %>% #summarise(ads = n()) %>% ggplot(aes(year, ads, color = ad_type)) + geom_line(aes(linetype = ad_type), stat = "smooth", size = 3, se = F, show.legend = FALSE) + geom_text_repel(data = ad_labels, aes(label = c("Shows Product Quickly", "Funny", "Animals")), nudge_x = 0.10, nudge_y = -0.8, size = 5, show.legend = FALSE) + scale_color_manual(values = c("#56B4E9", "#E69F00", "#009E73"), labels = c("Shows Product Quickly", "Funny", "Animals"), guide = guide_legend(reverse = TRUE)) + scale_y_continuous(breaks = seq(0, 13, 2)) + scale_x_continuous(breaks = seq(2000, 2020, 4)) + theme_minimal() + theme(plot.title = element_text(face = "bold", size = 18, hjust = 0.5, vjust = 1), plot.subtitle = element_text(size = 16, hjust = 0.5, vjust = 1), plot.caption = element_text(size = 12, hjust = 1), axis.title.x = element_text(vjust = -2, size = 16), axis.title.y = element_text(vjust = 1, size = 16), axis.text = element_text(size = 13, lineheight = 1)) + labs(y = "Ad Occurrences", x = "Year", color = "Ad Type", caption = "Source: FiveThirtyEight", title = "Ad Occurrences by Type", subtitle = "From 2000 - 2020")
In the bar plot, it is clear that the most common ad types are ”Animals” and “Funny.” All of these companies aired ads in these categories (funny, animals, and shows product quickly), which is indicative of their effectiveness in the advertisement industry. It is clear that beer companies, such as Bud Light or Budwiser, have more advertisements in all three categories when compared to car companies such as Kia or Toyota. One interesting observation is that the only company with no ads in the “Shows Product Quickly” category is the NFL. This makes sense because the NFL hosts the Super Bowl and does not have to worry about paying for extra runtime. One limitation with our plot is that the categories “double count” ads. If an ad is both funny and contains animals, then it will look as if there were two different ads, one funny and one with animals. Fortunately, the main goal of this plot was to focus on which ad types were popular and the colored stacked bar plots accomplish this goal.
In the line plot, there are three smooth curves that graph each ad type and show the trend of their yearly occurrences over time. Lines are defined by different colors, labels, and type so it is easier to distinguish between them. The smooth curves allow the viewer to see overall trends more easily. Funny ads peak in occurrence in the mid-2000s, but then drop off all the way into the present. Both ads that show the product quickly and have animals were on an uptrend, but dropped off in the late 2000s/early 2010s. Just like the bar plot, it is clear that animal ads and funny ads were the most common types of ads. One limitation of this plot is that we plotted an estimate of the conditional mean function for each of the three ad types. This means that we are not able to look directly at a year and see the exact number occurrences of a specific ad-type. However, the original line plot had issues with overplotting; the lines for each of the three line types overlapped, and it was extremely difficult to see the existence of trends over time. The smooth lines used in this visualization are more appropriate as they allow us to accomplish our goal of seeing overall trends over time, which is what we aimed to visualize.
2. What is the relationship between popularity of a Superbowl Ad and how much interaction it got?
The relevant variables are all numerical and include:
view_count: number of views
like_count: number of likes
dislike_count: number of dislikes
comment_count: number of comments
view_count is the measure for popularity, while the other
variables define interaction with the ad. We were interested in
examining the relationship between different popularity and interaction
metrics of the advertisements because we want to more clearly understand
what a “successful ad” entails. Some agencies could prefer more comments
to an ad to show that consumers are engaging with the content, but that
might be more heavily associated with the like to dislike ratio instead
of the view count. If so, that agency would create a commercial that
fosters more likes than pure views.
For the following two plots, the range of
view count was so large
(from 10 views to more than 10 million views) that it was nearly
impossible to visualize any interaction with just one plot. We decided
to facet our plot by 6 categories according to the quantiles of the
standard distribution for view count: few, some, moderate, many, high,
and viral views. In the first plot, we examined the relationship between
the view count and the ratio of likes. In order to plot the like
percentage, we created a new variable
ratio which is the
divided by the sum of
like_count and dislike_count, and plotted the
x-axis labels to show percentage (since the like ratio on Youtube is
shown in percentage form). We chose a ridge plot because it clearly
shows trends in like percentage while comparing these changes by the
view count category, and the peaks of the density ridges stacked
vertically make comparisons between view categories simple to track.
For the second plot, we wanted to look more specifically at the
relationship between view count and the number of interactions a video
gets. We defined interactions as the number of likes, dislikes and
comments a video got. We wanted to plot both
interactions as continuous variables. In order to find the best plot,
we experimented with scatter plots, line plots, step plots, and area
plots to see which plot was the most informative. The view counts were
clustered in the thousands of views but contained very few videos with
millions of views, so the scatter plot was not helpful. We found that a
line plot or step by itself was too fragmented and the connection
between points zig-zagged a lot and made it difficult to see the overall
trend. The combined plot shows each point but the area also makes it
much easier to trace the trend for each view category. We then added
labels, captions and made a few styling choices to get to our final
superbowl <- superbowl %>% mutate(ratio = like_count / (like_count + dislike_count), interactions = like_count + dislike_count + comment_count, view_category = case_when( view_count < 4000 ~ "Few\nLess than 4K", view_count >= 4000 & view_count < 30000 ~ "Some\n4K to 30K", view_count >= 30000 & view_count < 90000 ~ "Moderate\n30K to 90K", view_count >= 90000 & view_count < 500000 ~ "Many\n90K to 500K", view_count >= 500000 & view_count < 10000000 ~ "High\n500K to 10M", TRUE ~ "Viral\n10M+")) superbowl %>% mutate(Views = fct_relevel( view_category, "Viral\n10M+", "High\n500K to 10M", "Many\n90K to 500K", "Moderate\n30K to 90K","Some\n4K to 30K","Few\nLess than 4K" )) %>% ggplot(aes(x = ratio, y = Views, fill = Views)) + geom_density_ridges(scale = 0.9, show.legend = FALSE) + coord_cartesian(xlim = c(0.5, 1.0)) + scale_fill_discrete_sequential(palette = "Greens", order = c(6:1)) + scale_x_continuous(expand = c(0, 0.1), labels = scales::percent_format(accuracy = 1)) + scale_y_discrete(expand = expand_scale(mult = c(0, .2))) + labs( title = "Youtube Superbowl Ads and Like Percentages", x = "Like Percentage", y = "Views", caption = "(Percentage of likes to sum of likes and dislikes on the video)") + theme_ridges() + theme( plot.title.position = "plot", plot.title = element_text(face = "bold", size = 16, hjust = 0.36), plot.caption = element_text(size = 12), axis.title.y = element_text(face = "bold", size = 14, hjust = 0.54, vjust = 1), axis.title.x = element_text(face = "bold", size = 14, hjust = 1, vjust = 0), axis.text = element_text(size = 12, lineheight = 1), axis.text.x = element_text(vjust = -1))
# Problem with the order of the facet again! superbowl <- superbowl %>% mutate(view_category = paste(view_category, "Views")) superbowl <- superbowl %>% mutate(view_category = fct_relevel( view_category, "Few\nLess than 4K Views", "Some\n4K to 30K Views", "Moderate\n30K to 90K Views","Many\n90K to 500K Views", "High\n500K to 10M Views", "Viral\n10M+ Views")) ggplot(data = superbowl, aes(x = view_count, y = interactions)) + geom_area(aes(fill = view_category), show.legend = NULL) + geom_line() + facet_wrap(vars(view_category), scales ="free", nrow = 3, strip.position = "top", labeller = labeller("label_both")) + scale_y_continuous(label = label_number_si()) + scale_x_continuous(label = label_number_si()) + scale_fill_discrete_sequential(palette = "Greens") + labs(x = "View Count", y ="Number of Interactions", title ="Assessing Superbowl Ad Interactions", caption ="Interaction is defined by the sum of \nlikes, dislikes and comments on a video")+ theme_minimal() + theme(panel.grid.minor = element_blank(), strip.background = element_rect(colour = "black"), strip.placement = "inside", plot.title = element_text(face = "bold", size = 16, hjust = 0.5, vjust = 1), plot.caption = element_text(size = 12, hjust = 0), panel.spacing = unit(1.4, "lines"), axis.text.x = NULL, axis.title.x = element_text(size = 14, vjust = -.5), axis.title.y = element_text(size = 14, hjust = 0.5, vjust = 1), axis.text = element_text(size = 14, lineheight = 1), strip.text.x = element_text(size = 12))
In the ridge plot, most videos had a like percentage of over 60%, showing that most of the viewers who interacted with the ad enjoyed the video. An interesting finding was that there was generally a negative relationship between views and like percentage. That is, the category with the fewest video views peaked close to 100% likes and was more precisely close to 100 compared to higher view categories. Interestingly, the viral videos peaked at close to 75% likes (with a smaller peak at 90%) and is the most spread out with like percentages. The other categories from some to high views, peak at similar percentages with high views having slightly lower like percentages. Overall, this plot shows that having more views does not necessarily mean that the video is on average more likable. Instead it shows the opposite: higher view counts would predict a lower like percentage, meaning that companies who value video satisfaction rates (and want to avoid high dislikes ratios) would not want a viral ad video.
The area plot of view count by number of interactions reveals quite a few interesting results. The first is that on average, the number of views for ads is larger by a factor of about 100 that the number of interactions on that ad. On average, about 100 views resulted in 1 interaction with the video. Our guess prior to seeing the results of our visualization was that as the number of views increased, the number of proportions would also increase on a statistically significant rate. In terms of our data that means that we expected each higher category of views would have a significantly high level of interaction. As an example, we expected that proportionally, the number of view count to interactions would be higher for ads with “Viral Views” compared to “Mega Views” and ads with “Mega Views” would be higher that graphs with “Many Views”. Our results turned out to be a bit more complicated than that. We found that only ads on either extreme end followed this trend. So ads with a “few views” had much a smaller views to interaction ratio than videos in all other categories and ads with “viral views” had much bigger views to interaction ratio than videos in all other “categories.” For videos in the middle categories(“Some Views”,“Moderate Views”,“Many Views” and “High Views”), proportion of view counts to interactions seemed to be much more random and didn’t follow any discernible patterns. Each panel also had no discernible trend in the way they were plotted. Only videos that are in the viral category show a discernibly linear relationship between the number of views and number of interactions. The caveat for that seemingly linear trend is that the viral category has the least number of data points so it is possible that with more data points, the linear relationship may not be present.
Our presentation can be found here.
Best R and Mehta D, 2021, Super Bowl Ads, Electronic Dataset, Github Repository, Retrieved September 13, 2021 https://github.com/fivethirtyeight/superbowl-ads.
- Data Source: Our data source is from The Tidy Tuesday Project by way of FiveThirtyEight’s Github Repository.
- Viral Reference: https://www.forbes.com/sites/robertwynne/2018/03/09/there-are-no-guarantees-or-exact-statistics-for-going-viral/?sh=6224f76a5e8c
- Title image: https://www.pinterest.com/pin/746542075709663730/
- Youtube logo: Youtube
- Tv image: FiveThirtyEight
- Logos directly from: Budweiser, Bud Light, Pepsi, Doritos, Coca-Cola, Hyundai, E Trade, Kia, Toyota, and NFL
- Youtube interaction bar: https://www.tubefilter.com/2021/03/30/youtube-experiment-hiding-public-dislikes-creator-well-being/