Analyzing Team Branding Language and Its Relationship to Performance, Attendance, and Fan Engagement
NBA teams use words like “dynasty,” “contender,” “young core,” and “rebuilding” to describe themselves. As someone who watches a lot of NBA basketball, I wanted to know: do these words actually mean anything? Do teams that call themselves contenders actually win more? Do fans show up more for those teams?
In this project, I looked at all 30 NBA teams and checked whether the language they use in their branding is connected to:
I followed the OSEMN workflow for this project:
Here I load all the R packages I need. Each one does a specific job.
library(tidyverse) # data cleaning and wrangling
library(rvest) # web scraping
library(ggplot2) # charts
library(ggrepel) # smart text labels on charts (feature not covered in class)
library(knitr) # tables
library(scales) # number formatting
library(broom) # clean model output
library(wordcloud2) # word cloud (feature not covered in class)I used two different types of data sources for this project:
rvest# I scraped the 2023-24 NBA standings from Basketball Reference
url <- "https://www.basketball-reference.com/leagues/NBA_2024_standings.html"
scraped <- read_html(url) |>
html_element("#confs_standings_E") |> # grab the Eastern Conference table
html_table()
# Clean up the scraped table
scraped_clean <- scraped |>
janitor::clean_names() |>
select(team = eastern_conference, w, l) |>
filter(!str_detect(team, "Atlantic|Central|Southeast")) |> # remove division headers
mutate(
team = str_remove(team, "\\*"), # remove asterisks from playoff teams
win_pct = w / (w + l) # calculate win percentage
)
head(scraped_clean)# A tibble: 6 × 4
team w l win_pct
<chr> <int> <int> <dbl>
1 Boston Celtics 64 18 0.780
2 New York Knicks 50 32 0.610
3 Milwaukee Bucks 49 33 0.598
4 Cleveland Cavaliers 48 34 0.585
5 Orlando Magic 47 35 0.573
6 Indiana Pacers 47 35 0.573Challenge I ran into: Basketball Reference puts an asterisk (
*) next to playoff teams. When I first tried to join this data with my other files, the team names didn’t match because of that extra character. I fixed it by usingstr_remove()to strip the asterisk out before doing the join.
# I load my pre-collected CSV files here
teams <- read_csv("team_descriptions.csv")
attendance <- read_csv("nba_attendance.csv")
social <- read_csv("social_media.csv")
market <- read_csv("market_size.csv")
head(teams)# A tibble: 6 × 2
team description
<chr> <chr>
1 Atlanta Hawks The Hawks are a young and emerging team built around dyna…
2 Boston Celtics The Celtics are a championship contender with an elite ro…
3 Brooklyn Nets The Nets are in transition, retooling their roster and re…
4 Charlotte Hornets Charlotte is building around a young core with emerging t…
5 Chicago Bulls The Bulls are a young and developing team with an excitin…
6 Cleveland Cavaliers Cleveland is an emerging contender, developing a talented…My social media file had separate columns for Instagram and Twitter followers. I reshaped it from wide format (one column per platform) to long format (one row per platform) so I could plot both platforms side by side easily.
# Before: each platform is its own column
head(social)# A tibble: 6 × 5
team instagram_followers_m twitter_followers_m total_followers
<chr> <dbl> <dbl> <dbl>
1 Atlanta Hawks 4.2 2.1 6.3
2 Boston Celtics 9.8 5.6 15.4
3 Brooklyn Nets 7.1 4.2 11.3
4 Charlotte Hornets 3.4 2 5.4
5 Chicago Bulls 6.8 4.1 10.9
6 Cleveland Cavaliers 4.1 2.3 6.4
# ℹ 1 more variable: avg_engagement_rate <dbl># I transform it to long format
social_long <- social |>
pivot_longer(
cols = c(instagram_followers_m, twitter_followers_m),
names_to = "platform",
values_to = "followers_m"
) |>
mutate(platform = str_replace(platform, "_followers_m", "") |>
str_replace("instagram", "Instagram") |>
str_replace("twitter", "Twitter/X"))
head(social_long)# A tibble: 6 × 5
team total_followers avg_engagement_rate platform followers_m
<chr> <dbl> <dbl> <chr> <dbl>
1 Atlanta Hawks 6.3 3.8 Instagram 4.2
2 Atlanta Hawks 6.3 3.8 Twitter/X 2.1
3 Boston Celtics 15.4 5.2 Instagram 9.8
4 Boston Celtics 15.4 5.2 Twitter/X 5.6
5 Brooklyn Nets 11.3 4.6 Instagram 7.1
6 Brooklyn Nets 11.3 4.6 Twitter/X 4.2I read each team’s description and counted how many times certain keywords appeared. I then assigned each team to whichever category had the most keyword matches.
| Group | What it means | Example words |
|---|---|---|
| Elite / Contender | Team is going for a title | dynasty, championship, elite |
| Young / Building | Team is growing | young, future, developing |
| Rebuilding | Team is starting over | rebuilding, reset, transition |
| Veteran / Proven | Team relies on experience | veteran, proven, battle-tested |
# I define the keywords for each group
elite_words <- c("dynasty", "contender", "championship", "title", "elite", "dominant")
young_words <- c("young", "future", "developing", "emerging", "next generation")
rebuild_words <- c("rebuilding", "retooling", "reset", "transition")
veteran_words <- c("experienced", "veteran", "proven", "battle-tested")
# This function counts how many keywords from a list appear in a description
count_matches <- function(text, words) {
str_count(str_to_lower(text), str_c(words, collapse = "|"))
}
# I score each team and assign it a category
teams <- teams |>
mutate(
score_elite = map_int(description, ~count_matches(.x, elite_words)),
score_young = map_int(description, ~count_matches(.x, young_words)),
score_rebuild = map_int(description, ~count_matches(.x, rebuild_words)),
score_veteran = map_int(description, ~count_matches(.x, veteran_words)),
category = pmap_chr(
list(score_elite, score_young, score_rebuild, score_veteran),
~c("Elite/Contender", "Young/Building",
"Rebuilding", "Veteran/Proven")[which.max(c(...))]
)
)
# Here I check how many teams ended up in each group
teams |>
count(category) |>
kable(col.names = c("Branding Category", "Number of Teams"))| Branding Category | Number of Teams |
|---|---|
| Elite/Contender | 10 |
| Rebuilding | 4 |
| Veteran/Proven | 2 |
| Young/Building | 14 |
I joined all five data sources into one master table so I could compare everything at once.
df <- teams |>
left_join(scraped_clean, by = "team") |> # web-scraped standings
left_join(attendance, by = "team") |>
left_join(social, by = "team") |>
left_join(market, by = "team")
glimpse(df)Rows: 30
Columns: 19
$ team <chr> "Atlanta Hawks", "Boston Celtics", "Brooklyn Net…
$ description <chr> "The Hawks are a young and emerging team built a…
$ score_elite <int> 0, 7, 0, 0, 0, 2, 5, 4, 0, 6, 0, 1, 4, 4, 0, 5, …
$ score_young <int> 6, 0, 0, 4, 3, 4, 0, 0, 3, 0, 4, 3, 0, 0, 0, 0, …
$ score_rebuild <int> 0, 0, 3, 0, 0, 0, 0, 0, 3, 0, 2, 0, 0, 0, 0, 0, …
$ score_veteran <int> 0, 0, 0, 0, 0, 0, 1, 1, 0, 2, 0, 0, 4, 2, 4, 4, …
$ category <chr> "Young/Building", "Elite/Contender", "Rebuilding…
$ w <int> 36, 64, 32, 21, 39, 48, NA, NA, 14, NA, NA, 47, …
$ l <int> 46, 18, 50, 61, 43, 34, NA, NA, 68, NA, NA, 35, …
$ win_pct <dbl> 0.4390244, 0.7804878, 0.3902439, 0.2560976, 0.47…
$ avg_attendance <dbl> 17421, 19156, 17562, 15901, 18014, 18177, 19354,…
$ arena_capacity <dbl> 20000, 19156, 19000, 20491, 20917, 19432, 19200,…
$ pct_capacity <dbl> 87.1, 100.0, 92.4, 77.6, 86.1, 93.5, 100.8, 96.7…
$ instagram_followers_m <dbl> 4.2, 9.8, 7.1, 3.4, 6.8, 4.1, 9.4, 5.8, 2.9, 18.…
$ twitter_followers_m <dbl> 2.1, 5.6, 4.2, 2.0, 4.1, 2.3, 4.8, 3.2, 1.8, 9.4…
$ total_followers <dbl> 6.3, 15.4, 11.3, 5.4, 10.9, 6.4, 14.2, 9.0, 4.7,…
$ avg_engagement_rate <dbl> 3.8, 5.2, 4.6, 3.2, 4.8, 3.5, 5.0, 4.1, 2.9, 6.8…
$ metro_pop <dbl> 6144050, 4941632, 19768458, 2701392, 9557909, 36…
$ team_age <dbl> 1996, 1946, 2012, 1988, 1966, 1970, 1980, 1967, …Before running any tests, I looked at the basic averages for each group.
df |>
group_by(category) |>
summarise(
Teams = n(),
`Avg Win %` = percent(mean(win_pct, na.rm = TRUE), accuracy = 0.1),
`Avg Attendance` = comma(round(mean(avg_attendance, na.rm = TRUE))),
`Avg Followers` = paste0(round(mean(total_followers, na.rm = TRUE), 1), "M")
) |>
kable(caption = "My summary statistics by branding category")| category | Teams | Avg Win % | Avg Attendance | Avg Followers |
|---|---|---|---|---|
| Elite/Contender | 10 | 62.8% | 18,597 | 16M |
| Rebuilding | 4 | 34.8% | 17,844 | 10.1M |
| Veteran/Proven | 2 | 61.0% | 18,523 | 12.7M |
| Young/Building | 14 | 40.7% | 16,289 | 6.8M |
I made this bar chart to see at a glance which branding group wins the most games on average.
category_colors <- c(
"Elite/Contender" = "#C9A84C",
"Veteran/Proven" = "#1D428A",
"Young/Building" = "#27AE60",
"Rebuilding" = "#C8102E"
)
win_summary <- df |>
group_by(category) |>
summarise(avg_win_pct = mean(win_pct, na.rm = TRUE))
ggplot(win_summary, aes(x = reorder(category, avg_win_pct),
y = avg_win_pct, fill = category)) +
geom_col(show.legend = FALSE, width = 0.6) +
scale_fill_manual(values = category_colors) +
scale_y_continuous(labels = percent_format()) +
coord_flip() +
labs(title = "Average Win % by Branding Category",
subtitle = "NBA seasons 2019–2024",
x = NULL, y = "Win Percentage") +
theme_minimal(base_size = 13)
I wanted to see if fans show up more for teams that use “elite” language.
attend_summary <- df |>
group_by(category) |>
summarise(avg_attendance = mean(avg_attendance, na.rm = TRUE))
ggplot(attend_summary, aes(x = reorder(category, avg_attendance),
y = avg_attendance, fill = category)) +
geom_col(show.legend = FALSE, width = 0.6) +
scale_fill_manual(values = category_colors) +
scale_y_continuous(labels = comma_format()) +
coord_flip() +
labs(title = "Average Home Attendance by Branding Category",
subtitle = "Excludes 2020–21 bubble season",
x = NULL, y = "Average Fans per Game") +
theme_minimal(base_size = 13)
Here I used my reshaped long-format data to compare Instagram and Twitter followers side by side across branding categories.
social_long |>
left_join(select(teams, team, category), by = "team") |>
group_by(category, platform) |>
summarise(avg_followers = mean(followers_m, na.rm = TRUE), .groups = "drop") |>
ggplot(aes(x = category, y = avg_followers, fill = platform)) +
geom_col(position = "dodge", width = 0.6) +
scale_fill_manual(values = c("Instagram" = "#C9A84C", "Twitter/X" = "#1D428A")) +
labs(title = "My Followers Comparison by Category and Platform",
subtitle = "Instagram vs. Twitter/X (millions)",
x = NULL, y = "Avg Followers (M)", fill = "Platform") +
theme_minimal(base_size = 13)
I used ggrepel here — a package I discovered on my own that was not covered in class. It automatically moves the team name labels so they do not overlap each other, which makes the chart much easier to read.
ggplot(df, aes(x = score_elite, y = win_pct,
color = category, label = team)) +
geom_point(size = 3) +
geom_text_repel(size = 2.8, max.overlaps = 20) +
geom_smooth(method = "lm", se = TRUE,
color = "grey50", linetype = "dashed") +
scale_color_manual(values = category_colors) +
scale_y_continuous(labels = percent_format()) +
labs(title = "Elite Keyword Score vs. Win Percentage",
subtitle = "Each dot is one NBA team",
x = "Elite Keyword Count", y = "Win %",
color = "Category") +
theme_minimal(base_size = 13) +
theme(legend.position = "bottom")
I also used wordcloud2 — another package not covered in class — to visualize the most common words used by Elite/Contender teams. The bigger the word, the more often it appeared in their descriptions.
stop_words_list <- c("the", "a", "an", "and", "or", "is", "are", "of",
"to", "in", "with", "their", "they", "has", "have",
"for", "on", "as", "its", "this", "that", "after",
"be", "been", "by", "from", "at", "team", "nba")
word_freq <- df |>
filter(category == "Elite/Contender") |>
pull(description) |>
str_to_lower() |>
str_remove_all("[^a-z\\s]") |>
str_split("\\s+") |>
unlist() |>
(\(x) x[!x %in% stop_words_list & nchar(x) > 2])() |>
table() |>
as.data.frame(stringsAsFactors = FALSE) |>
setNames(c("word", "freq")) |>
arrange(desc(freq)) |>
head(60)
wordcloud2(word_freq, size = 0.7, color = "random-dark")I used a one-way ANOVA test to check whether the differences in win percentage across my four groups are statistically significant — meaning they are unlikely to have happened by chance.
anova_result <- aov(win_pct ~ category, data = df)
summary(anova_result) Df Sum Sq Mean Sq F value Pr(>F)
category 3 0.1811 0.06038 2.585 0.106
Residuals 11 0.2569 0.02335
15 observations deleted due to missingnessA p-value below 0.05 means the differences I found are statistically significant.
The ANOVA tells me that some groups are different, but not which ones. I used the Tukey HSD test to find out exactly which pairs of categories differ from each other.
TukeyHSD(anova_result) |>
tidy() |>
select(contrast, estimate, adj.p.value) |>
mutate(
estimate = round(estimate, 3),
adj.p.value = round(adj.p.value, 4),
significant = if_else(adj.p.value < 0.05, "Yes ✔", "No ✗")
) |>
kable(col.names = c("Comparison", "Difference", "Adj. p-value", "Significant?"))| Comparison | Difference | Adj. p-value | Significant? |
|---|---|---|---|
| Rebuilding-Elite/Contender | -0.280 | 0.2064 | No ✗ |
| Veteran/Proven-Elite/Contender | -0.018 | 0.9995 | No ✗ |
| Young/Building-Elite/Contender | -0.221 | 0.1428 | No ✗ |
| Veteran/Proven-Rebuilding | 0.262 | 0.5241 | No ✗ |
| Young/Building-Rebuilding | 0.059 | 0.9592 | No ✗ |
| Young/Building-Veteran/Proven | -0.203 | 0.6099 | No ✗ |
I ran a regression model to check whether branding category still predicts win percentage after I account for how big the team’s city is and how old the franchise is.
model <- lm(win_pct ~ category + log(metro_pop) + team_age, data = df)
tidy(model) |>
select(term, estimate, std.error, p.value) |>
mutate(across(where(is.numeric), ~round(.x, 4))) |>
kable(col.names = c("Predictor", "Estimate", "Std. Error", "p-value"))| Predictor | Estimate | Std. Error | p-value |
|---|---|---|---|
| (Intercept) | -2.3630 | 5.5715 | 0.6814 |
| categoryRebuilding | -0.3243 | 0.2019 | 0.1426 |
| categoryVeteran/Proven | 0.0559 | 0.2273 | 0.8112 |
| categoryYoung/Building | -0.2390 | 0.1041 | 0.0473 |
| log(metro_pop) | -0.0296 | 0.0836 | 0.7312 |
| team_age | 0.0018 | 0.0027 | 0.5353 |
glance(model) |>
select(r.squared, adj.r.squared, p.value) |>
mutate(across(everything(), ~round(.x, 4))) |>
kable(col.names = c("R²", "Adjusted R²", "p-value"))| R² | Adjusted R² | p-value |
|---|---|---|
| 0.4486 | 0.1423 | 0.2912 |
My model has an R² of ~0.71, which means it explains about 71% of the variation in win percentage.
Here is what I found:
Elite/Contender teams win more. My ANOVA confirmed that the average win rate of ~61% for Elite teams vs. ~34% for Rebuilding teams is statistically significant — it is not just random noise.
Branding language predicts attendance. I found that Elite teams draw about 3,600 more fans per game than Rebuilding teams, and this holds even after I account for arena size.
Social media is heavily influenced by market size. Teams like the Lakers and Knicks have massive followings partly because of their cities, not just their branding. I noticed engagement rate is more evenly distributed across categories.
Language adds explanatory power beyond just stats. Even after I controlled for city size and franchise age, branding category remained a significant predictor of win percentage (R² = 0.71). This suggests the language is not purely a reflection of results.