This project integrates qualitative media analysis with quantitative game data to explore the psychological drivers behind basketball comeback success in the NBA and WNBA. By examining how resilience, team cohesion, momentum and other aspects of major comebacks are framed in sports media narratives and reflected in on-court performance, the project investigates the between public perception and in game reality. The study draws on coded analysis of over 1,800 article quotes and performance metrics from 12 major comeback games, offering an initial framework for understanding how psychological attributes are constructed through both storytelling and measurable game dynamics.

Qualitative Dataset: Article Selection and Coding Procedures

To explore how psychological factors are portrayed in basketball comeback narratives, I curated a data set of media articles covering notable NBA and WNBA games from the 2023-2024 seasons. Each selected game involved a successful comeback after halftime, defined as overcoming a second quarter deficit of at least 18 points for NBA or 11 points for WNBA, followed by the comeback team tying the score and ultimately winning in regulation.

Game Selection Criteria

Games were selected chronologically from scraped data of NBA and WNBA games from basketball-reference.com across the 2023-2024 seasons, each meeting the following criteria:

- The comeback team trailed by ≥18 points (NBA) or ≥11 (WNBA) at halftime
- The comeback team tied the score at any point after the second half
- The comeback team ultimately won without entering overtime
- Complete play-by-play data was publicly available
- Media coverage existed from at least one major outlet

This process resulted in a final sample of 12 games: 6 from the NBA and 6 from the WNBA.

Article Collection and Snippet Coding

For each of the 12 games, I manually searched for articles published by major outlets (ESPN, NYT Athletic, Sports Illustrated, etc) within 48 hours of the game. The initial goal was to collect 4 articles per game across comparable outlets, but limited coverage of some games meant that in a few cases only 2-3 articles were available.

Snippet Extraction and Coding

Each article was copied into Google Gemini’s interface to extract direct quotes that matched predefined psychological and narrative categories, which were chosen after reviewing academic literature and community written sources on how comebacks are built by athletes and perceived by the public:

- Individual Resilence
- Self Efficacy
- Team Cohesion
- Emotional Regulation
- Verbal Persuasion
- Home Court Advantage/Crowd Influence
- Aftermath Language (Momentum, Draining, Performance Outcomes)
- Gendered Language

Coding was conducted with strict literal adherence to code definitions, following definitions with clear inclusion and exclusion criteria for each category. Look to qualitative_data_codebook.pdf to see these criteria. Snippets were marked by type (like author-written, headline, player quote, coach quote) and tagged with team perspective (comeback team = CBT, opponent team = OPP). Multi-code attribution was permitted when snippets matched multiple categories. All snippets were reviewed for strict literal adherence to the codebook definitions.

Data Aggregation

The full data set was initially stored in two separate files:

- `nba_raw_text_data.csv`
- `wnba_raw_text_data.csv`

Each file contained coded article snippets for it’s respective league that were then combined into a single file, bothleague_aggregated_text.csv. This was completed using an R script to verify and count code occurrences by league, followed by manually copying and pasting to merge the data sets while preserving all metadata.

The resulting, final, data set includes over 1,800 coded snippets across 40 unique articles.

Exploratory Analysis

Article Distribution by League

Before looking into the frequencies of coded content, this figure shows the distribution of articles across the two leagues. The final data set includes a few more NBA articles (n = 22) than WNBA articles (n = 18). This gap reflects both the initial availability of media coverage and the relative ease of locating multiple articles for NBA games within 48 hours of play.

While the goal was to maintain balanced representation across leagues, WNBA games often had fewer qualifying articles available from major outlets like ESPN, NYT Athletic, and Sports Illustrated. This disparity is acknowledged as a limitation of the data set, but the inclusion of 18 WNBA articles still provides a substantial basis for gendered comparison of halftime deficit comebacks and media portrayals of comeback narratives.

Top Media Sources

This figure shows the distribution of articles by media outlet across the full dataset. ESPN was the most represented source by a significant margin, followed then by NYT Athletic and Sports Illustrated. These three outlets collectively accounted for the majority of articles analyzed, which may reflect their dominant presence in mainstream sports media. A range of smaller outlets, including Swish Appeal, Yahoo Sports, and several team-specific or regional sources, helped fill in coverage gaps particularly for WNBA games.

Top 10 Attributes by Team Role

Comeback teams overwhelmingly dominate code frequency, especially in psychological and aftermath-related codes, suggesting that narratives tend to center resilience and performance of winners while minimizing detail about opponents. This figure shows total code mentions broken down by comeback team versus opponent. Comeback teams were the clear narrative center, with 8 of the 10 most frequent codes attributed to comeback teams. Especially telling is the fact that c_aft_perf_outcome_lang alone appeared over 550 times, more than any other code by a wide margin.

Codes linked to psychological strength and team dynamics (c_att_ind_resilience, c_att_team_efficacy_cohesion) were overwhelmingly more common for comeback teams, while opponent team coverage focused primarily on aftermath or consequence (o_aft_perf_outcome_lang, o_aft_draining_lang).

This suggests a clear asymmetry in narrative attention when turning focus away from neutral performance metrics: comeback teams seem to be framed as triumphant, while opponents are framed as tired, overwhelmed, or faltering, reinforcing a winner-loser binary even in the language used to describe effort and emotion.

Average Code Frequency Across All Articles

This figure provides a baseline view of which psychological and aftermath related codes appeared most frequently across all 40 articles in the data set. Performance outcome language, especially for the comeback teams (c_aft_perf_outcome_lang), was by far the most prevalent with ~14 mentions per article on average. This suggests a strong narrative emphasis on outcome framing in comeback scenarios.

Other comeback-oriented codes such as c_att_ind_resilience, c_att_team_efficacy_cohesion, and c_aft_momentum_lang followed closely, reinforcing a pattern of attributing comeback success to internal psychological strength and team dynamics. Codes associated with the opponent team (such as o_aft_draining_lang, o_att_momentum_lang) were less frequent overall which could reflect an asymmetry in narrative attention.

Here’s a closer look at the average mentions per article:
Table
code avg_mentions_per_article
c_aft_perf_outcome_lang 14.600
o_aft_perf_outcome_lang 7.850
c_att_ind_resilience 5.600
c_att_team_efficacy_cohesion 5.250
c_aft_momentum_lang 3.700
o_aft_draining_lang 2.150
c_att_verbal_persuasion 2.025
o_aft_momentum_lang 1.150
o_att_team_efficacy_cohesion 0.925
c_att_self_efficacy 0.825
o_att_ind_resilience 0.650
c_att_emotional_reg 0.575
o_att_verbal_persuasion 0.225
o_att_emotional_reg 0.075
o_att_self_efficacy 0.050
c_aft_draining_lang 0.050
c_gendered_lang 0.000
o_gendered_lang 0.000

Gendered language codes were virtually nonexistent across both leagues, with zero occurrences in the data set. While no literal gendered codes were detected, this absence may reflect a genuine reduction in overt narrative gendering, or maybe a shift toward more subtle forms of gendered framing not captured by this method, like subtle differences in tone, emphasis, or omission that fall outside the scope of literal keyword-based coding. The disparity in article availability between NBA and WNBA games might reveal a deeper layer of representational bias. The consistent difficulty in locating timely, in-depth WNBA coverage could reflect the ongoing marginalization of women’s sports in mainstream media. In this way, the absence of gendered language should not be read as neutrality, but rather perhaps as part of a broader ecosystem in which male athletes’ stories could be told more often, more vividly, and with greater narrative investment.

Average Code Frequency Among Top Media Outlets

This plot isolates code frequency averages across articles published by the three most represented media outlets: ESPN, NYT Athletic, and Sports Illustrated. While the overall order of codes remains largely consistent, subtle nuances emerge. NYT Athletic used performance outcome language surrounding comeback teams at a notably higher rate than ESPN or Sports Illustrated, which could suggest a more results driven framing in their coverage (though it’s important to note, the average snippets per article for NYT were about 14 occurrences higher than ESPN and about 22 occurrences higher than Sports Illustrated). Sports Illustrated had higher mentions of c_att_team_efficacy_cohesion, possibly indicating a greater emphasis on team-based narratives. ESPN, while high in both regards, appeared more moderate across the board.

This reveals that although performance outcomes dominate across outlets, narrative tone and emphasis may differ by publication, shaping the reader’s perception of what drives a successful comeback.

Average Code Frequency By League and Team Role

The pattern observed in this figure holds similarly across both NBA and WNBA games, with comeback teams receiving more narrative attention across nearly every code. However, WNBA articles displayed a slightly flatter curve with lower overall averages and narrower gaps between comeback and opponent team mentions. This could reflect not only the shorter average article length in WNBA coverage but also the smaller total number of articles available for those games. It may also point to a more balanced or restrained narrative style in how WNBA comebacks are reported. Additionally, these patterns could hint at subtle differences in how media outlets frame the emotional and psychological states of male versus female athletes, particularly in codes like c_att_ind_resilience and c_att_team_efficacy_cohesion, which, while still prominent, appear slightly less frequently in WNBA narratives than NBA ones.

To begin to survey whether specific psychological or aftermath related attributes differ significantly (α = .05) between NBA and WNBA, two-sample t-tests were conducted for the five most frequent coded categories, including:

  1. c_aft_perf_outcome_lang (comeback team performance framing),
  2. o_aft_perf_outcome_lang (opponent team performance framing),
  3. c_att_ind_resilience (comeback team individual resilience),
  4. c_att_team_efficacy_cohesion (comeback team team efficacy/cohesion), and
  5. c_aft_momentum_lang (comeback team momentum-related language).

Each test compared the mean frequency of the code between NBA and WNBA articles. While this does not account for article length or media outlet, it serves as a quick check on whether leagues differ systematically in how these key narrative elements are used. Notably, four of the five most frequently coded attributes were comeback team–focused, further supporting the idea that sports media narratives tend to emphasize the psychological qualities and aftermath experiences of winning teams, rather than offering balanced framing of both sides.

Code Frequency Differences Between NBA and WNBA Comeback Narratives
code mean_nba mean_wnba t_statistic p_value ci_lower ci_upper
c_aft_perf_outcome_lang 14.909 14.222 0.252 0.803 -4.846 6.220
o_aft_perf_outcome_lang 8.000 7.667 0.149 0.882 -4.194 4.860
c_att_ind_resilience 6.227 4.833 1.344 0.189 -0.728 3.515
c_att_team_efficacy_cohesion 6.136 4.167 1.456 0.154 -0.770 4.709
c_aft_momentum_lang 3.955 3.389 0.864 0.393 -0.762 1.893

NBA and WNBA articles used neutral aftermath performance language almost equally (14.91 vs 14.22 mentions on average). The p-value of 0.80 shows that there was no significant difference between leagues. For opponent team aftermath performance language, there was again nearly identical use (8.00 vs 7.67). The difference is tiny and non-significant (p = 0.88), however there remains a large disparity between how often this type of language is used for comeback teams compared to opponent teams. This reinforces that opponent teams are framed similarly in both leagues, and that comeback narratives are often centered on team that made the comeback.

NBA articles showed slightly more focus on individual resilience (6.23 vs 4.83) when compared to WNBA, but the difference isn’t statistically significant (p = 0.19). There’s a hint of difference, but it’s not strong enough to call conclusive. As for team cohesion, there is a similar pattern between WNBA and NBA mentions, with NBA having a higher mean mention (6.14 vs 4.17), but again the difference was not statistically significant (p = 0.15). And for momentum language in the aftermath of the comeback, there are slightly more mentions in NBA articles (3.95 vs 3.39), but this is not statistically significant (p = 0.39). Both leagues use momentum language moderately (compared to the ~14 mentions of neutral comeback language) and similarly.

To further confirm the suspicion of whether comeback teams are framed more heavily in aftermath-oriented language than their opponents, a paired t-test was conducted comparing c_aft_perf_outcome_lang and o_aft_perf_outcome_lang counts within each article. This test evaluates whether media narratives (combining both NBA and WNBA) systematically emphasize the performance outcomes of winning teams more than those of the teams that lost the lead.

T-Test Comparing Aftermath Performance Language Between CBTs and Opponents
attribute mean_diff t_statistic p_value ci_lower ci_upper
aft_perf_outcome_lang 6.75 3.469 0.001 2.815 10.685

The estimated mean difference in aftermath performance language was 6.75 mentions (95% CI: 2.81, 10.69), indicating significantly more narrative emphasis on comeback teams compared to their opponents (p = 0.0013).

Testing League-Level Narrative Differences

Mixed Effect Modeling using lme4::lmer()

To further test whether narrative emphasis differed by league, linear mixed-effects models were fit with league as a fixed effect and random intercepts for game_id and media_outlet. Models were run separately for comeback team psychological and aftermath framing to account for variation in game context and outlet style.

This approach helps to control for two major sources of variability:

- Game Context, some games may naturally prompt more psychological or emotional framing due 
to their stakes, players, or timing
- Media Outlet, some vary in editorial tone, audience, and writing style, which
all shape narrative emphasis

To reduce noise from sparsely represented sources, analysis was restricted to the top three most frequent outlets: ESPN, NYT Athletic, and Sports Illustrated. This was intended to improve model stability while still allowing for outlet-level variation to be modeled.

Due to persistent singular fit warnings in the opponent models, those instances were excluded. These warnings indicated zero or near-zero variance in opponent scores across games or media outlets, suggesting little meaningful clustering or variability to control for.

# Define column groups 
cbt_psych <- c("c_att_ind_resilience", "c_att_self_efficacy", "c_att_team_efficacy_cohesion", 
               "c_att_emotional_reg", "c_att_verbal_persuasion")

opp_psych <- c("o_att_ind_resilience", "o_att_self_efficacy", "o_att_team_efficacy_cohesion", 
               "o_att_emotional_reg", "o_att_verbal_persuasion")

cbt_after <- c("c_aft_perf_outcome_lang", "c_aft_momentum_lang", "c_aft_draining_lang")

opp_after <- c("o_aft_perf_outcome_lang", "o_aft_momentum_lang", "o_aft_draining_lang")

# Identify top 3 outlets 
top_outlets <- text_data %>%
  count(media_outlet) %>%
  arrange(desc(n)) %>%
  slice(1:3) %>%
  pull(media_outlet)

# Filter to top 3 and calculate scores
top_data <- text_data %>%
  filter(media_outlet %in% top_outlets) %>%
  rowwise() %>%
  mutate(
    cbt_psych_score = sum(c_across(all_of(cbt_psych)), na.rm = TRUE),
    opp_psych_score = sum(c_across(all_of(opp_psych)), na.rm = TRUE),
    cbt_after_score = sum(c_across(all_of(cbt_after)), na.rm = TRUE),
    opp_after_score = sum(c_across(all_of(opp_after)), na.rm = TRUE)
  ) %>%
  ungroup()

# Fit Models
cbt_psych_model <- lmer(cbt_psych_score ~ league + (1 | game_id) + (1 | media_outlet), data = top_data)
cbt_after_model <- lmer(cbt_after_score ~ league + (1 | game_id) + (1 | media_outlet), data = top_data)
# opp_psych_model <- lmer(opp_psych_score ~ league + (1 | game_id) + (1 | media_outlet), data = top_data)
# opp_after_model <- lmer(opp_after_score ~ league + (1 | game_id) + (1 | media_outlet), data = top_data)

summary(cbt_psych_model)
## Linear mixed model fit by REML. t-tests use Satterthwaite's method [
## lmerModLmerTest]
## Formula: cbt_psych_score ~ league + (1 | game_id) + (1 | media_outlet)
##    Data: top_data
## 
## REML criterion at convergence: 163.5
## 
## Scaled residuals: 
##     Min      1Q  Median      3Q     Max 
## -1.6748 -0.4428 -0.2256  0.4565  2.5109 
## 
## Random effects:
##  Groups       Name        Variance Std.Dev.
##  game_id      (Intercept) 19.58    4.425   
##  media_outlet (Intercept) 65.73    8.107   
##  Residual                 50.54    7.109   
## Number of obs: 24, groups:  game_id, 12; media_outlet, 3
## 
## Fixed effects:
##             Estimate Std. Error     df t value Pr(>|t|)  
## (Intercept)   16.134      5.330  2.476   3.027   0.0721 .
## leagueWNBA    -7.722      4.288 13.830  -1.801   0.0936 .
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Correlation of Fixed Effects:
##            (Intr)
## leagueWNBA -0.275
summary(cbt_after_model)
## Linear mixed model fit by REML. t-tests use Satterthwaite's method [
## lmerModLmerTest]
## Formula: cbt_after_score ~ league + (1 | game_id) + (1 | media_outlet)
##    Data: top_data
## 
## REML criterion at convergence: 165.7
## 
## Scaled residuals: 
##      Min       1Q   Median       3Q      Max 
## -1.30090 -0.60138 -0.07390  0.07159  2.26272 
## 
## Random effects:
##  Groups       Name        Variance Std.Dev.
##  game_id      (Intercept)  9.181   3.030   
##  media_outlet (Intercept) 40.272   6.346   
##  Residual                 68.326   8.266   
## Number of obs: 24, groups:  game_id, 12; media_outlet, 3
## 
## Fixed effects:
##             Estimate Std. Error     df t value Pr(>|t|)  
## (Intercept)   17.976      4.391  2.338   4.093   0.0418 *
## leagueWNBA    -3.656      4.284 12.167  -0.853   0.4099  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Correlation of Fixed Effects:
##            (Intr)
## leagueWNBA -0.300

Psychological Framing (Comeback Team)

- Intercept (NBA baseline):  On average, articles include 16.13 comeback team 
psychological codes for NBA games.

- League effect: 
WNBA articles include ~7.72 fewer psychological codes on average than NBA articles.
This difference is not statistically significant (p = 0.0936), but suggests a 
trend toward reduced psychological attribution in WNBA comeback coverage.

- Random effects:
Media outlet SD = 8.11 — substantial outlet-level variability in psychological framing.
Game ID SD = 4.43 — moderate variation based on the game being covered.

Aftermath Framing (Comeback Team)

- Intercept (NBA baseline): On average, articles include 17.98 comeback team
aftermath codes for NBA games.

- League effect:
WNBA articles use ~3.66 fewer aftermath codes on average compared to NBA articles.
This difference is again not statistically significant (p = 0.4099).

- Random effects:
Media outlet SD = 6.35 — almost a substainal variation in aftermath framing across outlets.
Game ID SD = 3.03 — lower game-level variability than in the psychological model.

Refitting the Model

To further examine whether the narrative differences by league are significant or not, a refitted linear mixed-effects model was constructed to try and include more granular controls. Two aspects of the previous model that stuck out to me was the fact that the amount of snippets present from each article where not controlled for, and that WNBA was definitely lacking in representation from the top three media outlets. So the updated model includes league and media_outlet(adding WNBA.com as the fourth top media outlet to broaden sample size and WNBA representation) as fixed effects, while retaining game_id as a random intercept to account for game-specific contexts. By treating media_outlet as a fixed effect, the model directly estimates narrative variation across ESPN, NYT Athletic, Sports Illustrated, and WNBA.com by allowing for outlet level differences to be interpreted rather than modeled as random effects. To address the variation in article length between media outlets a new variable, snippet_count, was introduced and used to standardize the scores that were used as response variables. This allowed for comparisons of narrative tone that are not confounded by differences in article size.

# Define code groups
cbt_psych <- c("c_att_ind_resilience", "c_att_self_efficacy", "c_att_team_efficacy_cohesion", 
               "c_att_emotional_reg", "c_att_verbal_persuasion")

opp_psych <- c("o_att_ind_resilience", "o_att_self_efficacy", "o_att_team_efficacy_cohesion", 
               "o_att_emotional_reg", "o_att_verbal_persuasion")

cbt_after <- c("c_aft_perf_outcome_lang", "c_aft_momentum_lang", "c_aft_draining_lang")

opp_after <- c("o_aft_perf_outcome_lang", "o_aft_momentum_lang", "o_aft_draining_lang")

snippet_cols <- c(cbt_psych, opp_psych, cbt_after, opp_after)

# Calculate scores and snippet count
updated_data <- text_data %>%
  rowwise() %>%
  mutate(
    cbt_psych_score = sum(c_across(all_of(cbt_psych)), na.rm = TRUE),
    opp_psych_score = sum(c_across(all_of(opp_psych)), na.rm = TRUE),
    cbt_after_score = sum(c_across(all_of(cbt_after)), na.rm = TRUE),
    opp_after_score = sum(c_across(all_of(opp_after)), na.rm = TRUE),
    snippet_count = sum(c_across(all_of(snippet_cols)), na.rm = TRUE)
  ) %>%
  ungroup()

# Group media outlets
top_4_outlets <- c("ESPN", "NYT Athletic", "SI", "WNBA")

updated_data <- updated_data %>%
  filter(media_outlet %in% top_4_outlets) %>%
  mutate(media_outlet_grouped = media_outlet)

# Standardize outcome variables (per-snippet)
updated_data <- updated_data %>%
  mutate(
    cbt_psych_score_std = cbt_psych_score / snippet_count,
    opp_psych_score_std = opp_psych_score / snippet_count,
    cbt_after_score_std = cbt_after_score / snippet_count,
    opp_after_score_std = opp_after_score / snippet_count
  )

# Fit the models
cbt_psych_model_std <- lmer(
  cbt_psych_score_std ~ league + media_outlet_grouped + (1 | game_id),
  data = updated_data
)

opp_psych_model_std <- lmer(
  opp_psych_score_std ~ league + media_outlet_grouped + (1 | game_id),
  data = updated_data
)
## boundary (singular) fit: see help('isSingular')
cbt_after_model_std <- lmer(
  cbt_after_score_std ~ league + media_outlet_grouped + (1 | game_id),
  data = updated_data
)

opp_after_model_std <- lmer(
  opp_after_score_std ~ league + media_outlet_grouped + (1 | game_id),
  data = updated_data
)

# View model summaries
summary(cbt_psych_model_std)
## Linear mixed model fit by REML. t-tests use Satterthwaite's method [
## lmerModLmerTest]
## Formula: cbt_psych_score_std ~ league + media_outlet_grouped + (1 | game_id)
##    Data: updated_data
## 
## REML criterion at convergence: -9.2
## 
## Scaled residuals: 
##     Min      1Q  Median      3Q     Max 
## -1.6740 -0.4163 -0.1072  0.3633  1.6892 
## 
## Random effects:
##  Groups   Name        Variance Std.Dev.
##  game_id  (Intercept) 0.004023 0.06343 
##  Residual             0.022082 0.14860 
## Number of obs: 26, groups:  game_id, 12
## 
## Fixed effects:
##                                  Estimate Std. Error       df t value Pr(>|t|)
## (Intercept)                       0.34411    0.06160 18.11732   5.586  2.6e-05
## leagueWNBA                       -0.05684    0.08069 14.56483  -0.704    0.492
## media_outlet_groupedNYT Athletic  0.05269    0.07374 15.69567   0.715    0.485
## media_outlet_groupedSI           -0.10873    0.08744 14.04709  -1.243    0.234
## media_outlet_groupedWNBA         -0.01697    0.12463 18.98655  -0.136    0.893
##                                     
## (Intercept)                      ***
## leagueWNBA                          
## media_outlet_groupedNYT Athletic    
## media_outlet_groupedSI              
## media_outlet_groupedWNBA            
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Correlation of Fixed Effects:
##             (Intr) lgWNBA m__NYA md__SI
## leagueWNBA  -0.645                     
## md_tlt_NYTA -0.538  0.189              
## md_tlt_grSI -0.581  0.360  0.376       
## md_tlt_WNBA -0.065 -0.278  0.121  0.046
summary(opp_psych_model_std)
## Linear mixed model fit by REML. t-tests use Satterthwaite's method [
## lmerModLmerTest]
## Formula: opp_psych_score_std ~ league + media_outlet_grouped + (1 | game_id)
##    Data: updated_data
## 
## REML criterion at convergence: -67.5
## 
## Scaled residuals: 
##     Min      1Q  Median      3Q     Max 
## -1.1226 -0.8880 -0.2656  0.7954  1.7861 
## 
## Random effects:
##  Groups   Name        Variance Std.Dev.
##  game_id  (Intercept) 0.000000 0.00000 
##  Residual             0.001588 0.03986 
## Number of obs: 26, groups:  game_id, 12
## 
## Fixed effects:
##                                   Estimate Std. Error        df t value
## (Intercept)                       0.035391   0.014901 21.000000   2.375
## leagueWNBA                        0.009349   0.018939 21.000000   0.494
## media_outlet_groupedNYT Athletic -0.023242   0.019384 21.000000  -1.199
## media_outlet_groupedSI            0.003081   0.023232 21.000000   0.133
## media_outlet_groupedWNBA          0.030750   0.031879 21.000000   0.965
##                                  Pr(>|t|)  
## (Intercept)                        0.0271 *
## leagueWNBA                         0.6267  
## media_outlet_groupedNYT Athletic   0.2439  
## media_outlet_groupedSI             0.8958  
## media_outlet_groupedWNBA           0.3457  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Correlation of Fixed Effects:
##             (Intr) lgWNBA m__NYA md__SI
## leagueWNBA  -0.635                     
## md_tlt_NYTA -0.591  0.209              
## md_tlt_grSI -0.641  0.408  0.379       
## md_tlt_WNBA -0.090 -0.297  0.152  0.058
## optimizer (nloptwrap) convergence code: 0 (OK)
## boundary (singular) fit: see help('isSingular')
summary(cbt_after_model_std)
## Linear mixed model fit by REML. t-tests use Satterthwaite's method [
## lmerModLmerTest]
## Formula: cbt_after_score_std ~ league + media_outlet_grouped + (1 | game_id)
##    Data: updated_data
## 
## REML criterion at convergence: -13.5
## 
## Scaled residuals: 
##     Min      1Q  Median      3Q     Max 
## -1.8748 -0.5705 -0.2317  0.5241  1.9550 
## 
## Random effects:
##  Groups   Name        Variance Std.Dev.
##  game_id  (Intercept) 0.005092 0.07136 
##  Residual             0.016713 0.12928 
## Number of obs: 26, groups:  game_id, 12
## 
## Fixed effects:
##                                   Estimate Std. Error        df t value
## (Intercept)                       0.342908   0.056815 16.827755   6.036
## leagueWNBA                        0.102280   0.075278 13.660439   1.359
## media_outlet_groupedNYT Athletic  0.083582   0.064800 14.920344   1.290
## media_outlet_groupedSI           -0.003904   0.076391 13.129899  -0.051
## media_outlet_groupedWNBA         -0.090907   0.110757 18.059647  -0.821
##                                  Pr(>|t|)    
## (Intercept)                       1.4e-05 ***
## leagueWNBA                          0.196    
## media_outlet_groupedNYT Athletic    0.217    
## media_outlet_groupedSI              0.960    
## media_outlet_groupedWNBA            0.422    
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Correlation of Fixed Effects:
##             (Intr) lgWNBA m__NYA md__SI
## leagueWNBA  -0.650                     
## md_tlt_NYTA -0.510  0.179              
## md_tlt_grSI -0.549  0.337  0.376       
## md_tlt_WNBA -0.055 -0.265  0.107  0.040
summary(opp_after_model_std)
## Linear mixed model fit by REML. t-tests use Satterthwaite's method [
## lmerModLmerTest]
## Formula: opp_after_score_std ~ league + media_outlet_grouped + (1 | game_id)
##    Data: updated_data
## 
## REML criterion at convergence: -1.6
## 
## Scaled residuals: 
##      Min       1Q   Median       3Q      Max 
## -2.01659 -0.45937 -0.07388  0.41680  2.47686 
## 
## Random effects:
##  Groups   Name        Variance Std.Dev.
##  game_id  (Intercept) 0.003304 0.05748 
##  Residual             0.033769 0.18376 
## Number of obs: 26, groups:  game_id, 12
## 
## Fixed effects:
##                                  Estimate Std. Error       df t value Pr(>|t|)
## (Intercept)                       0.27957    0.07283 18.93914   3.839  0.00111
## leagueWNBA                       -0.06430    0.09434 15.00178  -0.682  0.50588
## media_outlet_groupedNYT Athletic -0.11521    0.09043 16.09163  -1.274  0.22074
## media_outlet_groupedSI            0.11344    0.10773 14.56665   1.053  0.30947
## media_outlet_groupedWNBA          0.09607    0.15124 19.53565   0.635  0.53266
##                                    
## (Intercept)                      **
## leagueWNBA                         
## media_outlet_groupedNYT Athletic   
## media_outlet_groupedSI             
## media_outlet_groupedWNBA           
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Correlation of Fixed Effects:
##             (Intr) lgWNBA m__NYA md__SI
## leagueWNBA  -0.640                     
## md_tlt_NYTA -0.560  0.198              
## md_tlt_grSI -0.606  0.379  0.377       
## md_tlt_WNBA -0.075 -0.287  0.133  0.050

Psychological Framing (Comeback Team)

- Intercept (NBA, ESPN baseline): On average, NBA comeback articles from ESPN contain 0.34 CBT 
psychological codes per snippet.

- League effect: 
WNBA articles show a non-significant (p = 0.492) decrease of ~0.057 codes per snippet
compared to NBA articles, for the ESPN baseline.
There is no statistically significant difference in the rate of comeback team psychological 
framing between WNBA and NBA articles in top outlets.

- Media outlet effects:
NYT Athletic shows a slgightly higher code rate (+0.053 codes per snippet), not significant (p = 0.485). 
SI shows a slightly lower code rate (–0.109), not significant (p = 0.234).
WNBA.com is close to the ESPN baseline (–0.017), also not significant (p = 0.893).

- Random effects:
Game ID SD = 0.063 — modest article-level variability based on the game being covered.

Psychological Framing (Opponent Team)

- Intercept (NBA, ESPN baseline): Articles average 0.035 opponent psychological codes
per snippet, a very low frequency.

- League and media effects:
None of the fixed effects were statistically significant, meaning opponent teams are 
rarely described using psychological attributes, and this pattern holds consistently 
across both leagues and all major outlets.

- Random effects:
The model returned a boundary fit, meaning it detected no meaningful variation across games.

Aftermath Framing (Comeback Team)

-  Intercept (NBA, ESPN baseline): Articles contain ~0.34 aftermath codes per snippet, on average.

- League effect:
WNBA articles published by ESPN tend to use ~0.10 more aftermath codes per snippet than NBA articles
(p = 0.196), so this is not statistically significant.

- Media effects:
NYT Athletic shows a higher inclusion (0.084) of aftermath codes per snippet on average.
Sports Illustrated inclused 0.004 fewer aftermath codes per snipper on average.
WNBA.com included 0.091 fewer aftermath codes per snippet on average. 
Like psychological framing, none of these were statistically significant (all p > 0.2).

- Random effects: 
Game ID SD = 0.071 — slightly higher variability based on specific games compared to
psychological framing codes for comeback teams.

Aftermath Framing (Opponent Team)

- Intercept (NBA, ESPN baseline): Articles contain ~0.28 aftermath codes per 
snippet for opponent teams.

- League effect: 
ESPN WNBA articles use ~0.06 fewer opponent aftermath codes per snippet than NBA ones, 
but this is not significant (p = 0.51).

- Media effects:
NYT Athletic articles averaged 0.12 fewer opponent aftermath codes per snippet.
Sports Illustrated articles included 0.11 more aftermath codes per snipper on average. 
On average, WNBA.com articles included 0.096 more opponent aftermath codes per snippet. 

- Random effects; 
Game ID SD = 0.057 — slightly lower game level variability in opponent aftermath framing
compared to psychological framing codes for comeback teams.

Qualitative Analysis Summary

The qualitative analysis of comeback narratives reveals a consistent asymmetry in how media articles frame NBA and WNBA performances. Across both leagues, comeback teams were more frequently associated with psychological and aftermath-related codes, such as individual resilience, team cohesion, and performance outcomes, while opponent teams were more often linked to draining or fading momentum. Opponent psychology was rarely described, so much so that several models failed due to lack of variance.

The most frequent code by far was c_aft_perf_outcome_lang, reinforcing a tendency to emphasize what happened rather than why. Codes like c_att_ind_resilience and c_att_team_efficacy_cohesion were also prevalent, particularly in NBA coverage, suggesting a narrative emphasis on internal strength and team unity. Differences across media outlets were relatively small and not statistically significant once snippet count was controlled for.

Differences across media outlets were present but subtle. NYT Athletic tended to emphasize performance outcomes more heavily, while Sports Illustrated leaned slightly more on team-based codes. Gendered language was completely absent from all articles, which could reflect either a shift away from overtly gendered framing or a limitation of the codebook’s focus on literal phrasing. Notably, WNBA games were associated with shorter articles and fewer snippets, suggesting reduced narrative investment or representational disparities. Mixed-effects models offered partial support for these narrative trends. In absolute terms, NBA articles included more comeback team psychological and aftermath codes than WNBA articles, but these league-level differences were not statistically significant. Standardized models (per snippet) similarly showed no reliable differences by league or outlet. Opponent-focused narratives were sparse across all conditions.

These results suggest a clear narrative emphasis on comeback teams, especially in the NBA, but limited statistical evidence of league-level differences in code usage frequency. The findings may be shaped in part by sample size, outlet representation, and the constrained nature of literal snippet coding. Still, they offer a meaningful first step toward quantifying how sports media constructs psychological narratives around comebacks.

Quantitative Data: Collection and Analysis

To explore whether the comeback narratives correspond to differences in game performance, a quantitative data set was compiled using play by play logs from basketball-reference.com the 12 selected major comeback games. For each game, a maximum six-minute window was analyzed beginning at the moment of the first tie following the second half deficit. This narrow frame was selected to isolate the immediate aftermath of the tie, where shifts momentum and team execution are often most decisive.

The data set (pbp_wnba_nba_data.csv) includes manually collected performance based metrics for both the comeback team (c_) and their opponent (o_), using the following variables:

- points_scored
- points_allowed
- fg_made
- fg_attempts
- turnovers
- fouls_committed
- full_timeouts

These variables provide a more objective glimpse into which team performed more effectively, gained or lost momentum, the possible psychological drivers that lead to a sustained win after a tie following a deficit, and the overall control of the game during the minutes following the comeback tie. Since narratives were mostly focused on the comeback team out performing the opponent team by all metrics, the following exploration and testing will be centered around that.

Exporatory Analysis

Points Scored vs. Points Allowed

## Warning: There was 1 warning in `summarise()`.
## ℹ In argument: `across(c(c_points_scored, c_points_allowed), mean, na.rm =
##   TRUE)`.
## ℹ In group 1: `league = "NBA"`.
## Caused by warning:
## ! The `...` argument of `across()` is deprecated as of dplyr 1.1.0.
## Supply arguments directly to `.fns` through an anonymous function instead.
## 
##   # Previously
##   across(a:b, mean, na.rm = TRUE)
## 
##   # Now
##   across(a:b, \(x) mean(x, na.rm = TRUE))

In both leagues, comeback teams (CBT) outscored their opponents in the aftermath window among this sample. WNBA teams averaged a score of 11.75 points, and NBA teams averaged a score 10.08 points within the six minute max aftermath, which supports the heavy narrative focus on comeback team success and performance outcome framing (c_aft_perf_outcome_lang). Opponents were also coded for performance outcomes (o_aft_perf_outcome_lang), but far less frequently which is also mirrored in the lower points allowed and reduced coverage of opponent psychological code.

Other Metrics of Performance by League

In the six-minute aftermath window, WNBA comeback teams average over three times as many fouls as their NBA counterparts (≈2.5 vs. 0.7 per game) and turn the ball over more often (≈1.2 vs. 0.2), yet call fewer timeouts (≈0.5 vs. 0.7). WNBA opponents commit more fouls than NBA opponents (≈1.83 vs. 1.00), struggle more with ball security (≈2.33 vs. 1.50 turnovers), and call fewer timeouts (≈0.67 vs. 1.33), reflecting a more physical game with less reliance on stoppages. These sample-based averages paint NBA comebacks as low-error affairs managed with strategic timeouts, whereas WNBA comebacks unfold as higher-error battles with less reliance on clock management where WNBA opponents tend to succumb more to turnovers than to fouls or timeout tactics.

Average FG Percentage Comparison
league CBT FG Made CBT FG Attempts OPP FG Made OPP FG Attempts CBT FG % OPP FG %
NBA 4.000 5.833 1.667 6.333 0.686 0.263
WNBA 3.833 7.167 3.000 7.500 0.535 0.400

In this sample, NBA comeback teams shot with 68.6% (4.00 / 5.83) accuracy in field goals in the six-minute aftermath of their ties, while their opponents managed just 26.3% (1.67 / 6.33), creating a dramatic gap that likely fuels narratives of dominant resilience and momentum. By contrast, WNBA comebacks were more tightly contested, with comeback teams shooting a 53.5% (3.83 / 7.17) accuracy on average versus a 40.0% (3.00 / 7.50) average accuracy by their opponents, suggesting a more balanced exchange of execution and offering subtler grounds for storytelling around psychological edge.

Testing League Level Comeback Aftermath Differences

To evaluate the statistical significance of the post-comeback performance differences, three hypothesis tests were conducted comparing comeback-team and opponent metrics within the six-minute aftermath window.

Do comeback teams statistically outperform their opponents in the aftermath, regardless of league?

Across the 12 comeback games in the sample, comeback teams (CBTs) significantly outperformed their opponents during the six-minute window following the second-half tie:
Table
Metric Mean_Diff t_stat p_value CI_lower CI_upper
Points Scored 5.083 4.818 0.001 2.761 7.405
FG % 0.367 4.112 0.002 0.170 0.563
Turnovers -1.250 -3.191 0.009 -2.112 -0.388
Fouls Committed 0.167 0.394 0.701 -0.765 1.099
Full Timeouts -0.417 -1.603 0.137 -0.989 0.155

Overall, CBTs generally displayed superior performance in the immediate aftermath of tying the score, especially in terms of scoring, efficiency, and ball control. On average, CBTs scored 5.08 more points than their opponents (p = 0.0005), demonstrating clear offensive superiority. They also shot more efficiently, with a field goal percentage 0.367 higher than their opponents (p = 0.0017), and committed 1.25 fewer turnovers on average (p = 0.0086), highlighting stronger ball control during this critical stretch. There was no significant difference in fouls committed between CBTs and opponents (mean difference = +0.17, p = 0.701), and while CBTs used slightly fewer timeouts (mean difference = –0.42), this difference was not statistically significant (p = 0.137). Together, these results indicate that comeback teams when not separated by league tended to play more effectively and with greater composure in the immediate aftermath of tying the game.

Do NBA vs WNBA comeback performances differ?

Not significantly. When comparing NBA and WNBA CBTs in the 6-minute aftermath window across the sample, differences in performance were modest and statistically non-significant, though small trends emerged:

Table
Metric Mean_Diff t_stat p_value CI_lower CI_upper
Points Scored -1.167 -0.372 0.718 -8.202 5.869
FG % 0.199 1.724 0.115 -0.058 0.456
Turnovers -1.000 -1.978 0.094 -2.227 0.227
Fouls Committed -1.833 -1.752 0.129 -4.381 0.715
Full Timeouts 0.167 0.415 0.688 -0.753 1.086

When comparing NBA and WNBA comeback team (CBT) performance during the aftermath period across the 12-game sample, no statistically significant differences emerged. WNBA CBTs scored slightly more points than NBA CBTs on average (mean difference = –1.17, p = 0.718), but this gap was negligible. NBA CBTs shot somewhat more efficiently, with a field goal percentage advantage of +0.199, a difference that approached significance (p = 0.115). Notably, WNBA CBTs committed more turnovers than NBA CBTs (mean difference = –1.00, p = 0.094), a trend visible in the chart where WNBA comeback teams show higher mean turnovers. WNBA CBTs also committed more fouls (mean difference = –1.83, p = 0.129), and the chart reflects this with a visibly higher bar for WNBA fouls compared to NBA. Timeout usage appeared similar across leagues (mean difference = +0.17, p = 0.688). Overall, while the performance patterns in non-shooting metrics (fouls, turnovers, timeouts) differ somewhat between leagues, these differences did not reach conventional levels of statistical significance in this sample.

Is the performance gap between comeback teams and their opponents larger in NBA vs. WNBA?

Not statistically, but the 12-game sample suggests that NBA comebacks may involve larger momentum swings, especially in points and shooting:
Table
Metric Mean_Diff t_stat p_value CI_lower CI_upper
Points Diff 3.500 1.826 0.099 -0.796 7.796
FG% Diff 0.330 2.129 0.062 -0.020 0.681
Turnovers Diff 0.167 0.203 0.843 -1.679 2.012
Fouls Diff 1.000 1.205 0.268 -0.971 2.971
Timeouts Diff 0.500 0.958 0.361 -0.668 1.668

On average, NBA comeback teams outscored their opponents by 3.50 more points than WNBA comeback teams did (p = 0.099), a difference that approached significance. Similarly, the NBA showed a slightly larger advantage in field goal percentage over opponents, with a mean gap of +0.33 (p = 0.062), again nearly significant. Differences in other metrics were smaller and not meaningful: the turnover gap (p = 0.843) and timeout gap (p = 0.361) were minimal, while NBA games showed a modestly larger foul margin (p = 0.268) that did not reach significance. Overall, these patterns hint that NBA comebacks may feature somewhat larger aftermath performance margins, but the evidence in this sample is not strong enough to conclude that league consistently shapes the size of the performance gap.

Discussion, Limitations, and Final Remarks

This study explored how psychological narratives in sports media align and diverge from objective measures of performance during major NBA and WNBA comebacks. The qualitative analysis revealed a clear asymmetry in narrative focus: media articles overwhelmingly emphasized the psychological resilience, team cohesion, and aftermath success of comeback teams while largely neglecting the psychology and performance framing of opponents. Codes defined by neutral aftermath performance language dominated the narratives, reflecting the broader tendency of sports journalism to focus on results over underlying drivers. Mixed-effects models and t-tests suggested that this pattern held across leagues and outlets, with no reliable statistical evidence of consistent league-level differences once article length and outlet were accounted for.

The quantitative results partially mirrored the qualitative findings by reinforcing the central narrative emphasis on comeback team dominance. Just as media articles disproportionately highlighted the psychological strength, resilience, and performance outcomes of comeback teams, especially in the NBA, the statistical data showed that comeback teams did, in fact, significantly outperform their opponents in key metrics like points scored, shooting efficiency, and turnovers during the aftermath window. This alignment suggests that the media’s focus on comeback team superiority is at least partially grounded in measurable performance realities.

However, while the qualitative analysis revealed a clear disparity in narrative investment—NBA comebacks were covered in greater depth and with richer psychological framing—the quantitative results showed that actual performance differences between NBA and WNBA comeback teams were limited. NBA comeback teams displayed slightly larger performance gaps in points and shooting margin, but these differences were not statistically significant in this sample. This suggests that, despite comparable effectiveness in the aftermath window, WNBA comeback teams received far less narrative attention and psychological attribution than their NBA counterparts. The imbalance in media coverage may therefore reflect broader representational biases rather than differences in on-court performance.

Several limitations constrain these findings. First, the small sample size, 12 games, with just 6 from each league, limits statistical power and generalization. The restricted availability of WNBA articles, and their shorter length on average, may have influenced both narrative coding and model stability. Furthermore, the qualitative coding was based on literal snippet matching, which may miss more subtle or contextual gendered and psychological framing that does not rely on direct phrasing. In the quantitative domain, the narrow aftermath window (maximum of six minutes) was chosen to isolate momentum shifts but may not capture longer-term patterns that also shape narrative construction.

Despite these limitations, this study serves as a valuable pilot effort to explore the psychological drivers behind major comebacks—both as constructed by media narratives and as reflected in objective performance. By linking how comebacks are perceived through story lines of resilience, cohesion, and momentum with measurable on-court behaviors, this project offers an initial framework for understanding how perception and reality interact in defining comeback success. Future research should build on this foundation by expanding the sample, refining coding to capture implicit or contextual narrative patterns, and examining longer or alternative performance windows to better illuminate the psychological dynamics that fuel comebacks. Such work can help deepen our understanding of how cultural narratives of mental toughness, teamwork, and emotional regulation are formed and reinforced through both storytelling and statistical realities in sport.