Players

We selected quaterbacks using the following criteria and ended up with a final sample size of 20.

1. Started in at least 7 games (22 QBs)
2. Has a public twitter account (21 QBs)
3. Tweeted at least once since 9/1/20 (20 QBs)

Tweets

Starting on September 1, 2020, NFL player tweets were web scraped and sorted by week. For example, Pittsburg Steelers starting quarterback Ben Roethlisberger had a Week 2 tweet count (tweetcount) of 25 because he tweeted or retweeted posts 25 times between September 14 (his last game) and September 20 (his Week 2 game).

Performance Statistics

Quarterbacks’ on-field performance was wrangled through ProFootballReference.com game logs and measured weekly by the following categories:

1. Total Touchdowns (total_tds): Sum of Passing and Rushing Touchdowns
2. Total Yards (total_yards): Sum of Passing and Rushing Yards
3. Quarterback Rating: (Rate): A measure of game situation-standardized quarterback performance based on values including passing yards, passing touchdowns, and turnovers.
4. Fantasy Points (fpts): ESPN standard league fantasy points

\[ Fantasy{\:}Points = (4\times Passing{\;}Touchdowns){\;} + (6\times Rushing{\;}Touchdowns){\;} \\ + (.04\times Pasing{\;}Yards){\;} + (.1\times Rushing{\;}Yards){\;} \\- (2\times Fumbles){\;} - (2\times Interceptions){\;} \]

Sample Quarterback Wrangling

The 20 quarterbacks’ data consisted of both tweet counts and the four performance statistics. Ben Roethlisberger’s dataframe provides an example of the key measures we tracked.

Twitter Usage Distribution

While the 20 qualifying quarterbacks tweeted on average about 7 times per week, the median weekly tweet count was about 2 per week, with an IQR of 8 tweets. The tweet count distribution was skewed by outlier Ben Roethlisberger, who tweeted 31 tweets per week on average. Only Roethlisberger, Patrick Mahomes, Philip Rivers, Teddy Bridgewater and Russell Wilson tweeted more than 10 times per week.

Tweet Content

NFL quarterbacks are often leaders on and off the field; consequently, their tweets can cover a wide range of topics and emotions.

Performance Measures Summary

Average fantasy points, total yards, total touchdowns, and quarterback rating were all fairly normally distributed. Quarterbacks averaged 19.22 fantasy points, with a standard deviation of 4.99 points. An average outing for a quarterback consisted of 286.52 (standard deviation: 40.08) total yards with 2.18 (standard deviation: 0.73) touchdowns and a 99.32 (standard deviation: 14.09) QBR.

Overall Trends

To gain an overview of the link between Twitter usage and on-field performance, we graphically analyzed the relationships between the log of a player’s tweets that week and their corresponding playing statistic using all weeks from all eligible players. All four performance measurements displayed positive but weak relationships with the log of tweet counts, as shown in the correlation matrix below created using chart. Correlation function of the PerformanceAnalytics package. The asterisks above the correlation coefficients correspond to p-values following the significance levels below.

1. 0.0001: ***
2. 0.01: **
3. 0.05: *

library(PerformanceAnalytics)
chart.Correlation(cordata, histogram=TRUE, pch=19)

Individual Correlation Comparison

To see which quarterbacks displayed the highest magnitudes of association between their performance and their social media usage, we compared their personal correlation coefficients relating their on-field statistics and tweet counts. We used fantasy points to investigate these values, as fantasy points are understandable for non-sports fans but still encapsulate overall performance.

According to the below chart of individual correlation coefficients, about 75 percent of our sample of 17 quarterbacks (3 were omitted due to lack of sample size for a correlation coefficient) falls in the small to medium positive relationship category between fantasy points and tweet counts. Ryan Tannehill had the highest correlation coefficient at 0.91, suggesting that 82 percent of the differences in his fantasy points can be associated with differences in his pre-game week tweet volumes.

Individual Linear Regression Model Comparison

We used simple linear regression to measure both the strength and validitiy of the fantasy point and tweet count relationship. Coefficients represent the number of additional fantasy points associated with one tweet during the week prior to a game. We used fantasy points to investigate these models, as shown in this Russell Wilson example regression model.

ryan_tannehill_model <- lm(fpts ~ tweetcount, data=ryan_tannehill_data)
summary(ryan_tannehill_model)

## 
## Call:
## lm(formula = fpts ~ tweetcount, data = ryan_tannehill_data)
## 
## Residuals:
##     1     2     3     4 
## -4.78  1.68 -1.68  4.78 
## 
## Coefficients:
##             Estimate Std. Error t value Pr(>|t|)  
## (Intercept)   -3.280      8.011  -0.409   0.7219  
## tweetcount     7.900      2.533   3.118   0.0893 .
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 5.067 on 2 degrees of freedom
## Multiple R-squared:  0.8294, Adjusted R-squared:  0.7441 
## F-statistic: 9.725 on 1 and 2 DF,  p-value: 0.08928

Only two models were significant at a significance level of 0.10 (Patrick Mahomes and Ryan Tannehill), with the average p value being 0.58. The average coefficient was 1.01 with a standard deviation of 2.00; because zero is within one standard deviation of the mean coefficient value, is it difficult to point to any major associations between fantasy points and tweet counts for these individual quarterback models.

Case Study Prediction: Patrick Mahomes

To truly test our models, legimitate testing and training would be completed; however, for a sample of that testing process, we used our most statistically significant linear regression model (Patrick Mahomes) and predicted his Week 8 fantasy point totals.

This model predicted Patrick Mahomes to score 24.45 points during Week 8, which makes sense, as his tweet count of 18 was just below his mean of 18.50, and 24.45 points falls just shy of his average fantasy score of 25.11.

Patrick Mahomes ended up scoring 36.64 points during Week 8, giving a high level of error at 49 percent. While this result is not enticing, it displays both the limitations of our small sample size and the multitude of variables affecting NFL performance data (Mahomes played the New York Jets during Week 8, who have given up the 3rd most fantasy points to quarterbacks of all defenses this year).

\[ \hat{fpts} = \beta_1\times(tweetcount) + \beta_0 \] \[ Patrick\;Mahomes\;Predicted\;Week\;8\;Fantasy\;Points = 1.02\times(18) + 6.09 = 24.45 \] \[ Percent\;Error = \frac{36.64 - 24.45}{24.45} = 49.85\%\]

Findings

We found that there was a weak, positive, and linear relationship between the log of number of tweets in a week and fantasy points. In addition, the log of number of tweets also showed a similar relationship with total touchdowns, total yards, and quarterback rating. Simple linear regression models for the four categories confirmed these findings, with all but total yards being statistically significant at a level of 0.05.

Limitations

Due to the small sample size of quarterbacks and weeks chosen for the model, our results are not generalizable to all quarterbacks in the NFL. The Patrick Mahomes and Ryan Tannehill linear regression models between the log of tweet count and fantasy points were significant at the 0.05 significance level. Individual linear regression models like these had at most 7 data points to fit a line of best fit. More data is needed to produce more accurate results, and we cannot confidently say that the results of these tests are accurate.

Another limitation was that the sentiment analysis included tweets and retweets by each player. It is possible that a player could have retweeted tweets with a higher percentage of positive words than tweets that they wrote. This could skew the results by that player (or players) having a higher positive word ratio than expected.

The word cloud that was created is limited by the scale of quarterback-specific tweet counts. For example, Ben Roesthlisberger had the most tweets by a quarterback and we see that some of the most commonly tweeted words relate to him and his team. We expect that if all the quarterbacks had similar tweet counts, then the weight of these words related to specific quarterbacks would decrease.

There are a multitude of ways that this project could be expanded to make its results more generalizable. First, we could gather data from Instagram, Facebook, and TikTok in addition to Twitter to get a truer picture of a player’s social media activity. In addition, we could analyze total social media output from a player’s entire career, dramatically boosting our sample size.

Second, we could adjust for the strength of the opponent for every game. While many of our data points are skewed by this lurking variable (for example, our model predicted Patrick Mahomes to play poorly after his lack of Tweets, but he excelled, facing a low-ranked New York Jets defense). This standardization would help capture the true impact of social media on athletic performance.

Third, we could measure not just the number of posts, but the time spent on social media to see if that is a more significant predictor of on-field performance.