Introduction
Many teams throughout the history of sports have achieved success by taking advantage of market inefficiencies, a result of teams having different information. The Oakland Athletics, under Billy Beane, are most famous for this, focusing their attention to players who had a high on-base percentage instead of batting average, the conventional metric for evaluating hitters in baseball for many years. Every sport has examples of teams taking advantage of better information and using it to maximize the quality of their team. In the NFL, the market inefficiency was throwing the ball. The old-fashioned way of thinking in the NFL was that establishing the run was the most important task for an offense to be successful. Recently, teams have found success by emphasizing the passing game and being more aggressive, a result of data showing that the passing game is more efficient than the running game.
In football, in addition to showing the advantages of a prolific passing attack, analytics have provided evidence that play-action passes are also very efficient. Data on passing plays during the 2018 NFL season shows that the expected points added (EPA) per play-action play, of which there were 3,271,922 during 2018, was .2605. The EPA per non-play-action play, of which there were 14,425,129 during 2018, was -.0347. With the increased usage of play-action over the past several years and the significant boost in EPA/play generated from its use, it has become of critical importance that teams figure out how to defend play-action, or at least mitigate its success.
Using simple linear regression, play-action has a statistically significant effect on EPA/play, but what is it about play-action that makes it so successful? Is it the physical act of faking a hand-off, or are there other factors? This report will begin to analyze what makes play-action successful and if these findings can be used to help defenses neutralize the play.
Using Separation Data to Analyze Play-Action
In this report, one of the metrics that will be used to analyze play-action is separation, the distance between a receiver and the defender covering him. The tracking data provided for this competition included x- and y-coordinates for each potential receiver and defender in coverage every tenth of a second during a play. In order to calculate the separation for each matchup, the data frame was filtered to only include the frame at which the quarterback threw the ball because the original data frame was too large to complete the next steps. This event, when the ball is thrown, was chosen to analyze separation because when the quarterback throws the ball, he is likely going to throw to the receiver who is the most open, who has the most separation. Although they will often make a throw anticipating their receiver getting open after, this seemed like one of the most consequential frames for separation (along with when the ball arrives).
After splitting the data frame into two separate data frames for offensive and defensive players, they were joined using the left_join function in R in order to get an observation for each individual potential matchup (a row for each defensive and offensive player for every frame of a play), and separation was calculated using the distance formula. The data was then filtered to include the smallest separation for each defensive player, since the analysis is from their perspective, in order to generate matchups with receivers. This resulted in data on the closest receiver to each defender, which could be skewed if there is blown coverage (two defenders follow the same receiver and let another go free). A similar method was used to generate a column in the data for play-action (see code in the Appendix for more).
Separation is a valuable metric because it allows us to evaluate defenders even if they were not targeted on the play. If we only use EPA, a defender could be penalized if one of his teammates messes up in coverage, resulting in a high EPA even if he had great coverage on his receiver. Below are two graphs that illustrate which defensive backs (DBs) and linebackers (LBs) were the best in coverage in 2018 for every pass (minimum 200 plays) and for play-action passes only (minimum 50 plays).
Byron Jones had the lowest separation among DBs on all passes and play-action passes at 2.821 yards and 2.810 yards, respectively. Roquan Smith had the lowest separation among LBs on all passes at 5.278 yards, but was 20th on play-action passes at 6.421 yards. Jatavis Brown was first among LBs against play-action with an average separation of 5.712 yards.
Now, the report aims to answer the question: what makes play-action successful? It will start by examining what affects separation and if play-action inflates those effects.
The first factor that this report examines is matchups; does play-action result in favorable matchups, and what are favorable matchups? Below is a graph that shows the separation generated by the targeted receiver in each positional matchup and the effect of play-action. Against wide receivers (WRs), there is a relatively large difference in separation between DBs and LBs. However, there is not much of a difference in separation against running backs (RBs) and tight ends (TEs) between DBs and LBs.
Every matchup has a higher separation in play-action than non-play-action, but the difference is not that large. However, the graph shows that the LB-WR matchup is advantageous for the offense whether or not they run play-action, especially compared to the DB-WR matchup, so the offense would benefit from generating this matchup more often. The DB-WR matchup was targeted on 44.78% play-action plays and 45.67% of non-play-action plays, which were both the highest rates. The LB-WR matchup was targeted the second most often but at only 15.68% and 16.61% on play-action and non-play-action, respectively, so it does not appear that play-action success is a result of the offense taking advantage of beneficial matchups.
Another factor that could affect separation on play-action plays is the route that the receiver runs. Below is a graph that shows the average separation for each route on play-action and non-play-action plays created by the targeted receiver, grouped by defender’s position.
The graph illustrates that most of the routes create more separation on play-action compared to non-play-action regardless of the position of the defender, although that difference is not always significant. The routes with the largest difference in separation between play-action and non-play-action were wheel (1.95 yards), go (1.30 yards), corner (1.20 yards), and hitch (0.96 yards). This difference is also larger against LBs than DBs for some routes, such as hitch, post, and go.
Using EPA to Analyze Play-Action
The rest of the report will focus on what makes play-action successful. Play-action is effective, simply, because the defense thinks the offense is running the ball. However, is the physical fake hand-off the factor that makes play-action successful or are there other factors? Is it enough for the offense to be in a running formation in order for play-action to work? We will use EPA to attempt to answer these questions, as opposed to separation data, because we are interested in the success of the play as whole, not just individual matchups.
Whether or not the offense is in shotgun is a factor that can make a defense believe that the offense is running the ball. In 2018, 80% of passing plays were shotgun. The graph below illustrates how play-action and shotgun affects the success of passing plays.
This graph shows a couple of important ideas. One, regardless of whether or not it’s play-action, the EPA/play is higher when the offense is not in shotgun. Two, EPA/play is higher for both shotgun and non-shotgun plays when play-action is used. This shows that both the use of play-action and a run formation results in a more successful pass.
Two linear regression models were created to observe the effect that play-action and whether or not the offense is in shotgun has on EPA/play. The first model was a simple linear regression with only one predictor: play-action. According to this model, play-action increases predicted EPA/play by .207, on average. The second model took shotgun into account, and the effect of play-action on EPA/play decreased to .108. Both models showed that play-action has a statistically significant effect on EPA/play, but the second model illustrates that there are other factors that aid in its success; the simple linear regression model overestimated play-action’s effect.
Personnel also impacts what kind of play the defense thinks the offense is running. The idea that this part of the report is looking to test is whether or not personnel impacts the success of play-action due to the fact that certain personnel will make the defense believe that the offense is running the ball even before they fake the hand-off. A linear regression model was used to examine the effects that the number of defensive backs, the number of linebackers, and play-action have on the success of passing plays. (The number of defensive linemen was removed from the model due to concerns of multicollinearity).
The number of defensive backs and play-action have a statistically significant effect on EPA/play at the .001 significance level, although the number of linebackers does not, but the effect due to play-action is smaller when taking personnel into account (.207 to .152). If the defense thinks the offense is more likely to run the ball, they will have few defensive backs on the field, which leads to a more successful passing attack, whether or not they run play-action. This is shown in the graph below which illustrates how the success of a passing play is affected by the number of DBs and whether or not it is play-action. (The number of DBs is limited to less than or equal to 6 to include data for play-action and non-play-action).
The graph shows that passing with three or fewer DBs is successful whether or not the offense uses play-action. (It would be interesting to research further to see what causes such a large and opposite difference between play-action and non-play-action for two and three defensive backs). On the other hand, with four or more DBs, play-action is always more successful although passing in general is not.
Conclusion
This report aimed to shed light on what makes play-action successful; however, it did not take every potential factor into account. It showed that play-action, and defending play-action, can be aided by formation and personnel, but the actual act of faking a hand-off is important as well. A potential follow-up report could examine how defenses react at the moment the quarterback fakes the hand-off. Which defenders react the best to play-action? How does that reaction affect the success of the play?
My research showed that defenses may be able to mitigate the success of play-action if they can anticipate it. More defensive backs will make play-action less successful, and the same logic is already being applied to defending the pass in general. However, teams are concerned about how it affects their run defense, so another follow-up project could attempt to determine how often and when a defense should focus on stopping the pass vs. the run.
Here is the link to the Kaggle Notebook which includes the code used to complete this analysis hidden at the bottom: https://www.kaggle.com/jacobbleiweis/defending-play-action.
Comentarios