Fourth Down Aggressiveness: The new trend, but does it work?
Until recently, fourth down in the NFL was the special teams down. When the down marker on the sideline switched to “4,” the offense came off the field. Depending on field position, either the kicker would come on to try for three points, or the punter would come on to willingly give possession of the football to the other team, under the guise of “playing the field position game.” But in today’s NFL, many coaches have started to realize that the game that actually matters is the points game, not the field position one — and one thing I hope we can all agree on is that a team is far more likely to score if they’re the ones in possession of the football.
However, not every coach has embraced the analytics, as they say. Pete Carroll in particular has come under fire for many questionable decisions in recent seasons, choosing to “go ahead and play defense,” despite having Russell Wilson at quarterback and a subpar defense. Yet, Carroll’s Seahawks have gone 11–5 and 12–4 in the past two years, while Doug Pederson (one of the pioneers of the aggressive movement) and his Eagles have gone 9–7 and 4–11–1 in the same time frame, leading to Pederson’s dismissal as Eagles Head Coach. This begs the question — while we as fans generally enjoy more offense, is being aggressive a sound strategy for long term success?
To answer this question, I used the nflfastR package to build my own fourth down model, and then simulated a season’s worth of drives twice, once using the model to make fourth down decisions, and once by always treating fourth down as the special teams down. The model was based on the expected points (EP) metric in the package. EP is calculated from a number of different factors, which are: seconds remaining in the half, yard line, whether the home team is the possession team, roof type, down, yards to go, era, and timeouts remaining for each team.
Since the Eagles are my favorite team (and Doug Pederson’s aggressive tendencies led me to learn more about advanced analytics in the NFL), I decided to use their play by play data for the simulation. The Eagles had, on average, 178 drives in those seasons, and 23 of those ended in turnovers. Rather than accounting for the probability of a turnover occurring on any drive, I simply subtracted those 23 drives from the simulation, leaving me with 155 drives to simulate (an interested reader could improve my project by building drives ending in a turnover into the simulation).
The fourth down decision model was fairly simple to build. Although there are some yard lines where a field goal, a punt, or a fourth down attempt are all reasonable decisions, I decided to set a hard cut off so the model never had to decide between the three. At the opponent’s 43 yard line, the decision to punt becomes the decision to kick a field goal, so we’re only trying field goals from 60 yards and within in the simulation. This is another area where the project could be expanded upon, by randomizing the cut off line and/or the decision. The decisions were made by calculating expected points for each situation and comparing them. For field goals, this was simply 3 times the probability of making the kick, and these probabilities are shown in the figure containing R code below. Punts were very tricky to do, because I had to try to account for the fact that we don’t know where the punt will land when we decide to punt the ball. My solution was to have two constant EP values for punts, based on the locations of Cameron Johnston’s 199 punts for the Eagles at the time of the simulation. The only difference between the two is that one had the Eagles as the home team, and the other had the Eagles as the away team (and these values were -4.23 and -4.55, respectively). To decide whether or not to go for it, I calculated the expected points for the Eagles given 1st&10 at the yard line needed for a first down, the expected points for the opponent given 1st&10 at the current spot, and the probability the Eagles would convert. The total EP was the weighted average of the two EP values.
Now, it’s time for the results of the simulation. The aggressive coach scored 290 points on 155 drives, good for 1.87 points per drive, whereas the conservative coach scored 230 points his 155 drive, good for 1.48 points per drive. However, it is important to note that both of these numbers are not great. By raw point totals, 290 points is the third worst offense of the 2020 NFL season, ahead of the two New York teams, and 230 points finishes 13 points behind the New York Jets for the worst offense in the NFL. Points per drive puts these two fictitious teams in a slightly better light. 1.87 points per drive would be 26th in the NFL this year, just behind the Los Angeles Rams, but 1.48 points per drive would be ahead of only those same New York Jets this season. We’ve established that neither offense would be particularly good, but is there a significant difference between the two?
To answer this question, I ran a two sample t-test in R to calculate the difference in the means of the two datasets (or the points per drive values). The most interesting part is the 95% confidence interval, which is (-0.186,0.961). Because this interval includes 0, we cannot say that either approach scores more points per drive than the other. The results of the t-test back this up, with a t score of 1.327 and a p-value of 0.185.
However, it is important to note that there are a number of limitations to this model. The biggest limitation, in my opinion, is that none of this accounts for game situation. Every drive was simulated independently of the others (and independently of any potential opponent), with the game setting always halfway through the first quarter. Even the most conservative coach would go for it on fourth down if punting the ball meant certain defeat. Other limitations are the absence of penalties (and every penalty goes into the play by play as zero yards gained, so there were an abnormally high number of zero yard plays in the simulation), and a small issue with the EP for a fourth down conversion. I calculated it as simply the EP if the Eagles gained exactly enough yards, but it is obviously possible to gain more than just the necessary yardage for a first down. This issue would have made the model slightly less likely to go for it on fourth down, but it is small enough that I’m not concerned about it. The issue I most want to correct in a future update of this model is the expected points for a punt, which I discussed earlier. Ideally, I would get this calculation by getting the expected points for the opposition at every possible place where a punt could land, and weighting it with the probability of the punt actually landing there. But once that issue is fixed, I could see this being relatively useful as a tool for me to be an armchair expert while watching NFL games from the comfort of my own home.
Author’s note: The R code used to both create the fourth down model and simulate NFL drives can be found at the following GitHub repository: https://github.com/jkatz22/fourth_down_model_KSA
