Is Controlling the Rushing or Passing Game the Key to NFL Victories?

### Abstract

#### Purpose

To evaluate whether controlling the running game or the passing game contributes more to winning in the NFL.

#### Methods

This analysis uses regression analysis to dispel the myth that controlling the rushing game wins NFL games. Final-game rushing and passing statistics are endogenous because teams that are ahead will rush more in order to protect the ball and run the clock down. To address this issue, I use first-half statistics (essentially stripping the endogenous component from the statistics), with the justification that the halftime leader wins 78 percent of the time. The data are the 256 regular season games for 2005. I use logistic models to model the probability of winning a game based on differences in rushing success and passing success in the first half.

#### Results

I find that having a first-half passing-yard advantage increases the probability of winning, but having a first-half rushing-yard advantage has no significant effect.

#### Conclusions and Applications

The results suggest that the common belief that controlling the running game is the key to winning in the NFL may be a misguided belief. Coaches and teams may have greater success by focusing on the passing games, both offensively and defensively.

**Keywords:** Football, NFL, passing, rushing, coaching

### Introduction

A common assessment of the key to winning professional football games is to control the running game. And a very common statistic used to support this claim is that teams are much more likely to win if they have a 100-yard rusher. This is often used in recapping games and when analysts describe the keys to victory. For example, the recap of a 2005 victory for the St. Louis Rams over the Houston Texans indicated: “[Steven] Jackson finished with 25 carries for 110 yards, improving the Rams’ record when having a 100-yard rusher to 38-0 since moving to St. Louis in 1995” (1). This implies that rushing 100 yards was the catalyst for the victory. Likewise, many analysts say that establishing the running game is a key to victory. For example, one analyst argued that a key to winning for the Tampa Bay Buccaneers over the Oakland Raiders in Superbowl XXXVII was to “contain Raiders’ [running back] Charlie Garner,” citing evidence that: “In the past five seasons, the Bucs are 1-12 when opponents have a 100-yard rusher” (11). In a closely tied statistic to rushing dominance, analysts also argue that teams that control the time of possession are more likely to win. These assessments imply that controlling the passing game is of much less importance.

A set of articles on espn.com using data from the NFL’s 2003 and 2004 regular season games supports these contentions by arguing that preventing an opposing runner from gaining 100 yards and winning the time-of-possession battle increased a team’s chances of winning (5, 6). At the same time, these articles imply that the passing game is insignificant, citing as evidence:

1. Teams having a 100-yard rusher win 75 percent of games;
2. Teams winning the time of possession battles win 67 percent of games;
3. Having a 300-yard passer has no advantage, as those teams only win 46 percent of games.

A problem with these simple assessments is that teams that are winning will rush the ball more to run out the clock and reduce the chance of turnovers and will often wait until the clock runs down before starting a play. So, if a team is heading towards victory, they are likely to increase their rushing yards while boosting their time of possession. Likewise, a team that is behind will pass more for potentially higher-gaining plays and to preserve the clock. Thus, in statistical terms, we could say that rushing yards, passing yards, and time of possession are endogenous, or partly a product of the outcome rather than just a contributor to the outcome. This makes it difficult to attribute any advantages in rushing yards or time of possession to the winner as causal impacts. In fact, what happens in the first half or even the first quarter can dictate the outcome of the game, as teams leading after just the first quarter (in 2003 and 2004 games) won 75 percent of the time (5).

This paper presents an empirical test of these issues with econometric analysis. Primarily, this analysis tests whether controlling the rushing or passing game was more likely to contribute to a victory in NFL games in the 2005 season. Rushing and passing advantages represent efficiency both on offense and defense. In addition, the model examines the relative contribution of turnovers, penalties, and sacks allowed to the probability of winning. These models could represent a more accurate picture of the effects of certain factors on winning, as they hold other factors constant.

The twist in this analysis is that the model corrects for endogeneity by using the first-half statistics. This essentially strips a large portion of the endogenous component from these statistics, as teams are not likely to change strategies to “ball preservation” or “speedy catch up” until the second half. Given that 78.5 percent of teams leading at halftime in 2005 games ended up winning the game, having a halftime advantage in many of these statistics should contribute to a higher probability of winning.

Determining the key contributions to winning in the NFL is important as teams, subject to the college draft and salary caps, attempt to obtain the best allocation of talent among various positions. If it does turn out that big passing games are the keys to victory, then investing relatively more on players in passing-related offensive and defensive positions than on players in rushing-related positions may be wiser.

Research on football issues has been very limited in the academic literature. There have been some interesting analyses on optimal 4th-down strategies (8, 9). Some research has attempted to predict the outcome of a game based on betting markets or power scores (4, 10). Other research has examined the success of teams over the course of a season (2, 3, 12). However, to my knowledge, this is the first analysis attempting to predict outcomes of games in a multivariate framework based on in-game statistics.

The most similar prior research examined how certain factors contributed to the number of wins NFL teams had (7). This article examined how first downs, average rushing yards per carry and passing yards per completion, interceptions, fumbles, and other factors affected the number of wins a team had, and then used the results to judge coaching efficiency. The models use full-game statistics so the results are subject to the biases mentioned above.

In this study, I first present a simple breakdown of the descriptive statistics for the first and second half, which clearly demonstrates the likely existence of endogeneity using the full-game statistics, as the eventual winner or the half-time leader clearly changes strategy in run-pass mix in the second half. The stark contrast found between the models using full-game vs. first-half statistics further corroborates how endogeneity affects the models using the full-game statistics. In particular, while the models using full-game statistics show a connection between controlling the rushing game and the probability of winning and no connection for controlling the passing game, the models using first-half statistics show the opposite: that controlling the passing game matters, but controlling the running game does not. Given that the analyses based on the first-half statistics should be free of biases from endogeneity, it appears that controlling the passing game is the key to winning. In addition, both full-game and first-half models show that the time of possession has no effect on the probability of winning, after controlling for other factors.

### Methods

#### Data

The sample includes all 256 regular season games from the 2005 NFL season. Each of the 32 teams has 16 games in the sample. The data come from the “Gamebooks” that are available on the NFL’s website (nfl.com). These data were used with permission granted from the National Football League’s Licensing Office. The advantage of these data is that they provide both final and first-half statistics, while a disadvantage is that the relevant statistics need to be manually extracted from each game report, which is roughly 10 pages long for each game. The descriptive statistics are presented in the Appendix. Table A1 shows the average team-level first-half and full-game statistics for the 512 team-game observations. Table A2 shows the average game-level statistics used in the econometric models for the 256 regular season games, with the key variables being “moderate” and “great” control of the rushing and passing games.

What is useful to show here are the differences that exist between first-half and second-half statistics for the eventual winners versus the losers and for the first-half leaders versus the trailers. These demonstrate how the second-half strategies can be dictated by first-half success, which is the basis for the argument that full-game statistics are endogenous to the outcome. Table 1 shows these results for the 243 games that did not go into overtime, as the second-half statistics cannot be calculated for the 13 games going into overtime because of how the NFL Gamebooks are set up. The first two columns, based on which team wins the game, show that, whereas the winner had an average of a 22-passing-yard advantage in the first half (119 versus 97), it had 26 fewer yards passing than the loser in the second half.

The next set of columns makes the comparison based on which team had the lead at halftime. There were 227 games in which one team led at halftime and the game did not go into overtime. Table 1 shows that there was little difference between the first half and second half in the rushing advantage for the halftime leader. However, that difference for the passing advantage is much greater. The halftime leader had a 34-yard passing advantage (126 vs. 92) in the first half and a 39-yard passing disadvantage (85 vs. 124) in second-half passing yards. Furthermore, the advantage for the leader in terms of fewer sacks allowed increased from 0.43 to 0.76. The differences are even starker in the final two columns for the 141 games that had a team leading by 7 or more at halftime. The 49-yard first-half passing advantage for the leader turned to a 50-yard disadvantage in the second half. And the 0.49 first-half advantage for the leader in fewer sacks allowed turned to a 0.90 second-half advantage. Note that the lack of much difference between the leader and trailer in first-half versus second-half rushing yards does not indicate that strategy does not shift, as the ratio of passing-to-rushing yards does increase for the trailer and decrease for the leader.

These results offer strong statistical evidence that the halftime leader passes less (probably to help protect the ball and run the clock down) and is more careful with the ball (with fewer turnovers). In addition, the results indicate that the halftime trailer passes more. The implication for statistical analysis is that many full-game statistics are likely endogenous to the eventual outcome. This includes rushing yards, passing yards, turnovers, and the number of sacks allowed. Thus, any comparison of full-game or final statistics for the winner versus the loser would be biased indicators of a causal effect.

#### Econometric Models

Given the likely bias that would come from using full-game statistics, the primary model will use first-half statistics, while still basing the outcome on the eventual game winner. As mentioned above, the justification for this is that 78.5 percent of the teams that led at halftime ended up winning the game. In order to provide a comparison so that readers can gauge the level of bias in using full-game statistics, an initial set of models will show the results from models using the full-game statistics.

The econometric model is the following:


where Y_i, the dependent variable, is a dichotomous indicator for whether the home team won game i, R_i and P_i represent measures of the rushing and passing advantages of the home team relative to the visiting team, Xi is a vector of three other statistics for the home team relative to the visiting team, including penalty yards, turnovers, and sacks allowed, and Hi and Ai are vectors of 31 indicator variables for which team is the home team and away team in game i, with one team excluded. Thus, all statistical variables are created in terms of the home-team statistic minus the visiting-team’s statistic or, in a few cases, the advantage of the home team over the away team. For example, the variable for rushing-yards advantage would be the number of rushing yards for the home team minus the number of rushing yards for the visiting team. The results would be the same regardless of whether the model predicts the probability of the home team or the visiting team winning.

For both sets of models with final statistics and first-half statistics, three sets of rushing and passing variables are created. The first set has the raw difference in rushing and passing yards, measured as the advantages the home team has over the visiting team. The second set has a variable indicating whether one of the teams had “moderate” control of the rushing or passing yards, with the threshold being 25 yards for the models with first-half statistics and 50 yards for models with full-game statistics. For the models with first-half statistics, this variable is coded as “1” if the home team had at least 25 more rushing (or passing) yards than the visiting team at halftime, “-1” if the visiting team had at least 25 more yards than the home team, and “0” if the absolute difference in yards between the two teams was less than 25. The third set of variables, constructed similarly to the second set, has variables indicating whether one of the teams had “great” control of the rushing or passing game. The thresholds are 50 yards for the models with first-half statistics and 100 yards for models with full-game statistics. Note that these variables taking on the values of (-1, 0, 1) essentially constrains the absolute values of the following two effects to be the same: (a) the effect of home-team control of the rushing/passing game on the probability of the home team winning and (b) the negative of the effect of visiting-team control of the rushing/passing game on the probability of the home team winning. This helps to give greater power to the model.

The models include three other statistical variables: the difference in penalty yards, the difference in turnovers, and the difference in the number of times the team is sacked. Including the number of penalties had a very small effect, so it was excluded so that the full effect of penalty yards could be estimated.

Finally, the model includes team fixed effects for both being the home team and being the visitor. That is, it includes 31 dummy variables for the home team and 31 dummy variables for the away team, excluding one team as the reference category. They account for differences in team-specific factors, such as the quality of coaching and the strength of home-field advantage (e.g., from fan enthusiasm and weather conditions). In addition, the team fixed effects account for differences in the strength and weakness of the passing vs. rushing games for teams and for opponents.

These team fixed effects are included to help avoid unobserved team heterogeneity affecting the results. For example, one of the better teams in 2005 was the Indianapolis Colts, which had a very strong passing game. Thus, without team fixed effects, the general success of the Colts could contribute to a positive correlation between passing yards and winning that could be due to other unobserved factors. By including team fixed effects, the estimates represent within-team variation across games in winning attributable to within-team variation across games in control of the rushing and passing game. The coefficients on these (not reported) generally reflect differences across teams in both home and away winning percentages, after taking into account the other variables included in the model.

Equation (1) is estimated with logit models. The models have a final sample of 212 games because 44 games were dropped by the model due to perfect prediction of the outcome—e.g., 8 observations were dropped for Seattle home games because they won all those games. In estimation, it turned out that that the marginal effects were highly dependent on the home and visiting teams used for the prediction. Some teams that won (or lost) nearly all their home or away games were too close to a predicted probability of winning of one (or zero), so that the marginal effect of the variables would be close to zero for them. To correct for this, the reported marginal effects are calculated as the averages for all team combinations that played in the 2005 season.

The model presented here is fairly simple. One reason for this is that the home- and away-team fixed effects account for a wide set of team-specific factors (some unobservable and some observable), such as the quality of coaching, having artificial turf, and generally favoring either passing or rushing. The other reason why the model is kept simple is that it is designed to estimate the full effects of having advantages in the rushing game and the passing game. As it turns out, this simple model tells an interesting story.

The model could be made more complex by including such factors as the run-pass mix, time-of-possession, and return yards off of kick-offs and punts. These other factors are excluded because they could themselves be products of running and passing success in the game. For example, having a higher time-of-possession is an indicator of rushing the ball successfully. And, having a rush-pass mix favoring passing may be an indicator of success in the passing game. Controlling for these variables would cause the model to factor out part of why having rushing or passing advantages helps win games, so that the coefficient estimates on the rushing and passing advantages would represent partial effects rather than the full effects the model aims to estimate. Separate analyses below do test whether time-of-possession matters, after controlling for rushing and passing yards, as well as the other factors that are in our primary set of models.

Another factor excluded from the model for similar reasons is the number of return yards from kick-offs and punts. Return-yard success (or more generally, special-teams success) could be representative of other factors. Indeed, one of the ESPN articles notes that teams returning a punt or kick-off for a TD win only 42 percent of the time (6). One confounding factor is that teams have a greater chance of return success on kick-offs than on punts, but having more kick-off returns is an indication that the other team has scored more often. Given these complexities, we exclude return-yardage indications. Given that we use team fixed effects, this should not be a problem to our analysis, as within-team variation in special-teams success relative to the other team (holding constant special-teams’ opportunities) should be mostly uncorrelated with the within-team variation in rushing and passing success relative to the other team.

### Results and Discussion

#### Is controlling the rushing or passing game more important to winning?

Table 2 presents the results of the econometric models that examine the relationship between full-game statistics and the probability of winning. These results are subject to biases created by the endogeneity described above, so they are meant to be compared to the results of the preferred model, in Table 3, which is based on the relationship between first-half statistics and the probability of winning.

The results in Table 2 are consistent with the widely held belief that controlling the rushing game is the key to winning and that great passing success is not important. The coefficient estimate on the rushing-yard difference is positive and significant at the one-percent level. The corresponding marginal effect, in brackets, indicates that each 10-yard advantage in rushing yards is associated with a 2.3-percentage-points higher probability of winning (p < 0.01). The coefficient estimate on passing-yards advantage is small and insignificant. Considering the indicators for “moderate” control of the rushing and passing game, having a 50-yard advantage in rushing yards is associated with an estimated 17.2-percentage-points higher probability of winning (p < 0.01). The estimate on having a 50-yard advantage in passing yards is again insignificant. Having “great” control of the rushing game (100-yard advantage) is associated with an estimated 31.4-percentage-points higher probability of winning (p < 0.01). Having “great” control of the passing game is still statistically insignificant.

As for other results, each turnover is associated with a decrease in the probability of winning of about 16 percentage points (p < 0.01), while each sack is associated with an 11-percentage-points decrease in the probability of winning (p < 0.01). These seemingly large effects could be indicative of the extra chances that teams take when they are behind late in the game. Penalty yards do not appear to make a difference, after controlling for other factors.

The main point from the models using full-game statistics is that total rushing yards or controlling the rushing game is positively correlated with the probability of winning, while passing yards and controlling the passing game has little correlation with the probability of winning.

The results from models using first-half statistics give the opposite conclusion. The estimates indicate that controlling the passing game is the key to winning, not controlling the rushing game. In contrast to the results in Table 2, those in Table 3, for the coefficient estimates on first-half statistics, arguably represent causal effects because most teams probably do not start the strategy of protecting the ball and running out the clock to end the game while still in the first half.

All three of the coefficient estimates on the passing yard advantage are positive and significant (p < 0.01). The estimates on rushing yard advantage are still positive, but smaller than those for the passing-yard advantage and statistically insignificant. The estimated marginal effects indicate that each 10 yards of passing gained increase the probability of winning by 2.6 percentage points, while having a 25- or 50-yard-passing advantage in the first half increases the probability of winning by about 21 percentage points. Thus, these estimates indicate that controlling the passing game in the first half increases a team’s probability of winning the game by about 12 percentage points, while controlling the rushing game in the first half has no significant effect on the probability of winning.

Among the other factors, first-half penalty yards again do not affect the probability of winning. Each turnover is estimated to reduce the chance of winning by about 10 percentage points (p < 0.01), while each sack allowed reduced the probability of winning by about 5 percentage points (p < 0.10). The estimated marginal effects of turnovers and sacks allowed are smaller for the first-half model than for the full-game model. This could indicate that, like rushing and passing yards, the full-game statistics on the number of turnovers and sacks allowed are endogenous and reflective of the outcome of the game, as the teams that are behind will be susceptible to more turnovers and sacks as they pass more and take more chances to try to catch up.

#### Does time of possession matter?

Another commonly-held belief is that having a greater time-of-possession is a major key to winning, as 67 percent of the teams that won the time-of-possession battle in 2003 and 2004 had won their games(5). This suggests that winning the time-of-possession battle increases a team’s chances of winning by about one-third. However, this statistic is also a product of a team’s success (or endogenous) and thus subject to biases. For example, teams that are ahead will let the clock run down further between plays.

Table 4 presents the coefficient estimates on variables representing time of possession from models similar to column (1) in Tables 2 and 3—i.e., models that use the rushing- and passing-yard advantage. It includes estimates using the full-game and first-half statistics. The first row has the estimates on the actual time-of-possession advantage; the second row has the estimates on indicators for whether the team had a higher time-of-possession, and the last two rows have estimates on indicators for having advantages of 7 minutes (for the full game) and 5 minutes (for the first half), which are roughly the average mean absolute differences. For the full-game statistics, none of the time-of-possession variables is statistically significant. For the models based on first-half statistics, all of the coefficient estimates on time of possession are negative, with the first one being statistically significant (p < 0.10). These results suggest that time-of-possession is not important to winning, holding constant other factors.

### Conclusions

This paper is the first analysis to model a production function for winning an NFL game based on in-game statistics. This carefully constructed framework, which models victories based on home-team over away-team statistics, can be used for other models for winning games in the NFL or in other sports leagues.

The results of this analysis cast doubt on the contention that the key to winning games in the NFL is to control the rushing game. The results do indicate that having a rushing advantage for the full game is positively correlated with the probability of winning and having a passing advantage for the full game is not correlated with winning, holding other factors constant. However, these correlations are likely due to endogeneity, in that full-game rushing and passing yards are partly products of a team’s success during the game. In other words, as demonstrated in this paper, the strategy for second-half rushing-passing mix depends on where a team stands at halftime. This means that we cannot label these correlations as causal influences.

The econometric strategy in this analysis is to identify a causal effect of various factors by using first-half statistics. These first-half statistics should be exogenous because strategies to run the clock down and to take extra precautions of preserving the ball (and to play catch-up by passing the ball so that incompletions stop the clock) arguably do not start until sometime in the second half. Of course, there could be cases in which teams build such a huge lead early in the first half that they start such a strategy at some point in the second quarter. But, typically teams that are ahead would want to build on their momentum in the first half before shifting strategy at some point in the second half.

One other key result is that having a time-of-possession advantage does not matter, after controlling for other factors (e.g., rushing and passing yards). However, the major findings from models using first-half statistics are that, on average, controlling the passing game contributes significantly to the probability of winning and controlling the rushing game has little impact. Having some level of control over the passing game in the first half is estimated to increase a team’s chance of winning by 21 percentage points. It is not that rushing success does not matter, as many would argue that having the threat of a potent running attack is key to a successful passing game. In addition, a strong running game may help with ball preservation for holding a second-half lead. But, in contrast to conventional thought, holding other things constant, it appears that a big passing day is more important to victory than a big running game. It is important to keep in mind here that passing advantage and control incorporates both how strong a team’s passing game is and how strong its pass defense is.

### Applications in Sport

The results in this analysis suggest that NFL coaches may be more successful if they were to place more emphasis on the passing game than on the running game. This result may translate to lower levels of football (e.g., high school and college). In this case, for professional football or something lower, obtaining and developing premier players for passing-related offensive and defensive positions may be more important than obtaining and developing premier players in rushing-related positions.

### Tables

#### Table 1
A comparison of first-half and second-half statistics for the eventual winner versus the loser and the halftime leader versus the trailer.

**NOTE:** These statistics exclude the 13 games that go into overtime because second-half
statistics cannot be determined.

#### Table 2
Logistic regression model for the relationship between full-game statistics and the probability of winning (N=212)

**NOTE:** *, **, and *** indicate statistical significance at the five- and one-percent level. The models also include dummy variables for each visiting team and home team. Standard errors are in parentheses and marginal effects are in brackets.

#### Table 3
Logistic regression model for the relationship between first-half statistics and the probability of winning (N=212)

**NOTE:** *, **, and *** indicate statistical significance at the five- and one-percent level. The models also include dummy variables for each visiting team and home team. Standard errors are in parentheses and marginal effects are in brackets.

#### Table 4
Logistic regression model for coefficient estimates on time-of-possession variables

NOTE: *, **, and *** indicate statistical significance at the ten-, five- and one-percent level. The models also include dummy variables for each visiting team and home team. Each coefficient estimate is based on a separate regression. These regressions include, for either full-game and first-half statistics, the same regressors represented in column (1) of Tables 2 and 3. Standard errors are in parentheses and marginal effects are in brackets.

#### Table A.1.
Average team statistics in key categories for 2005 regular season games (N=512)

NOTE: Standard deviations are in parentheses. The final-game statistics include 13 overtimes (or 26 observations), all of which lasted less than the full 15 minutes allowed. Thus, the differences do not exactly represent second half statistics.

#### Table A.2.
Average game statistics in key categories for 2005 regular season games (N=256)

### References

Associated Press (2005). Rams Rally to Down Texans in Overtime. <http://www.tsn.ca/nfl/teams/news_story/?ID=144796&hubname=nfl-rams>, accessed August 28, 2006.

Berri, D. (2007). Back to back evaluations on the gridiron. In Statistical Thinking in Sports. Albert, J., and Konig, R.H. eds. CRC Press, Ann Arbor, MI. pp. 235-56.

Berri, D., Schmidt, M., and Brook, S. (2006). The Wages of Wins. Stanford University Press, Stanford, MI.

Boulier, B. L., and Stekler, H.O. (2003). Predicting the outcomes of National Football League games International Journal of Forecasting, 19, 257−70.

Garber, G. (2005a). Turnovers, early deficits lead to losses. <http://sports.espn.go.com/nfl/news/story?id=2241121>, 2005a, accessed December 2, 2005.

Garber, G. (2005b). Penalties hurt but aren’t indicator of failure. <http://sports.espn.go.com/nfl/news/story?id=2241159>, 2005b, accessed December 2, 2005.

Hadley, L., Poitras, M., Ruggiero, J., and Knowles, S. (2000). Performance Evaluation of National Football League Teams. Managerial and Decision Economics, 2000, 21, 63-70.

Kreider, B. (2006). To Punt or Not to Punt. The UMAP Journal, 17, 353-63.

Romer, D. (2006). Do Firms Maximize? Evidence from Professional Football. The Journal of Political Economy, 114, 340-65.

Song, C, Boulier, B. L., and Stekler, H. O. (2007). The comparative accuracy of judgmental and model forecasts of American football games. International Journal of Forecasting, 23, 405–13.

Stroud, R. (2003). Keys to Victory, St. Petersburg Times, January 26, 2003, <http://www.sptimes.com/2003/01/26/Bucs/Keys_to_victory.2.shtml>, accessed August 25, 2006.

Terry, N. (2007). Investing in NFL Prospects: Factors Influencing Team Winning Percentage. International Advances in Economic Research 13, 117.

### Corresponding Author

**Jeremy Arkes**
Associate Professor of Economics
Graduate School of Business and Public Policy
Naval Postgraduate School
555 Dyer Rd.
Monterey, CA 93943
<[email protected]>
831-656-2646

### Author Biography

Dr. Jeremy Arkes is an Associate Professor of Economics in the Graduate School of Business and Public Policy at the Naval Postgraduate School.