Portal de Periódicos da CAPES

Econometric tests of American college football's conventional wisdom

2010; Taylor & Francis; Volume: 43; Issue: 20 Linguagem: Inglês

10.1080/00036840903286331

1466-4283

Trevon D. Logan,

Sports, Gender, and Society

Abstract College football fans, coaches and observers have adopted a set of beliefs about how college football poll voters behave. I document three pieces of conventional wisdom in college football regarding the timing of wins and losses, the value of playing strong opponents and the value of winning by wide margins. Using a unique data set with 25 years of Associated Press (AP) poll results, I use a hedonic regression to test college football's conventional wisdom. In particular, I test (1) whether it is better to lose early or late in the season, (2) whether teams benefit from playing stronger opponents and (3) whether teams are rewarded for winning by large margins. Contrary to conventional wisdom, I find that (1) it is better to lose later in the season than earlier, (2) AP voters do not pay attention to the strength of a defeated opponent and (3) the benefit of winning by a large margin is negligible. I conclude by noting how these results inform debates about a potential playoff in college football. Acknowledgements I thank Rodney Andrews, Lisa D. Cook, Travis D. Logan, Thomas Logan, Jr., Anthony T. Logan, Vu Nguyen, Michael Sinkey and Michael Stengel for numerous helpful conversations. Paul J. Healy, Dan Levin, Matthew S. Lewis, James Peck, and participants at the 2008 AEA Pipeline Conference provided helpful suggestions. I am greatly indebted to my team of superb undergraduate research assistants who did the heavy lifting of raw data collection: Gregory Barson, Donald Butler, Mark Byrnes, Megan Collins, Collin Davis, Matthew Dodovich, Ashley Higgins, Michael Kuch, Patrick Sprinkle, and John Taylor. Jun Nakabayashi provided excellent help in auditing the data based upon an algorithm suggested by Travis D. Logan. The usual disclaimer applies. Notes 1 For example, Michigan's Michigan Stadium has seated more than 100 000 spectators since 1956, and Nebraska's Memorial Stadium (current capacity above 80 000) has been sold out for every home game since 1962. On average, a team will play at home only six games per season – streaks of 200 or more imply home sellouts for more than 30 years. 2 For example, more than half of all US televisions were tuned in to the Texas–Arkansas game in 1969, and the Army–Navy game of 1926 attracted a crowd of more than 100 000 (MacCambridge, Citation2005). In some locations, politics take a back seat to college football. In 2006, the recount of the contested election in Ohio's 15th congressional district was delayed by 1 day so that county election officials could watch the Ohio State–Michigan game. 3 The National Collegiate Athletic Association (NCAA) lists 10 organizations that bestow national championships in college football, some of them retroactively. 4 See Logan (Citation2007) for specific examples of the conventional wisdom. 5 For the purposes here, I am combining discussions of strength of schedule, opponent's strength and quality wins since the conclusions of these three lines of argument are the same. 6 Two important caveats to the discussion of the strength of schedule are the roles of schedules themselves and of conferences. First, it should be noted that college football schedules are decided many years in advance. For example, the full schedules for the next three seasons are already posted for most major college football programs. Since players only have a set amount of eligibility, and because the number of scholarships is limited, the quality of a future opponent, particularly well into the future, is unknown. This also means that the quality of opponents in a given year may be weakly correlated with the scheduler's intent – one cannot predict the future quality of a team when they do not know who the majority of the players will be. To that end, a team that devised to create a weak schedule may inherit a strong one through no fault of their own. Second, teams play a significant number of their games within a conference, whose strength in any given year is not known. 7 The original BCS formula incorporated margin of victory in the 1998–2001 seasons. 8 See Appendix A, available from the author, for further details. The 10 most winning programs of this time period were chosen, and the remaining 15 teams were chosen based on a stratification by quality (the likelihood of a team being consistently ranked over the entire time period) and geography (so as not to include a supermajority of teams from a particular conference, and to capture teams from all regions). One additional requirement was that a team had to be ranked in the final poll at least 25% of the available years. See the limitations of the data in this section. 9 Previous studies have looked at how rankings evolve over a season and not what features poll voters take as most important. Goff (Citation1996) looks at final season ranking as a function of mid-and preseason ranking, and Lebovic and Sigelman (Citation2001) also look at inertia in the polls. See Campbell et al. (Citation2007), Frechette et al. (Citation2007) and Paul et al. (Citation2007) for examples of work that use smaller samples of AP poll statistics. 10 Of the five teams not included in the data that won a national championship, Southern Methodist University (SMU) was beset with an NCAA ‘death penalty’ in 1986 that devastated the football program. Since SMU was already on NCAA probation fore recruiting violations (1985–1988), the second set of charges that surfaced in 1986 led the NCAA to issue the most severe sanctions allowed. The penalty was severe, not only the loss of scholarships (55 scholarships over a 4-year period), but the 1987 season was cancelled and the 1988 season called for only a limited number of away games to be played (SMU cancelled this season as well). All television games and bowl appearances for SMU football were disallowed during the 1988 and 1989 seasons. In addition, SMU players could transfer to another school without having to wait a probationary season to play (which is the procedure under normal NCAA rules). Since the sanctions by the NCAA, SMU has had one winning season (1997–1998, going 6–5), and several commentators believe that the SMU ‘death penalty’ led to the collapse of the Southwestern Conference due to lost revenues (and, through its demise, the emergence of superconferences such as the Southeastern Conference and Big 12). 11 The Coaches’ Poll is not used because the ballots were not released publicly until the last regular season ballot of the 2005 season. Similarly, the Coaches’ Poll has been criticized because there is evidence that the coaches themselves do not fill out the ballots (Barnhart, Citation1998) and that coaches have a strong bias towards their own teams, the teams in their conference, and their other opponents (Mandel, Citation2005). But, as recently stressed by Buchanan and Yoon (Citation2006) it is not possible to assert that there is no bias in the AP poll, particularly when pollsters achieve a surprising amount of consensus with regards to the number one team so often. 12 Campbell et al. (Citation2007) and Paul et al. (Citation2007) are recent studies of football ranking points, but they use a small number of variables and are not concerned with game characteristics themselves. 13 See the Appendix, available from the author, for ranking results. 14 It is very rare, but in the beginning of some seasons teams may play two games before the first updated ranking is released. This is so rare, however, that it does not affect the results discussed further. 15 For example, teams will usually play their eighth or ninth game by the 10th poll week. Considering that teams now play 11 or 12 games, this implies that three-fourth or more of the schedule has been played. 16 Although it varies from year to year, the average number of AP voters from 1980 to 2004 was 65. 17 Additional analysis, available in an Appendix from the author, shows that the main results are robust to a variety of alternative specifications and other checks. For example, the late loss result hold for alternative definitions of ‘late’ weeks, the margin of victory result holds for continuous measures of margin of victory, and the opponent strength result holds when looking at ranked teams only. For these specification checks and supplementary analysis, see the Appendix available from the author by request. 18 This was done in Table 3 but the results not shown due to space limitations. 19 These effects give each team an indicator and each season an indicator that is included in the regression. 20 As an example of this sort of effect that may be correlated over week of the season for successive seasons, opponent strength early in the season may be partially due to the fact that teams tend to play weaker opponents early in the season, and this effect may be the same one season to the next. 21 The results presented in the Appendix show that the failure of the conventional wisdom is robust to a number of specification checks and alternative definitions of key variables. 22 The AP polls are for each week are listed at http://www.appollarchive.com/football/index.cfm. 23 By NCAA rule, BCS division teams can only play one opponent per season from a non-BCS division. These games are almost always played early in the season. 24 When I replicated the results of Table 3 with opponent strength defined as wins minus losses at the time the game was played, the results were unchanged – this is likely due to the fact that the current win/loss record is highly correlated with the season win/loss record, particularly after the first few games of the season are played, and as such is not a strong check of the results. 25 In regression similar to those of Table A3 where I restrict the sample to be of ranked teams playing ranked teams (and where I do not include measures of the opponent's rank), the coefficient on opponent strength is 1.64 [0.14] the coefficient on winning * opponent strength is 0.42 [0.04], and the coefficient on losing * opponent strength is 12.89 [1.08] – t-statistics in brackets.