As an avid Illinois football fan, I have been suitably dismayed by the tenure of Tim Beckman thus far. After a resoundingly awful 2-10 first year, Beckman lost the fanbase from the outset and hasn't been able to win it back. Still, after a 16-14 win over Penn State on Saturday, the Illini have moved to 5-6 on the season and are just one win away from bowl eligibility. Speculation is rampant that Beckman needs to beat Northwestern this Saturday to save his job, but I wondered whether or not that was really the case.
In an effort to objectively assess the matter, I created a logistic regression model to predict whether or not a coach would be fired at the end of the season. Using data from Sports-Reference and Stassen, I looked at a total of 531 coach-seasons - the entire tenure to date of any coach hired by a Power 5 school during the BCS era. The goal: to predict whether a coach would be fired at the end of any given season.
After a fair amount of tinkering, I've developed a pretty solid model to predict whether a coach will get the ax. Although no model will fully capture the individual variables that go into the decision of whether or not to fire a coach, I found that the following 7 variables were all highly statistically significant in predicting the downfall (or lack thereof) for any given coach:
1. Current season win percentage
2. Historical program performance - I used program winning percentage over the last 25 years leading up to the coach's current season. This helps to account for situations like Duke not firing Carl Franks after he went 0-11 in his third year, but Florida firing Ron Zook after he went 7-5, also in his third year.
3. Is the coach in his 2nd season? - Historically, most coaches in their 2nd years are safe, no matter what they do. There are some recent exceptions like Turner Gill at Kansas and Jon Embree at Colorado that give some credence to the idea this trend could be changing, however. (Note: 1st-year coaches were excluded from the data set altogether because exactly zero of the coaches I collected data for were fired in year 1).
4. Is the coach in his 3rd season? - Same idea as variable #3, 3rd-year coaches are often given the benefit of the doubt - however, firings do become more prevalent in year 3 so this variable isn't as strong as for year 2 coaches. I tested a variable for 4th-year coaches as well but it was no longer statistically significant, indicating that by year 4, the coach can no longer rely on the excuse of still building the program.
5. Is there a new athletic director? - This proved to be a very significant variable - new athletic directors are more likely to make a splash and fire coaches who they didn't bring on board themselves.
6. Is the athletic director in his 2nd year? - The coach is still not out of danger if he makes it through his new AD's 1st season, as there were several instances in the data set where a coach was fired in the 2nd year of his AD's regime.
7. 'Approval rating' - I wanted to account for the entire tenure of a coach to reflect that a coach can build up good will with several strong seasons in a row (or bad will, as the case may be). For each coach, I measured the expected winning percentage for that season (set as the average of the previous year's winning percentage and the program's winning percentage over the previous 25 years), and then measured how their winning percentage compared to that expectation. I then looked at the coach's cumulative results over or below expectation during their tenure, the result being their 'approval rating' at season's end.
Overall, the model is certainly not perfect, but is in the right ballpark for the majority of coaches. Since I like to show my work, here's the output from R:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -1.5575 0.8247 -1.889 0.058950 .
W. -6.2276 1.5934 -3.908 9.29e-05 ***
ProgramWL25 5.1321 1.8918 2.713 0.006671 **
Year.2. -3.5507 0.7106 -4.997 5.83e-07 ***
Year.3. -1.4684 0.4502 -3.261 0.001109 **
AD.Year.1. 1.8624 0.4706 3.958 7.57e-05 ***
AD.Year.2. 1.2859 0.5723 2.247 0.024647 *
APPROVAL -2.8100 0.7587 -3.704 0.000213 ***
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
(Dispersion parameter for binomial family taken to be 1)
Null deviance: 382.77 on 530 degrees of freedom
Residual deviance: 221.85 on 523 degrees of freedom
With the model in place, I of course had to take a look at current coaches and their likelihood of being fired. Here are the ignominious leaders in the clubhouse so far for 2014:
1. Bill Blankenship, Tulsa - 83.3%
Blankenship inherited a decent program from Todd Graham but has gone just 3-9 and 2-9 the last two seasons, making him a prime candidate to be let go.
2. Brady Hoke, Michigan - 80.8%
Unsurprisingly, the much-maligned Hoke may not be long for Michigan. While his overall win percentage isn't horrible by the standards of many programs, 5-6 doesn't cut it at Michigan, which had won 70.5% of its games over the past 25 years.
3. Bobby Hauck, UNLV - 76.6%
UNLV is not a particularly strong program, but after Hauck saved his job in year 4 with a 7-6 season after 3 straight 2-win seasons to open his career, he's gone right back to 2 wins in year 5 and seems a likely candidate to be fired accordingly.
4. Al Golden, Miami (FL) - 70.8%
The model doesn't account for NCAA sanctions, so Golden may get the benefit of the doubt, but similar to Hoke, 6-5 as a 4th year coach doesn't cut it when your school has won 73.2% of its games over the past 25 years.
5. Mike London, Virginia - 51.0%
The model predicted London was a goner last year after a 2-10 year 4 (80.9% chance of firing) which would almost certainly be the death knell for most coaches. Yet possibly due to a large buyout, London made it back for year 5. Although he's had a stronger season (5-6 to date), he still is a candidate to be let go, as he has yet to turn around the Virginia program after 5 years.
6. Will Muschamp, Florida - 48.1%
Muschamp's resignation has already been announced, but the model viewed him as a coin-flip case to be let go. At 6-5, Florida's season hasn't been completely awful but he's failed to erase the bad memories of his 4-8 season a year ago.
7. Kevin Wilson, Indiana - 38.4%
Indiana is a historically weak program and Wilson may get another chance based on that alone. However, with 0 winning seasons in his tenure, it wouldn't be a big surprise if Indiana decided to make the change.
8. Bo Pelini, Nebraska - 31.8%
Pelini is an interesting test case as he's won 9 or 10 games in every year at Nebraska and is on track to do so this year as well. The question is whether or not Nebraska is still a top 10 program, or perhaps more pertinently, if Nebraska's administration still sees them that way.
9. Dana Holgorsen, West Virginia - 26.4%
Holgorsen survived a bad year 3 (4-8) and West Virginia has been much improved this year so he may be safe. However, the Mountaineers are still only 6-5 so it's still possible that a change could be made.
10. Paul Rhoads, Iowa State - 24.9%
Rhoads has built up some goodwill for making three bowl games at a bad Iowa State program, but he's now 5-16 over his last 2 seasons, opening up the possibility that he could be let go.
Other coaches of note:
Charlie Weis, Kansas - 16.8%
Since Weis was still only in his 3rd year and inherited a bad situation, the model predicted he was unlikely to get fired. However, Kansas's administration has proved to have a quick trigger finger, also cutting Turner Gill loose after just 2 seasons.
Tim Beckman, Illinois - 5.7%
And of course, after I built a model to gauge my chances of seeing a new coach at Illinois, each new iteration brought Beckman's expected chances of being fired down more and more. Whether Illinois fans like it or not (and I certainly don't), Illinois rates as a historically weak program, and he's still only in his third year, along with the AD (Mike Thomas) that hired him. I suspect Beckman's actual chances of being let go are much higher than this, but there could well be an unwelcome surprise in store for Illini fans who've been waiting to see Beckman get the boot.
Have any questions? Any other coaches you want to see the percentages for? Please comment below!