Estimating NCAA Football Coaches’ Abilities An Application of Item Response Theory

Brandon LeBeau

2016-07-31

Background

  • Began after Tim Brewster was fired at the University of Minnesota.
    • Now they have a new coach again!
  • Wanted to try to predict next great coach.
  • Proceeded to explore data available to answer this question.

Data

Goals

  • Predict the ‘ability’ of the coaches.
  • Find other variables that explain variation in the ‘ability’ of the coaches.
  • Predict the next ‘great’ coach.

Model

Model 2

  • Not a traditional IRT model as the game ID is not included.
    • The model does allow for a coaches ability to vary with years.
    • The team effect is constant.
  • This model was fitted using R: lme4 and rstan.
  • Results shown throughout are from lme4.
  • ‘Modern’ era data, 1998 onward and coaches with at least 6 games per year.

Team Effect

Ability Estimates

Iowa State made good hire?

Fire Tim Brewster?

Does a coach overperform compared to average team ability?

  • Value-added like logic.
  • If coaches over/under perform compared to a team average
    • are they more likely to be retained?
    • if fired, are they a good fit for another team?
  • How quickly does team expectation change?

6 Coaches on the Hot Seat

Which of these 6 should be fired?

Back to Minnesota

Next Steps

  • Add covariates to model shown above
    • or use the ability estimates obtained above as an outcome.
  • Explore moving average for team expectation.
  • Explore good vs bad coaching hires.

Connect