Last week I examined how one can use logistic regression to estimate the value of a timeout, with minimal success. I promised a better way to estimate this value which frees us from some of the inherent limitations to logistic regression, mainly that the value of timeouts are linear (e.g. the difference in win probability from 3 to 2 timeouts is the same as from 2 to 1) and constant (e.
On December 27, the Pittsburgh Steelers were down 13-3 against their rivals, the Baltimore Ravens, after the first half. The Ravens opened the second half with the ball, but quickly punted the ball following a three-and-out drive. On the following possession, the Steelers took over the ball on their own 21 yard line, and made a short gain followed by an incomplete pass. Facing second and 5 from their 26 yard line, Steelers quarterback Ben Roethlisberger had a problem: the play clock was winding down and the team was not yet set.